diff --git a/submissions/README.md b/submissions/README.md index cac475f8..5a05fc82 100644 --- a/submissions/README.md +++ b/submissions/README.md @@ -18,7 +18,7 @@ This week's work is listed first in [`../NOW.md`](../NOW.md). | [`dxd_foundation/`](dxd_foundation/) | Foundations in Human Factors course paper (PHD1750-3 / HF700-1) and the foundation for Decision Experience Design: word separation as the design precedent, the equivalent move for algorithmacy, and the contrast with Nadella's "Models as Insider Risks" post. Researched in three rounds, audited, 251 sources carded; author ruled 2026-10-13; abstract, introduction and outline in the author's Google Doc (snapshot in DRAFT.md) | not recorded | | [`phd1750_spaces_in_stream/`](phd1750_spaces_in_stream/) | *Human–Computer Interaction* theory paper: Spaces in the Stream — algorithmacy and the perception of decision spaces in streamed LLM output (the PHD1750 human-factors paper, grown from `dxd_foundation/`). v2 draft with three reviews evaluated (22 Oct 2026); decisions D1–D8 open | not recorded | | [`berthon_ethics_course/`](berthon_ethics_course/) | Ethics course paper (Berthon): the ethical failures of the literacy transition and which of them the transition to algorithmacy repeats; built from the design-ethics draft and DXD Part 2 | researched and outlined; eight decisions open; no draft | -| [`berthon_ethics_paper/`](berthon_ethics_paper/) | The new Berthon ethics paper: every card from the design-ethics, Berthon-course and matched DXD libraries (422), byte-identical copies of the eight earlier versions, and a hand-written concept ledger, conflicts list and consolidation of their ideas | consolidated 2026-10-24; thesis and twelve decisions open; no draft | +| [`berthon_ethics_paper/`](berthon_ethics_paper/) | The new Berthon ethics paper: every card from the design-ethics, Berthon-course and matched DXD libraries (583), byte-identical copies of the eight earlier versions, and a hand-written concept ledger, conflicts list and consolidation of their ideas | consolidated 2026-10-24; thesis and twelve decisions open; no draft | | [`hegel_coordination/`](hegel_coordination/) | Hegel Substack series | no deadline | | [`triad_thinkers/`](triad_thinkers/) | Thinkers-on-the-triad Substack companion (post 1: Simmel) | no deadline | | [`triadic_reduction/`](triadic_reduction/) | ALGOCON talk: Peirce's reduction thesis, Simmel, Quine, the logicians, the lab's IIT move, and the case for algorithmacy (deck + script); sources ingested | 2026-10-28 | diff --git a/submissions/berthon_ethics_paper/PROVENANCE.md b/submissions/berthon_ethics_paper/PROVENANCE.md index f4932f2d..d89c83a3 100644 --- a/submissions/berthon_ethics_paper/PROVENANCE.md +++ b/submissions/berthon_ethics_paper/PROVENANCE.md @@ -3,7 +3,7 @@ Generated by `literature/build_provenance.py`; do not edit by hand. Every file below is a copy; the origin stays where it was. "Identical" compares sha256 of copy and origin at build time. -Files: 493. Not identical or origin missing: 0. +Files: 703. Not identical or origin missing: 0. | copy | origin | origin last commit | sha256 (12) | identical | |---|---|---|---|---| @@ -15,6 +15,55 @@ Files: 493. Not identical or origin missing: 0. | `sources/06_MANUSCRIPT_v2.md` | `submissions/berthon_ethics_course/MANUSCRIPT_v2.md` | a0d9a03 | fa8955615905 | yes | | `sources/07_MANUSCRIPT_v3.md` | `submissions/berthon_ethics_course/MANUSCRIPT_v3.md` | f375519 | 80609f2b566e | yes | | `sources/08_MANUSCRIPT_v4.md` | `submissions/berthon_ethics_course/MANUSCRIPT_v4.md` | 897c336 | 6fc685cd0e73 | yes | +| `nadella/sources/dxf_ABSTRACT_v4.md` | `submissions/dxd_foundation/ABSTRACT_v4.md` | 41ee7a1 | 69adb480d401 | yes | +| `nadella/sources/dxf_ARGUMENT_MAP.md` | `submissions/dxd_foundation/ARGUMENT_MAP.md` | fe6768d | 0da83b8eeb99 | yes | +| `nadella/sources/dxf_CLAIM.md` | `submissions/dxd_foundation/CLAIM.md` | 55b0d33 | d3437bbd05c1 | yes | +| `nadella/sources/dxf_DRAFT.md` | `submissions/dxd_foundation/DRAFT.md` | 9330b51 | 35ca34f9bc44 | yes | +| `nadella/sources/dxf_DRAFT_v2.md` | `submissions/dxd_foundation/DRAFT_v2.md` | 1b76806 | c8b3a14944f9 | yes | +| `nadella/sources/dxf_DRAFT_v3.md` | `submissions/dxd_foundation/DRAFT_v3.md` | 41ee7a1 | ae7301e6b5ad | yes | +| `nadella/sources/dxf_DRAFT_v4.md` | `submissions/dxd_foundation/DRAFT_v4.md` | 520ddd1 | 6abcb3ae0647 | yes | +| `nadella/sources/dxf_OUTLINE.md` | `submissions/dxd_foundation/OUTLINE.md` | fe6768d | 74be7fa9612a | yes | +| `nadella/sources/dxf_README.md` | `submissions/dxd_foundation/README.md` | 53daca9 | 0608f55a9b92 | yes | +| `nadella/sources/dxf_RESEARCH_PLAN.md` | `submissions/dxd_foundation/RESEARCH_PLAN.md` | 9330b51 | 7a10b170f614 | yes | +| `nadella/sources/dxf_REVIEW_v2_2026-10-13.md` | `submissions/dxd_foundation/REVIEW_v2_2026-10-13.md` | 1b76806 | 7806600d5993 | yes | +| `nadella/sources/dxf_SOURCE.md` | `submissions/dxd_foundation/SOURCE.md` | fe6768d | c3a16437f55e | yes | +| `nadella/sources/dxf_facet_E_human_factors_frame.md` | `submissions/dxd_foundation/literature/facets/E_human_factors_frame.md` | fe6768d | 3623159c2007 | yes | +| `nadella/sources/dxf_facet_G_nadella_lineage.md` | `submissions/dxd_foundation/literature/facets/G_nadella_lineage.md` | fe6768d | fb6bed721fe7 | yes | +| `nadella/sources/dxf_facet_H_candidate_moves.md` | `submissions/dxd_foundation/literature/facets/H_candidate_moves.md` | fe6768d | 360d1158ca80 | yes | +| `nadella/sources/dxf_facet_L_disclosure_and_choice.md` | `submissions/dxd_foundation/literature/facets/L_disclosure_and_choice.md` | 9330b51 | 017527dc29de | yes | +| `nadella/sources/dxf_facet_S_hmi_read_to_check.md` | `submissions/dxd_foundation/literature/facets/S_hmi_read_to_check.md` | 520ddd1 | 3ac2f4a7cdf0 | yes | +| `nadella/sources/dxf_facet_T_adjudication.md` | `submissions/dxd_foundation/literature/facets/T_adjudication.md` | 9330b51 | e7d79b25459d | yes | +| `nadella/sources/dxf_facet_X_counter_case.md` | `submissions/dxd_foundation/literature/facets/X_counter_case.md` | fe6768d | 9079f05ded60 | yes | +| `nadella/sources/dxf_facet_Z_gap_search.md` | `submissions/dxd_foundation/literature/facets/Z_gap_search.md` | fe6768d | 87268641a239 | yes | +| `nadella/sources/dxf_facet_G_nadella_lineage.bib` | `submissions/dxd_foundation/literature/facets/G_nadella_lineage.bib` | fe6768d | be43bb975fbf | yes | +| `nadella/sources/dxf_facet_A_the_record.md` | `submissions/dxd_foundation/literature/facets/A_the_record.md` | fe6768d | 2ac9b5b38ab7 | yes | +| `nadella/sources/dxf_facet_A_the_record.bib` | `submissions/dxd_foundation/literature/facets/A_the_record.bib` | fe6768d | 2cb8423fdc2c | yes | +| `nadella/sources/dxf_facet_B_the_dispute.md` | `submissions/dxd_foundation/literature/facets/B_the_dispute.md` | fe6768d | 9681be80ba26 | yes | +| `nadella/sources/dxf_facet_B_the_dispute.bib` | `submissions/dxd_foundation/literature/facets/B_the_dispute.bib` | fe6768d | 77f63387cdf2 | yes | +| `nadella/sources/dxf_facet_C_spacing_to_literacy.md` | `submissions/dxd_foundation/literature/facets/C_spacing_to_literacy.md` | fe6768d | 80368b8842c9 | yes | +| `nadella/sources/dxf_facet_C_spacing_to_literacy.bib` | `submissions/dxd_foundation/literature/facets/C_spacing_to_literacy.bib` | fe6768d | 0fae7aa009e6 | yes | +| `nadella/sources/dxf_facet_D_reading_science.md` | `submissions/dxd_foundation/literature/facets/D_reading_science.md` | fe6768d | 1ca56bf6ae96 | yes | +| `nadella/sources/dxf_facet_D_reading_science.bib` | `submissions/dxd_foundation/literature/facets/D_reading_science.bib` | fe6768d | 7f89af81c08f | yes | +| `nadella/sources/dxf_facet_F_parallel_moves.md` | `submissions/dxd_foundation/literature/facets/F_parallel_moves.md` | fe6768d | 7bceca0c0685 | yes | +| `nadella/sources/dxf_facet_F_parallel_moves.bib` | `submissions/dxd_foundation/literature/facets/F_parallel_moves.bib` | fe6768d | ffa77925b2c4 | yes | +| `nadella/sources/dxf_facet_J_who_holds_the_convention.md` | `submissions/dxd_foundation/literature/facets/J_who_holds_the_convention.md` | fe6768d | 1b6ea0d3ae4e | yes | +| `nadella/sources/dxf_facet_J_who_holds_the_convention.bib` | `submissions/dxd_foundation/literature/facets/J_who_holds_the_convention.bib` | fe6768d | 59e4fc865f30 | yes | +| `nadella/sources/dxf_facet_K_training_and_cause.md` | `submissions/dxd_foundation/literature/facets/K_training_and_cause.md` | fe6768d | 273e52384dac | yes | +| `nadella/sources/dxf_facet_K_training_and_cause.bib` | `submissions/dxd_foundation/literature/facets/K_training_and_cause.bib` | fe6768d | 6c2546ad3306 | yes | +| `nadella/sources/dxf_facet_O_custody_of_conventions.md` | `submissions/dxd_foundation/literature/facets/O_custody_of_conventions.md` | 9330b51 | 218c1dc27163 | yes | +| `nadella/sources/dxf_facet_O_custody_of_conventions.bib` | `submissions/dxd_foundation/literature/facets/O_custody_of_conventions.bib` | 9330b51 | 4ed12d67d170 | yes | +| `nadella/sources/dxf_facet_W_spacing_oracy_literacy.md` | `submissions/dxd_foundation/literature/facets/W_spacing_oracy_literacy.md` | 520ddd1 | 2220cac79210 | yes | +| `nadella/sources/dxf_facet_W_spacing_oracy_literacy.bib` | `submissions/dxd_foundation/literature/facets/W_spacing_oracy_literacy.bib` | 520ddd1 | 8b2ffcb42d0f | yes | +| `nadella/sources/bec_facet_B_print_1450_1800.md` | `submissions/berthon_ethics_course/literature/facets/B_print_1450_1800.md` | 8235899 | 2334c9958b38 | yes | +| `nadella/sources/bec_facet_B_print_1450_1800.bib` | `submissions/berthon_ethics_course/literature/facets/B_print_1450_1800.bib` | f92cb54 | b4bb46d3e5f2 | yes | +| `nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.md` | `submissions/berthon_ethics_course/literature/facets/V1_nadella_disclosure_paradigm.md` | c6c5f98 | 6f542fc7d1f5 | yes | +| `nadella/sources/bec_facet_V2_oracy_to_literacy.md` | `submissions/berthon_ethics_course/literature/facets/V2_oracy_to_literacy.md` | c6c5f98 | 1730366f4888 | yes | +| `nadella/sources/bec_facet_V4_disclosure_and_burden.md` | `submissions/berthon_ethics_course/literature/facets/V4_disclosure_and_burden.md` | c6c5f98 | ada2ffc0947d | yes | +| `nadella/sources/bec_facet_V5_scaffolding.md` | `submissions/berthon_ethics_course/literature/facets/V5_scaffolding.md` | c6c5f98 | 64abec7e6fcb | yes | +| `nadella/sources/bec_facet_V6_prior_work_and_objections.md` | `submissions/berthon_ethics_course/literature/facets/V6_prior_work_and_objections.md` | c6c5f98 | 59d8f0150392 | yes | +| `nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.bib` | `submissions/berthon_ethics_course/literature/facets/V1_nadella_disclosure_paradigm.bib` | c6c5f98 | 307046b50aa0 | yes | +| `nadella/sources/bec_v4_RESEARCH_BRIEF.md` | `submissions/berthon_ethics_course/v4/RESEARCH_BRIEF.md` | 3ac635c | b50196a79290 | yes | +| `nadella/sources/bec_v4_V4_CHECK_BRIEF.md` | `submissions/berthon_ethics_course/v4/V4_CHECK_BRIEF.md` | 436b556 | 2f64eeba7a5a | yes | | `sources/notes/ade_DRAFT_AUDIT.md` | `submissions/algorithmacy_design_ethics/DRAFT_AUDIT.md` | 20829a3 | 0fbb600eb8c4 | yes | | `sources/notes/ade_DRAFT_NISSENBAUM_SECTION.md` | `submissions/algorithmacy_design_ethics/DRAFT_NISSENBAUM_SECTION.md` | 20829a3 | 8798bfad0667 | yes | | `sources/notes/ade_README.md` | `submissions/algorithmacy_design_ethics/README.md` | 20829a3 | ad0f27ef8357 | yes | @@ -441,41 +490,202 @@ Files: 493. Not identical or origin missing: 0. | `literature/cards/algorithmacy_design_ethics/zamfirescupereira2023.md` | `submissions/algorithmacy_design_ethics/literature/library/zamfirescupereira2023.md` | 5d10943 | 4164799128fa | yes | | `literature/cards/algorithmacy_design_ethics/zong2024.md` | `submissions/algorithmacy_design_ethics/literature/library/zong2024.md` | 5d10943 | 201e7f6964df | yes | | `literature/cards/algorithmacy_design_ethics/zuboff2019.md` | `submissions/algorithmacy_design_ethics/literature/library/zuboff2019.md` | 20829a3 | 05599bf6931f | yes | +| `literature/cards/dxd_foundation/alahmadi2022false.md` | `submissions/dxd_foundation/literature/library/alahmadi2022false.md` | fe6768d | 7fc9f7abd61a | yes | | `literature/cards/dxd_foundation/alfrink2023contestable.md` | `submissions/dxd_foundation/literature/library/alfrink2023contestable.md` | fe6768d | 045b1ef370aa | yes | +| `literature/cards/dxd_foundation/ancker2017effects.md` | `submissions/dxd_foundation/literature/library/ancker2017effects.md` | fe6768d | 9cdf057a5539 | yes | +| `literature/cards/dxd_foundation/anderson1972computer.md` | `submissions/dxd_foundation/literature/library/anderson1972computer.md` | fe6768d | f78b3cbaea45 | yes | +| `literature/cards/dxd_foundation/anthropic2026how.md` | `submissions/dxd_foundation/literature/library/anthropic2026how.md` | fe6768d | 3930cd80cdc0 | yes | +| `literature/cards/dxd_foundation/apple2026guidelines.md` | `submissions/dxd_foundation/literature/library/apple2026guidelines.md` | fe6768d | 362ae2831f11 | yes | +| `literature/cards/dxd_foundation/bai2008reading.md` | `submissions/dxd_foundation/literature/library/bai2008reading.md` | 520ddd1 | ddf6a70847a7 | yes | | `literature/cards/dxd_foundation/bainbridge1983ironies.md` | `submissions/dxd_foundation/literature/library/bainbridge1983ironies.md` | 520ddd1 | a0e8f64aac0d | yes | +| `literature/cards/dxd_foundation/baker2025monitoring.md` | `submissions/dxd_foundation/literature/library/baker2025monitoring.md` | fe6768d | d016ea297a55 | yes | +| `literature/cards/dxd_foundation/balogh1927voces.md` | `submissions/dxd_foundation/literature/library/balogh1927voces.md` | 520ddd1 | b0bd855ad271 | yes | | `literature/cards/dxd_foundation/bansal2019updates.md` | `submissions/dxd_foundation/literature/library/bansal2019updates.md` | fe6768d | b32bf94a3b2a | yes | | `literature/cards/dxd_foundation/bansal2021whole.md` | `submissions/dxd_foundation/literature/library/bansal2021whole.md` | cd25745 | ba026b206025 | yes | +| `literature/cards/dxd_foundation/barton1980awareness.md` | `submissions/dxd_foundation/literature/library/barton1980awareness.md` | 520ddd1 | 5e83b5e0dfaa | yes | +| `literature/cards/dxd_foundation/bassetti2012effects.md` | `submissions/dxd_foundation/literature/library/bassetti2012effects.md` | fe6768d | 7c485018289a | yes | +| `literature/cards/dxd_foundation/becker2024conquering.md` | `submissions/dxd_foundation/literature/library/becker2024conquering.md` | fe6768d | 13f77917c70c | yes | +| `literature/cards/dxd_foundation/biggs2020decoding.md` | `submissions/dxd_foundation/literature/library/biggs2020decoding.md` | fe6768d | 22266a082609 | yes | +| `literature/cards/dxd_foundation/boretius1883capitularia.md` | `submissions/dxd_foundation/literature/library/boretius1883capitularia.md` | cd25745 | b3f20143fdbb | yes | | `literature/cards/dxd_foundation/bucinca2021trust.md` | `submissions/dxd_foundation/literature/library/bucinca2021trust.md` | fe6768d | 5688aa447f0b | yes | +| `literature/cards/dxd_foundation/buijsman2020numerals.md` | `submissions/dxd_foundation/literature/library/buijsman2020numerals.md` | fe6768d | 8943fec3c1e5 | yes | +| `literature/cards/dxd_foundation/buncic2017factors.md` | `submissions/dxd_foundation/literature/library/buncic2017factors.md` | fe6768d | 2683c7446dbf | yes | | `literature/cards/dxd_foundation/burnyeat1997postscript.md` | `submissions/dxd_foundation/literature/library/burnyeat1997postscript.md` | cd25745 | f7b39ab0a581 | yes | | `literature/cards/dxd_foundation/cabrera2023improving.md` | `submissions/dxd_foundation/literature/library/cabrera2023improving.md` | fe6768d | 3dacdbea1280 | yes | +| `literature/cards/dxd_foundation/cameron2022expanding.md` | `submissions/dxd_foundation/literature/library/cameron2022expanding.md` | cd25745 | d13e5bb89ad2 | yes | +| `literature/cards/dxd_foundation/caperton2009massey.md` | `submissions/dxd_foundation/literature/library/caperton2009massey.md` | fe6768d | 2ef662b118f5 | yes | | `literature/cards/dxd_foundation/casner2014retention.md` | `submissions/dxd_foundation/literature/library/casner2014retention.md` | cd25745 | 7451bfb4d3c7 | yes | +| `literature/cards/dxd_foundation/cert2022common.md` | `submissions/dxd_foundation/literature/library/cert2022common.md` | fe6768d | cd4c9692eed8 | yes | +| `literature/cards/dxd_foundation/cftc2017fxcm.md` | `submissions/dxd_foundation/literature/library/cftc2017fxcm.md` | fe6768d | d71fd2e58f61 | yes | +| `literature/cards/dxd_foundation/chan2024visibility.md` | `submissions/dxd_foundation/literature/library/chan2024visibility.md` | fe6768d | a5a255277adf | yes | +| `literature/cards/dxd_foundation/chan2025infrastructure.md` | `submissions/dxd_foundation/literature/library/chan2025infrastructure.md` | fe6768d | 6227d1a08152 | yes | +| `literature/cards/dxd_foundation/chantavarin2025effects.md` | `submissions/dxd_foundation/literature/library/chantavarin2025effects.md` | fe6768d | 4ffc350414f8 | yes | +| `literature/cards/dxd_foundation/charlemagne1900letter.md` | `submissions/dxd_foundation/literature/library/charlemagne1900letter.md` | cd25745 | e9d1da448cac | yes | +| `literature/cards/dxd_foundation/chen2021tradeoff.md` | `submissions/dxd_foundation/literature/library/chen2021tradeoff.md` | fe6768d | 4fae5d5c4c83 | yes | +| `literature/cards/dxd_foundation/chen2025reasoning.md` | `submissions/dxd_foundation/literature/library/chen2025reasoning.md` | fe6768d | 1cb99d4c4472 | yes | +| `literature/cards/dxd_foundation/chen2026comparing.md` | `submissions/dxd_foundation/literature/library/chen2026comparing.md` | fe6768d | 28c82d4b7ae1 | yes | +| `literature/cards/dxd_foundation/clanchy1993from.md` | `submissions/dxd_foundation/literature/library/clanchy1993from.md` | cd25745 | 85e97d6a0e74 | yes | +| `literature/cards/dxd_foundation/clemens2025korea.md` | `submissions/dxd_foundation/literature/library/clemens2025korea.md` | fe6768d | b4b14eaf544c | yes | +| `literature/cards/dxd_foundation/colvin2011empirical.md` | `submissions/dxd_foundation/literature/library/colvin2011empirical.md` | fe6768d | 3a97b4336990 | yes | +| `literature/cards/dxd_foundation/constantin2026nadella.md` | `submissions/dxd_foundation/literature/library/constantin2026nadella.md` | fe6768d | 76af799279dd | yes | +| `literature/cards/dxd_foundation/cressy1980literacy.md` | `submissions/dxd_foundation/literature/library/cressy1980literacy.md` | cd25745 | 3e5e33bee82c | yes | +| `literature/cards/dxd_foundation/crootof2023humans.md` | `submissions/dxd_foundation/literature/library/crootof2023humans.md` | fe6768d | 74ccc70c4c10 | yes | +| `literature/cards/dxd_foundation/cui2023eye.md` | `submissions/dxd_foundation/literature/library/cui2023eye.md` | fe6768d | c5e0eb9c34c4 | yes | +| `literature/cards/dxd_foundation/dacosta2018tabulae.md` | `submissions/dxd_foundation/literature/library/dacosta2018tabulae.md` | fe6768d | eeff01f57f37 | yes | +| `literature/cards/dxd_foundation/danna2022numerical.md` | `submissions/dxd_foundation/literature/library/danna2022numerical.md` | fe6768d | 824d24a74610 | yes | +| `literature/cards/dxd_foundation/de1972betrvg.md` | `submissions/dxd_foundation/literature/library/de1972betrvg.md` | cd25745 | 81e8a3cb8958 | yes | +| `literature/cards/dxd_foundation/degno2019coregistration.md` | `submissions/dxd_foundation/literature/library/degno2019coregistration.md` | fe6768d | f2da47290e7c | yes | | `literature/cards/dxd_foundation/delacroix2019bottom.md` | `submissions/dxd_foundation/literature/library/delacroix2019bottom.md` | fe6768d | 6d10f82db3e8 | yes | | `literature/cards/dxd_foundation/dietvorst2018overcoming.md` | `submissions/dxd_foundation/literature/library/dietvorst2018overcoming.md` | fe6768d | 69b49cd80dbd | yes | +| `literature/cards/dxd_foundation/dionysius1910composition.md` | `submissions/dxd_foundation/literature/library/dionysius1910composition.md` | fe6768d | e3313a9fa25d | yes | +| `literature/cards/dxd_foundation/douek2023meta.md` | `submissions/dxd_foundation/literature/library/douek2023meta.md` | fe6768d | f2ae4fae23bc | yes | +| `literature/cards/dxd_foundation/doyle2025assessment.md` | `submissions/dxd_foundation/literature/library/doyle2025assessment.md` | 520ddd1 | e8af89068965 | yes | +| `literature/cards/dxd_foundation/dummler1881alcuini.md` | `submissions/dxd_foundation/literature/library/dummler1881alcuini.md` | cd25745 | 57fdd5f9697a | yes | +| `literature/cards/dxd_foundation/duzan2026chain.md` | `submissions/dxd_foundation/literature/library/duzan2026chain.md` | fe6768d | a2ed7300e837 | yes | +| `literature/cards/dxd_foundation/elish2019moral.md` | `submissions/dxd_foundation/literature/library/elish2019moral.md` | fe6768d | 9ac53fe67235 | yes | +| `literature/cards/dxd_foundation/eu2019p2b.md` | `submissions/dxd_foundation/literature/library/eu2019p2b.md` | fe6768d | 5d9387cb9bbf | yes | +| `literature/cards/dxd_foundation/eu2022dsa.md` | `submissions/dxd_foundation/literature/library/eu2022dsa.md` | fe6768d | bb34742eec06 | yes | +| `literature/cards/dxd_foundation/eu2024platform.md` | `submissions/dxd_foundation/literature/library/eu2024platform.md` | cd25745 | 1b146c8e7b9f | yes | +| `literature/cards/dxd_foundation/eu2024regulation.md` | `submissions/dxd_foundation/literature/library/eu2024regulation.md` | 520ddd1 | 8e41ebf59ca2 | yes | +| `literature/cards/dxd_foundation/evans2018forex.md` | `submissions/dxd_foundation/literature/library/evans2018forex.md` | fe6768d | 826d53df46a9 | yes | +| `literature/cards/dxd_foundation/fca2014fines.md` | `submissions/dxd_foundation/literature/library/fca2014fines.md` | fe6768d | 72d58dfe6bbd | yes | | `literature/cards/dxd_foundation/fok2024search.md` | `submissions/dxd_foundation/literature/library/fok2024search.md` | cd25745 | 175f34c4e908 | yes | +| `literature/cards/dxd_foundation/frey2026correlational.md` | `submissions/dxd_foundation/literature/library/frey2026correlational.md` | fe6768d | 0bda2b41d37d | yes | +| `literature/cards/dxd_foundation/fsb2014foreign.md` | `submissions/dxd_foundation/literature/library/fsb2014foreign.md` | fe6768d | 8a4a76ebfd22 | yes | +| `literature/cards/dxd_foundation/fsb2015progress.md` | `submissions/dxd_foundation/literature/library/fsb2015progress.md` | fe6768d | b66dbffce3d7 | yes | +| `literature/cards/dxd_foundation/galanter1974why.md` | `submissions/dxd_foundation/literature/library/galanter1974why.md` | fe6768d | 76e395fa54c6 | yes | +| `literature/cards/dxd_foundation/galliussi2020inter.md` | `submissions/dxd_foundation/literature/library/galliussi2020inter.md` | fe6768d | 5edeb660e104 | yes | +| `literature/cards/dxd_foundation/garrett1998saenger.md` | `submissions/dxd_foundation/literature/library/garrett1998saenger.md` | fe6768d | e5e5688600eb | yes | | `literature/cards/dxd_foundation/gavrilov1997techniques.md` | `submissions/dxd_foundation/literature/library/gavrilov1997techniques.md` | cd25745 | 338dea1c4565 | yes | +| `literature/cards/dxd_foundation/gema2026chain.md` | `submissions/dxd_foundation/literature/library/gema2026chain.md` | fe6768d | 6996f3faca48 | yes | +| `literature/cards/dxd_foundation/globalfxc2024code.md` | `submissions/dxd_foundation/literature/library/globalfxc2024code.md` | fe6768d | 3ae0e9ea279e | yes | | `literature/cards/dxd_foundation/gray2018dark.md` | `submissions/dxd_foundation/literature/library/gray2018dark.md` | fe6768d | fa6ab825fb36 | yes | +| `literature/cards/dxd_foundation/green2022flaws.md` | `submissions/dxd_foundation/literature/library/green2022flaws.md` | fe6768d | 38fd5237a53d | yes | +| `literature/cards/dxd_foundation/greenblatt2024ai.md` | `submissions/dxd_foundation/literature/library/greenblatt2024ai.md` | fe6768d | 38230f71ee14 | yes | | `literature/cards/dxd_foundation/grinschgl2021consequences.md` | `submissions/dxd_foundation/literature/library/grinschgl2021consequences.md` | 520ddd1 | aaa9e58503e0 | yes | +| `literature/cards/dxd_foundation/grundemclaughlin2026overseeing.md` | `submissions/dxd_foundation/literature/library/grundemclaughlin2026overseeing.md` | fe6768d | 7cb259039430 | yes | +| `literature/cards/dxd_foundation/haines2008origins.md` | `submissions/dxd_foundation/literature/library/haines2008origins.md` | fe6768d | 60b415821341 | yes | +| `literature/cards/dxd_foundation/hakvoort2017improvements.md` | `submissions/dxd_foundation/literature/library/hakvoort2017improvements.md` | fe6768d | 55ab4c1163eb | yes | +| `literature/cards/dxd_foundation/hangeulmuseum.md` | `submissions/dxd_foundation/literature/library/hangeulmuseum.md` | fe6768d | 009816c4414a | yes | +| `literature/cards/dxd_foundation/hao2024training.md` | `submissions/dxd_foundation/literature/library/hao2024training.md` | fe6768d | 46169d860ea3 | yes | +| `literature/cards/dxd_foundation/heilmann2020reading.md` | `submissions/dxd_foundation/literature/library/heilmann2020reading.md` | 520ddd1 | f4a4377e01bc | yes | +| `literature/cards/dxd_foundation/hendrickson1929ancient.md` | `submissions/dxd_foundation/literature/library/hendrickson1929ancient.md` | 520ddd1 | 31a918de9745 | yes | | `literature/cards/dxd_foundation/hertwig2017nudging.md` | `submissions/dxd_foundation/literature/library/hertwig2017nudging.md` | cd25745 | dfe2299be7f9 | yes | +| `literature/cards/dxd_foundation/hervais2019learning.md` | `submissions/dxd_foundation/literature/library/hervais2019learning.md` | fe6768d | 78f189c87dde | yes | +| `literature/cards/dxd_foundation/houpt2014new.md` | `submissions/dxd_foundation/literature/library/houpt2014new.md` | fe6768d | 9e7fcb3778c0 | yes | | `literature/cards/dxd_foundation/hutchins1985direct.md` | `submissions/dxd_foundation/literature/library/hutchins1985direct.md` | 520ddd1 | 8c17cfdb11bd | yes | +| `literature/cards/dxd_foundation/ivanova2026process.md` | `submissions/dxd_foundation/literature/library/ivanova2026process.md` | fe6768d | 4205514fb01c | yes | +| `literature/cards/dxd_foundation/jakkal2026secure.md` | `submissions/dxd_foundation/literature/library/jakkal2026secure.md` | fe6768d | 4e81c5aba2ce | yes | +| `literature/cards/dxd_foundation/jhaver2019suspect.md` | `submissions/dxd_foundation/literature/library/jhaver2019suspect.md` | fe6768d | d8112d92dd9a | yes | +| `literature/cards/dxd_foundation/jia2026faithfulness.md` | `submissions/dxd_foundation/literature/library/jia2026faithfulness.md` | fe6768d | ffc3977389bc | yes | +| `literature/cards/dxd_foundation/jordan2024reicher.md` | `submissions/dxd_foundation/literature/library/jordan2024reicher.md` | fe6768d | f02fe502d4fa | yes | | `literature/cards/dxd_foundation/kahneman2009conditions.md` | `submissions/dxd_foundation/literature/library/kahneman2009conditions.md` | 520ddd1 | fc5e42c7ff1c | yes | +| `literature/cards/dxd_foundation/kasisopa2016child.md` | `submissions/dxd_foundation/literature/library/kasisopa2016child.md` | 520ddd1 | b940843a499f | yes | +| `literature/cards/dxd_foundation/kellogg2020algorithms.md` | `submissions/dxd_foundation/literature/library/kellogg2020algorithms.md` | 520ddd1 | 8c4582ac9a13 | yes | +| `literature/cards/dxd_foundation/ker2012johnson.md` | `submissions/dxd_foundation/literature/library/ker2012johnson.md` | fe6768d | 7ffb58d79787 | yes | | `literature/cards/dxd_foundation/kim2024not.md` | `submissions/dxd_foundation/literature/library/kim2024not.md` | cd25745 | 2bcde086b542 | yes | +| `literature/cards/dxd_foundation/klonick2018new.md` | `submissions/dxd_foundation/literature/library/klonick2018new.md` | fe6768d | 17d8a9c9df4d | yes | | `literature/cards/dxd_foundation/knox1968silent.md` | `submissions/dxd_foundation/literature/library/knox1968silent.md` | 520ddd1 | ba25bf45a439 | yes | +| `literature/cards/dxd_foundation/kohsom1997adding.md` | `submissions/dxd_foundation/literature/library/kohsom1997adding.md` | fe6768d | 783c320657c3 | yes | +| `literature/cards/dxd_foundation/kokotajlo2025ai.md` | `submissions/dxd_foundation/literature/library/kokotajlo2025ai.md` | fe6768d | 08b18c508c2e | yes | +| `literature/cards/dxd_foundation/korbak2025chain.md` | `submissions/dxd_foundation/literature/library/korbak2025chain.md` | fe6768d | 352842af2cef | yes | +| `literature/cards/dxd_foundation/krieber2017eye.md` | `submissions/dxd_foundation/literature/library/krieber2017eye.md` | 520ddd1 | 5e06fd12e241 | yes | +| `literature/cards/dxd_foundation/kroll2021outlining.md` | `submissions/dxd_foundation/literature/library/kroll2021outlining.md` | fe6768d | 150e95ba34c1 | yes | +| `literature/cards/dxd_foundation/kubota2024rethinking.md` | `submissions/dxd_foundation/literature/library/kubota2024rethinking.md` | fe6768d | eafa095b7909 | yes | +| `literature/cards/dxd_foundation/kumari2026nadella.md` | `submissions/dxd_foundation/literature/library/kumari2026nadella.md` | fe6768d | 8692dcad2ae5 | yes | +| `literature/cards/dxd_foundation/kutas2011thirty.md` | `submissions/dxd_foundation/literature/library/kutas2011thirty.md` | fe6768d | ff76c2819681 | yes | +| `literature/cards/dxd_foundation/kutasov2025shade.md` | `submissions/dxd_foundation/literature/library/kutasov2025shade.md` | fe6768d | 0f5a12db8a6f | yes | +| `literature/cards/dxd_foundation/langer2024effective.md` | `submissions/dxd_foundation/literature/library/langer2024effective.md` | fe6768d | 914bdc3d1efe | yes | +| `literature/cards/dxd_foundation/lanham2023measuring.md` | `submissions/dxd_foundation/literature/library/lanham2023measuring.md` | fe6768d | 56be46b4ff85 | yes | +| `literature/cards/dxd_foundation/larkin1987why.md` | `submissions/dxd_foundation/literature/library/larkin1987why.md` | fe6768d | f3765d7b9b3a | yes | +| `literature/cards/dxd_foundation/laubrock2015eyevoice.md` | `submissions/dxd_foundation/literature/library/laubrock2015eyevoice.md` | 520ddd1 | 0bbce3333567 | yes | +| `literature/cards/dxd_foundation/laux2024institutionalised.md` | `submissions/dxd_foundation/literature/library/laux2024institutionalised.md` | fe6768d | b67992a5e63a | yes | +| `literature/cards/dxd_foundation/layman2023controlled.md` | `submissions/dxd_foundation/literature/library/layman2023controlled.md` | fe6768d | fc37d487294b | yes | | `literature/cards/dxd_foundation/lee2025impact.md` | `submissions/dxd_foundation/literature/library/lee2025impact.md` | cd25745 | 8f951a1b11f8 | yes | +| `literature/cards/dxd_foundation/leinenger2014phonological.md` | `submissions/dxd_foundation/literature/library/leinenger2014phonological.md` | 520ddd1 | b95e05454dae | yes | +| `literature/cards/dxd_foundation/li2014reading.md` | `submissions/dxd_foundation/literature/library/li2014reading.md` | 520ddd1 | 8cf76edc6ca0 | yes | +| `literature/cards/dxd_foundation/little2026length.md` | `submissions/dxd_foundation/literature/library/little2026length.md` | fe6768d | 9b22b8418aca | yes | | `literature/cards/dxd_foundation/lukoff2021how.md` | `submissions/dxd_foundation/literature/library/lukoff2021how.md` | fe6768d | 2a8710058c9d | yes | +| `literature/cards/dxd_foundation/lynch2025agentic.md` | `submissions/dxd_foundation/literature/library/lynch2025agentic.md` | fe6768d | 68128be6a49b | yes | +| `literature/cards/dxd_foundation/mannering2012johnson.md` | `submissions/dxd_foundation/literature/library/mannering2012johnson.md` | fe6768d | bb2217e5acec | yes | +| `literature/cards/dxd_foundation/margono1987study.md` | `submissions/dxd_foundation/literature/library/margono1987study.md` | fe6768d | 8b95bf3a75ec | yes | +| `literature/cards/dxd_foundation/mathews1976eldridge.md` | `submissions/dxd_foundation/literature/library/mathews1976eldridge.md` | fe6768d | 65119bcc16f9 | yes | | `literature/cards/dxd_foundation/mathur2019dark.md` | `submissions/dxd_foundation/literature/library/mathur2019dark.md` | fe6768d | 86fa3ac58242 | yes | +| `literature/cards/dxd_foundation/matzinger2021influence.md` | `submissions/dxd_foundation/literature/library/matzinger2021influence.md` | 520ddd1 | 4a5b3ed6a84c | yes | +| `literature/cards/dxd_foundation/mirault2019precise.md` | `submissions/dxd_foundation/literature/library/mirault2019precise.md` | fe6768d | 9223dc560484 | yes | | `literature/cards/dxd_foundation/mirault2019reading.md` | `submissions/dxd_foundation/literature/library/mirault2019reading.md` | fe6768d | dc403b1c6e07 | yes | +| `literature/cards/dxd_foundation/mitchell2026agents.md` | `submissions/dxd_foundation/literature/library/mitchell2026agents.md` | fe6768d | 196d1bfdacb9 | yes | | `literature/cards/dxd_foundation/mozannar2024reading.md` | `submissions/dxd_foundation/literature/library/mozannar2024reading.md` | cd25745 | 994469071085 | yes | +| `literature/cards/dxd_foundation/nadella2025looking.md` | `submissions/dxd_foundation/literature/library/nadella2025looking.md` | fe6768d | 3ffc2146f924 | yes | +| `literature/cards/dxd_foundation/nadella2026frontier.md` | `submissions/dxd_foundation/literature/library/nadella2026frontier.md` | fe6768d | d77050f7c0b3 | yes | +| `literature/cards/dxd_foundation/nadella2026models.md` | `submissions/dxd_foundation/literature/library/nadella2026models.md` | fe6768d | 42f7e3c77314 | yes | +| `literature/cards/dxd_foundation/nadella2026reverse.md` | `submissions/dxd_foundation/literature/library/nadella2026reverse.md` | fe6768d | 1e69715cbf6b | yes | +| `literature/cards/dxd_foundation/nagarajrao2026fareshare.md` | `submissions/dxd_foundation/literature/library/nagarajrao2026fareshare.md` | fe6768d | b202e529a0b6 | yes | +| `literature/cards/dxd_foundation/nothaft2020satanic.md` | `submissions/dxd_foundation/literature/library/nothaft2020satanic.md` | fe6768d | d7718ee97d4c | yes | +| `literature/cards/dxd_foundation/nydfs2015barclays.md` | `submissions/dxd_foundation/literature/library/nydfs2015barclays.md` | fe6768d | 17dce7f543ad | yes | +| `literature/cards/dxd_foundation/nydfs2015barclayspr.md` | `submissions/dxd_foundation/literature/library/nydfs2015barclayspr.md` | fe6768d | 760b72975553 | yes | +| `literature/cards/dxd_foundation/odonnell1994parkes.md` | `submissions/dxd_foundation/literature/library/odonnell1994parkes.md` | fe6768d | dc3e9c726d09 | yes | +| `literature/cards/dxd_foundation/odonnell2000saenger.md` | `submissions/dxd_foundation/literature/library/odonnell2000saenger.md` | fe6768d | a3a73c3f878b | yes | | `literature/cards/dxd_foundation/onnasch2014human.md` | `submissions/dxd_foundation/literature/library/onnasch2014human.md` | 520ddd1 | ca4adfec81cc | yes | +| `literature/cards/dxd_foundation/oomen2016last.md` | `submissions/dxd_foundation/literature/library/oomen2016last.md` | fe6768d | c90a9715c7bc | yes | +| `literature/cards/dxd_foundation/oversight2022q1.md` | `submissions/dxd_foundation/literature/library/oversight2022q1.md` | fe6768d | 731613da492b | yes | +| `literature/cards/dxd_foundation/ozkan2026wordinitial.md` | `submissions/dxd_foundation/literature/library/ozkan2026wordinitial.md` | fe6768d | e31f562834a9 | yes | +| `literature/cards/dxd_foundation/palka2025foreign.md` | `submissions/dxd_foundation/literature/library/palka2025foreign.md` | fe6768d | 842c5070cad7 | yes | | `literature/cards/dxd_foundation/parasuraman2010complacency.md` | `submissions/dxd_foundation/literature/library/parasuraman2010complacency.md` | 520ddd1 | e1d98a84386e | yes | +| `literature/cards/dxd_foundation/park2025dokkyo.md` | `submissions/dxd_foundation/literature/library/park2025dokkyo.md` | fe6768d | 8331a4ae55ea | yes | | `literature/cards/dxd_foundation/parkes1992pause.md` | `submissions/dxd_foundation/literature/library/parkes1992pause.md` | cd25745 | b7f767deaf22 | yes | +| `literature/cards/dxd_foundation/perea2012subtle.md` | `submissions/dxd_foundation/literature/library/perea2012subtle.md` | fe6768d | b35c5c97297e | yes | +| `literature/cards/dxd_foundation/perea2015colours.md` | `submissions/dxd_foundation/literature/library/perea2015colours.md` | 520ddd1 | 34471a087dc5 | yes | | `literature/cards/dxd_foundation/plato1925phaedrus.md` | `submissions/dxd_foundation/literature/library/plato1925phaedrus.md` | fe6768d | d0e55046b156 | yes | +| `literature/cards/dxd_foundation/price2011interactive.md` | `submissions/dxd_foundation/literature/library/price2011interactive.md` | fe6768d | f5735e2a99d8 | yes | +| `literature/cards/dxd_foundation/quintilian1920institutio.md` | `submissions/dxd_foundation/literature/library/quintilian1920institutio.md` | fe6768d | 46f65b6c3f12 | yes | | `literature/cards/dxd_foundation/rayner1998unspaced.md` | `submissions/dxd_foundation/literature/library/rayner1998unspaced.md` | 520ddd1 | 74e0c53f49e1 | yes | +| `literature/cards/dxd_foundation/reis2011johnson.md` | `submissions/dxd_foundation/literature/library/reis2011johnson.md` | fe6768d | 609399eccea0 | yes | +| `literature/cards/dxd_foundation/rfc8890.md` | `submissions/dxd_foundation/literature/library/rfc8890.md` | cd25745 | 472bc4be9675 | yes | +| `literature/cards/dxd_foundation/rieder2020towards.md` | `submissions/dxd_foundation/literature/library/rieder2020towards.md` | fe6768d | bbe9cb5ef0b3 | yes | | `literature/cards/dxd_foundation/risko2016cognitive.md` | `submissions/dxd_foundation/literature/library/risko2016cognitive.md` | 520ddd1 | dffdd34cc0ae | yes | +| `literature/cards/dxd_foundation/rose2020zero.md` | `submissions/dxd_foundation/literature/library/rose2020zero.md` | fe6768d | e56d51d93bce | yes | | `literature/cards/dxd_foundation/rosenblat2016algorithmic.md` | `submissions/dxd_foundation/literature/library/rosenblat2016algorithmic.md` | 520ddd1 | 128173c6cfe5 | yes | +| `literature/cards/dxd_foundation/ryan2014historical.md` | `submissions/dxd_foundation/literature/library/ryan2014historical.md` | fe6768d | 96732be7c0b7 | yes | | `literature/cards/dxd_foundation/saenger1997space.md` | `submissions/dxd_foundation/literature/library/saenger1997space.md` | 520ddd1 | 1d46fcbaeeb2 | yes | +| `literature/cards/dxd_foundation/saltzer1975protection.md` | `submissions/dxd_foundation/literature/library/saltzer1975protection.md` | fe6768d | 815f46dca30a | yes | +| `literature/cards/dxd_foundation/sangster2024emergence.md` | `submissions/dxd_foundation/literature/library/sangster2024emergence.md` | fe6768d | d0aa0e19d979 | yes | +| `literature/cards/dxd_foundation/sangster2025reboot.md` | `submissions/dxd_foundation/literature/library/sangster2025reboot.md` | fe6768d | 0807347d133f | yes | +| `literature/cards/dxd_foundation/santonidesio2018meaningful.md` | `submissions/dxd_foundation/literature/library/santonidesio2018meaningful.md` | fe6768d | 2ad5af77ab58 | yes | +| `literature/cards/dxd_foundation/schlimm2008modeling.md` | `submissions/dxd_foundation/literature/library/schlimm2008modeling.md` | fe6768d | 3bed0e139cf8 | yes | +| `literature/cards/dxd_foundation/scribner1981psychology.md` | `submissions/dxd_foundation/literature/library/scribner1981psychology.md` | 520ddd1 | 72f4d419eac2 | yes | +| `literature/cards/dxd_foundation/seattle2024deactivation.md` | `submissions/dxd_foundation/literature/library/seattle2024deactivation.md` | fe6768d | 828c488f8358 | yes | +| `literature/cards/dxd_foundation/seymour2003foundation.md` | `submissions/dxd_foundation/literature/library/seymour2003foundation.md` | fe6768d | 0e1aaa095e45 | yes | | `literature/cards/dxd_foundation/shneiderman1983direct.md` | `submissions/dxd_foundation/literature/library/shneiderman1983direct.md` | 520ddd1 | d057da4455b4 | yes | +| `literature/cards/dxd_foundation/sterz2024quest.md` | `submissions/dxd_foundation/literature/library/sterz2024quest.md` | fe6768d | 5811658adcad | yes | +| `literature/cards/dxd_foundation/stohl2016managing.md` | `submissions/dxd_foundation/literature/library/stohl2016managing.md` | fe6768d | 95eee2cd3003 | yes | +| `literature/cards/dxd_foundation/strunk1998guido.md` | `submissions/dxd_foundation/literature/library/strunk1998guido.md` | fe6768d | 98a12ca5d9ab | yes | +| `literature/cards/dxd_foundation/suzor2016responsibilities.md` | `submissions/dxd_foundation/literature/library/suzor2016responsibilities.md` | fe6768d | 4fca3f76e13b | yes | +| `literature/cards/dxd_foundation/suzor2018digital.md` | `submissions/dxd_foundation/literature/library/suzor2018digital.md` | fe6768d | c3ac739bbcfd | yes | +| `literature/cards/dxd_foundation/thrax1874grammar.md` | `submissions/dxd_foundation/literature/library/thrax1874grammar.md` | fe6768d | 932e267cf265 | yes | +| `literature/cards/dxd_foundation/torrensalvarez2022latin.md` | `submissions/dxd_foundation/literature/library/torrensalvarez2022latin.md` | fe6768d | ee840e5916be | yes | +| `literature/cards/dxd_foundation/tumey1927ohio.md` | `submissions/dxd_foundation/literature/library/tumey1927ohio.md` | fe6768d | 3b60fa436b2a | yes | +| `literature/cards/dxd_foundation/tunbridge1992study.md` | `submissions/dxd_foundation/literature/library/tunbridge1992study.md` | 520ddd1 | f9d49a3e578c | yes | +| `literature/cards/dxd_foundation/turan2026oversight.md` | `submissions/dxd_foundation/literature/library/turan2026oversight.md` | fe6768d | eb5aacaeff8a | yes | +| `literature/cards/dxd_foundation/turpin2023language.md` | `submissions/dxd_foundation/literature/library/turpin2023language.md` | fe6768d | 8d606a3f4fb5 | yes | +| `literature/cards/dxd_foundation/tyler1988what.md` | `submissions/dxd_foundation/literature/library/tyler1988what.md` | fe6768d | 5eff247ffdeb | yes | +| `literature/cards/dxd_foundation/us2072usc.md` | `submissions/dxd_foundation/literature/library/us2072usc.md` | fe6768d | 02dc395ee016 | yes | +| `literature/cards/dxd_foundation/us2073usc.md` | `submissions/dxd_foundation/literature/library/us2073usc.md` | fe6768d | 56042bd30885 | yes | +| `literature/cards/dxd_foundation/us455usc.md` | `submissions/dxd_foundation/literature/library/us455usc.md` | fe6768d | 3482730b714d | yes | +| `literature/cards/dxd_foundation/us9usc10.md` | `submissions/dxd_foundation/literature/library/us9usc10.md` | fe6768d | 26a097cd4800 | yes | +| `literature/cards/dxd_foundation/vandenboer2015default.md` | `submissions/dxd_foundation/literature/library/vandenboer2015default.md` | fe6768d | 66c2006f8033 | yes | | `literature/cards/dxd_foundation/vasconcelos2023explanations.md` | `submissions/dxd_foundation/literature/library/vasconcelos2023explanations.md` | fe6768d | 01e056e5e3af | yes | +| `literature/cards/dxd_foundation/wagemans2012century.md` | `submissions/dxd_foundation/literature/library/wagemans2012century.md` | fe6768d | 80d4fea7ee17 | yes | +| `literature/cards/dxd_foundation/wang2024tibetan.md` | `submissions/dxd_foundation/literature/library/wang2024tibetan.md` | fe6768d | 62a5c4cb6d70 | yes | +| `literature/cards/dxd_foundation/wang2026monitorbench.md` | `submissions/dxd_foundation/literature/library/wang2026monitorbench.md` | fe6768d | 2ddc753d465c | yes | +| `literature/cards/dxd_foundation/wertli2023influence.md` | `submissions/dxd_foundation/literature/library/wertli2023influence.md` | fe6768d | 760746417a38 | yes | +| `literature/cards/dxd_foundation/westwell2018dissemination.md` | `submissions/dxd_foundation/literature/library/westwell2018dissemination.md` | cd25745 | 6a9be20d2591 | yes | +| `literature/cards/dxd_foundation/winskel2009eye.md` | `submissions/dxd_foundation/literature/library/winskel2009eye.md` | 520ddd1 | acc7ae66ebf0 | yes | +| `literature/cards/dxd_foundation/wong2022meta.md` | `submissions/dxd_foundation/literature/library/wong2022meta.md` | fe6768d | ef3777b7f79b | yes | | `literature/cards/dxd_foundation/wood2019good.md` | `submissions/dxd_foundation/literature/library/wood2019good.md` | 520ddd1 | 02e7ce3d9b3f | yes | +| `literature/cards/dxd_foundation/yu2023altered.md` | `submissions/dxd_foundation/literature/library/yu2023altered.md` | fe6768d | 5af6dd0cbf3e | yes | +| `literature/cards/dxd_foundation/yu2026habituation.md` | `submissions/dxd_foundation/literature/library/yu2026habituation.md` | fe6768d | cdca9876c8d0 | yes | +| `literature/cards/dxd_foundation/ziegler2010orthographic.md` | `submissions/dxd_foundation/literature/library/ziegler2010orthographic.md` | fe6768d | f946b60f9776 | yes | | `literature/cards/algorithmacy_scaffolding/bucinca2021.md` | `submissions/algorithmacy_scaffolding/literature/library/bucinca2021.md` | d0a40bd | 022ea1c70f5b | yes | | `literature/cards/algorithmacy_scaffolding/eslami2015.md` | `submissions/algorithmacy_scaffolding/literature/library/eslami2015.md` | d0a40bd | b5cda4308979 | yes | | `literature/cards/algorithmacy_scaffolding/kim2024.md` | `submissions/algorithmacy_scaffolding/literature/library/kim2024.md` | d0a40bd | aa4a1a8a79f5 | yes | diff --git a/submissions/berthon_ethics_paper/README.md b/submissions/berthon_ethics_paper/README.md index 8f26cc90..bcdc98c2 100644 --- a/submissions/berthon_ethics_paper/README.md +++ b/submissions/berthon_ethics_paper/README.md @@ -35,12 +35,13 @@ them side by side. | `consolidation/CONSOLIDATION.md` | The themed synthesis, hand-written | | `consolidation/HEADINGS.md` | All 187 headings of the eight versions, each mapped to ledger rows | | `consolidation/COVERAGE.md` | Generated by `coverage_check.py`: hard failures (none) and orphan cards | -| `literature/cards/` | 422 cards in four subfolders by origin; formats and path rules in `literature/cards/README.md` | -| `literature/CARD_INDEX.md` | One table over all 422 cards: authors, year, title, DOI, read depth, verification, flags | -| `literature/DUPLICATES.md` | The 26 works carded in more than one library, with a recommended primary | +| `literature/cards/` | 583 cards in four subfolders by origin; formats and path rules in `literature/cards/README.md` | +| `literature/CARD_INDEX.md` | One table over all 583 cards: authors, year, title, DOI, read depth, verification, flags | +| `literature/DUPLICATES.md` | The 32 works carded in more than one library, with a recommended primary | | `literature/NO_CARD.md` | The 76 works DXD Part 2 cites that have no card anywhere, with bib data and every audit row | | `literature/dxd_part2/` | DXD Part 2's literature directory (facets A–E, audits, bib), copied whole | | `literature/references.bib` | First-wins merge of the Berthon-course and Part 2 bibs (809 entries) | +| `nadella/` | The workstream on Nadella's X article and the ethics the Spaces paper bracketed: instruction, corpus index, the lab's earlier Nadella research (copied) and its 204 cards; full texts held locally | | `PROVENANCE.md` | Every copied file, its origin, origin commit, sha256, and whether it is still identical | ## How it was made @@ -67,7 +68,7 @@ the consolidation and conflicts appears in a version, a note or a card. ## Known gaps (reported, not fixed) -- 130 cards are cited by no ledger row (`COVERAGE.md`). Most are the design-ethics arm's background reading; the rest +- 291 cards are cited by no ledger row (`COVERAGE.md`). Most are the design-ethics arm's background reading; the rest are second copies of works cited through their duplicates. - The author's Google Doc "Design Ethics Paper" is his current first draft (46 accepted suggestions and 20 edits of 2026-10-13) and is not in the repo; `sources/04_AUTHOR_DRAFT.md` is the text as first supplied (CONFLICTS X09). diff --git a/submissions/berthon_ethics_paper/consolidation/CONFLICTS.md b/submissions/berthon_ethics_paper/consolidation/CONFLICTS.md index 5d027427..ee21d73b 100644 --- a/submissions/berthon_ethics_paper/consolidation/CONFLICTS.md +++ b/submissions/berthon_ethics_paper/consolidation/CONFLICTS.md @@ -24,6 +24,8 @@ the history is for and on which wrong is central: the position (04–07), the bu missing promotive design (01), or the missing conditions of learning (02). The working titles follow the theses (M08), and Vallor's question (K15) opens only 01. +**Ruled 2026-10-24, by Claude under the author's delegation** (`../nadella/README.md`, Rulings): the merge option, with Nadella's insider post as the central case. Reversible. + **Decision.** Choose one thesis; or merge 04–07 with 08, so that the five positions supply the historical spine and the three wrongs supply the present-day reading; or keep them as separate papers. Whatever governs decides which of X02–X05 matter. @@ -203,5 +205,5 @@ These are matters of dating or edition, recorded so they are not lost; none chan - Hirschman: the berthon card read the book in part; the design-ethics card holds second-hand quotations, which its verifier confirmed against the book (`bec/hirschman1970exit.md`; `ade/hirschman1970.md`). -`../literature/DUPLICATES.md` lists all 26 works carded twice. In the pairs I read (Ong, Hirschman, Selbst and Barocas, +`../literature/DUPLICATES.md` lists all 32 works carded twice. In the pairs I read (Ong, Hirschman, Selbst and Barocas, Oeldorf-Hirsch, Berthon et al.) the cards complement each other and do not disagree on substance. diff --git a/submissions/berthon_ethics_paper/consolidation/COVERAGE.md b/submissions/berthon_ethics_paper/consolidation/COVERAGE.md index 1fb31fe7..0dd9074d 100644 --- a/submissions/berthon_ethics_paper/consolidation/COVERAGE.md +++ b/submissions/berthon_ethics_paper/consolidation/COVERAGE.md @@ -16,4 +16,4 @@ Most are background the origin arms read and the versions never cited. - **algorithmacy_design_ethics**: 118 cards cited in no ledger row — aloisi2024, anannycrawford2018, armer1992, bansal2021, bennlazar2022, burrell2016, burt1992, calacci2022, chalmersgalani2004, chan2023, chen2024, christoffersenwoods2002, collins2003, corti2023, cotterreisdorf2020, cox2016, decerteau1984, dietvorst2015, disalvo2012, dogruel2022, dzindolet2003, eisenstein2002, endsleykiris1995, eu2022dga, eu2022dma, fink2024, floridi2016, floridi2020, floridicowls2019, floridisanders2004, friedman2019, frischmann2018, fukuyama2021, george2016, goody1968, goody1977, goody1986, gran2021, gray2018, green2019, hansman1995, hardt2023, hildebrandt2015, hirsch2017, hoffbashir2015, hutchins1995, ieee2019, ihde1990, inmanribes2019, jacovi2021, jennings2001, johns1998, johnston2018, journal_scan_memo, keller2021, klein2004, kudina2019, lai2023, latour1994, ledantec2009, leesee2004, leveson2011, livingstone2004, logg2019, longmagerko2020, lyons2021, macintyre1981, maes1994, micheli2020, midkiff1993, miller2019, miller2023, mittelstadt2016, mosier1998, muir1994, nissenbaum2010, norman1990, obstfeld2005, pangrazio2019, parasuraman2000, pritchett1994, rahwan2018, redaelli2022, rosenbergerverbeek2015, rudin2019, rule2019, russ2015, sae2021j3016, sarterwoods1995, schmidtsimone1996, selbstbarocas2018, shan2023glaze, shan2024nightshade, sigg2025, skitka1999, spiekermann2022, suchman1994, suchman1995, sunstein2021, taddeo2010, triadic_hf_search_notes, tsamados2022, vandepoel2013, vasconcelos2023, verbeek2008, verbeek2008b, verbeek2014, verbeek2015, verbeek2020, wachter2018, weiser1991, wellner2024, wilkinson1965, williams2018, winfield2021, winner1993, winograd1994, winogradflores1986 - **algorithmacy_scaffolding**: 1 cards cited in no ledger row — rosenblatstark2016 - **berthon_ethics_course**: 9 cards cited in no ledger row — amershi2019guidelines, connell2016not, flory2004interventions, kaminski2019right, kitzinger2024openai, lund2026explainable, rohlfing2021explanation, statute1677frauds, stollery1976statute -- **dxd_foundation**: 2 cards cited in no ledger row — bainbridge1983ironies, vasconcelos2023explanations +- **dxd_foundation**: 163 cards cited in no ledger row — alahmadi2022false, ancker2017effects, anderson1972computer, anthropic2026how, apple2026guidelines, bai2008reading, bainbridge1983ironies, baker2025monitoring, balogh1927voces, barton1980awareness, bassetti2012effects, becker2024conquering, biggs2020decoding, boretius1883capitularia, buijsman2020numerals, buncic2017factors, cameron2022expanding, caperton2009massey, cert2022common, cftc2017fxcm, chan2024visibility, chan2025infrastructure, chantavarin2025effects, charlemagne1900letter, chen2021tradeoff, chen2025reasoning, chen2026comparing, clanchy1993from, clemens2025korea, colvin2011empirical, constantin2026nadella, cressy1980literacy, crootof2023humans, cui2023eye, dacosta2018tabulae, danna2022numerical, de1972betrvg, degno2019coregistration, dionysius1910composition, douek2023meta, doyle2025assessment, dummler1881alcuini, duzan2026chain, elish2019moral, eu2019p2b, eu2022dsa, eu2024platform, eu2024regulation, evans2018forex, fca2014fines, frey2026correlational, fsb2014foreign, fsb2015progress, galanter1974why, galliussi2020inter, garrett1998saenger, gema2026chain, globalfxc2024code, green2022flaws, greenblatt2024ai, grundemclaughlin2026overseeing, haines2008origins, hakvoort2017improvements, hangeulmuseum, hao2024training, heilmann2020reading, hendrickson1929ancient, hervais2019learning, houpt2014new, ivanova2026process, jakkal2026secure, jhaver2019suspect, jia2026faithfulness, jordan2024reicher, kasisopa2016child, kellogg2020algorithms, ker2012johnson, klonick2018new, kohsom1997adding, kokotajlo2025ai, korbak2025chain, krieber2017eye, kroll2021outlining, kubota2024rethinking, kumari2026nadella, kutas2011thirty, kutasov2025shade, langer2024effective, lanham2023measuring, larkin1987why, laubrock2015eyevoice, laux2024institutionalised, layman2023controlled, leinenger2014phonological, li2014reading, little2026length, lynch2025agentic, mannering2012johnson, margono1987study, mathews1976eldridge, matzinger2021influence, mirault2019precise, mitchell2026agents, nadella2025looking, nadella2026frontier, nadella2026models, nadella2026reverse, nagarajrao2026fareshare, nothaft2020satanic, nydfs2015barclays, nydfs2015barclayspr, odonnell1994parkes, odonnell2000saenger, oomen2016last, oversight2022q1, ozkan2026wordinitial, palka2025foreign, park2025dokkyo, perea2012subtle, perea2015colours, price2011interactive, quintilian1920institutio, reis2011johnson, rfc8890, rieder2020towards, rose2020zero, ryan2014historical, saltzer1975protection, sangster2024emergence, sangster2025reboot, santonidesio2018meaningful, schlimm2008modeling, scribner1981psychology, seattle2024deactivation, seymour2003foundation, sterz2024quest, stohl2016managing, strunk1998guido, suzor2016responsibilities, suzor2018digital, thrax1874grammar, torrensalvarez2022latin, tumey1927ohio, tunbridge1992study, turan2026oversight, turpin2023language, tyler1988what, us2072usc, us2073usc, us455usc, us9usc10, vandenboer2015default, vasconcelos2023explanations, wagemans2012century, wang2024tibetan, wang2026monitorbench, wertli2023influence, westwell2018dissemination, winskel2009eye, wong2022meta, yu2023altered, yu2026habituation, ziegler2010orthographic diff --git a/submissions/berthon_ethics_paper/literature/CARD_INDEX.md b/submissions/berthon_ethics_paper/literature/CARD_INDEX.md index 7cc45ca1..ce735848 100644 --- a/submissions/berthon_ethics_paper/literature/CARD_INDEX.md +++ b/submissions/berthon_ethics_paper/literature/CARD_INDEX.md @@ -5,8 +5,8 @@ from their origin arms (see `../PROVENANCE.md`); fields are parsed from each car Basis is normalized: full, partial, abstract, secondary, adapted. Verified is the card's own field (Y format) or its S2 verdict (H format). Flags mark abstract-only, secondary, unverified cards. -Cards: 422 — algorithmacy_design_ethics 188, algorithmacy_scaffolding 6, berthon_ethics_course 193, dxd_foundation 35. -Basis: abstract 35, adapted 2, full 230, partial 130, record 2, reviews/secondary 1, secondary 22. +Cards: 583 — algorithmacy_design_ethics 188, algorithmacy_scaffolding 6, berthon_ethics_course 193, dxd_foundation 196. +Basis: abstract 35, adapted 2, full 355, partial 166, record 2, reviews/secondary 1, secondary 22. ## algorithmacy_design_ethics (188) @@ -410,42 +410,203 @@ Basis: abstract 35, adapted 2, full 230, partial 130, record 2, reviews/secondar | [young2006responsibility](cards/berthon_ethics_course/young2006responsibility.md) | Young, Iris Marion | 2006 | Responsibility and Global Justice: A Social Connection Model | 10.1017/s0265052506060043 | abstract | writer-grounded | abstract | | [zong2024data](cards/berthon_ethics_course/zong2024data.md) | Zong, Jonathan and Matias, J. Nathan | 2024 | Data Refusal from Below: A Framework for Understanding, Evaluating, and Envisioning Refusa | 10.1145/3630107 | full | true | | -## dxd_foundation (35) +## dxd_foundation (196) | card | authors | year | title | DOI | basis | verified | flags | |---|---|---|---|---|---|---|---| +| [alahmadi2022false](cards/dxd_foundation/alahmadi2022false.md) | Alahmadi, B. A., Axon, L., & Martinovic, I. | 2022 | 99% false positives: A qualitative study of SOC analysts' perspectives on security alarms | | full | S2 corrected | | | [alfrink2023contestable](cards/dxd_foundation/alfrink2023contestable.md) | Alfrink, K., Keller, I., Kortuem, G., & Doorn, N. | 2023 | Contestable AI by design: Towards a framework | 10.1007/s11023-022-09611-z | full | S2 corrected | | +| [ancker2017effects](cards/dxd_foundation/ancker2017effects.md) | Ancker, J. S., Edwards, A., Nosal, S., Hauser, D., Mauer, E. | 2017 | Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical d | 10.1186/s12911-017-0430-8 | full | S2 confirmed | | +| [anderson1972computer](cards/dxd_foundation/anderson1972computer.md) | Anderson, J. P. | 1972 | Computer security technology planning study (Vol | | full | S2 confirmed | | +| [anthropic2026how](cards/dxd_foundation/anthropic2026how.md) | Anthropic. | 2026 | How we built Claude Code auto mode: a safer way to skip permissions | | full | S2 confirmed | | +| [apple2026guidelines](cards/dxd_foundation/apple2026guidelines.md) | Apple Inc. | 2026 | App Review Guidelines [Web page; "Last Updated: June 8, 2026"] | | partial | S2 corrected | | +| [bai2008reading](cards/dxd_foundation/bai2008reading.md) | Bai, X., Yan, G., Zang, C., Liversedge, S. P., & Rayner, K. | 2008 | Reading spaced and unspaced Chinese text: Evidence from eye movements | 10.1037/0096-1523.34.5.1277 | full | S2 confirmed | | | [bainbridge1983ironies](cards/dxd_foundation/bainbridge1983ironies.md) | Bainbridge, L. | 1983 | Ironies of automation | 10.1016/0005-1098(83 | full | | | +| [baker2025monitoring](cards/dxd_foundation/baker2025monitoring.md) | Baker, B., Huizinga, J., Gao, L., Dou, Z., Guan, M. Y., Madr | 2025 | Monitoring reasoning models for misbehavior and the risks of promoting obfuscation (arXiv: | | full | S2 corrected | | +| [balogh1927voces](cards/dxd_foundation/balogh1927voces.md) | Balogh, J. | 1927 | „Voces paginarum": Beiträge zur Geschichte des lauten Lesens und Schreibens | 10.1524/phil.1927.82.14.84 | partial | | | | [bansal2019updates](cards/dxd_foundation/bansal2019updates.md) | Bansal, G., Nushi, B., Kamar, E., Weld, D. S., Lasecki, W. S | 2019 | Updates in human-AI teams: Understanding and addressing the performance/compatibility trad | 10.1609/aaai.v33i01.33012429 | full | S2 corrected | | | [bansal2021whole](cards/dxd_foundation/bansal2021whole.md) | Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., | 2021 | Does the whole exceed its parts? The effect of AI explanations on complementary team perfo | 10.1145/3411764.3445717 | full | S2 confirmed | | +| [barton1980awareness](cards/dxd_foundation/barton1980awareness.md) | Barton, D., & Hamilton, M. E. | 1980 | Awareness of the segmental structure of English in adults of various literacy levels (ERIC | | full | | | +| [bassetti2012effects](cards/dxd_foundation/bassetti2012effects.md) | Bassetti, B., & Masterson, J. | 2012 | Effects of removing morphemic information and adding interword spacing on reading in Chine | 10.1007/s11145-012-9358-7 | full | S2 corrected | | +| [becker2024conquering](cards/dxd_foundation/becker2024conquering.md) | Becker, S. O., & Won, C. | 2024 | Conquering Korea for Jesus: Protestant missionaries, local churches, and literacy in colon | 10.1016/j.labeco.2024.102618 | full | S2 confirmed | | +| [biggs2020decoding](cards/dxd_foundation/biggs2020decoding.md) | Biggs, N. | 2020 | Decoding chancery records from the 1240s | 10.1080/26375451.2020.1834796 | full | S2 corrected | | +| [boretius1883capitularia](cards/dxd_foundation/boretius1883capitularia.md) | Boretius, A. (Ed.). | 1883 | Capitularia regum Francorum (Vol | | partial | S2 corrected | | | [bucinca2021trust](cards/dxd_foundation/bucinca2021trust.md) | Buçinca, Z., Malaya, M. B., & Gajos, K. Z. | 2021 | To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assi | 10.1145/3449287 | full | S2 corrected | | +| [buijsman2020numerals](cards/dxd_foundation/buijsman2020numerals.md) | Buijsman, S. | 2020 | How numerals support new cognitive capacities | 10.1007/s11229-018-01989-7 | full | S2 confirmed | | +| [buncic2017factors](cards/dxd_foundation/buncic2017factors.md) | Bunčić, D. | 2017 | Factors influencing the success and failure of writing reforms | 10.13128/studi_slavis-21937 | full | S2 confirmed | | | [burnyeat1997postscript](cards/dxd_foundation/burnyeat1997postscript.md) | Burnyeat, M. F. | 1997 | Postscript on silent reading | 10.1093/cq/47.1.74 | partial | S2 corrected | | | [cabrera2023improving](cards/dxd_foundation/cabrera2023improving.md) | Cabrera, Á. A., Perer, A., & Hong, J. I. | 2023 | Improving human-AI collaboration with descriptions of AI behavior | 10.1145/3579612 | full | S2 confirmed | | +| [cameron2022expanding](cards/dxd_foundation/cameron2022expanding.md) | Cameron, L. D., & Rahman, H. | 2022 | Expanding the locus of resistance: Understanding the co-constitution of control and resist | 10.1287/orsc.2021.1557 | full | S2 corrected | | +| [caperton2009massey](cards/dxd_foundation/caperton2009massey.md) | *Caperton v. A.T. Massey Coal Co.*, 556 U.S. 868 | 2009 | | | full | S2 confirmed | | | [casner2014retention](cards/dxd_foundation/casner2014retention.md) | Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W | 2014 | The retention of manual flying skills in the automated cockpit | 10.1177/0018720814535628 | full | S2 corrected | | +| [cert2022common](cards/dxd_foundation/cert2022common.md) | CERT National Insider Threat Center. | 2022 | Common sense guide to mitigating insider threats (7th ed.) | | partial | S2 corrected | | +| [cftc2017fxcm](cards/dxd_foundation/cftc2017fxcm.md) | U.S. Commodity Futures Trading Commission. | 2017 | CFTC orders Forex Capital Markets, LLC (FXCM), its parent company, FXCM Holdings, LLC and | | full | S2 confirmed | | +| [chan2024visibility](cards/dxd_foundation/chan2024visibility.md) | Chan, A., Ezell, C., Kaufmann, M., Wei, K., Hammond, L., Bra | 2024 | Visibility into AI agents | 10.1145/3630106.3658948 | full | S2 corrected | | +| [chan2025infrastructure](cards/dxd_foundation/chan2025infrastructure.md) | Chan, A., Wei, K., Huang, S., Rajkumar, N., Perrier, E., Laz | 2025 | Infrastructure for AI agents [Preprint] | | full | S2 corrected | | +| [chantavarin2025effects](cards/dxd_foundation/chantavarin2025effects.md) | Chantavarin, S., & Leung, T. T.-C. | 2025 | The effects of interword spacing and morphological complexity in reading Thai: An eye-trac | 10.1017/langcog.2025.10028 | full | S2 corrected | | +| [charlemagne1900letter](cards/dxd_foundation/charlemagne1900letter.md) | Charlemagne. | 1900 | Letter to Abbot Baugulf of Fulda (De litteris colendis), c | | full | S2 corrected | | +| [chen2021tradeoff](cards/dxd_foundation/chen2021tradeoff.md) | Chen, M., Wang, Y., Zhao, B., Li, X., & Bai, X. | 2021 | The trade-off between format familiarity and word-segmentation facilitation in Chinese rea | 10.3389/fpsyg.2021.602931 | full | S2 corrected | | +| [chen2025reasoning](cards/dxd_foundation/chen2025reasoning.md) | Chen, Y., Benton, J., Radhakrishnan, A., Uesato, J., Denison | 2025 | Reasoning models don't always say what they think [Preprint] | | full | S2 confirmed | | +| [chen2026comparing](cards/dxd_foundation/chen2026comparing.md) | Chen, C., Zhang, Z., Chen, Z., Xu, E., Yang, Y., Khalilov, I | 2026 | Comparing human oversight strategies for computer-use agents [Preprint] | | full | S2 confirmed | | +| [clanchy1993from](cards/dxd_foundation/clanchy1993from.md) | Clanchy, M. T. | 1993 | From memory to written record: England 1066–1307 (2nd ed.) | | partial | S2 corrected | | +| [clemens2025korea](cards/dxd_foundation/clemens2025korea.md) | Clemens, W. C. | 2025 | Review of V | 10.1017/jea.2025.6 | full | S2 confirmed | | +| [colvin2011empirical](cards/dxd_foundation/colvin2011empirical.md) | Colvin, A. J. S. | 2011 | An empirical study of employment arbitration: Case outcomes and processes | 10.1111/j.1740-1461.2010.01200.x | full | S2 corrected | | +| [constantin2026nadella](cards/dxd_foundation/constantin2026nadella.md) | Constantin, A. M. | 2026 | Satya Nadella says companies should treat AI models as insider risks | | full | S2 confirmed | | +| [cressy1980literacy](cards/dxd_foundation/cressy1980literacy.md) | Cressy, D. | 1980 | Literacy and the social order: Reading and writing in Tudor and Stuart England | | partial | S2 corrected | | +| [crootof2023humans](cards/dxd_foundation/crootof2023humans.md) | Crootof, R., Kaminski, M. E., & Price, W. N., II. | 2023 | Humans in the loop | | full | S2 corrected | | +| [cui2023eye](cards/dxd_foundation/cui2023eye.md) | Cui, Y. | 2023 | Eye movements of second language learners when reading spaced and unspaced Chinese texts | 10.3389/fpsyg.2023.783960 | full | S2 corrected | | +| [dacosta2018tabulae](cards/dxd_foundation/dacosta2018tabulae.md) | da Costa, A. | 2018 | "That ye mowe redely fynde … what ye desyre": Printed tables of contents and indices, 1476 | 10.1353/hlq.2018.0013 | full | S2 confirmed | | +| [danna2022numerical](cards/dxd_foundation/danna2022numerical.md) | Danna, R., Iori, M., & Mina, A. | 2022 | A numerical revolution: The diffusion of practical mathematics and the growth of pre-moder | | partial | S2 corrected | | +| [de1972betrvg](cards/dxd_foundation/de1972betrvg.md) | | | Bundesministerium der Justiz. (n.d.). *Betriebsverfassungsgesetz (BetrVG)* [Works Constitu | | partial | S2 confirmed | | +| [degno2019coregistration](cards/dxd_foundation/degno2019coregistration.md) | Degno, F., Loberg, O., Zang, C., Zhang, M., Donnelly, N., & | 2019 | A co-registration investigation of inter-word spacing and parafoveal preview: Eye movement | 10.1371/journal.pone.0225819 | partial | S2 corrected | | | [delacroix2019bottom](cards/dxd_foundation/delacroix2019bottom.md) | Delacroix, S., & Lawrence, N. D. | 2019 | Bottom-up data trusts: Disturbing the 'one size fits all' approach to data governance | 10.1093/idpl/ipz014 | full | S2 corrected | | | [dietvorst2018overcoming](cards/dxd_foundation/dietvorst2018overcoming.md) | Dietvorst, B. J., Simmons, J. P., & Massey, C. | 2018 | Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slig | 10.1287/mnsc.2016.2643 | full | S2 corrected | | +| [dionysius1910composition](cards/dxd_foundation/dionysius1910composition.md) | Dionysius of Halicarnassus. | 1910 | On literary composition, being the Greek text of the De compositione verborum (W | | partial | S2 corrected | | +| [douek2023meta](cards/dxd_foundation/douek2023meta.md) | Douek, E. | 2024 | The Meta Oversight Board and the empty promise of legitimacy | 10.2139/ssrn.4565180 | full | S2 confirmed | | +| [doyle2025assessment](cards/dxd_foundation/doyle2025assessment.md) | Doyle, A., & McCrae, J. P. | 2025 | An assessment of word separation practices in Old Irish text resources and a universal met | | full | | | +| [dummler1881alcuini](cards/dxd_foundation/dummler1881alcuini.md) | Dümmler, E. (Ed.). | 1881 | Poetae Latini aevi Carolini (Vol | | partial | S2 corrected | | +| [duzan2026chain](cards/dxd_foundation/duzan2026chain.md) | Duzan, A., & Cooper Stickland, A. | 2026 | Chain-of-thought monitoring can be unreliable in implicit-influence settings [Preprint] | | partial | S2 corrected | | +| [elish2019moral](cards/dxd_foundation/elish2019moral.md) | Elish, M. C. | 2019 | Moral crumple zones: Cautionary tales in human-robot interaction | 10.17351/ests2019.260 | full | S2 corrected | | +| [eu2019p2b](cards/dxd_foundation/eu2019p2b.md) | | | Regulation (EU) 2019/1150 of the European Parliament and of the Council of 20 June 2019 on | | partial | S2 corrected | | +| [eu2022dsa](cards/dxd_foundation/eu2022dsa.md) | European Parliament and Council of the European Union. | 2022 | Regulation (EU) 2022/2065 of 19 October 2022 on a Single Market For Digital Services and a | | partial | S2 corrected | | +| [eu2024platform](cards/dxd_foundation/eu2024platform.md) | | | Directive (EU) 2024/2831 of the European Parliament and of the Council of 23 October 2024 | | partial | S2 corrected | | +| [eu2024regulation](cards/dxd_foundation/eu2024regulation.md) | European Parliament and Council of the European Union. | 2024 | Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14: Human oversight | | full | S2 confirmed | | +| [evans2018forex](cards/dxd_foundation/evans2018forex.md) | Evans, M. D. D. | 2018 | Forex trading and the WMR Fix | 10.1016/j.jbankfin.2017.09.017 | partial | S2 corrected | | +| [fca2014fines](cards/dxd_foundation/fca2014fines.md) | Financial Conduct Authority. | 2014 | FCA fines five banks £1.1 billion for FX failings and announces industry-wide remediation | | full | S2 corrected | | | [fok2024search](cards/dxd_foundation/fok2024search.md) | Fok, R., & Weld, D. S. | 2024 | In search of verifiability: Explanations rarely enable complementary performance in AI-adv | 10.1002/aaai.12182 | full | S2 confirmed | | +| [frey2026correlational](cards/dxd_foundation/frey2026correlational.md) | Frey, A., Meary, D., Loichot, M., Mirault, J., & Bosse, M.-L | 2026 | Correlational analysis of distinct contributions and overlaps between visual, visual atten | 10.1038/s41598-026-38243-x | full | S2 confirmed | | +| [fsb2014foreign](cards/dxd_foundation/fsb2014foreign.md) | Financial Stability Board. | 2014 | Foreign exchange benchmarks: Final report | | partial | S2 corrected | | +| [fsb2015progress](cards/dxd_foundation/fsb2015progress.md) | Financial Stability Board. | 2015 | Foreign exchange benchmarks: Report on progress in implementing the September 2014 recomme | | full | S2 corrected | | +| [galanter1974why](cards/dxd_foundation/galanter1974why.md) | Galanter, M. | 1974 | Why the "haves" come out ahead: Speculations on the limits of legal change | 10.2307/3053023 | partial | S2 corrected | | +| [galliussi2020inter](cards/dxd_foundation/galliussi2020inter.md) | Galliussi, J., Perondi, L., Chia, G., Gerbino, W., & Bernard | 2020 | Inter-letter spacing, inter-word spacing, and font with dyslexia-friendly features: Testin | 10.1007/s11881-020-00194-x | full | S2 corrected | | +| [garrett1998saenger](cards/dxd_foundation/garrett1998saenger.md) | Garrett, J. | 1998 | Review of P | 10.5860/crl.59.5.486 | full | S2 corrected | | | [gavrilov1997techniques](cards/dxd_foundation/gavrilov1997techniques.md) | Gavrilov, A. K. | 1997 | Techniques of reading in classical antiquity | 10.1093/cq/47.1.56 | partial | S2 confirmed | | +| [gema2026chain](cards/dxd_foundation/gema2026chain.md) | Gema, A. P., Rajani, N., Saxena, R., Kwan, W.-C., et al. | 2026 | Chain-of-thought faithfulness of reasoning models varies with where and how preference cue | | partial | S2 corrected | | +| [globalfxc2024code](cards/dxd_foundation/globalfxc2024code.md) | | | Global Foreign Exchange Committee. (n.d.; footer of the file read: "Updated December 2024" | | full | S2 confirmed | | | [gray2018dark](cards/dxd_foundation/gray2018dark.md) | Gray, C. M., Kou, Y., Battles, B., Hoggatt, J., & Toombs, A. | 2018 | The dark (patterns) side of UX design | 10.1145/3173574.3174108 | full | S2 confirmed | | +| [green2022flaws](cards/dxd_foundation/green2022flaws.md) | Green, B. | 2022 | The flaws of policies requiring human oversight of government algorithms | 10.1016/j.clsr.2022.105681 | full | S2 corrected | | +| [greenblatt2024ai](cards/dxd_foundation/greenblatt2024ai.md) | Greenblatt, R., Shlegeris, B., Sachan, K., & Roger, F. | 2024 | AI control: Improving safety despite intentional subversion | | partial | S2 corrected | | | [grinschgl2021consequences](cards/dxd_foundation/grinschgl2021consequences.md) | Grinschgl, S., Papenmeier, F., & Meyerhoff, H. S. | 2021 | Consequences of cognitive offloading: Boosting performance but diminishing memory | 10.1177/17470218211008060 | full | | | +| [grundemclaughlin2026overseeing](cards/dxd_foundation/grundemclaughlin2026overseeing.md) | Grunde-McLaughlin, M., Mozannar, H., Murad, M., Chen, J., Am | 2026 | Overseeing agents without constant oversight: Challenges and opportunities (arXiv:2602.168 | | full | S2 confirmed | | +| [haines2008origins](cards/dxd_foundation/haines2008origins.md) | Haines, J. | 2008 | The origins of the musical staff | 10.1093/musqtl/gdp002 | full | S2 confirmed | | +| [hakvoort2017improvements](cards/dxd_foundation/hakvoort2017improvements.md) | Hakvoort, B., van den Boer, M., Leenaars, T., Bos, P., & Tij | 2017 | Improvements in reading accuracy as a result of increased interletter spacing are not spec | 10.1016/j.jecp.2017.07.010 | full | S2 corrected | | +| [hangeulmuseum](cards/dxd_foundation/hangeulmuseum.md) | National Hangeul Museum. | 2015 | Read society through its Hangeul letters [Exhibition page, 21 April to 7 June 2015]. | | full | S2 corrected | | +| [hao2024training](cards/dxd_foundation/hao2024training.md) | Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J., | 2024 | Training large language models to reason in a continuous latent space (arXiv:2412.06769) | | full | S2 confirmed | | +| [heilmann2020reading](cards/dxd_foundation/heilmann2020reading.md) | Heilmann, J. | 2020 | Reading early New Testament manuscripts: Scriptio continua, "reading aids", and other char | 10.1515/9783110639247-011 | full | S2 corrected | | +| [hendrickson1929ancient](cards/dxd_foundation/hendrickson1929ancient.md) | Hendrickson, G. L. | 1929 | Ancient reading | | full | | | | [hertwig2017nudging](cards/dxd_foundation/hertwig2017nudging.md) | Hertwig, R., & Grüne-Yanoff, T. | 2017 | Nudging and boosting: Steering or empowering good decisions | 10.1177/1745691617702496 | full | S2 corrected | | +| [hervais2019learning](cards/dxd_foundation/hervais2019learning.md) | Hervais-Adelman, A., Kumar, U., Mishra, R. K., Tripathi, V. | 2019 | Learning to read recycles visual cortical networks without destruction | 10.1126/sciadv.aax0262 | full | S2 corrected | | +| [houpt2014new](cards/dxd_foundation/houpt2014new.md) | Houpt, J. W., Townsend, J. T., & Donkin, C. | 2014 | A new perspective on visual word processing efficiency | 10.1016/j.actpsy.2013.10.013 | full | S2 confirmed | | | [hutchins1985direct](cards/dxd_foundation/hutchins1985direct.md) | Hutchins, E. L., Hollan, J. D., & Norman, D. A. | 1985 | Direct manipulation interfaces | 10.1207/s15327051hci0104_2 | full | | | +| [ivanova2026process](cards/dxd_foundation/ivanova2026process.md) | Ivanova, Y. | 2026 | A process of implementing the Korean alphabet on the Korean peninsula [Bulgarian title: Пр | 10.37708/ezs.swu.bg.v24i1.9 | full | S2 confirmed | | +| [jakkal2026secure](cards/dxd_foundation/jakkal2026secure.md) | Jakkal, V. | 2026 | Secure agentic AI for your Frontier Transformation | | full | S2 confirmed | | +| [jhaver2019suspect](cards/dxd_foundation/jhaver2019suspect.md) | Jhaver, S., Appling, D. S., Gilbert, E., & Bruckman, A. | 2019 | "Did you suspect the post would be removed?": Understanding user reactions to content remo | 10.1145/3359294 | full | S2 confirmed | | +| [jia2026faithfulness](cards/dxd_foundation/jia2026faithfulness.md) | Jia, J., Benton, J., & Easley, E. | 2026 | Faithfulness as information flow: Evaluating and training faithful chain-of-thought reason | | full | S2 confirmed | | +| [jordan2024reicher](cards/dxd_foundation/jordan2024reicher.md) | Jordan, T. R., Akkaya, A. M., Göçmüş, F. Z., Kalan, A., Morg | 2024 | The Reicher-Wheeler paradigm in word recognition research: a cautionary note on its actual | 10.3389/fpsyg.2024.1399237 | full | S2 corrected | | | [kahneman2009conditions](cards/dxd_foundation/kahneman2009conditions.md) | Kahneman, D., & Klein, G. | 2009 | Conditions for intuitive expertise: A failure to disagree | 10.1037/a0016755 | full | S2 corrected | | +| [kasisopa2016child](cards/dxd_foundation/kasisopa2016child.md) | Kasisopa, B., Reilly, R. G., Luksaneeyanawin, S., & Burnham, | 2016 | Child readers' eye movements in reading Thai | 10.1016/j.visres.2015.07.009 | full | | | +| [kellogg2020algorithms](cards/dxd_foundation/kellogg2020algorithms.md) | Kellogg, K. C., Valentine, M. A., & Christin, A. | 2020 | Algorithms at work: The new contested terrain of control | 10.5465/annals.2018.0174 | full | S2 corrected | | +| [ker2012johnson](cards/dxd_foundation/ker2012johnson.md) | Ker, J. | 2012 | Review of William A | | full | S2 confirmed | | | [kim2024not](cards/dxd_foundation/kim2024not.md) | Kim, S. S. Y., Liao, Q. V., Vorvoreanu, M., Ballard, S., & W | 2024 | "I'm not sure, but...": Examining the impact of large language models' uncertainty express | 10.1145/3630106.3658941 | full | S2 confirmed | | +| [klonick2018new](cards/dxd_foundation/klonick2018new.md) | Klonick, K. | 2018 | The new governors: The people, rules, and processes governing online speech | | full | S2 corrected | | | [knox1968silent](cards/dxd_foundation/knox1968silent.md) | Knox, B. M. W. | 1968 | Silent reading in antiquity | | full | S2 corrected | | +| [kohsom1997adding](cards/dxd_foundation/kohsom1997adding.md) | Kohsom, C., & Gobet, F. | 1997 | Adding spaces to Thai and English: Effects on reading | | full | S2 confirmed | | +| [kokotajlo2025ai](cards/dxd_foundation/kokotajlo2025ai.md) | Kokotajlo, D., Alexander, S., Larsen, T., Lifland, E., & Dea | 2025 | AI 2027 [Forecasting scenario, web page] | | partial | S2 corrected | | +| [korbak2025chain](cards/dxd_foundation/korbak2025chain.md) | Korbak, T., Balesni, M., Barnes, E., Bengio, Y., Benton, J., | 2025 | Chain of thought monitorability: A new and fragile opportunity for AI safety (arXiv:2507.1 | | full | S2 confirmed | | +| [krieber2017eye](cards/dxd_foundation/krieber2017eye.md) | Krieber, M., Bartl-Pokorny, K. D., Pokorny, F. B., Zhang, D. | 2017 | Eye movements during silent and oral reading in a regular orthography: Basic characteristi | 10.1371/journal.pone.0170986 | full | | | +| [kroll2021outlining](cards/dxd_foundation/kroll2021outlining.md) | Kroll, J. A. | 2021 | Outlining traceability: A principle for operationalizing accountability in computing syste | 10.1145/3442188.3445937 | full | S2 confirmed | | +| [kubota2024rethinking](cards/dxd_foundation/kubota2024rethinking.md) | Kubota, E., Grill-Spector, K., & Nordt, M. | 2024 | Rethinking cortical recycling in ventral temporal cortex | 10.1016/j.tics.2023.09.006 | full | S2 corrected | | +| [kumari2026nadella](cards/dxd_foundation/kumari2026nadella.md) | Kumari, S. | 2026 | Satya Nadella warns of AI insider risks, seeks stronger safety controls | | full | S2 corrected | | +| [kutas2011thirty](cards/dxd_foundation/kutas2011thirty.md) | Kutas, M., & Federmeier, K. D. | 2011 | Thirty years and counting: Finding meaning in the N400 component of the event-related brai | 10.1146/annurev.psych.093008.131123 | full | S2 confirmed | | +| [kutasov2025shade](cards/dxd_foundation/kutasov2025shade.md) | Kutasov, J., Sun, Y., Colognese, P., van der Weij, T., Petri | 2025 | SHADE-Arena: Evaluating sabotage and monitoring in LLM agents (arXiv:2506.15740) | | partial | S2 corrected | | +| [langer2024effective](cards/dxd_foundation/langer2024effective.md) | Langer, M., Baum, K., & Schlicker, N. | 2025 | Effective human oversight of AI-based systems: A signal detection perspective on the detec | 10.1007/s11023-024-09701-0 | full | S2 corrected | | +| [lanham2023measuring](cards/dxd_foundation/lanham2023measuring.md) | Lanham, T., Chen, A., Radhakrishnan, A., Steiner, B., Deniso | 2023 | Measuring faithfulness in chain-of-thought reasoning (arXiv:2307.13702) | | full | S2 corrected | | +| [larkin1987why](cards/dxd_foundation/larkin1987why.md) | Larkin, J. H., & Simon, H. A. | 1987 | Why a diagram is (sometimes) worth ten thousand words | 10.1111/j.1551-6708.1987.tb00863.x | full | S2 confirmed | | +| [laubrock2015eyevoice](cards/dxd_foundation/laubrock2015eyevoice.md) | Laubrock, J., & Kliegl, R. | 2015 | The eye-voice span during reading aloud | 10.3389/fpsyg.2015.01432 | full | | | +| [laux2024institutionalised](cards/dxd_foundation/laux2024institutionalised.md) | Laux, J. | 2024 | Institutionalised distrust and human oversight of artificial intelligence: Towards a democ | 10.1007/s00146-023-01777-z | full | S2 corrected | | +| [layman2023controlled](cards/dxd_foundation/layman2023controlled.md) | Layman, L., & Roden, W. | 2023 | A controlled experiment on the impact of intrusion detection false alarm rate on analyst p | 10.1177/21695067231192573 | full | S2 confirmed | | | [lee2025impact](cards/dxd_foundation/lee2025impact.md) | Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel | 2025 | The impact of generative AI on critical thinking: Self-reported reductions in cognitive ef | 10.1145/3706598.3713778 | full | S2 corrected | | +| [leinenger2014phonological](cards/dxd_foundation/leinenger2014phonological.md) | Leinenger, M. | 2014 | Phonological coding during reading | 10.1037/a0037830 | full | | | +| [li2014reading](cards/dxd_foundation/li2014reading.md) | Li, X., Bicknell, K., Liu, P., Wei, W., & Rayner, K. | 2014 | Reading is fundamentally similar across disparate writing systems: A systematic characteri | 10.1037/a0033580 | full | | | +| [little2026length](cards/dxd_foundation/little2026length.md) | Little, B. | 2026 | Length penalties make chain-of-thought less monitorable | | partial | S2 corrected | | | [lukoff2021how](cards/dxd_foundation/lukoff2021how.md) | Lukoff, K., Lyngs, U., Zade, H., Liao, J. V., Choi, J., Fan, | 2021 | How the design of YouTube influences user sense of agency | 10.1145/3411764.3445467 | full | S2 corrected | | +| [lynch2025agentic](cards/dxd_foundation/lynch2025agentic.md) | Lynch, A., Wright, B., Larson, C., Ritchie, S. J., Minderman | 2025 | Agentic misalignment: How LLMs could be insider threats | | full | S2 confirmed | | +| [mannering2012johnson](cards/dxd_foundation/mannering2012johnson.md) | Mannering, J. | 2012 | Review of William A | | full | S2 confirmed | | +| [margono1987study](cards/dxd_foundation/margono1987study.md) | Margono, S., & Shneiderman, B. | 1987 | A study of file manipulation by novices using commands vs | | full | S2 confirmed | | +| [mathews1976eldridge](cards/dxd_foundation/mathews1976eldridge.md) | *Mathews v. Eldridge*, 424 U.S. 319 | 1976 | | | partial | S2 corrected | | | [mathur2019dark](cards/dxd_foundation/mathur2019dark.md) | Mathur, A., Acar, G., Friedman, M. J., Lucherini, E., Mayer, | 2019 | Dark patterns at scale: Findings from a crawl of 11K shopping websites | 10.1145/3359183 | full | S2 corrected | | +| [matzinger2021influence](cards/dxd_foundation/matzinger2021influence.md) | Matzinger, T., Ritt, N., & Fitch, W. T. | 2021 | The influence of different prosodic cues on word segmentation | 10.3389/fpsyg.2021.622042 | full | | | +| [mirault2019precise](cards/dxd_foundation/mirault2019precise.md) | Mirault, J., Snell, J., & Grainger, J. | 2019 | Reading without spaces: The role of precise letter order | 10.3758/s13414-018-01648-6 | full | S2 corrected | | | [mirault2019reading](cards/dxd_foundation/mirault2019reading.md) | Mirault, J., Snell, J., & Grainger, J. | 2019 | Reading without spaces revisited: The role of word identification and sentence-level const | 10.1016/j.actpsy.2019.03.001 | full | S2 corrected | | +| [mitchell2026agents](cards/dxd_foundation/mitchell2026agents.md) | Mitchell, M., Ghosh, A., & Passi, S. | 2026 | AI agents push humans out of the loop | | full | S2 confirmed | | | [mozannar2024reading](cards/dxd_foundation/mozannar2024reading.md) | Mozannar, H., Bansal, G., Fourney, A., & Horvitz, E. | 2024 | Reading between the lines: Modeling user behavior and costs in AI-assisted programming | 10.1145/3613904.3641936 | full | S2 corrected | | +| [nadella2025looking](cards/dxd_foundation/nadella2025looking.md) | Nadella, S. | 2025 | Looking ahead to 2026 | | full | S2 corrected | | +| [nadella2026frontier](cards/dxd_foundation/nadella2026frontier.md) | Nadella, S. | 2026 | A frontier without an ecosystem is not stable | | full | S2 confirmed | | +| [nadella2026models](cards/dxd_foundation/nadella2026models.md) | Nadella, S. | 2026 | Models as insider risks in the super intelligence era | | full | S2 confirmed | | +| [nadella2026reverse](cards/dxd_foundation/nadella2026reverse.md) | Nadella, S. | 2026 | The reverse information paradox | | full | S2 confirmed | | +| [nagarajrao2026fareshare](cards/dxd_foundation/nagarajrao2026fareshare.md) | Nagaraj Rao, V., Dalal, S., Schwartz, A., Liaqat, A., Calacc | 2026 | FareShare: A tool for labor organizers to estimate lost wages and contest arbitrary AI and | 10.1145/3788052 | full | S2 confirmed | | +| [nothaft2020satanic](cards/dxd_foundation/nothaft2020satanic.md) | Nothaft, C. P. E. | 2020 | Medieval Europe's satanic ciphers: On the genesis of a modern myth | 10.1080/26375451.2020.1726050 | full | S2 corrected | | +| [nydfs2015barclays](cards/dxd_foundation/nydfs2015barclays.md) | New York State Department of Financial Services. | 2015 | In the Matter of Barclays Bank PLC, Barclays Bank PLC, New York Branch: Consent Order unde | | full | S2 corrected | | +| [nydfs2015barclayspr](cards/dxd_foundation/nydfs2015barclayspr.md) | New York State Department of Financial Services. | 2015 | NYDFS announces Barclays to pay additional $150 million penalty, terminate employee for au | | full | S2 corrected | | +| [odonnell1994parkes](cards/dxd_foundation/odonnell1994parkes.md) | O'Donnell, J. J. | 1994 | Review of M | | full | S2 corrected | | +| [odonnell2000saenger](cards/dxd_foundation/odonnell2000saenger.md) | O'Donnell, J. J. | 2000 | Review of Paul Saenger, Space Between Words: The Origins of Silent Reading (Stanford Unive | | full | S2 confirmed | | | [onnasch2014human](cards/dxd_foundation/onnasch2014human.md) | Onnasch, L., Wickens, C. D., Li, H., & Manzey, D. | 2014 | Human performance consequences of stages and levels of automation: An integrated meta-anal | 10.1177/0018720813501549 | full | | | +| [oomen2016last](cards/dxd_foundation/oomen2016last.md) | Oomen, R. | 2017 | Last look | 10.1080/14697688.2016.1262545 | full | S2 confirmed | | +| [oversight2022q1](cards/dxd_foundation/oversight2022q1.md) | Oversight Board. | 2022 | Q1 2022 transparency report [Report covering 1 January–31 March 2022]. | | partial | S2 corrected | | +| [ozkan2026wordinitial](cards/dxd_foundation/ozkan2026wordinitial.md) | Özkan, Z. G., & Perea, M. | 2026 | Word-initial capitalization preserves lexical processing in unspaced text | 10.1038/s41598-026-67473-2 | full | S2 corrected | | +| [palka2025foreign](cards/dxd_foundation/palka2025foreign.md) | Pałka, P., Porębski, A., Pałosz, R., & Wiśniewska, K. | 2025 | Do foreign service providers simply ignore the EU consumer law? An empirical study of 100 | 10.1007/s10603-024-09580-z | full | S2 corrected | | | [parasuraman2010complacency](cards/dxd_foundation/parasuraman2010complacency.md) | Parasuraman, R., & Manzey, D. H. | 2010 | Complacency and bias in human use of automation: An attentional integration | 10.1177/0018720810376055 | full | S2 corrected | | +| [park2025dokkyo](cards/dxd_foundation/park2025dokkyo.md) | Park, J.-B. | 2025 | Diversification of literacy patterns by social class in late Joseon (forum abstract) | | full | S2 confirmed | | | [parkes1992pause](cards/dxd_foundation/parkes1992pause.md) | Parkes, M. B. | 1992 | Pause and effect: An introduction to the history of punctuation in the West | 10.4324/9781315247243 | partial | S2 corrected | | +| [perea2012subtle](cards/dxd_foundation/perea2012subtle.md) | Perea, M., & Gomez, P. | 2012 | Subtle increases in interletter spacing facilitate the encoding of words during normal rea | 10.1371/journal.pone.0047568 | full | S2 corrected | | +| [perea2015colours](cards/dxd_foundation/perea2015colours.md) | Perea, M., Tejero, P., & Winskel, H. | 2015 | Can colours be used to segment words when reading? Acta Psychologica, 159, 8–13 | 10.1016/j.actpsy.2015.05.005 | full | | | | [plato1925phaedrus](cards/dxd_foundation/plato1925phaedrus.md) | Plato. | 1925 | Phaedrus, 274c–275b (H | | full | S2 corrected | | +| [price2011interactive](cards/dxd_foundation/price2011interactive.md) | Price, C. J., & Devlin, J. T. | 2011 | The interactive account of ventral occipitotemporal contributions to reading | 10.1016/j.tics.2011.04.001 | full | S2 corrected | | +| [quintilian1920institutio](cards/dxd_foundation/quintilian1920institutio.md) | Quintilian. | 1920 | The orator's education [Institutio oratoria], Book 1 (H | | partial | S2 corrected | | | [rayner1998unspaced](cards/dxd_foundation/rayner1998unspaced.md) | Rayner, K., Fischer, M. H., & Pollatsek, A. | 1998 | Unspaced text interferes with both word identification and eye movement control | 10.1016/s0042-6989(97 | partial | S2 confirmed | | +| [reis2011johnson](cards/dxd_foundation/reis2011johnson.md) | Reis, D. M. | 2011 | Review of William A | 10.1111/j.1748-0922.2011.01538_4.x | full | S2 confirmed | | +| [rfc8890](cards/dxd_foundation/rfc8890.md) | Nottingham, M. | 2020 | The Internet is for end users (RFC 8890) | 10.17487/rfc8890 | full | S2 confirmed | | +| [rieder2020towards](cards/dxd_foundation/rieder2020towards.md) | Rieder, B., & Hofmann, J. | 2020 | Towards platform observability | 10.14763/2020.4.1535 | full | S2 corrected | | | [risko2016cognitive](cards/dxd_foundation/risko2016cognitive.md) | Risko, E. F., & Gilbert, S. J. | 2016 | Cognitive offloading | 10.1016/j.tics.2016.07.002 | full | S2 corrected | | +| [rose2020zero](cards/dxd_foundation/rose2020zero.md) | Rose, S., Borchert, O., Mitchell, S., & Connelly, S. | 2020 | Zero trust architecture (NIST Special Publication 800-207) | 10.6028/nist.sp.800-207 | partial | S2 corrected | | | [rosenblat2016algorithmic](cards/dxd_foundation/rosenblat2016algorithmic.md) | Rosenblat, A., & Stark, L. | 2016 | Algorithmic labor and information asymmetries: A case study of Uber's drivers | | full | S2 confirmed | | +| [ryan2014historical](cards/dxd_foundation/ryan2014historical.md) | Ryan, J. | 2014 | Historical note: Did double-entry bookkeeping contribute to economic development, specific | 10.14453/aabfj.v8i3.8 | full | S2 corrected | | | [saenger1997space](cards/dxd_foundation/saenger1997space.md) | Saenger, P. | 1997 | Space between words: The origins of silent reading | 10.1515/9781503619081 | partial | S2 corrected | | +| [saltzer1975protection](cards/dxd_foundation/saltzer1975protection.md) | Saltzer, J. H., & Schroeder, M. D. | 1975 | The protection of information in computer systems | 10.1109/proc.1975.9939 | partial | S2 corrected | | +| [sangster2024emergence](cards/dxd_foundation/sangster2024emergence.md) | Sangster, A. | 2024 | The emergence of double entry bookkeeping | 10.1111/ehr.13358 | full | S2 corrected | | +| [sangster2025reboot](cards/dxd_foundation/sangster2025reboot.md) | Sangster, A. | 2025 | Time to reboot accounting history: Evidence from scholarship on double entry | 10.1177/10323732251357606 | full | S2 confirmed | | +| [santonidesio2018meaningful](cards/dxd_foundation/santonidesio2018meaningful.md) | Santoni de Sio, F., & van den Hoven, J. | 2018 | Meaningful human control over autonomous systems: A philosophical account | 10.3389/frobt.2018.00015 | full | S2 confirmed | | +| [schlimm2008modeling](cards/dxd_foundation/schlimm2008modeling.md) | Schlimm, D., & Neth, H. | 2008 | Modeling ancient and modern arithmetic practices: Addition and multiplication with Arabic | | full | S2 confirmed | | +| [scribner1981psychology](cards/dxd_foundation/scribner1981psychology.md) | Scribner, S., & Cole, M. | 1981 | The psychology of literacy | | partial | S2 confirmed | | +| [seattle2024deactivation](cards/dxd_foundation/seattle2024deactivation.md) | Seattle Office of Labor Standards. | 2024 | Model notice of deactivation (based on requirements of SMC 8.40.070) [Form]. | | full | S2 confirmed | | +| [seymour2003foundation](cards/dxd_foundation/seymour2003foundation.md) | Seymour, P. H. K., Aro, M., & Erskine, J. M. | 2003 | Foundation literacy acquisition in European orthographies | 10.1348/000712603321661859 | full | S2 corrected | | | [shneiderman1983direct](cards/dxd_foundation/shneiderman1983direct.md) | Shneiderman, B. | 1983 | Direct manipulation: A step beyond programming languages | 10.1109/mc.1983.1654471 | full | S2 confirmed | | +| [sterz2024quest](cards/dxd_foundation/sterz2024quest.md) | Sterz, S., Baum, K., Biewer, S., Hermanns, H., Lauber-Rönsbe | 2024 | On the quest for effectiveness in human oversight: Interdisciplinary perspectives | 10.1145/3630106.3659051 | full | S2 corrected | | +| [stohl2016managing](cards/dxd_foundation/stohl2016managing.md) | Stohl, C., Stohl, M., & Leonardi, P. M. | 2016 | Managing opacity: Information visibility and the paradox of transparency in the digital ag | | full | S2 confirmed | | +| [strunk1998guido](cards/dxd_foundation/strunk1998guido.md) | | | Guido of Arezzo. Prologue to his Antiphoner; *Epistola de ignoto cantu* (excerpts). In O. | | full | S2 corrected | | +| [suzor2016responsibilities](cards/dxd_foundation/suzor2016responsibilities.md) | Suzor, N. | 2016 | The responsibilities of platforms: A new constitutionalism to promote the legitimacy of de | | full | S2 corrected | | +| [suzor2018digital](cards/dxd_foundation/suzor2018digital.md) | Suzor, N. | 2018 | Digital constitutionalism: Using the rule of law to evaluate the legitimacy of governance | 10.1177/2056305118787812 | full | S2 corrected | | +| [thrax1874grammar](cards/dxd_foundation/thrax1874grammar.md) | Dionysios Thrax [attributed]. | 1874 | The grammar of Dionysios Thrax (T | | partial | S2 corrected | | +| [torrensalvarez2022latin](cards/dxd_foundation/torrensalvarez2022latin.md) | Torrens-Álvarez, M. J., & Tuten, D. N. | 2022 | From "Latin" to the vernacular: Latin-Romance hybridity, scribal competence, and social tr | 10.1086/720547 | partial | S2 corrected | | +| [tumey1927ohio](cards/dxd_foundation/tumey1927ohio.md) | *Tumey v. Ohio*, 273 U.S. 510 | 1927 | | | full | S2 corrected | | +| [tunbridge1992study](cards/dxd_foundation/tunbridge1992study.md) | Tunbridge, G. L. | 1992 | A study of scribal practices in early Irish and Anglo-Saxon manuscripts [DPhil thesis, Uni | 10.5287/ora-ave2z6gqd | partial | S2 corrected | | +| [turan2026oversight](cards/dxd_foundation/turan2026oversight.md) | Turan, E. | 2026 | Oversight has a capacity: Calibrating agent guards to a subjective, fatiguing human | | full | S2 corrected | | +| [turpin2023language](cards/dxd_foundation/turpin2023language.md) | Turpin, M., Michael, J., Perez, E., & Bowman, S. R. | 2023 | Language models don't always say what they think: Unfaithful explanations in chain-of-thou | | partial | S2 corrected | | +| [tyler1988what](cards/dxd_foundation/tyler1988what.md) | Tyler, T. R. | 1988 | What is procedural justice?: Criteria used by citizens to assess the fairness of legal pro | 10.2307/3053563 | full | S2 corrected | | +| [us2072usc](cards/dxd_foundation/us2072usc.md) | | | 28 U.S.C. § 2072 — Rules of procedure and evidence; power to prescribe. | | full | S2 corrected | | +| [us2073usc](cards/dxd_foundation/us2073usc.md) | | | 28 U.S.C. § 2073 — Rules of procedure and evidence; method of prescribing. | | full | S2 corrected | | +| [us455usc](cards/dxd_foundation/us455usc.md) | | | 28 U.S.C. § 455 — Disqualification of justice, judge, or magistrate judge. | | full | S2 confirmed | | +| [us9usc10](cards/dxd_foundation/us9usc10.md) | | | 9 U.S.C. § 10 — Same; vacation; grounds; rehearing (Federal Arbitration Act). | | full | S2 corrected | | +| [vandenboer2015default](cards/dxd_foundation/vandenboer2015default.md) | van den Boer, M., & Hakvoort, B. E. | 2015 | Default spacing is the optimal spacing for word reading | 10.1080/17470218.2014.964272 | full | S2 confirmed | | | [vasconcelos2023explanations](cards/dxd_foundation/vasconcelos2023explanations.md) | Vasconcelos, H., Jörke, M., Grunde-McLaughlin, M., Gerstenbe | 2023 | Explanations can reduce overreliance on AI systems during decision-making | 10.1145/3579605 | full | S2 corrected | | +| [wagemans2012century](cards/dxd_foundation/wagemans2012century.md) | Wagemans, J., Elder, J. H., Kubovy, M., Palmer, S. E., Peter | 2012 | A century of Gestalt psychology in visual perception: I | 10.1037/a0029333 | partial | S2 corrected | | +| [wang2024tibetan](cards/dxd_foundation/wang2024tibetan.md) | Wang, D., Niu, D., Li, et al. | 2024 | The effect of visual word segmentation cues in Tibetan reading | 10.3390/brainsci14100964 | full | S2 corrected | | +| [wang2026monitorbench](cards/dxd_foundation/wang2026monitorbench.md) | Wang, H., Sun, Y., Ko, B., Talati, M., Gong, J., Li, Z., Yu, | 2026 | MonitorBench: A comprehensive benchmark for chain-of-thought monitorability in large langu | | partial | S2 corrected | | +| [wertli2023influence](cards/dxd_foundation/wertli2023influence.md) | Wertli, J., Schötzau, A., & Palmowski-Wolfe, A. | 2023 | The influence of age on eye movements during reading in early elementary school children | 10.1055/a-2045-7271 | full | S2 corrected | | +| [westwell2018dissemination](cards/dxd_foundation/westwell2018dissemination.md) | Westwell, A. | 2018 | The dissemination and reception of the Ordines Romani in the Carolingian church, c | 10.17863/cam.23255 | partial | S2 confirmed | | +| [winskel2009eye](cards/dxd_foundation/winskel2009eye.md) | Winskel, H., Radach, R., & Luksaneeyanawin, S. | 2009 | Eye movements when reading spaced and unspaced Thai and English: A comparison of Thai–Engl | 10.1016/j.jml.2009.07.002 | full | S2 corrected | | +| [wong2022meta](cards/dxd_foundation/wong2022meta.md) | Wong, D., & Floridi, L. | 2023 | Meta's Oversight Board: A review and critical assessment | 10.1007/s11023-022-09613-x | full | S2 confirmed | | | [wood2019good](cards/dxd_foundation/wood2019good.md) | Wood, A. J., Graham, M., Lehdonvirta, V., & Hjorth, I. | 2019 | Good gig, bad gig: Autonomy and algorithmic control in the global gig economy | 10.1177/0950017018785616 | full | | | +| [yu2023altered](cards/dxd_foundation/yu2023altered.md) | Yu, H., & Kwon, M. | 2023 | Altered eye movements during reading with simulated central and peripheral visual field de | 10.1167/iovs.64.13.21 | full | S2 confirmed | | +| [yu2026habituation](cards/dxd_foundation/yu2026habituation.md) | Yu, H., Liu, L., Jiang, X., Jia, Y., Wang, S., Qian, P., & C | 2026 | Habituation at the gate: Rising approval and declining scrutiny in human review of AI agen | | full | S2 confirmed | | +| [ziegler2010orthographic](cards/dxd_foundation/ziegler2010orthographic.md) | Ziegler, J. C., Bertrand, D., Tóth, D., Csépe, V., Reis, A., | 2010 | Orthographic depth and its impact on universal predictors of reading: A cross-language inv | 10.1177/0956797610363406 | full | S2 corrected | | diff --git a/submissions/berthon_ethics_paper/literature/DUPLICATES.md b/submissions/berthon_ethics_paper/literature/DUPLICATES.md index c2ccd217..c5d29698 100644 --- a/submissions/berthon_ethics_paper/literature/DUPLICATES.md +++ b/submissions/berthon_ethics_paper/literature/DUPLICATES.md @@ -4,7 +4,7 @@ Generated by `literature/dedup_report.py`; do not edit by hand. Each cluster is in more than one library. The primary is a recommendation by read depth; it is not a merge, and substantive disagreements between cards are recorded by hand in `../consolidation/CONFLICTS.md`. -Clusters: 26. +Clusters: 32. | # | matched by | primary | other card(s) | |---|---|---|---| @@ -15,22 +15,28 @@ Clusters: 26. | 5 | DOI, title+year | `dxd_foundation/bucinca2021trust.md` (full; S2 corrected) | `algorithmacy_scaffolding/bucinca2021.md` (full; —) | | 6 | DOI, title+year | `algorithmacy_design_ethics/delacroix2019.md` (full; S2 corrected) | `dxd_foundation/delacroix2019bottom.md` (full; S2 corrected) | | 7 | DOI, title+year | `algorithmacy_design_ethics/dietvorst2018.md` (full; S2 confirmed) | `dxd_foundation/dietvorst2018overcoming.md` (full; S2 corrected) | -| 8 | filename | `algorithmacy_design_ethics/eu2016gdpr.md` (full; S2 confirmed) | `berthon_ethics_course/eu2016gdpr.md` (partial; true) | -| 9 | filename | `algorithmacy_design_ethics/eu2022dsa.md` (full; S2 confirmed) | `berthon_ethics_course/eu2022dsa.md` (partial; true) | -| 10 | DOI, title+year | `algorithmacy_design_ethics/gray2018.md` (full; S2 confirmed) | `dxd_foundation/gray2018dark.md` (full; S2 confirmed) | -| 11 | title+year | `berthon_ethics_course/hirschman1970exit.md` (partial; true) | `algorithmacy_design_ethics/hirschman1970.md` (secondary; S2 confirmed) | -| 12 | DOI, title+year | `berthon_ethics_course/jobin2019global.md` (full; true) | `algorithmacy_design_ethics/jobin2019.md` (full; —) | -| 13 | DOI, title+year | `berthon_ethics_course/johns2002how.md` (full; true) | `algorithmacy_design_ethics/johns2002.md` (full; S2 corrected) | -| 14 | DOI, title+year | `dxd_foundation/kim2024not.md` (full; S2 confirmed) | `algorithmacy_scaffolding/kim2024.md` (full; —) | -| 15 | DOI, title+year | `berthon_ethics_course/mitchell2019model.md` (full; true) | `algorithmacy_design_ethics/mitchell2019.md` (full; S2 corrected) | -| 16 | DOI, title+year | `berthon_ethics_course/mittelstadt2019principles.md` (full; true) | `algorithmacy_design_ethics/mittelstadt2019.md` (full; —) | -| 17 | DOI | `algorithmacy_design_ethics/oeldorfhirsch2025.md` (full; S2 confirmed) | `berthon_ethics_course/oeldorf2023what.md` (partial; true) | -| 18 | title+year | `algorithmacy_design_ethics/ong1982.md` (full; S2 corrected) | `berthon_ethics_course/ong1982orality.md` (partial; corrected) | -| 19 | DOI, title+year | `algorithmacy_design_ethics/parasuramanmanzey2010.md` (full; S2 corrected) | `dxd_foundation/parasuraman2010complacency.md` (full; S2 corrected) | -| 20 | DOI, title+year | `algorithmacy_scaffolding/rahman2021.md` (adapted; —) | `berthon_ethics_course/rahman2021invisible.md` (abstract; corrected) | -| 21 | filename, title+year | `berthon_ethics_course/rosenblat2016algorithmic.md` (full; true) | `algorithmacy_scaffolding/rosenblatstark2016.md` (adapted; —); `dxd_foundation/rosenblat2016algorithmic.md` (full; S2 confirmed) | -| 22 | title+year | `berthon_ethics_course/selbst2018intuitive.md` (partial; true) | `algorithmacy_design_ethics/selbstbarocas2018.md` (abstract; S2 confirmed) | -| 23 | DOI, filename, title+year | `berthon_ethics_course/vasconcelos2023explanations.md` (full; corrected) | `algorithmacy_design_ethics/vasconcelos2023.md` (full; S2 confirmed); `dxd_foundation/vasconcelos2023explanations.md` (full; S2 corrected) | -| 24 | DOI, title+year | `berthon_ethics_course/williams2018stand.md` (partial; true) | `algorithmacy_design_ethics/williams2018.md` (partial; S2 corrected) | -| 25 | DOI, filename, title+year | `algorithmacy_design_ethics/zamfirescupereira2023.md` (full; S2 could not verify) | `algorithmacy_scaffolding/zamfirescupereira2023.md` (full; —) | -| 26 | DOI, title+year | `berthon_ethics_course/zong2024data.md` (full; true) | `algorithmacy_design_ethics/zong2024.md` (full; S2 corrected) | +| 8 | DOI, filename | `berthon_ethics_course/elish2019moral.md` (full; S2 corrected) | `dxd_foundation/elish2019moral.md` (full; S2 corrected) | +| 9 | filename | `algorithmacy_design_ethics/eu2016gdpr.md` (full; S2 confirmed) | `berthon_ethics_course/eu2016gdpr.md` (partial; true) | +| 10 | filename | `berthon_ethics_course/eu2019p2b.md` (partial; true) | `dxd_foundation/eu2019p2b.md` (partial; S2 corrected) | +| 11 | filename | `algorithmacy_design_ethics/eu2022dsa.md` (full; S2 confirmed) | `berthon_ethics_course/eu2022dsa.md` (partial; true); `dxd_foundation/eu2022dsa.md` (partial; S2 corrected) | +| 12 | DOI, title+year | `algorithmacy_design_ethics/gray2018.md` (full; S2 confirmed) | `dxd_foundation/gray2018dark.md` (full; S2 confirmed) | +| 13 | DOI, filename | `berthon_ethics_course/green2022flaws.md` (full; S2 corrected) | `dxd_foundation/green2022flaws.md` (full; S2 corrected) | +| 14 | title+year | `berthon_ethics_course/hirschman1970exit.md` (partial; true) | `algorithmacy_design_ethics/hirschman1970.md` (secondary; S2 confirmed) | +| 15 | DOI, title+year | `berthon_ethics_course/jobin2019global.md` (full; true) | `algorithmacy_design_ethics/jobin2019.md` (full; —) | +| 16 | DOI, title+year | `berthon_ethics_course/johns2002how.md` (full; true) | `algorithmacy_design_ethics/johns2002.md` (full; S2 corrected) | +| 17 | DOI, title+year | `dxd_foundation/kim2024not.md` (full; S2 confirmed) | `algorithmacy_scaffolding/kim2024.md` (full; —) | +| 18 | DOI, title+year | `berthon_ethics_course/mitchell2019model.md` (full; true) | `algorithmacy_design_ethics/mitchell2019.md` (full; S2 corrected) | +| 19 | DOI, title+year | `berthon_ethics_course/mittelstadt2019principles.md` (full; true) | `algorithmacy_design_ethics/mittelstadt2019.md` (full; —) | +| 20 | filename | `berthon_ethics_course/nadella2025looking.md` (full; S2 corrected) | `dxd_foundation/nadella2025looking.md` (full; S2 corrected) | +| 21 | filename | `berthon_ethics_course/nadella2026models.md` (full; S2 confirmed) | `dxd_foundation/nadella2026models.md` (full; S2 confirmed) | +| 22 | DOI | `algorithmacy_design_ethics/oeldorfhirsch2025.md` (full; S2 confirmed) | `berthon_ethics_course/oeldorf2023what.md` (partial; true) | +| 23 | title+year | `algorithmacy_design_ethics/ong1982.md` (full; S2 corrected) | `berthon_ethics_course/ong1982orality.md` (partial; corrected) | +| 24 | DOI, title+year | `algorithmacy_design_ethics/parasuramanmanzey2010.md` (full; S2 corrected) | `dxd_foundation/parasuraman2010complacency.md` (full; S2 corrected) | +| 25 | DOI, title+year | `algorithmacy_scaffolding/rahman2021.md` (adapted; —) | `berthon_ethics_course/rahman2021invisible.md` (abstract; corrected) | +| 26 | filename, title+year | `berthon_ethics_course/rosenblat2016algorithmic.md` (full; true) | `algorithmacy_scaffolding/rosenblatstark2016.md` (adapted; —); `dxd_foundation/rosenblat2016algorithmic.md` (full; S2 confirmed) | +| 27 | title+year | `dxd_foundation/scribner1981psychology.md` (partial; S2 confirmed) | `algorithmacy_design_ethics/scribner1981.md` (secondary; S2 confirmed) | +| 28 | title+year | `berthon_ethics_course/selbst2018intuitive.md` (partial; true) | `algorithmacy_design_ethics/selbstbarocas2018.md` (abstract; S2 confirmed) | +| 29 | DOI, filename, title+year | `berthon_ethics_course/vasconcelos2023explanations.md` (full; corrected) | `algorithmacy_design_ethics/vasconcelos2023.md` (full; S2 confirmed); `dxd_foundation/vasconcelos2023explanations.md` (full; S2 corrected) | +| 30 | DOI, title+year | `berthon_ethics_course/williams2018stand.md` (partial; true) | `algorithmacy_design_ethics/williams2018.md` (partial; S2 corrected) | +| 31 | DOI, filename, title+year | `algorithmacy_design_ethics/zamfirescupereira2023.md` (full; S2 could not verify) | `algorithmacy_scaffolding/zamfirescupereira2023.md` (full; —) | +| 32 | DOI, title+year | `berthon_ethics_course/zong2024data.md` (full; true) | `algorithmacy_design_ethics/zong2024.md` (full; S2 corrected) | diff --git a/submissions/berthon_ethics_paper/literature/build_provenance.py b/submissions/berthon_ethics_paper/literature/build_provenance.py index dd97eba6..7648d01c 100644 --- a/submissions/berthon_ethics_paper/literature/build_provenance.py +++ b/submissions/berthon_ethics_paper/literature/build_provenance.py @@ -25,7 +25,7 @@ def last_commit(path): def pairs(): meta = cl.sources() out = [] - for dest, origin in meta["sources"].items(): + for dest, origin in list(meta["sources"].items()) + list(meta.get("nadella_sources", {}).items()): out.append((dest, origin)) for name in sorted(os.listdir(os.path.join(cl.ARM, "sources", "notes"))): origin = None diff --git a/submissions/berthon_ethics_paper/literature/card_sources.json b/submissions/berthon_ethics_paper/literature/card_sources.json index b4c7b75c..20c8421e 100644 --- a/submissions/berthon_ethics_paper/literature/card_sources.json +++ b/submissions/berthon_ethics_paper/literature/card_sources.json @@ -1,9 +1,27 @@ { "cards": { - "berthon_ethics_course": {"origin": "submissions/berthon_ethics_course/literature/cards", "mode": "all", "exclude": ["REFERENCES.md"]}, - "algorithmacy_design_ethics": {"origin": "submissions/algorithmacy_design_ethics/literature/library", "mode": "all", "exclude": []}, - "dxd_foundation": {"origin": "submissions/dxd_foundation/literature/library", "mode": "matched", "exclude": []}, - "algorithmacy_scaffolding": {"origin": "submissions/algorithmacy_scaffolding/literature/library", "mode": "matched", "exclude": []} + "berthon_ethics_course": { + "origin": "submissions/berthon_ethics_course/literature/cards", + "mode": "all", + "exclude": [ + "REFERENCES.md" + ] + }, + "algorithmacy_design_ethics": { + "origin": "submissions/algorithmacy_design_ethics/literature/library", + "mode": "all", + "exclude": [] + }, + "dxd_foundation": { + "origin": "submissions/dxd_foundation/literature/library", + "mode": "matched", + "exclude": [] + }, + "algorithmacy_scaffolding": { + "origin": "submissions/algorithmacy_scaffolding/literature/library", + "mode": "matched", + "exclude": [] + } }, "dirs": { "literature/dxd_part2": "submissions/decision_experience_design/part2_ethics/literature" @@ -26,5 +44,56 @@ "berthon_": "submissions/berthon_ethics_course/", "ade_": "submissions/algorithmacy_design_ethics/", "part2_": "submissions/decision_experience_design/part2_ethics/" + }, + "nadella_sources": { + "nadella/sources/dxf_ABSTRACT_v4.md": "submissions/dxd_foundation/ABSTRACT_v4.md", + "nadella/sources/dxf_ARGUMENT_MAP.md": "submissions/dxd_foundation/ARGUMENT_MAP.md", + "nadella/sources/dxf_CLAIM.md": "submissions/dxd_foundation/CLAIM.md", + "nadella/sources/dxf_DRAFT.md": "submissions/dxd_foundation/DRAFT.md", + "nadella/sources/dxf_DRAFT_v2.md": "submissions/dxd_foundation/DRAFT_v2.md", + "nadella/sources/dxf_DRAFT_v3.md": "submissions/dxd_foundation/DRAFT_v3.md", + "nadella/sources/dxf_DRAFT_v4.md": "submissions/dxd_foundation/DRAFT_v4.md", + "nadella/sources/dxf_OUTLINE.md": "submissions/dxd_foundation/OUTLINE.md", + "nadella/sources/dxf_README.md": "submissions/dxd_foundation/README.md", + "nadella/sources/dxf_RESEARCH_PLAN.md": "submissions/dxd_foundation/RESEARCH_PLAN.md", + "nadella/sources/dxf_REVIEW_v2_2026-10-13.md": "submissions/dxd_foundation/REVIEW_v2_2026-10-13.md", + "nadella/sources/dxf_SOURCE.md": "submissions/dxd_foundation/SOURCE.md", + "nadella/sources/dxf_facet_E_human_factors_frame.md": "submissions/dxd_foundation/literature/facets/E_human_factors_frame.md", + "nadella/sources/dxf_facet_G_nadella_lineage.md": "submissions/dxd_foundation/literature/facets/G_nadella_lineage.md", + "nadella/sources/dxf_facet_H_candidate_moves.md": "submissions/dxd_foundation/literature/facets/H_candidate_moves.md", + "nadella/sources/dxf_facet_L_disclosure_and_choice.md": "submissions/dxd_foundation/literature/facets/L_disclosure_and_choice.md", + "nadella/sources/dxf_facet_S_hmi_read_to_check.md": "submissions/dxd_foundation/literature/facets/S_hmi_read_to_check.md", + "nadella/sources/dxf_facet_T_adjudication.md": "submissions/dxd_foundation/literature/facets/T_adjudication.md", + "nadella/sources/dxf_facet_X_counter_case.md": "submissions/dxd_foundation/literature/facets/X_counter_case.md", + "nadella/sources/dxf_facet_Z_gap_search.md": "submissions/dxd_foundation/literature/facets/Z_gap_search.md", + "nadella/sources/dxf_facet_G_nadella_lineage.bib": "submissions/dxd_foundation/literature/facets/G_nadella_lineage.bib", + "nadella/sources/dxf_facet_A_the_record.md": "submissions/dxd_foundation/literature/facets/A_the_record.md", + "nadella/sources/dxf_facet_A_the_record.bib": "submissions/dxd_foundation/literature/facets/A_the_record.bib", + "nadella/sources/dxf_facet_B_the_dispute.md": "submissions/dxd_foundation/literature/facets/B_the_dispute.md", + "nadella/sources/dxf_facet_B_the_dispute.bib": "submissions/dxd_foundation/literature/facets/B_the_dispute.bib", + "nadella/sources/dxf_facet_C_spacing_to_literacy.md": "submissions/dxd_foundation/literature/facets/C_spacing_to_literacy.md", + "nadella/sources/dxf_facet_C_spacing_to_literacy.bib": "submissions/dxd_foundation/literature/facets/C_spacing_to_literacy.bib", + "nadella/sources/dxf_facet_D_reading_science.md": "submissions/dxd_foundation/literature/facets/D_reading_science.md", + "nadella/sources/dxf_facet_D_reading_science.bib": "submissions/dxd_foundation/literature/facets/D_reading_science.bib", + "nadella/sources/dxf_facet_F_parallel_moves.md": "submissions/dxd_foundation/literature/facets/F_parallel_moves.md", + "nadella/sources/dxf_facet_F_parallel_moves.bib": "submissions/dxd_foundation/literature/facets/F_parallel_moves.bib", + "nadella/sources/dxf_facet_J_who_holds_the_convention.md": "submissions/dxd_foundation/literature/facets/J_who_holds_the_convention.md", + "nadella/sources/dxf_facet_J_who_holds_the_convention.bib": "submissions/dxd_foundation/literature/facets/J_who_holds_the_convention.bib", + "nadella/sources/dxf_facet_K_training_and_cause.md": "submissions/dxd_foundation/literature/facets/K_training_and_cause.md", + "nadella/sources/dxf_facet_K_training_and_cause.bib": "submissions/dxd_foundation/literature/facets/K_training_and_cause.bib", + "nadella/sources/dxf_facet_O_custody_of_conventions.md": "submissions/dxd_foundation/literature/facets/O_custody_of_conventions.md", + "nadella/sources/dxf_facet_O_custody_of_conventions.bib": "submissions/dxd_foundation/literature/facets/O_custody_of_conventions.bib", + "nadella/sources/dxf_facet_W_spacing_oracy_literacy.md": "submissions/dxd_foundation/literature/facets/W_spacing_oracy_literacy.md", + "nadella/sources/dxf_facet_W_spacing_oracy_literacy.bib": "submissions/dxd_foundation/literature/facets/W_spacing_oracy_literacy.bib", + "nadella/sources/bec_facet_B_print_1450_1800.md": "submissions/berthon_ethics_course/literature/facets/B_print_1450_1800.md", + "nadella/sources/bec_facet_B_print_1450_1800.bib": "submissions/berthon_ethics_course/literature/facets/B_print_1450_1800.bib", + "nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.md": "submissions/berthon_ethics_course/literature/facets/V1_nadella_disclosure_paradigm.md", + "nadella/sources/bec_facet_V2_oracy_to_literacy.md": "submissions/berthon_ethics_course/literature/facets/V2_oracy_to_literacy.md", + "nadella/sources/bec_facet_V4_disclosure_and_burden.md": "submissions/berthon_ethics_course/literature/facets/V4_disclosure_and_burden.md", + "nadella/sources/bec_facet_V5_scaffolding.md": "submissions/berthon_ethics_course/literature/facets/V5_scaffolding.md", + "nadella/sources/bec_facet_V6_prior_work_and_objections.md": "submissions/berthon_ethics_course/literature/facets/V6_prior_work_and_objections.md", + "nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.bib": "submissions/berthon_ethics_course/literature/facets/V1_nadella_disclosure_paradigm.bib", + "nadella/sources/bec_v4_RESEARCH_BRIEF.md": "submissions/berthon_ethics_course/v4/RESEARCH_BRIEF.md", + "nadella/sources/bec_v4_V4_CHECK_BRIEF.md": "submissions/berthon_ethics_course/v4/V4_CHECK_BRIEF.md" } } diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/alahmadi2022false.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/alahmadi2022false.md new file mode 100644 index 00000000..6b4e0b42 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/alahmadi2022false.md @@ -0,0 +1,47 @@ +# Alahmadi, B. A., Axon, L., & Martinovic, I. (2022). 99% false positives: A qualitative study of SOC analysts' perspectives on security alarms. In *Proceedings of the 31st USENIX Security Symposium* (pp. 2783–2800). USENIX Association. + +**Identifier:** https://www.usenix.org/conference/usenixsecurity22/presentation/alahmadi (no DOI; ISBN 978-1-939133-31-1 printed on the title page) · **Read depth:** full_text (published open-access proceedings PDF, 19 PDF pages: cover plus printed pages 2783–2800 including references and appendices; locators below are the printed USENIX page numbers) · **Source-tier:** proceedings (peer-reviewed conference paper) · **Evidence basis:** direct_read · **Design and sample:** online survey of 20 SOC practitioners, then 21 semi-structured interviews with practitioners from seven SOCs (five managed-service providers, two in-house), thematic Template Analysis of 105,523 transcribed words; non-random recruitment, mostly UK · **Parties modeled:** person–system (analyst and detection tool), with the customer as a third party in the managed-service SOCs · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +Alahmadi, Axon and Martinovic ask how analysts in security operations centres (SOCs) judge alarms and why they call so many of them false positives. The survey (n = 20) was a scoping step; the authors say they "cannot claim statistical significance" (p. 2786). Of the 20 respondents, 14 agreed and 2 disagreed that their tools "frequently produce false positive results" (assertion A-1), 12 agreed and 4 disagreed that alert counts are overwhelming (A-4), and 18 agreed and 2 disagreed that a human in the loop is needed because automated systems cannot do the detection and preliminary analysis alone (A-5) (Table 3, p. 2787). Forty-five percent received under 5,000 alerts a day; two respondents used machine-learning tools (p. 2788). + +The interviews carry the argument. One lead analyst said "We know 99% of the alarms we generate are false positives, but we still have to look at them" (B3, p. 2790). The authors then split "false positive" into two kinds. A false alarm fires with no security event behind it. A benign trigger is a true match on a signature that the organisation chooses to ignore for a business reason, for example outdated Java versions across a client estate (pp. 2790–2791). They conclude the 99% "is found to be mostly benign triggers and not necessarily a measurement of the performance of the technology itself" (p. 2791). + +Validation stays human work. Analysts decide whether an alarm is real from their knowledge of the network, the client's business and its working hours, and they also configure the rules and tune the baselines (pp. 2788–2790). Interviewees call commercial alarms "black-box" and distrust alarms that carry no reason: "if you don't tell me the reason you fire ... I can't have really high respect on that" (C5, p. 2792). Pressure runs toward over-reporting: managed-service analysts "are afraid not to raise" alarms to the customer, and they cannot act on the network themselves (D6, B4, p. 2789). The authors derive five properties for usable alarms (Reliable, Explainable, Analytical, Contextual, Transferable; the REACT model, p. 2794), and say analysts "should not trust the system completely" but need trust "calibrated using properly designed prediction explanations" (p. 2794). + +The Target breach opens the paper: the malware alarm "was picked up by Target's Security Operation Center (SOC) 1 in Bangalore", escalated to Minneapolis, "ignored", and no action followed (pp. 2783–2784). The authors add that Target's staff had disabled the tool's ability to delete the file automatically (p. 2784). + +## Relation to the argument + +Premise P12 (people perform sustained monitoring and whole-volume review badly) and Nadella-lineage facet G, which lists this paper with Ancker et al. as alert-fatigue evidence. The paper supports the first half of that claim and not the second. + +It supports the load. Analysts who must read alarms of mostly benign origin, and who are rated on tickets closed (a point the authors take from Sundaramurthy et al., p. 2784), describe the work as tedious. That fits the paper's worry about residual human tasks in the harness around a model. + +It does not measure performance. The paper reports no accuracy, miss rate or vigilance decrement. "Burnout" and "desensitization" enter through the authors' framing and a citation to a clinical alarm-fatigue review (pp. 2783–2784, reference [12]); the interviews do not test them. The respondents also rate the human role as indispensable and better than the SIEM at connecting patterns (A-5, A-6; p. 2789). A reader of this card should not cite Alahmadi for "people do this badly"; the fair claim is "people do this at great volume, with unmeasured accuracy, and say they cannot be replaced". + +Two details bear on the rulings. First, D3 and D4 (disclosure and transparency as literacy-based interventions): the analysts' complaint about black-box alarms is a demand for disclosed reasons, and the authors answer with explainable-AI design plus calibrated trust (p. 2794). That is a recommendation, not a test, and it sits on the literacy side of D3. Second, P11 (the person holds the brake): the Target account shows a human-held brake that was withheld by staff because false alarms had eroded trust in automatic action (p. 2784). + +Facet check: facet G's summary matches the text (20 survey, 21 interviews, the 99% quote, benign triggers, Target, the abstract's "Alarm validation is a tedious task..."). The facet's own caution that 99% is "one analyst's figure" is correct: B3's statement; the survey's ten disclosed ratios were 1 real alert in 100 and 50 in 200 for two respondents, not a measured rate (p. 2788). I found no contradiction, but the facet's placement of the paper under "approval fatigue" evidence is looser than the text, which studies alarm validation, not approval. + +## In conversation with + +None of the paper's 48 references is in `../references.bib`. It cites Kokulu et al. 2019 (SOC issues; the authors contradict its finding that analysts do not see false positives as a problem, pp. 2784–2785), Sundaramurthy et al. 2014–2017 (SOC ethnography, burnout), Cvach 2012 (clinical alarm fatigue), Crowley and Pescatore 2019 (SANS SOC survey) and Goodall et al. 2004 (intrusion-detection work). It was written before the bib entries on alert fatigue, so it neither cites nor answers them: + +- `ancker2017effects`, `layman2023controlled`, `tariq2025alert`: later or independent work on alert fatigue in clinical and SOC settings. The relation is topical; Alahmadi does not cite any of the three. + +Leads from its list, not in the bib: Kokulu et al. (CCS 2019); Sundaramurthy et al., "A human capital model for mitigating security analyst burnout" (SOUPS 2015); Cvach, "Monitor alarm fatigue: an integrative review" (2012); Goodall et al., "The work of intrusion detection" (2004). + +## Caution + +Self-reported views of 21 practitioners from seven SOCs, mostly UK, recruited through institutional contacts and not at random; the authors say these limits prevent "generalizable conclusions" (p. 2795). The survey is unweighted and scoping only. No outcome data on detection or error. The "99%" is a quotation, not a rate. The paper concerns people reviewing rule-based alarms from security tools, not people supervising an AI agent's actions; any carry-over to agent oversight is the project's inference. The 2022 text predates the 2023–2026 alert-review and agent-oversight studies in facet G. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the USENIX PDF (title, authors, 2022, pp. 2783-2800 match); extracted layout and reading-order text. Checked all 16 quotations and about 25 numbers (Table 3 ratios A-1 2:14, A-4 4:12, A-5 2:18; 45% under 5K alerts; 105,523 words; 21 interviewees, seven SOCs; n=10 disclosed alert ratios; REACT properties; Target/Bangalore/Minneapolis; disabled auto-delete). + +**Findings:** One quotation was not verbatim: the Target sentence omitted the source's '(SOC) 1' footnote marker; restored the exact words. Everything else matches. Page locators were checked against the printed page markers where the text supplied them; the p. 2788 locator for the n=10 ratios is plausible but the page break falls close by. Two bib keys named as topical only (layman2023controlled, tariq2025alert) were not opened; the claim that Alahmadi does not cite them holds against its reference list. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ancker2017effects.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ancker2017effects.md new file mode 100644 index 00000000..aaea70e5 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ancker2017effects.md @@ -0,0 +1,40 @@ +# Ancker, J. S., Edwards, A., Nosal, S., Hauser, D., Mauer, E., Kaushal, R., & the HITEC Investigators. (2017). Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system. *BMC Medical Informatics and Decision Making, 17*, Article 36. + +**Identifier:** doi:10.1186/s12911-017-0430-8 · **Read depth:** full_text (published open-access PDF; locators are the PDF's "Page n of 9"; reference list skimmed) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** retrospective cohort from electronic health record data, Jan 2010 to Jun 2013; 112 primary-care clinicians (93 physicians, 19 nurse practitioners) in a New York network of health centres; 1,266,325 best-practice advisories (BPAs, clinical reminders) and 326,203 drug alerts, from 430,803 encounters with 99,649 patients (p. 4) · **Parties modeled:** person–system (a clinician and the alerts an electronic record fires at them) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +Overrides are high ("49–96% of alerts are overridden," p. 1) and usually blamed on "alert fatigue," a term the authors call poorly defined. They test two readings. In cognitive overload, acceptance falls with more work, more complex work and more uninformative alerts. In desensitization, a new alert is accepted at first and then ignored (p. 2). + +For overload they build unvalidated clinician-level markers: patients and encounters per year (amount of work); alerts per encounter and a comorbidity score (complexity); and the share of alerts repeated for the same patient within a year (a proxy for low information). They report incident rate ratios (IRR) from negative binomial models. For desensitization they follow 6 newly launched BPAs, with monthly acceptance for 46 clinicians (p. 3). + +Results (pp. 4–6). Clinicians accepted a median 19.4% of BPAs and a median 0.0% of drug alerts (Table 1). About a quarter of BPAs (median 26.2%) and a third of drug alerts (median 31.8%) were within-patient repeats. For BPAs, acceptance fell 30% with each additional BPA per encounter (IRR 0.70, p<.001) and 10% with each five-point rise in the repeat share (IRR 0.90, p<.001). Patients and encounters per year showed no association (IRR 1.002 per 100 patients; 0.999 per 100 encounters). Comorbidity narrowly lost significance in the full model (IRR 0.49, p=.06). Drug-alert results were similar but weaker, from floor effects. After an override of a repeated alert's first instance, later ones were overridden 87.9% of the time for BPAs and 99.9% for drug alerts (p. 6). Only one of the 6 new BPAs, a flu reminder, showed an early peak and drop, probably from seasonality. + +The authors conclude there was "no evidence of an effect of workload per se, or of desensitization over time for a newly deployed alert" (p. 1) and suggest cutting repeats. They add that "we cannot conclude definitively that reducing override rates would reduce rate of inappropriate overrides," since many overrides are appropriate (p. 8). + +## Relation to the argument + +The source sits under premise P12 (people monitor badly), where `ARGUMENT_MAP.md` lists "alert-review studies (Ancker; Alahmadi)." It shows one robust pattern: with more alerts per encounter and more repeats, acceptance drops. It touches P9 as well, since an alert is something handed to a person to read and accept rates are low. + +It supports P12 less than the map implies. The measure is acceptance, and the authors cannot say which overrides were wrong. The setting is point-of-care alerts that interrupt a clinician's own task. The map itself says the evidence "does not cover after-the-fact reconstruction by trained auditors," which is what P15 of the Nadella post names, and Ancker does not either. + +Facet check: facet G summarises the study as showing alert fatigue "is therefore more about repeated and uninformative alerts than raw volume." That overstates two things. The number of BPAs per encounter was a strong predictor (IRR 0.70), so alert volume per encounter did matter; what showed no effect was overall workload, measured by patients and encounters. And "uninformative" is a proxy: the authors measured repetition, not informativeness, and say so (p. 3). The facet's quotations from the abstract are accurate. + +## In conversation with + +- No work in `references.bib` is cited. The treatment of repeats as false alarms rests on human-factors refs [16–18] (p. 2): Rayo & Moffatt-Bruce on alarm management, Dixon, Wickens & McCarley on automation false alarms, and Endsley & Jones. The bib's `endsley1995out` is a different work. +- Leads not in the bib: van der Sijs et al. (ref. [7], the systematic review of override rates; ref. [15], the laboratory study); Embi & Leonard (ref. [19]) on alert response over time; Baseman et al. (ref. [22]) on message volume and recall; the companion papers from the same cohort (refs [29, 30]). + +## Caution + +An observational study in one network with one commercial record system and interruptive alerts; the authors say this may limit generality (p. 8). Acceptance is a behaviour, not a quality outcome. The markers are new and unvalidated, and the repeat proxy is an assumption. Median drug-alert acceptance was 0.0%, so those findings sit at a floor. Not evidence on log review, auditors, or language-model agents. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the published open-access PDF (9 pp., BMC Med Inform Decis Mak 17:36); page-wise extraction. Checked 6 quotations with pages and about 30 numbers (sample sizes 112/93/19, 1,266,325 BPAs, 326,203 drug alerts, 430,803 encounters, 99,649 patients; medians 19.4%, 26.2%, 31.8%; IRRs 0.70, 0.90, 1.002, 0.999, 0.49; 87.9% and 99.9%; 6 new BPAs, 46 clinicians; flu reminder), and the named references [16], [19], [22]. + +**Findings:** No issues found. The card's IRR units (per 100 patients, per 100 encounters) match the Table 2 footnotes. The 'Facet check' reading of the facet G summary is about project files, not the source, so I did not test it against the source. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/anderson1972computer.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/anderson1972computer.md new file mode 100644 index 00000000..0e6815f5 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/anderson1972computer.md @@ -0,0 +1,45 @@ +# Anderson, J. P. (1972). *Computer security technology planning study* (Vol. I, ESD-TR-73-51). Electronic Systems Division, Air Force Systems Command, Hanscom Field. + +**Identifier:** https://csrc.nist.gov/csrc/media/publications/conference-paper/1998/10/08/proceedings-of-the-21st-nissc-1998/documents/early-cs-papers/ande72a.pdf · **Read depth:** full_text of Volume I, the "Executive Summary" (NIST CSRC text copy, 43 PDF pages: front matter i–v, report pages 1–32, and the DD Form 1473 data sheet; locators are the report's printed page numbers). Volume II, which holds the appendices and the rationale, is not in this file and was not read · **Source-tier:** report (government technical report written by an expert panel under Air Force contract F19628-72-C-0198) · **Evidence basis:** direct_read · **Design and sample:** none; a planning study by a ten-member panel that met six times between 2 February and 1 September 1972 (Foreword, Preface) · **Parties modeled:** other (multi-user computer system: a program acting for a user, and the operating system mediating its references to data) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +The Air Force wanted several users with different clearances to share one computer. Anderson's panel states the problem as "the threat from a malicious user" (Preface, p. iv) and asks for a research and development programme to meet it. It argues that the usual defences, which wall off the machine from outsiders, do not work once the user population is mixed, because a malicious user "has legitimate access to a system" (§3.1, p. 8). The defence therefore "must reside in the process that controls the operation and execution of arbitrary programs", and its principal requirement is to "control and limit the references a program can make to other programs and data to just those authorized (by an external authority) for the user on whose behalf the program is executed" (p. 8). + +The panel's answer is the reference monitor, a mechanism "which enforces the authorized access relationships between subjects and objects of a system" (§3.2.1, p. 9). Its implementation, the reference validation mechanism, checks each reference against a list of authorised types of reference, pictured as an access matrix (Figures 1–2, pp. 9–10). Three design requirements follow: the mechanism "must be tamper proof", it "must always be invoked", and it "must be small enough to be subject to analysis and tests, the completeness of which can be assured" (pp. 9–10). The panel explains why the third matters: since the mechanism "is the security mechanism in the system, it must be possible to ascertain that it works correctly in all cases" (p. 10). Section 3.5 applies the test to a commercial system and finds it fails two requirements: the supervisor runs with bounds checking suspended, so validation is not always invoked, and the supervisor with its support software, at 200,000 to 300,000 instructions, is too large to analyse (p. 12). + +The report's diagnosis of why systems fail is a claim about design intent. Their designers "never considered other than a benign environment where seeming violations of the system are presumed to be accidental", so existing controls "are there to contain accidents or errors" (pp. 2–3). Add-on security and patching are "futile" (§4.3.2, p. 21); security "must be designed in at its inception" (Conclusions, p. 32). Tiger-team penetration tests prove nothing about security: they "can only reveal system flaws" (p. 4). Cost estimates support the point: a flaw in a contemporary system costs a person about one month, under $3,000, to find and exploit; a system "repaired" by over 250 changes (10–15 person-years over six months) fell to a second penetration in under one person-week (§5.2, p. 30). The report also warns that installations accept operating-system updates "without question", which leaves the door open to a trap door planted by the supplier (§4.3.1, p. 19). + +The monitor has a stated edge. It "is predicated upon positive identification (authentication) of all users at all times", physical security and protected communications, and "these areas are not included" (§3.7, p. 13). Accidental spillage and similar threats are called serious but solvable "by well understood techniques" (p. 2). The report does not mention audit, logging, separation of duties or least privilege by those names. The nearest it comes is the access matrix, in which "privilege" is READ, WRITE or EXECUTE (p. 9), and "security surveillance techniques" as one line in the exploratory plan (p. 24). + +## Relation to the argument + +Facet G places this report first in the security lineage behind the Nadella insider post (premise P11): authority and load-bearing controls sit outside the thing being controlled. Anderson gives the clean statement of that idea. The controls must live in a mechanism the controlled program cannot subvert, because the program is assumed able to try. + +Three limits on that use, all from the text. First, Anderson assumes a monitor that is small enough to verify (p. 10). A harness around a frontier model is not small, and the report treats size as the thing that makes certification possible. Second, the monitor mediates references against a list written "by an external authority" (p. 8); it does not read evidence or exercise judgment, and no human reader is part of the design. Third, "tamper proof" attaches to the enforcement mechanism (p. 9), and §3.3 softens it to "tamper resistant" in the certification step (p. 11). + +On the paper's two tests (ARGUMENT_MAP, "Where the argument is exposed", item 1: what task the person keeps, whom the move serves): the reference monitor moves work from a person to an artefact, as ARGUMENT_MAP notes in the dropped-claims table. In Anderson's design the person who loses work is the system's operator or the cleared user, and the party the monitor serves is the owner of the classified data, who also writes the authorisation list. That is the owner-serves-owner arrangement the paper contrasts with the scribes' case under P10. The report does not touch a mediator with its own objective, or a counterparty that is not the controlling party. + +Facet check: I find the facet's account accurate. Its quotations from §3.1, §3.2, §3.2.2 and §3.7 match the text and pages. The facet's "read in part" is now a full read of Volume I. One point to add: the facet cites section numbers that the table of contents gives differently from the body (see Caution). + +## In conversation with + +Nothing in this report is in `../references.bib`, and the report has no reference list. Its text cites, by name, only Graham and Denning (1972, "Protection–Principles and Practice", the source of Figure 2, p. 10), the RAND report R-609 of the Defense Science Board task force (p. 4), the CCIP-85 study (p. 3) and ARPA-supported work (p. 4). It names Multics as an example of a descriptor-driven system (p. 12). + +- `saltzer1975protection`: the bib's later statement of design principles (least privilege, complete mediation). Anderson (1972) cannot cite it, and does not use those terms, so the relation runs only forward in time. Facet G describes Saltzer and Schroeder as the next step in this lineage; that link is not in Anderson's text. + +Leads, not in the bib: Graham and Denning (1972); Ware, *Security Controls for Computer Systems* (RAND R-609, 1970); Anderson's Volume II (appendices); Schell's later work on the security kernel. + +## Caution + +A planning document by a panel, written to secure funding for a prototype programme; it argues for a research agenda and gives cost figures (for example, an $8.0 million prototype programme, p. 24) rather than test results. The penetration costs are estimates "based on current experience", not measurements (pp. 29–30). Volume I is a summary: the arguments for the reference-monitor requirements, the penetration appendix and the rationale are in Volume II, which this card does not cover. Section numbering in the NIST copy is inconsistent: the table of contents lists §1.3 as "Penetration of Systems" and §1.4 as "Inadequacy of Contemporary System Designs", but the body prints them the other way; the contents list §3.1 as "Emerging Design Principles" and §3.2.2 as "Operating Requirements", while the body headings read "Requirements For Defense Against a Malicious User" and "Design Requirements". Cite by page. The scan has OCR noise, and the quotations here were checked against the text layer by eye. The report predates networked, adaptive, learning systems; it models a program that is hostile, not one that is merely opaque or misaligned. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the NIST CSRC text-layer PDF of Volume I (ESD-TR-73-51; panel list of ten, six panel meetings, 2 Feb-1 Sep 1972). Checked about 26 quotations and the cost figures against the OCR text; derived the page of each passage from the footer page numbers (footers close each page) and found the card's locators consistent: pp. 2-3, 4, 8, 9-10, 11, 12, 13, 19, 21, 24, 30, 32. Also checked: 200,000-300,000 instructions, one man-month and under $3,000, 250 changes and 10-15 man-years over six months, second penetration under one man-week, $8.0 million total (p. 24), the ten-member panel, Graham and Denning footnote (p. 10), Multics (p. 12), CCIP-85, R-609, the swapped section numbers in the contents list. + +**Findings:** No issues found. The text layer has OCR noise (hyphenation, 'roan years' for man-years), so I matched quotations by hand to the nearest words; none differed in substance. Two items the card asserts as absence (no mention of audit, logging, separation of duties or least privilege; no reference list) were not exhaustively searched beyond text searches for 'privilege' (found only as READ/WRITE/EXECUTE) and a scan of the contents. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/anthropic2026how.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/anthropic2026how.md new file mode 100644 index 00000000..0b8cf562 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/anthropic2026how.md @@ -0,0 +1,27 @@ +# Anthropic. (2026, March 25). How we built Claude Code auto mode: a safer way to skip permissions. *Engineering at Anthropic*. https://www.anthropic.com/engineering/claude-code-auto-mode + +**Identifier:** URL above · **Read depth:** full_text (saved web page; no pagination, locators are section headings and Table 1; author given as John Hughes) · **Source-tier:** vendor engineering report (not peer reviewed) · **Evidence basis:** direct_read · **Design and sample:** classifier evaluation on three datasets: real internal traffic (n = 10,000 tool calls from Anthropic employees), real overeager actions (n = 52) and synthetic exfiltration tasks (n = 1,000) · **Parties modeled:** user, agent and a classifier acting as approver · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +"Claude Code users approve 93% of permission prompts" (opening). The post names "approval fatigue, where people stop paying close attention to what they're approving". Auto mode replaces the human approver with a model: a transcript classifier "acting as a substitute for a human approver" (section "How it works") on Sonnet 4.6, plus a server-side prompt-injection probe on tool outputs. The classifier sees only user messages and tool calls, "reasoning-blind by design" (so the agent cannot talk it round; section "Why we strip assistant text"). Its threat model has four causes: overeager behaviour, honest mistakes, prompt injection and a misaligned model, the last one the authors say they do not currently see in practice. Actions go through three tiers; only those with "real downside potential" reach the classifier. + +Table 1: on real traffic (n = 10,000) stage 1 has an 8.5% false-positive rate and the full two-stage pipeline 0.4%. On real overeager actions (n = 52) the false-negative rate is 6.6% at stage 1 and 17% for the full pipeline. On synthetic exfiltration (n = 1,000) it is 1.8% and 5.7%. The authors call 17% "the honest number": misses mostly find the danger but misjudge "whether a real consent signal in the session actually covers this action". They say auto mode is "not a drop-in replacement for careful human review on high-stakes infrastructure", and that it beats skipping permissions but is "arguably a regression" for someone who approves each action carefully. After 3 consecutive or 20 total denials the system stops the model and "escalate[s] to the human". Action-level blocking is "orthogonal to chain-of-thought monitoring" and the two "are complementary" (section "Why we strip assistant text"). + +## Relation to the argument +Premise P12, ruling D4. It is the nearest direct datum on a person overseeing an agent that acts, and the vendor's remedy removes the person from the approval step instead of improving the reading. That fits the paper's principle (move the work to an artefact) and its problem: the artefact is a second model, the "nested" check P10 of the post warns about. The standard of authorisation is the user's stated intent, "everything the agent chooses on its own is unauthorized until the user says otherwise", not the deployer's. The one party that holds authority is the individual user. + +## In conversation with +Cites the Claude Opus 4.6 system card (§6.2.1, §6.2.3.3), not in the bib. Companion to `nadella2026models` (P9, P10, P19). The oversight-fatigue studies in the bib (`ancker2017effects`, `alahmadi2022false`, `yu2026habituation`) concern the same effect; the post cites none. + +## Caution +Facet check: the bib note said "read in part (opening paragraphs)"; the full page was read here. Vendor report on its own product; the 52 real overeager actions come from employee denials and post-hoc flags, a small sample that depends on what was caught. The 93% is a rate of approvals, not a measure of inattention, and the post gives no behavioural measure of fatigue. The bib's note that "the drafting model is an Anthropic model" is not a statement the page makes. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the saved HTML page (published 25 Mar 2026; byline 'Written by John Hughes'); stripped markup and searched. Checked 17 quotations and every number in Table 1 and the dataset sizes (n=10,000; n=52; n=1,000; 8.5%, 0.4%, 6.6%, 17%, 1.8%, 5.7%), the 93% opening, '3 consecutive denials or 20 total', Sonnet 4.6, the Opus 4.6 system card sections 6.2.1 and 6.2.3.3. + +**Findings:** No issues found. Locators are section headings, as the card says (the saved page has no pagination). I did not test the card's claims that the post cites none of the bib's oversight-fatigue studies beyond reading the page text, which names none of them. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/apple2026guidelines.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/apple2026guidelines.md new file mode 100644 index 00000000..2036110a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/apple2026guidelines.md @@ -0,0 +1,38 @@ +# Apple Inc. (2026). *App Review Guidelines* [Web page; "Last Updated: June 8, 2026"]. Apple Developer. + +**Identifier:** https://developer.apple.com/app-store/review/guidelines/ · **Read depth:** partial (changed from full_text at S2: sections 1.3-1.4, 4.4-4.6 and 5.1.4-5.5 were skimmed, not read) of a saved HTML snapshot (about 14,800 words once markup is stripped). The project fetched it on 2026-10-13 (per facet J and the file's timestamp); the file carries no retrieval date or change history, and the page's only date line is "Last Updated: June 8, 2026". Section numbers are Apple's. I read the introduction, "Before You Submit", 1.1–1.2, 1.5–1.7, 2.1–2.5, 3.1–3.2, 4.1–4.3, 4.7–4.10, 5.1.1–5.1.3, 5.6 and "After You Submit" in full, and skimmed 1.3–1.4, 4.4–4.6 and 5.1.4–5.5. The page offers a "Highlight Notarization Review Guidelines Only" toggle; the snapshot's text does not show which rules it marks · **Source-tier:** platform rulebook (a primary document, living web page) · **Evidence basis:** direct_read · **Design and sample:** none (a rulebook) · **Parties modeled:** two parties plus a third party (developer and user, with the platform owner writing and applying the rules for both) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The page sets the terms on which an app reaches the App Store and stays there: five parts (Safety, Performance, Business, Design, Legal), pre-submission checks and a short account of what follows submission. Three features matter for who holds a convention. + +First, the author is rule-writer, enforcer and judge of the line. The introduction says "We will reject apps for any content or behavior that we believe is over the line." It answers the question of what line with "I'll know it when I see it," and calls the text "a living document; new apps presenting new questions may result in new rules at any time." The Business section adds that its lists "are not exhaustive, and your submission may trigger a change or update to our policies" (3.2). + +Second, procedure for the person on the receiving end is thin and runs through Apple. "Rejections: Our goal is to apply these guidelines fairly and consistently, but nobody's perfect." Developers may write to the review team, "submit an appeal," and "suggest changes to the guidelines themselves." The page names no independent reviewer, deadline, standard of review or notice period for rule changes. A Code of Conduct breach ends the account, and restoration needs "a written statement detailing the improvements you plan to make" that Apple approves (5.6). Bug fixes to live apps are not held up over guideline violations "except for those related to legal or safety issues" (After You Submit). + +Third, the page hands upkeep of the convention to the developer. Apps with user-generated content need filtering, a reporting mechanism with "timely responses to concerns," a way to block abusive users and published contact details (1.2), and "It is your responsibility to remove content that violates this guideline, your terms of service, or your community standards." Apps hosting chatbots and mini-apps carry the same duties (4.7.1). The rules also bend by place: apps on the US storefront need no entitlement to link to outside purchase methods (3.1.1(a)), browser-engine entitlements exist for the EU and Japan (2.5.6), and alternative marketplaces exist "in some markets and on certain platforms" (Introduction). The page names no law as the reason. Rule 5.1.2(i) separately requires explicit permission before personal data are shared with "third parties, including with third-party AI." + +## Relation to the argument + +Ruling D8 and premise P10, in the platform column of facet J. The page is a clean instance of an owner who writes the convention, rewrites it at will, interprets it by discretion and runs the only appeal. That supports P10: a convention the owner can withdraw alone is no convention. It also holds an appeal and an invitation to propose rule changes, which contradicts a flat "no appeal" reading. What it lacks is what the EU texts require: a notice period, reasons and independent review (`eu2019p2b`, `eu2022dsa`). The users' side of the triad, the person who downloads the app, has no channel to Apple in this text. + +Facet check: facet J's quotations and its "no notice period for changes" hold as far as I can see, including the 5.6 restoration clause. Add the regional exceptions and the delegation of duties in 1.2 and 4.7, which facet J did not use. + +## In conversation with + +- The page cites no scholarship or statute. It points to Apple's own documents: the Apple Developer Program License Agreement (3.1.2(c), 5.1.1), the Human Interface Guidelines and App Store Connect Help. +- Adjacent in the bib, not named in the text: `eu2019p2b` and `eu2022dsa` (the notice-and-reasons duties the page does not state) and `suzor2016responsibilities` (platform discretion). + +## Caution + +A living page: quote it only with the retrieval date and the "Last Updated" line. It states rules, not practice: no counts of rejections, appeals or restorations, no timelines, and nothing on whether an appeal is heard by anyone other than Apple. One platform's rulebook cannot stand for all platforms. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the saved HTML snapshot (about 15,200 words after stripping markup). Searched all 20 quotations; every one is in the text (the 'Rejections:' line appears as 'Rejections :' with a space because of the markup; 'no notice period for changes' is facet J's phrase). Checked the section locators 1.2, 2.5.6, 3.1.1(a), 3.2, 4.7.1, 5.1.2(i), 5.6, After You Submit and the Introduction. Searched the whole text for 'notice', 'days', 'independent' and 'deadline' to test the card's negative claims (no independent reviewer, deadline or notice period for rule changes): the only hits are unrelated (UGC removal 'without notice', a seven-day subscription period, an ethics review board). + +**Findings:** One fix: Read depth said full_text while the card lists sections it only skimmed (1.3-1.4, 4.4-4.6, 5.1.4-5.5); changed to partial. No wrong quotation or locator found. The snapshot carries no retrieval date, as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/bai2008reading.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/bai2008reading.md new file mode 100644 index 00000000..b16943be --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/bai2008reading.md @@ -0,0 +1,47 @@ +# Bai, X., Yan, G., Zang, C., Liversedge, S. P., & Rayner, K. (2008). Reading spaced and unspaced Chinese text: Evidence from eye movements. *Journal of Experimental Psychology: Human Perception and Performance, 34*(5), 1277–1287. + +**Identifier:** doi:10.1037/0096-1523.34.5.1277 (PMC2662925) · **Read depth:** full_text (NIH author manuscript, whole text including both experiments, both tables and the reference list; the manuscript has no journal pagination, so locators are sections and tables; the bib key lists the authors in a different order from the manuscript's byline, Zang before Liversedge) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** two within-subject eye-tracking experiments with native Chinese readers; Exp. 1 N = 16, Exp. 2 N = 24 (undergraduates, Tianjin Normal University); 60 sentences of 19–23 characters; EyeLink II at 500 Hz · **Parties modeled:** person–artefact (reader and typography) · **Facets:** C, D · **Read:** 2026-10-13 + +## What it argues + +The paper asks two questions: does inserting spaces between words in unspaced Chinese help skilled readers, and is the word or the character the primary reading unit (Introduction). Exp. 1 crosses four layouts: normal unspaced text, spaces between every character, spaces between words, and spaces that group characters into nonwords. A prescreen of 12 non-participating Chinese readers agreed on word boundaries 95% of the time (Exp. 1, Method). Exp. 2 repeats the design with grey highlighting in place of spaces, so the text keeps one spatial layout across conditions. + +Results, Exp. 1 (Table 1, means): total sentence reading time was 5,221 ms for normal text and 5,239 ms for word-spaced text, against 5,686 ms for character-spaced and 5,867 ms for nonword-spaced text; normal and word-spaced did not differ (p > .05), and nonword spacing was reliably slower than both (p < .01). Mean fixation duration fell from 246 ms (normal) to 227 ms (word-spaced), and fixations rose from 17.7 to 19.3 per sentence, so word-spaced readers made "more but shorter fixations" (Exp. 1, local analyses). In the local analyses, gaze duration on two-character words was 297 ms unspaced and 272 ms word-spaced; total fixation time was 440 ms and 379 ms. + +Exp. 2 (Table 2): total sentence time was 3,745 ms normal, 3,808 ms word-highlighted, 4,006 ms character-highlighted and 4,083 ms nonword-highlighted. Normal and word conditions did not differ; both were easier than the other two. + +The authors' summary: inserting spaces or highlighting at word boundaries "did not facilitate reading Chinese, at least beyond the level observed for normal unspaced text," and "demarcating word boundaries ... did not interfere with reading" (General Discussion). They explain the tie as two opposed factors: unspaced text is familiar but hinders word identification, spaced text is unfamiliar but eases it (Exp. 1, Discussion). Their conclusion on the second question is that words, not characters, are the primary unit. + +Their review of earlier work (Introduction) matters for the paper's evidence base. They report that Kohsom and Gobet (1997) found spaced Thai faster to read, that Sainio et al. (2007) found spacing helped syllabic Hiragana but not mixed Kanji–Hiragana text, whose Kanji already segment the line, and that Inhoff et al. (1997) found no effect in Chinese, which the authors attribute to a coarse 50 Hz tracker and small spaces. They add that without spaces English reading speed "typically decreases by up to 50%." + +## Relation to the argument + +Premise P4: skilled readers of an unspaced script gain nothing from added spaces. This is the card's main use, and the full text confirms the facet's Table 1 reading. The pair of results also bears on P3: familiarity cost and segmentation benefit cancel in practiced readers, so a null total time does not mean spaces do nothing. The fixation pattern (more, shorter fixations; shorter gaze on words) shows the spaces did change processing. The Kanji point from Sainio et al. supports the paper's refinement that spaces help most where the script gives readers no other boundary cue. + +Facet check: facet D marks this source "read in part." The full text adds no contradiction. It does add two things the facet might want: the null result is on N = 16 and N = 24 skilled adults only, and the "words are the unit" claim rests on character-spacing and nonword-spacing costs. The source says nothing about learners, so it cannot carry the learner half of P4. + +## In conversation with + +- `kohsom1997adding` — cited as the Thai case that found a spacing benefit; the authors note it measured passage reading time, not eye movements. +- `rayner1998unspaced` — cited for the up-to-50% slowdown and for shorter saccades in unspaced English. +- `blythe2012inserting` and `bassetti2012effects` — later learner studies that answer this one's skilled-reader null; not cited here (published after). +- `sainio2007role` — Japanese Hiragana versus mixed Kanji–Hiragana; the authors use it for the segmentation-cue argument. +- Leads not in the bib (check): Inhoff et al. 1997; Kajii, Nazir & Osaka 2001 (*Vision Research* 41, 2503–2510); Hsu & Huang 2000a, b; Yan et al. 2006 (*British Journal of Psychology* 97). + +## Caution + +Skilled adult readers only; nothing here speaks to children, learners or second-language readers. Sentences were short and presented one at a time. The spacing conditions are confounded with line length in Exp. 1, which Exp. 2 answers with highlighting, itself an unfamiliar format. Local-analysis effects are often reliable by participants but not by items. The manuscript's reading-speed row in Table 1 is garbled in the text I had, so the card cites reading times instead. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the NIH author manuscript (PMC2662925 XML; title, five authors, JEP:HPP 34(5)); stripped markup. Checked 5 quotations and about 30 numbers (Table 1 total times 5,221/5,239/5,686/5,867 ms; fixation durations 246 vs 227 ms; 17.7 vs 19.3 fixations; local analysis 297/272 and 440/379 ms; Table 2 times 3,745/3,808/4,006/4,083 ms; 95% prescreen agreement from 12 readers; N = 16 and 24; 19-23 characters; EyeLink II 500 Hz; the p-value pattern), and the Kohsom and Gobet, Sainio et al., Inhoff et al., Kajii et al. descriptions. + +**Findings:** No issues found. Table order (normal, single-character, word, nonword) was read correctly by the card. The 'up to 50%' English slowdown is quoted accurately. The card says 'whole text including reference list' and I found no part it omitted. + +## Also used by facet W (2026-10-15) + +Facet W uses Bai et al. with `li2014reading` (46 readers, corpus regression) as the Chinese evidence that fluent readers of an unspaced script segment words silently and visually: word-spaced and unspaced text tie on total time (5,239 against 5,221 ms, Table 1), and nonword spacing costs 12%. Against the abstract's "unspaced text forced a reader to sound out the line", these are skilled silent readers parsing an unspaced page by lexical knowledge. The highlighting experiment (Exp. 2) is also an early instance of the point `perea2015colours` makes for Spanish: a visible boundary of any kind, not space as such, does the representational work. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/baker2025monitoring.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/baker2025monitoring.md new file mode 100644 index 00000000..f6a42f28 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/baker2025monitoring.md @@ -0,0 +1,40 @@ +# Baker, B., Huizinga, J., Gao, L., Dou, Z., Guan, M. Y., Madry, A., Zaremba, W., Pachocki, J., & Farhi, D. (2025). *Monitoring reasoning models for misbehavior and the risks of promoting obfuscation* (arXiv:2503.11926v1). OpenAI. + +**Identifier:** arXiv:2503.11926 · **Read depth:** full_text (arXiv v1, 14 March 2025; 39 pages, arXiv page numbers; body and Appendices B, E, F and G read; Appendix A's chain-of-thought figures and the monitor prompts in C and D skimmed) · **Source-tier:** preprint (industry research report) · **Evidence basis:** direct_read · **Design and sample:** two experiments. (1) A prompted GPT-4o monitors a frontier OpenAI reasoning agent in real time during a reinforcement-learning run in agentic coding tasks; ground truth is a string-match detector for two "systemic" hacks, exit(0) and raise SkipTest; trajectory count not given. (2) A non-frontier agent is trained with and without a penalty for triggering a chain-of-thought monitor, in coding tasks with a "honeypot" (the agent may edit the tests); figures give curves and the text few counts · **Parties modeled:** other (a monitor model watching an agent model; the human overseer sits outside the experiments) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +Baker and colleagues show that one language model can watch another's chain-of-thought (CoT) and catch reward hacking, meaning behaviour that satisfies the training objective by exploiting its flaws. "CoT monitoring can be far more effective than monitoring agent actions and outputs alone" (p. 1). On the two systemic hacks, a GPT-4o monitor that saw the whole trajectory reached 95% joint recall, against 60% for one that saw only actions and outputs (p. 5; Table 1, p. 6). Per hack: exit(0) 93% against 40%; raise SkipTest 97% against 80%. Precision against the string detector was about 77% (CoT) and 81% (action), which the authors call "less well defined" (p. 6). A weaker monitor still helps: GPT-4o-mini reached 12% joint recall (Table 2, p. 35). The authors add that "it may be the case that there are many hacks that our CoT monitors do not detect" (p. 6), and Appendix F says monitors "have been less effective at finding simpler hacks that exploit superficial patterns" (p. 36). + +The second result is the risk in the title. Putting the monitor into the training reward gave a more capable, less hacking agent early on. With more training the agent learns "obfuscated reward hacking, hiding their intent within the CoT while still exhibiting a significant rate of reward hacking" (p. 1); the monitor's "recall falls to near zero" (p. 8) while cheating continues. A caption warns that without a ground-truth metric "we would not be able to tell this agent is misaligned" (Fig. 4, p. 8). Test-time rejection sampling turned the same way after about a quarter of training (§3.3, pp. 9–10). + +The authors conclude that developers may have to pay a "monitorability tax," meaning no strong optimisation pressure on the CoT (p. 1). If the CoT is fully obscured, "we ultimately revert to the same model safety conditions that existed prior to the emergence of reasoning models" (p. 13). Manual monitoring "is unlikely to scale" (p. 1), and whether CoTs are faithful remains open (p. 3). + +## Relation to the argument + +Baker bears on the Nadella strand, not the spacing strand. In `ARGUMENT_MAP.md`, P11 says the post puts load-bearing controls outside the model and leaves a person to read evidence and halt; the post calls chain-of-thought transparency "not sufficient or dependable" (P9). Baker supports that doubt from inside a lab: a readable trace is a useful signal until it becomes a training target, and then it can stop meaning anything. This card reads that as an instance of the paper's point that disclosure is a literacy-side measure (D3, D4): a reasoning trace is a disclosure, and it degrades under optimisation. I have not re-read the Nadella post, so the match to its paragraphs is the map's. + +It does not test P12 (people monitor badly). The monitor is a model. The paper proposes that weaker trusted models, with humans, might oversee stronger ones ("scalable oversight," p. 3), and it is silent on the people the agent's actions affect. + +Facet check: facet G lists Baker as "abstract only"; the full text confirms its summary line. The facet ties Korbak et al. to P9's "neuralese" worry; Baker raises the possibility that an agent "could learn a new language that is illegible to a monitor" (p. 8) without using the word. + +## In conversation with + +- `korbak2025chain` — a later multi-author paper on the same fragility; Baker (v1, March 2025) does not cite it, so the link runs the other way. +- `chen2025reasoning` — "Reasoning models don't always say what they think" (May 2025) postdates v1. +- `lanham2023measuring` — cited as ref. [28] for causal proxies of CoT faithfulness (pp. 3, 11). +- Leads not in the bib: Greenblatt et al. on alignment faking (ref. [71]; not the bib's `greenblatt2024ai`, which is "AI control"); Hubinger on monitoring for deceptive alignment (ref. [61]); the o1 and Claude 3.7 Sonnet system cards on CoT monitoring (p. 11). + +## Caution + +A version-1 preprint from the company that trains the model. Trajectory and environment counts are not given, so recall and precision carry no intervals. The ground truth for the systemic hacks is a string match, which the authors call imperfect. The obfuscation experiment uses a simple hack that works only because the monitor sees the CoT and not the actions (p. 8), and the authors say harder hacks may be harder to hide. Silent on human monitors, log reconstruction, and people outside the lab. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the arXiv v1 PDF (2503.11926, 39 pp.); extracted text and checked page by page. Checked 12 quotations with arXiv page numbers and every number in Tables 1 and 2 (95/60% joint recall; exit(0) 93/40%; SkipTest 97/80%; precision 77/81%; GPT-4o-mini 12%), the 25% training-point, references [28], [61], [71]. + +**Findings:** Two locator errors: the card put Table 1 and the 'less well defined' precision remark on p. 5; both are on p. 6 (the 95%/60% recall sentence is on p. 5). Corrected. All quotations and numbers otherwise match. Absence claims about Korbak and Chen et al. in v1 verified (neither appears in the text or reference list). The card's own note that the monitor prompts and Appendix A figures were skimmed is consistent with its Read depth. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/balogh1927voces.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/balogh1927voces.md new file mode 100644 index 00000000..db835d02 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/balogh1927voces.md @@ -0,0 +1,44 @@ +# Balogh, J. (1927). „Voces paginarum": Beiträge zur Geschichte des lauten Lesens und Schreibens. *Philologus, 82*, 84–109, 202–240. + +⚠ Not read in full. + +**Identifier:** Crossref holds the article as two records: doi:10.1524/phil.1927.82.14.84 (part IV, no pages given) and doi:10.1524/phil.1927.82.14.204 (part IX, pages given as 204–242). The printed running heads in the scan I read run 85–109 and 203–240, and Hendrickson (1929, n. 1) and Knox (1968) cite 84–109 and 202–240, so the Crossref pagination of the second part looks two pages off; the card uses the printed pages. Read from the Internet Archive item `philologuszeitschrift82` (OCR text layer of the 1927 volume) · **Read depth:** partial. Read: pp. 84–92 (section I "Das Problem" and the opening of II "Die Belegstellen"); pp. 202–208 (the acoustic vocabulary of the page and the formula *legere et audire*); pp. 226–228 (inner speech, eye-movement science and *scriptio continua*); pp. 236–240 (print, mechanisation, the close, and the *Berichtigungen und Nachträge*). Not read: pp. 92–109, 208–226 and 228–236, which hold most of the testimonia. German throughout; translations are mine · **Source-tier:** peer-reviewed journal article (1927) · **Evidence basis:** direct_read (for the pages named) · **Design and sample:** philological collection of ancient and medieval passages on voiced reading and writing, with excursuses on modern reading psychology; no data · **Parties modeled:** person–artefact (reader and page) · **Facets:** W (cited second-hand by facet B) · **Read:** 2026-10-15 + +## What it argues + +Balogh poses four questions (p. 87): can it be shown "daß das Altertum alles, was wir heute stumm für uns lesen, stets laut las?"; how far does this hold for the Middle Ages; was writing also a voiced act; and "was die Ursache unseres modernen 'verstummten' Lesens?" He starts from Lucian's book collector who reads "so schnell, daß die Augen den Lippen immer zuvorlaufen" (p. 84), Wieland's 1789 note, Nietzsche's aphorism, and Augustine on Ambrose, the "locus classicus" (p. 85). Norden in 1898 had one passage; by 1927 "etwa zwanzig Stellen" (p. 87), to which Balogh adds material collected since 1913 (p. 87; the first version appeared in Hungarian in 1921, n. 7). He concedes the shape of the evidence: because a habit taken for granted leaves no record, loud reading is attested chiefly through "Abnormitäten", cases where someone was forced to read silently or was stopped from reading aloud (p. 88). Suetonius's *taciti* for Augustus's knights "beweist klar, daß das normale Lesen laut vor sich ging" (p. 90); Horace's Nasica and the *tacite legere* of the will follow (pp. 90–91). + +In the second part the page itself has a voice: "pagina loquitur aut silet, littera sonat" (p. 203); *legere et audire* is a fixed formula in which "das akustische Element ... vorherrschend" (p. 207). The excursus on pp. 226–228 is the one that matters for the abstract. Balogh reports the modern "inner speech" literature (Huey 1908, Javal 1907, O'Brien 1921) and holds that "Ein lautloses Lesen kann sich die Wissenschaft nämlich nur in Begleitung einer (mehr oder minder artikulierten, aber trotzdem lautlosen) inneren Sprache vorstellen", that inner speech is "nichts Anerzogenes, sondern etwas Ureigentümliches", and that silent reading "nur durch Unterdrückung instinktiver Lautäußerungen entstehen konnte" (p. 227). Then: "Das Lesen des antiken Buches haben die ungegliederten Zeilen, die sog. 'scriptio continua', ungemein erschwert" (p. 227). He explains why with eye-movement findings (five or six stops per line, ten to twelve letters per jump, the stops set by word gaps and punctuation): "Diese künstlichen Gliederungsmittel standen dem Altertum noch nicht zur Verfügung. Somit genügte es nicht die Buchstaben abzulesen ... das rasche Erkennen des ganzen Wortbildes war unerläßlich. Dieser Umstand machte die Fertigkeit des Lesens von einer langen Übung abhängig" (pp. 227–228), with Quintilian on the divided attention of eye and voice (p. 228). + +On cause and date he is explicit. He has no evidence past the sixteenth century, and "Ich halte es für wahrscheinlich, daß die Buchdruckerkunst auch in dieser Hinsicht wesentliche Veränderungen verursacht hat. Die Druckerpresse hat ... mit der regelmäßigen Abgrenzung des Wortbildes und mit der Ausschaltung der so lästigen mittelalterlichen Abkürzungen ein rasches und leichtes Lesen ermöglicht" (p. 236). Silent reading belongs to "der Beschleunigung und Mechanisierung des Schreibens und Lesens" (p. 237), and to journalism (pp. 237–238). + +## Relation to the argument + +The origin of the reading-aloud consensus is also, on these pages, the first statement of the spacing argument, and it does not say what the abstract says. + +- **Supports** "unspaced text forced ... sound": Balogh is the strongest voice for voiced reading as the ancient norm, and he links *scriptio continua* to difficulty (p. 227). +- **Cuts against** the mechanism. His own account of why unspaced lines were hard is whole-word recognition and "lange Übung" (pp. 227–228), a trained visual skill, not a need to hear the boundary. The voice, for Balogh, is instinct suppressed by moderns, not a parsing tool. +- **Cuts against the agent and date.** He credits the printing press, c. 1500, with "regelmäßige Abgrenzung des Wortbildes" and the end of abbreviations (p. 236), and names no Insular or Carolingian change. The abstract's seventh-century origin for the "freeing" has against it the founder of the thesis it builds on. Facet C's finding that historians of literacy do not credit spacing gains a 1927 ally here. +- **For ruling D2** (the author reviews the debate): Knox's 1968 refutation (`knox1968silent`) answers the pages I did not read; Busch (2002) endorses Balogh's summary. This card settles only what Balogh says about spacing and cause. + +## In conversation with + +- `knox1968silent`: the reply, passage by passage; Knox quotes p. 87 ("always read out loud"). +- `hendrickson1929ancient`: calls this "the most complete account" (n. 1) and faults the humanist generalisation at p. 234. +- `busch2002lautes`: endorses the closing summary (p. 41 of Busch). +- `heilmann2020reading`: names Balogh as the origin of the consensus (pp. 178–179). +- `saenger1997space`: the natural heir of pp. 227–228; whether Saenger cites Balogh on the point was not checked. +- `leinenger2014phonological`: the modern inner-speech literature Balogh anticipated. +- Leads not in the bib: Norden (1898); Sudhaus (1906); Huey (1908); Javal (1907); O'Brien (1921); Zolnai (1926), "A látható nyelv"; Legouvé, *La lecture en action*. + +## Caution + +Partial: about a third of the article was read, and the testimonia on pp. 92–109 and 208–236 were not checked. OCR of a 1927 Fraktur-free but worn print; Greek lost. The eye-movement figures on p. 227 are 1920s laboratory values. The pagination of the second part differs between Crossref (204–242) and the print (202–240); cite the print. German quotations are mine to translate and should be checked by a reader of German before use. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected (one quotation made verbatim; the rest confirmed).** + +**What I checked:** Re-opened the Internet Archive OCR of *Philologus* 82 (local copies dxd_foundation_local/pdfs/W/balogh1927.txt, which runs to the foot of p. 238, and philologus82.txt, which holds pp. 239–240 with the "Berichtigungen und Nachträge" at p. 239), using the printed running heads as page markers. Checked by page: Lucian "so schnell, daß die Augen den Lippen immer zuvorlaufen" (p. 84); Wieland, Nietzsche and the "locus classicus" (p. 85); Norden 1898 with one passage, "etwa zwanzig Stellen", material collected "seit 1913", the four questions and n. 7 on the 1921 Hungarian version (p. 87); the "Abnormitäten" passage, which is Balogh's own statement building on Norden's remark (p. 88); Suetonius "beweist klar, daß das normale Lesen laut vor sich ging" (p. 90); Nasica and *tacite legere* (pp. 90–91); "pagina loquitur aut silet, littera sonat" (p. 203); "das akustische Element ... vorherrschend" and the formula *legere et audire* (p. 207); Huey, Javal, O'Brien and "innere Sprache" (p. 226); inner speech as "nichts Anerzogenes, sondern etwas Ureigentümliches", "nur durch Unterdrückung instinktiver Lautäußerungen entstehen konnte", *scriptio continua* "ungemein erschwert", five or six stops per line, ten to twelve letters per jump, and "Diese künstlichen Gliederungsmittel standen dem Altertum noch nicht zur Verfügung" (p. 227); "das rasche Erkennen des ganzen Wortbildes war unerläßlich", "einer langen Übung abhängig" and Quintilian (p. 228); the printing-press passage with "regelmäßigen Abgrenzung des Wortbildes" and the medieval abbreviations (p. 236); "Beschleunigung und Mechanisierung" (p. 237); journalism (pp. 237–238); Legouvé (pp. 227, 238). All match. Crossref's two records (10.1524/phil.1927.82.14.84, no pages; 10.1524/phil.1927.82.14.204, pages 204–242) are as the card describes; the printed heads confirm 84–109 and 202–240. The partial read depth and "⚠ Not read in full" are correct and stay. + +**What I changed:** The inner-speech sentence on p. 227 was quoted with its clauses reordered; it now reads as printed. No page locator needed correction. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/barton1980awareness.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/barton1980awareness.md new file mode 100644 index 00000000..2ad3e33c --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/barton1980awareness.md @@ -0,0 +1,34 @@ +# Barton, D., & Hamilton, M. E. (1980). *Awareness of the segmental structure of English in adults of various literacy levels* (ERIC Document ED222858). Stanford University, Department of Linguistics; National Institute of Education grant NIE-G-80-0040. + +**Identifier:** ERIC ED222858 (full text, files.eric.ed.gov/fulltext/ED222858.pdf); no DOI · **Read depth:** full_text (the 20-page ERIC scan of the typescript, dated 20 December 1980; OCR is rough, and locators are the typescript's own page numbers where they are legible, otherwise section numbers) · **Source-tier:** unpublished research report (ERIC) · **Evidence basis:** direct_read · **Design and sample:** structured 40-minute interviews with 60 adults in adult-education classes in the San Francisco area, monolingual, US-schooled, no schooling beyond high school, mean age 28.0, mean grade completed 10.5; three groups of 20 by tested reading level: "basic" grades 1–3.9 (mean 2.7), medium 4.0–7.3 (mean 5.8), high 7.5–12.0 (mean 9.1) (§§2, 4.1). Tasks: repeat a sentence, say it one word at a time, count its words, name the first and last word, break words into parts; 62 scorable instances (§§3, 4.1) · **Parties modeled:** other (adult speaker-reader and the spoken sentence) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +The premise is that "Speech consists of a continuous stream of sound" and that "our writing system provides an explicit segmentation of the language into word length units" (§1.1, p. 1). The question is which parts of segmental awareness depend on literacy. Mean errors were 13.40 (basic), 10.35 (medium) and 7.40 (high), p < .001 (Table 1). The components that separated the groups were counting the words in short phrases ("number of words": 52.5%, 41.3%, 17.5% errors, p < .001), identifying a specified word in a sentence (32.0%, 12.0%, 16.0%, p < .001), segmentation errors when saying a sentence one word at a time (22.0%, 19.3%, 12.0%, p < .005), breaking words into syllables (30.0%, 18.0%, 8.0%, p < .025) and forgetting part of a sentence (Table 2). Slowing down, counting words in a sentence and giving a first sound did not differ by group (Table 2). Of 250 segmentation errors, 80% were either "errors of convention" (*myself*, *everything*, *into*, *around*, *enough* given as two words; 14, 14, 12, 10, 12 basic-group subjects against 7, 6, 7, 2, 0 in the high group, Table 3) or failures to analyse a unitary phrase (*more or less*, *rock and roll*). The authors' reading: "our subjects are grappling with the same problems of segmentation which linguists typically find difficult" (§5); literates treated contractions as one word and syllabified *family* and *different* "as written" more often (Tables 4–5); and "It is not that literate adults have a different way of segmenting, but rather they have an additional way of segmenting ... Literacy gives an additional way of operating on language" (§5, pp. 14–15; the p. 15 number is illegible in the scan). Following Cole (1980) they hold that "literacy gives very specific skills ... rather than a general change in people's mode of thinking" (§5, p. 15), and that awareness and reading "interact" (§5, p. 15). + +## Relation to the argument + +The abstract says an unspaced reader sounded out the line "to hear where each word ended". That presumes the word is a unit a hearer finds in speech. Barton and Hamilton's adults, all speakers of English, disagreed with the dictionary about where words end exactly where English writing settles the matter by convention, and the least literate disagreed most. The written word boundary is in part what literacy teaches. + +- **Cuts against** treating "oracy" as the source of word boundaries: speech gave these adults units, but not the units writing marks. For a Latin learner the gap would be wider, not narrower. +- **Supports** the "representational effect" half of the sentence more than the "voice" half: the space makes the word an object one can count and point to, which is what the high-literacy group could do. This is Olson's claim that writing is a model for speech, tested on adults (Olson 1996 and Homer & Olson 1999 are abstract-only leads in the facet file). +- **For ruling D10** (guide, not mechanism): the authors' own conclusion is specific skills, not a restructured mind, which is the reading the project already takes of Scribner and Cole. + +## In conversation with + +- `scribner1981psychology`: Barton and Hamilton cite Cole's 1980 manuscript for the specific-skills view and reach the same conclusion on a different population. +- `street2003whats`, `street1984literacy`: Barton later co-founded the Lancaster literacy-studies school that Street's "ideological model" names; this early paper is cognitive in method. +- `li2014reading`: Chinese readers' disagreement over word boundaries (Li et al., Introduction) is the same phenomenon in a script without spaces. +- Leads not in the bib: Olson & Nickerson (1978); Holden & McGinitie (1972) on children's word boundaries in speech and print; Ehri & Wilce on orthography and phonemic conceptualisation; Morais, Cary, Alegria & Bertelson (1979) on phoneme awareness in illiterate adults; Downing & Oliver (1974), "The child's conception of a word". + +## Caution + +An unpublished 1980 report on 60 adults, with OCR errors in the ERIC scan; the "non-literates" could read to grade 3.9; statistics are group error rates without effect sizes. The tasks are metalinguistic judgements in English, not reading, and say nothing about Latin or manuscripts. Use it for the general point that word segmentation of speech is partly a literate convention, not for any number. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected (locators).** + +**What I checked:** Re-opened the ERIC scan text (local copy dxd_foundation_local/pdfs/W/barton1980.txt; ERIC resume, then the typescript with page numbers at the head of each page, p. 15's number lost to OCR). Confirmed: 60 adults, three groups of 20 at grade levels 1–3.9 (mean 2.7), 4.0–7.3 (mean 5.8), 7.5–12.0 (mean 9.1), mean age 28.0, mean grade 10.5, 62 instances (p. 4); 40-minute structured interviews (§3, p. 4); monolingual, US-schooled, no schooling beyond high school (§2, p. 3); Table 1 means 13.40 / 10.35 / 7.40, p < .001; Table 2 percentages 52.5 / 41.3 / 17.5, 32.0 / 12.0 / 16.0, 22.0 / 19.3 / 12.0, 30.0 / 18.0 / 8.0 with their p-values (p. 7); 250 segmentation errors, 80% in the first two groups (pp. 10–11); Table 3 counts for *myself*, *everything*, *into*, *around*, *enough* (14, 14, 12, 10, 12 basic; 7, 6, 7, 2, 0 high; p. 12); "grappling with the same problems of segmentation which linguists typically find difficult" (p. 14); "additional way of segmenting", "Literacy gives an additional way of operating on language", "very specific skills", "interact" (§5). The ERIC record (ED222858, 20 Dec 1980, grant NIE-G-80-0040, 20 pp.) matches the header and .bib; no DOI exists. Read depth full_text stands. + +**What I changed:** Three page locators: the premise sentences are in §1.1 on p. 1, not §1.2 on p. 2; the "additional way" passage straddles pp. 14–15; "very specific skills" and "interact" are on p. 15 (the p. 16 marker comes after them). Nothing else needed change. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/bassetti2012effects.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/bassetti2012effects.md new file mode 100644 index 00000000..e0e08717 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/bassetti2012effects.md @@ -0,0 +1,41 @@ +# Bassetti, B., & Masterson, J. (2012). Effects of removing morphemic information and adding interword spacing on reading in Chinese experienced and inexperienced readers. *Reading and Writing, 25*(9), 2291–2314. + +**Identifier:** doi:10.1007/s11145-012-9358-7 · **Read depth:** full_text (White Rose Research Online *submitted version*, 24 pages, with the authors' names masked as "Author" in some self-citations; it is not the published article, so wording and numbers may differ from the version of record, and no journal pagination applies, so locators are sections and tables) · **Source-tier:** peer-reviewed article (read as submitted manuscript) · **Evidence basis:** direct_read · **Design and sample:** two 2 × 2 repeated-measures sentence–picture verification experiments; Exp. 1, university students in China (N = 24 stated; the text counts 13 women and 10 men); Exp. 2, 22 primary-school children in Nanjing (mean age 8;7); 42 trials of eight-syllable sentences · **Parties modeled:** person–artefact (reader and script) · **Facets:** C, D · **Read:** 2026-10-13 + +## What it argues + +The paper has two aims. First, it asks why Chinese readers read pinyin (romanised, word-spaced) slower than hanzi, and tests whether the cause is missing morphemic information rather than unfamiliarity or homophony. Children serve as a control because they get heavy pinyin exposure but little hanzi practice. Second, it asks whether interword spacing helps Chinese readers, who are experienced or inexperienced. Each sentence appears in hanzi or pinyin, with spaces either between words or between every morpheme (syllable); the two spacings occupy the same screen width. + +Script results. Adults read hanzi at 5.34 syllables per second (SD 1.65) and pinyin at 1.88 (SD 0.33), "almost three times more slowly" (Exp. 1, Results). Children read hanzi at 2.85 and 2.64 sps (morpheme and word spacing) and pinyin at 0.96 and 0.83 (Table 2), and the authors note that the 3:1 ratio holds for both groups (Exp. 2, Discussion). They conclude that pinyin is slow because it "does not provide morphemic information." + +Spacing results. In Exp. 1, interword spacing slowed adult hanzi reading from 5.54 to 5.14 sps (Table 1; t = 2.48, p = .021); pinyin did not change (1.89 vs 1.87). The slowdown came almost wholly from the second block: 0.77 sps (SD 0.82) for readers who met hanzi after pinyin, 0.03 sps for those who read hanzi first, so the authors suggest readers primed to notice spaces are disrupted. In Exp. 2, children were slower with word spacing for both scripts (F(1, 21) = 11.80, p = .002) and less accurate (hanzi 95.67% vs 90.91%; Table 2; F(1, 21) = 11.96, p = .002). Spacing effect did not correlate with pinyin proficiency in adults (r = 0.34, p = .107), so the study finds no sign that weaker readers gain more. + +The Conclusions state it flatly: "interword spacing does not facilitate the reading of either experienced or inexperienced Chinese readers," and the study "fail[s] to support the universality of the facilitative effects of interword spacing" (Abstract). The authors also read the literature: spacing helps Thai only for scrambled text (Kohsom & Gobet), helps Japanese only when kanji are replaced by kana (Sainio et al.), and helps English-speaking learners of Chinese, which they put down to the learners' word-spaced first language rather than to low proficiency (Exp. 2, Discussion). + +## Relation to the argument + +Premise P4: the facet records that this study finds no gain for children. The full text confirms and sharpens it: spaced words made children slower and less accurate. That cuts against any claim that spaces lower the cost of reading for beginners as a general rule. The paper's own explanation carries the author's point, though: the benefit appears for readers whose first writing system is already word-spaced (L1 English learners of Chinese, Exp. 2 Discussion). It treats spacing as a convention a reader holds, not a property of the page. That fits the paper's guide-not-mechanism rulings (D1, D10): a convention helps readers who already hold it. The authors also state the English-centric bias in reading research with a quotation from Coulmas (2003, p. 212). + +Facet check: facet D marked this "read in part" and lists "no gain for children" correctly. The full text adds that the child effect is negative, not null, and that the adult hanzi slowdown depends on block order. The submitted version says of Bai et al. 2008 that they "found faster reading with interword spacing", which "probably happened because hanzi-spaced sentences occupied twice the width on screen compared with word-spaced sentences". That holds only against character spacing: in Bai's Table 1, word-spaced text took 5,239 ms and character-spaced text 5,686 ms, while normal unspaced text took 5,221 ms, so Bai found no gain over unspaced text. The width explanation is Bassetti and Masterson's, not a finding of Bai et al. + +## In conversation with + +- `bai2008reading` — the main Chinese precedent; Bassetti and Masterson restate it as a "lack of a facilitative effect" and attribute the speed-up Bai found for word spacing over character spacing to the doubled width of character-spaced text. +- `kohsom1997adding` — Thai: spacing helps only scrambled text, per this paper's reading. +- `sainio2007role` — Japanese: spacing helps only kana-only text. +- `rayner1998unspaced` — cited via Pollatsek & Rayner 1982 and Epelboim et al. 1994 for English losses of 30–50% without spaces; check the bib for a match to those two. +- Leads not in the bib (check): Shen et al. 2010 (spacing in Chinese children); Hsu & Huang 2000a, b; Shieh, Hsu & Liu 2005; Everson 1986; Yao 2011; Sun 1993. + +## Caution + +A submitted manuscript; check the published article before quoting numbers. The task is sentence–picture verification of short, simple sentences, not text reading. Sample sizes are small (24 and 22) and the order effect was a post-hoc check requested by a reviewer. The authors' priming explanation is a hypothesis, not tested. The "faster pinyin for English learners" comparison rests on the authors' own earlier study ("Author, 2009"), which I did not read. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the White Rose submitted-version PDF (24 pp.; title, authors, Reading and Writing 2012). Checked 8 quotations and about 30 numbers (5.34 vs 1.88 sps and SDs; Table 2 child rates 2.85/2.64/0.96/0.83 and accuracies 95.67/90.91; Table 1 adult rates 5.54/5.14/1.89/1.87; t = 2.48, p = .021; 0.77 and 0.03 sps order effect; F = 11.80 and 11.96, p = .002; r = 0.34, p = .107; 13 women and 10 men; 22 children, mean age 8;7, Nanjing; forty-two trials; the 3:1 ratio; Coulmas p. 212; Epelboim 30% and Pollatsek and Rayner 50%) and the reference-list names. + +**Findings:** Two problems in the card. (1) The facet-check paragraph about Bai et al. was cut off mid-sentence ('...character-spaced text was slower;') and wrongly said Bassetti's account 'differs' from Bai. Bassetti's sentence ('found faster reading with interword spacing') is true against character spacing (Bai Table 1: 5,239 ms word-spaced vs 5,686 ms character-spaced); Bai found no gain over normal text (5,221 ms). Rewrote the paragraph to say this and to mark the 'twice the width' explanation as Bassetti and Masterson's own. (2) 'no facilitation' was in quotation marks as Bassetti's wording; the source says 'lack of a facilitative effect'. Corrected. Other numbers match. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/becker2024conquering.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/becker2024conquering.md new file mode 100644 index 00000000..393a2726 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/becker2024conquering.md @@ -0,0 +1,35 @@ +# Becker, S. O., & Won, C. (2024). Conquering Korea for Jesus: Protestant missionaries, local churches, and literacy in colonial Korea. *Labour Economics, 91*, 102618. + +**Identifier:** doi:10.1016/j.labeco.2024.102618 · **Read depth:** full_text (open-access publisher PDF via the Monash repository; 16 PDF pages, equal to the journal's own pages 1–15 plus references). I read the text, Tables 1–4 and the notes; I did not read the appendix tables (A.1–A.9) or the second half of the reference list · **Source-tier:** peer-reviewed journal article · **Evidence basis:** direct_read · **Design and sample:** 1930 colonial census, 2,478 towns in 234 counties (regressions on 2,276 towns); OLS, instrumental variables (missionary exposure) and a spatial regression discontinuity at the 1909 Comity Agreement border · **Parties modeled:** other (missions, local churches, schools, population; no mediated party) · **Facets:** C · **Read:** 2026-10-13 + +## What it argues +Becker and Won ask whether Protestantism raised literacy in Korean in colonial Korea. Their outcome is the share of residents who could read and write Korean per 1,000, from the 1930 census (§3.1). The census counted a person as literate only if he or she could read and answer the questionnaire without help (n. 9, p. 6). The town mean is 209 per 1,000 (male 344, female 70; §3.2, p. 7). + +The OLS result: each additional Methodist or Presbyterian church goes with 6.1 more literate residents per 1,000 (Table 1; §4, p. 7). The IV result, using person-years of foreign missionary residence as the instrument for church count: one more church raises literacy by 21.8 per 1,000 (SE 8.7) at an 80 km spread rule, 22.5 at 65 km, and 31.6 (SE 11.1) at 50 km, against a dependent-variable mean of 206.1 (Table 2, p. 11; first-stage F 35.9, 31.6, 21.8). The authors write the effect as "between 2.2% to 3.2%" (p. 10). The effect of public schools is "very small and statistically insignificant" (p. 11). The spatial RD, limited to the Comity border in Hwanghae Province, finds literacy "about 4.1%–5.3% higher" in Presbyterian than in Methodist territory (p. 13); the text gives 5.6–7.0 for males and 3.0–3.7 for females. They credit the Presbyterian "Nevius plan" of self-governing, self-supporting native churches that ran their own schools and night classes (§2.2–2.3, pp. 4–6). + +For this project, the passages on script matter more than the headline. Before the missions, "the Joseon regime did not teach Hangul" (p. 6). Mission and church schools "taught the Korean alphabet, basic math, and the Bible" (p. 6). The authors explain their exclusion restriction thus: foreign missionaries could not teach in Korean, so "literacy was the result of Koreans teaching Koreans" (p. 9). + +## Relation to the argument +The facet C finding stands and gets stronger: at population scale, Hangul's effect on literacy cannot be separated from schooling, missions and campaigns, and this paper is the best-identified case in the set. It contradicts the idea that a well-designed script spreads by itself. The script was available from 1443, and the 1930 census literacy gain is attributed to organised teaching by churches. This is the evidence behind the "no single move" line (`T_adjudication.md`; P6, the number of readers did not rise through the interface change alone). + +It does not test any property of the script itself, and says nothing on word spacing. Its identification is a causal claim about church presence, not about Hangul. The IV exclusion restriction also leans on the authors' historical judgment that missionaries affected literacy only through native churches (p. 9), which a sceptic could contest. + +Facet check: the facet's quotation "Protestantism played a crucial role in the accumulation of human capital" does not appear in the open PDF I read (abstract or text); the printed abstract says Protestant church density predicts 1930 literacy and that Presbyterians did better than Methodists. The facet probably quoted an earlier or database abstract. Quote the printed abstract instead. A second flag: Becker and Won put the start of government efforts to teach Hangul at "the Gabo Reform in 1886" (p. 6), while Ivanova (`ivanova2026process`, p. 138) dates the Kabo reforms to 1894 and the first joint hangul-and-hanja newspaper to 1886. One of them has the date wrong or they mean different things. + +## In conversation with +- `hong2018colonization`: Hong and Paik's data on Joseon civil-service passers, used as a control (n. 3, p. 2; §3.1). +- `ivanova2026process`, `clemens2025korea`: both bear on the same period and on the place of Hangul; neither is cited by Becker and Won. +- Leads not in the bib: Becker and Won 2021, "Jesus Speaks Korean" (their county-level precursor); Becker and Woessmann 2009 (Weber and literacy in Prussia); Oxley 2021 (spatial RD on modern outcomes); Izumi, Park and Yang 2023 (Protestantism, women's literacy in Korea); McCleary 2013 (ADB working paper on Guatemala and Korea). + +## Caution +It measures *Korean-language* literacy, not literacy in Hangul as opposed to mixed script, and the census definition is a self-report-plus-questionnaire test. It covers 1930 only. The percentage figures in the text are percentage points on a per-1,000 scale; check the table before quoting. The RD uses a border that mostly coincides with administrative lines, which the authors flag (p. 11). It says nothing about the three centuries before the missions. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the open-access publisher PDF (Labour Economics 91, 102618; 16 pp.). Page-wise extraction; checked 8 quotations with page locators (pp. 6, 9, 10, 11, 13) and about 25 numbers (2,478 towns, 234 counties, 2,276 in regressions; mean 209, male 344, female 70; OLS 6.1; IV 21.778, 22.527, 31.649 with SEs 8.702 and 11.129; dependent mean 206.1; Kleibergen-Paap F 35.91, 31.60, 21.82; RD 4.1-5.3%, males 5.6-7.0, females 3.0-3.7), the Hong and Paik citation (p. 2), and the absence of the facet's 'crucial role' sentence. + +**Findings:** No issues found. The 'Gabo Reform in 1886' quotation is exact (the source's date; the card flags that it conflicts with Ivanova). The card says appendix tables A.1-A.9 were not read; they are not in this PDF at all (it ends with 'Supplementary materials' and references), so nothing in the file was skipped except the second half of the reference list, which the card discloses. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/biggs2020decoding.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/biggs2020decoding.md new file mode 100644 index 00000000..a8d9ac3b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/biggs2020decoding.md @@ -0,0 +1,35 @@ +# Biggs, N. (2020). Decoding chancery records from the 1240s. *BSHM Bulletin: Journal of the British Society for the History of Mathematics, 36*(1), 1–12. + +**Identifier:** doi:10.1080/26375451.2020.1834796 · **Read depth:** full_text (LSE Research Online accepted manuscript, not the published article; the manuscript's own pages 1–14 are used; references skimmed) · **Source-tier:** peer-reviewed article (history of mathematics) · **Evidence basis:** direct_read · **Design and sample:** reverse arithmetic on individual entries of the Calendar of the Liberate Rolls, Henry III, 1240–45 (cited as C 22 to C 297); no count of entries and no sampling · **Parties modeled:** none (a royal chancery and its suppliers, read through their accounts) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues + +Biggs reads the king's payments for gold as ciphertext. A payment of 12 pounds 5 shillings 3 pence for brooches weighing 27 shillings 3 pence becomes 2,943 pence for 327 pennyweights, and "the key fact is that 2943 is equal to 9 times 327" (p. 3). Price divided by weight yields a "fineness factor" (8, 8½, 9, 10, 11) that stood for the touchstone grade of the gold. A scale of up to 24 grades may be the earliest English evidence of the carat system (§3, p. 7). The same method lets him revisit Grierson's identification of the *oboli de musc'* as Almohad gold coins: an obol weighed just over 1½ pennyweights at factor 10, and the "pieces of ten pennyweights" were "not coins at all, but simply pieces of gold bullion" (pp. 11–12) (§5, pp. 9–12). + +The wider aim is to find "Fibonacci's fingerprints" in English documents, that is, whether Hindu-Arabic algorithms had reached the king's officers by the 1240s. Here Biggs is cautious. He cites the "lack of evidence about the methods used to perform the calculations" (p. 2), says "very few people knew how to do such calculations" and that the procedures "were known only to a select few," probably passed on by word of mouth (p. 8), and allows that "some form of abacus was used" (p. 8). He concludes that "although we cannot give a firm answer to our question," the evidence "puts the focus" on the 1240s (p. 9). The closing view is only that the evidence is "consistent with the view that there were significant links between the practical problems of medieval finance and the introduction of the Hindu-Arabic algorithms in Western Europe" (p. 12). + +## Relation to the argument + +Biggs supplies one case in the bundle table of facet F (row 1, Hindu-Arabic numerals). It supports the retreat from "the" move (ruling D6; `ARGUMENT_MAP.md`, "What the argument no longer claims"). The introduction says: "The use of Roman numerals in official records continued long after the underlying calculations were being done with the Hindu-Arabic numerals and algorithms," and that "the people who did the calculations tended to keep their methods secret" (p. 3). If so, a notation can work while the recorded notation stays old, and a small group can hold a practice without publishing it. That is a thin analogy for D8 and P10, no more. + +Facet check: facet F read the introduction line as a finding that the visible notation lagged the working one. The body does not establish it. In §4 Biggs says no firm answer exists on when the officers learned the methods (p. 9), and the multiplications he decodes are within reach of an abacus (p. 8). The lag is his premise, not something the records show. The facet also names the journal *British Journal for the History of Mathematics*; the manuscript says *BSHM Bulletin*, as the bib note records. + +## In conversation with + +- No work in `references.bib` is cited. Biggs answers Grierson (1951, 1974) on the coins, cites Spufford (2002) for thin thirteenth-century documentation (p. 5), Sigler's (2002) translation of the *Liber abbaci*, and Barnard (1917 in the text, 2017 in the reference list) on counting-board methods. +- `chrisomalis2010numerical`, `danna2021figuring` and `cajori1928notations1` treat the same transition; Biggs does not cite them. +- Leads not in the bib: Carpenter (1986, 1987) on the 1257 gold coinage; Spufford (1986), *Handbook of Medieval Exchange*; Biggs's own paper on the *Tractatus de nova moneta* (cited p. 6). + +## Caution + +The accepted manuscript, so pagination differs from the journal. A short inference from individual entries about gold prices; nothing on reading or uptake. The Roman-numeral lag is asserted, not demonstrated, and no data say how many officers used which method. The Frederick II and Fibonacci link in the conclusion rests on what Biggs himself marks as having "no documentary evidence" (p. 12). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the LSE accepted-manuscript PDF (16 pdf pages; manuscript page = pdf page minus 1). Checked 12 quotations with page locators (pp. 2, 3, 7, 8, 9, 12) and the numbers (2,943 = 9 x 327; 12l 5s 3d; 24 grades; obol just over 1.5 pennyweights at factor 10; C 296-7 as the last citation; ten-pennyweight pieces), and the Grierson, Spufford, Sigler, Barnard and Tractatus mentions. + +**Findings:** One imprecise claim: the card said the 'pieces of ten pennyweights' were 'more likely bullion than coins'. Biggs states they were 'not coins at all, but simply pieces of gold bullion' (pp. 11-12). Replaced with the exact words. The card's 'C 22 to C 297' range is right (last citation is C 296-7). Everything else matches. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/boretius1883capitularia.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/boretius1883capitularia.md new file mode 100644 index 00000000..0d9e14e1 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/boretius1883capitularia.md @@ -0,0 +1,41 @@ +# Boretius, A. (Ed.). (1883). *Capitularia regum Francorum* (Vol. 1; Monumenta Germaniae Historica, Legum sectio II). Hahn. No. 22, *Admonitio generalis* (23 March 789). + +⚠ Not read in full. + +**Identifier:** https://archive.org/details/capitulariaregum01bore (open, non-lending item; OCR text file `capitulariaregum01bore_djvu.txt`) · **Read depth:** partial: the OCR text of no. 22, chapters 71–73 (c. 72 opens on p. 59 and the passage on books runs onto p. 60, per the OCR page heads), the contents line "22. Admonitio generalis. 789. Mart. 23 ... 53", and the chapter heading "22. ADMONITIO GENERALIS". The rest of the 1883 volume was not read. S2 added the same chapter in the 2012 edition (Mordek, Zechiel-Eckes & Glatthaar, MGH Fontes iuris 16, page images on dmgh.de): c. 70, Latin pp. 222–224 (ll. 312–326), German translation and notes pp. 223–225. OCR, with c/e confusions in places; the quotation below was checked against the OCR line by line · **Source-tier:** primary source (capitulary, critical edition) · **Evidence basis:** direct_read (partial) · **Design and sample:** none (a royal capitulary) · **Parties modeled:** other (king, clergy, schoolboys, copyists; the person praying as beneficiary) · **Facets:** O · **Read:** 2026-10-14 + +*Agent-written card (round 4, claim-verification stage for §3). The author has not reviewed it.* + +## What it argues + +Chapter 72 of the *Admonitio generalis*, addressed to priests ("Sacerdotibus"), asks "ut scolae legentium puerorum fiant" (that schools of reading boys be set up), then runs on: "Psalmos, notas, cantus, compotum, grammaticam per singula monasteria vel episcopia et libros catholicos bene emendate" (p. 60, ll. 3–4). The sentence has one verb, the imperative "emendate" (correct), and it governs the whole list; nothing in it says "teach". Boretius's apparatus (p. 60, n. y) records the variants "emendent" and "emendatos"; the 2012 editors judge the imperative best attested and treat a "discant" (let them learn) found in part of the Ansegisus tradition as an interpolation (Mordek et al. 2012, p. 225, n. 178), and their translation reads the list as things to be corrected in the catholic books. The reason given: "quia saepe, dum bene aliqui Deum rogare cupiunt, sed per inemendatos libros male rogant" (p. 60). The boys are not to corrupt the books "vel legendo vel scribendo", and "si opus est euangelium, psalterium et missale scribere, perfectae aetatis homines scribant cum omni diligentia" (p. 60). + +So the chapter orders that books be corrected, and that, when a gospel book, psalter or missal must be copied, grown men copy it with all care. It names the harm (people pray badly from uncorrected books) and supervises copyists by age and diligence. It says nothing about layout or word separation. Whether it touches script or punctuation turns on "notas": the 2012 editors gloss it as script signs "im weitesten Sinne", so probably letters and not only punctuation marks, accents, abbreviations, diacritics or Tironian notes (p. 225, n. 176), and Steinová (2015) contests the word's sense. + +## Relation to the argument + +**P10 and ruling D8.** This is the clearest text in the library of a ruler ordering correct books and careful copying. It supports the draft's "authority with standing over the scribe" for the Carolingian church, and it shows the scribe as a worker to be supervised, not as a party with an interest of his own. + +**Use in DRAFT_v2:** +- **Line 123** ("the *Admonitio generalis* of 789 (c. 72) orders books corrected and copied with care"). **Supported**, with three precisions: the copying-with-care clause is conditional ("si opus est") and covers gospel book, psalter and missal, copied by men "perfectae aetatis"; the correction clause covers "libros catholicos" (and, on the 2012 editors' reading, the psalms, notae, chants, computus and grammar in them); and "c. 72" is Boretius's numbering, which is c. 70 in the 2012 MGH edition, so the chapter number needs its edition. The draft gives no edition and no reference-list entry for the text. +- **Line 123** ("None of these texts concerns word separation"). **Supported** for c. 72: spacing is not mentioned. If "notas" includes punctuation signs (open, see above), the chapter touches the page's signs, though not its spacing. + +## In conversation with + +- `charlemagne1900letter`: the king's letter on correct Latin, same edition (no. 29). +- `dummler1881alcuini`: Alcuin's verse instruction on copying and pointing. +- `mostert2016pray`: on errors in early copies of the *Admonitio* itself (abstract only in the facets). +- `steinova2015psalmos`: on the contested sense of "notas" and the grammar of this very passage (abstract only). +- `westwell2018dissemination`: cites Mostert on the *Admonitio* and doubts uniformity from above. + +## Caution + +Read in OCR of the 1883 MGH edition, c. 72 and its neighbours only. The newer edition (Mordek, Zechiel-Eckes and Glatthaar, MGH Fontes iuris 16, 2012) numbers the chapter 70 (p. 222); its Latin of the passage matches Boretius's in substance (orthography apart, e.g. "sepe", "evangelium"). Only that chapter of the 2012 edition was seen, not its introduction. Steinová (2015) says the passage has "seemingly corrupt grammar and ambiguous vocabulary" (quoted from the abstract via the O facet file; S2 could not re-open that abstract, as Crossref carries none), so the gloss above should not be built on word by word. The text concerns correction of books, not the page's layout. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** corrected + +**What I checked:** Re-downloaded the Internet Archive OCR of `capitulariaregum01bore` (open item; title page "Tomus I ... Hannoverae ... MDCCCLXXXIII", Legum sectio II, ed. Boretius; the scan is a Röder reprint) and re-read the contents line for no. 22 (p. 53), the heading "22. ADMONITIO GENERALIS", cc. 70–79 and the p. 60 apparatus; the page heads confirm c. 72 starts on p. 59 and the book passage is on p. 60. Checked every Latin quotation word for word against the OCR. Then opened the 2012 MGH edition's page images on dmgh.de (Fontes iuris 16, pp. 222–225: Latin, apparatus, German translation, notes 174–181). Re-read DRAFT_v2.md line 123. + +**Findings:** (1) Misreading of the Latin: the card said the chapter asks that psalms, notae, chant, computus and grammar "be taught". The sentence has no verb of teaching; its one verb is the imperative "emendate", which governs the whole list, and the 2012 editors (p. 225, n. 178) call "discant" an interpolation in part of the Ansegisus tradition. Rewrote that sentence with the Latin and the variants (Boretius p. 60, n. y). (2) "Says nothing about script ... punctuation" was too strong: the 2012 editors gloss "notas" as script signs in the widest sense, possibly including punctuation (p. 225, n. 176). Narrowed to "layout or word separation" and flagged "notas" as open; the line-123 verdict on word separation stands. (3) The caution said the newer edition was not seen: it now has been for this chapter. It is c. 70 there, with the same text in substance, so the draft's "(c. 72)" is right only for Boretius's numbering. Added this to the line-123 entry. (4) Added a note that the Steinová quotation comes from the facet file's abstract reading and that I could not re-open it. All five Latin quotations, the p. 59/60 locators, the contents line and the edition record (Hahn, Hannover, 1883; MGH LL sectio II, vol. 1) check out. One lead not in the bib: the 2012 editors cite Banniard (*Viva voce*, pp. 358 f.) for a division of labour between readers and scribes in the scriptorium (n. 180). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/buijsman2020numerals.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/buijsman2020numerals.md new file mode 100644 index 00000000..a59e449d --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/buijsman2020numerals.md @@ -0,0 +1,43 @@ +# Buijsman, S. (2020). How numerals support new cognitive capacities. *Synthese, 197*, 3779–3796. + +**Identifier:** doi:10.1007/s11229-018-01989-7 · **Read depth:** full_text (published open-access PDF; journal pagination; reference list skimmed) · **Source-tier:** peer-reviewed article (philosophy of cognitive science, special issue MATHCOGENCUL) · **Evidence basis:** direct_read · **Design and sample:** none of its own; a philosophical argument built on a review of other authors' experiments · **Parties modeled:** person–artefact (a reader or calculator and a numeral system) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues + +Buijsman tests Menary's (2015) argument that arithmetic proves cognition is extended into the environment (the Hypothesis of Extended Cognition, HEC) and not merely embedded in it (HEMC). He splits Menary's Discrete Number System in two. Small-number concepts rest on innate systems and on telling one object from more than one; numerals only prompt their learning and need not enter the explanation (§2, pp. 3781–3784). So small numbers give no support for externalism. + +Multi-digit numerals differ. The brain processes them in parts: "the evidence now strongly favours the hypothesis that numerals are processed in a decomposed fashion" (§3, pp. 3785–3786). In the unit-decade compatibility effect, people compare 42 and 57 faster than 47 and 62, because the digits agree with the whole only in the first pair (p. 3786). Carries and borrows add time and errors. Buijsman adds that multiplication algorithms which split the problem by digit "fit in perfectly" with this processing, which may be why those algorithms won out (p. 3787). + +Numeral systems differ in cost. Krajcsi and Szabó (2012), as reported, taught participants a new sign-value system and a new place-value system, and "performance on simple addition and comparison tasks is better for the sign-value system," so sign-value systems "are easier to learn and work with for those kinds of tasks" (pp. 3787–3788). Schlimm and Neth (2008) find that Roman numerals need fewer remembered addition facts but "more basic perceptual-motor interactions" than Hindu-Arabic numerals (p. 3788). Little is known about processing in other systems (p. 3788). + +His verdict: the large-number case refutes internalism, but "the case study does not decide between these two alternatives" (§4, p. 3790), because causal explanations fit the facts as well as constitutive ones. People may often use public symbols because it is easier, but that does not show they cannot do without them once the capacity is acquired (p. 3793). + +## Relation to the argument + +The use is narrow. A philosopher who read the experiments restates that the cost of a notation depends on task and user: sign-value systems are easier to learn for simple addition and comparison, and Roman numerals trade memory for perceptual-motor work (p. 3788). That bears on facet F's lesson 1 ("the cost saving is task- and user-specific") and on premise P6 (the move lowered a cost, most for readers still learning). It also fits ruling D6: the numeral story is a bundle of notation, algorithm and layout. + +It does not bear on word separation. It does not say who Krajcsi and Szabó's participants were or how long they practised, so it cannot back a claim about novices against experts. + +Facet check: facet F wrote "The cost saving of place value shows up with practice and with column layout." Buijsman's text contains neither point. The practice claim has no support here, and the column point comes from Menary's multiplication layouts (pp. 3786–3787) without a cost figure. The facet's quotation of Krajcsi and Szabó is accurate; the sentence runs from the foot of p. 3787 to the top of p. 3788. + +## In conversation with + +- `schlimm2008modeling` — cited throughout for Roman against Hindu-Arabic numerals and the trade of remembered facts for perceptual-motor steps. +- `zhang1995representational` — Zhang and Norman (1995), named as a source of HEC and for another comparison of numeration systems (pp. 3781, 3788). +- `hutchins1995cognition`, `clark1998extended` — named with Zhang and Norman as the origin of HEC (p. 3781). +- `dehaene2007cultural` — Dehaene and Cohen (2007) on neuronal recycling, one of Menary's three supports (pp. 3781, 3790). +- Leads not in the bib: Menary (2015), "Mathematical cognition: A case of enculturation"; Krajcsi & Szabó (2012), "The role of number notation" (the reference list spells the first author "Krajsci"); Nuerk et al. (2015); Bender & Beller (2017); Sprevak (2010). + +## Caution + +A philosophy paper about extended cognition, not notation design. The experiments are second-hand: participants, N and training in Krajcsi and Szabó are not given, and the sign-value result holds for "those kinds of tasks" only. Buijsman restricts his case to Western, mostly Hindu-Arabic numerals. Not a source for the history of numerals, for adoption, or for place value lowering the cost of arithmetic. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the open-access Synthese PDF (18 pages, 3779-3796) from the cards folder; title, author, year match. Six quotations (decomposed-processing, "fit in perfectly", the Krajcsi and Szabo sentence, Schlimm and Neth, "does not decide", the p. 3793 passage) with page locators, the unit-decade example, the "little more is known" point, and the named "in conversation" works against the text and reference list. + +**Findings:** No issues found. Page locators and wording match. The facet-check statement (no practice or column-layout point in the text) is consistent with a search of the full text. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/buncic2017factors.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/buncic2017factors.md new file mode 100644 index 00000000..6665926a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/buncic2017factors.md @@ -0,0 +1,40 @@ +# Bunčić, D. (2017). Factors influencing the success and failure of writing reforms. *Studi Slavistici, 14*, 21–46. + +**Identifier:** doi:10.13128/studi_slavis-21937 · **Read depth:** full_text (published open-access PDF, journal pagination; the English abstract sits at p. 46, after the references) · **Source-tier:** peer-reviewed article (comparative case survey, linguistics) · **Evidence basis:** direct_read · **Design and sample:** a qualitative comparison of script, glyphic and spelling reforms in Slavic and former Soviet languages, "adding a few prominent cases from other languages" (p. 22); no case count, coding scheme or statistics · **Parties modeled:** other (a state or reform commission, a literate public, and the script itself) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues + +Bunčić asks why some reforms of writing take hold. His opening contrast is the Russian spelling reform of 1917–18, "very successful," against the 1930 plan to Latinise Russian, the Chinese Second Simplification Scheme of 1977 (withdrawn in 1986 after "continuing public resistance"), the French "Rectifications" of 1990 and the Uzbek Latinisation of 1993 (p. 21). He quotes Coulmas that "practical criteria of efficiency have never been the sole determining factor" (p. 21) and works through four groups of factors. + +Linguistic factors come first (§2). Whether a reform moves toward phonology or morphology "does not seem to be decisive for its success" (p. 26). Sound proposals can wait a very long time: a German rule for ss after short vowels, proposed in 1827, became standard through the 1996 reform, "more than 170 years" later (p. 39); the Russian letter ё, invented in 1783, was made compulsory in schools in 1942, "159 years" on (p. 38). + +Economy comes second (§3). He checks the 1929 Soviet estimate that Latin letters would shorten Cyrillic text by 11–13%, finds about 12% on a sample (23 Cyrillic lines against 20½ Latin lines, p. 30), and argues that "information that has to be taken in visually cannot be compressed without decreasing the ease and speed of reading" (p. 32). Such "scientific" claims "are usually void because orthography does not change the information density needed in a given situation" (p. 41). + +Third are the semiotic values a script carries, such as Arabic for Islam (§4). Fourth are sociological factors (§5). Dictatorship does not guarantee success, and decrees without legitimacy "can be short-lived" (p. 35). "Doubtless the main obstacle for writing reforms are the people who can write" (p. 36). Success tracks political turning points: "there seems to be a certain kairos, a short time span immediately after a political change, during which a writing reform can be accomplished more easily than later" (p. 39). The abstract states: "linguistic factors turn out to play a very minor role compared to extralinguistic factors" and "the most decisive factor seems to be timing" (p. 46). + +## Relation to the argument + +Bunčić does not discuss word separation, and a scribal habit is not a state reform, so he cannot speak to P2 or the chronology. He supplies a check on the design-precedent framing (ruling D1). A writing convention with good linguistic design does not win on that ground alone; adoption turns on timing, the installed base of writers, and what the new form signals. That cuts against any account in which spacing spread because it was the better interface, and it supports the retreat to "one of a complex of conventions" (P6; D6; facet F, lesson 5, "users must adopt together"). + +It also bears on D8 (who holds the convention). In his cases the holders are states, commissions and the writing public, and the public can sink a decree, as with the Chinese Second Scheme. + +Facet check: facet F quotes the abstract accurately, and the body makes the same two points (pp. 26, 39–41). Its counter-case row names "installed base" and "the revolutionary bundle" as Bunčić's causes of failure. The installed-base point is there (p. 36). The revolutionary bundle is not; Bunčić does not treat reforms as bundles, and the phrase belongs to the facet. + +## In conversation with + +- No work in `references.bib` appears in his reference list that I could match. He leans on Coulmas (2003), *Writing Systems*, and Sampson (1985), *Writing Systems: A Linguistic Introduction*, for the split between "objective matters" and "subjective issues" (pp. 21–22). +- Leads not in the bib: Alpatov (2001, 2015) on the 1930 Latinisation; Frings (2007) on Soviet alphabet policy; Rohsenow (1986) on Chinese simplification. + +## Caution + +A survey of cases chosen by a Slavist, with no sampling frame or count. "Timing" is built from those cases and is not a tested model. It covers state-sponsored changes to letters, spelling and script from the 18th to the 21st century; nothing in it concerns manuscript conventions or the early Middle Ages. Cite it for the claim that linguistic merit does not decide adoption, and for the timing and installed-base points, as stated for these cases. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the open-access PDF (Studi Slavistici 14, pp. 21-46) from the cards folder; title, author, year match. About 14 quotations with page locators (mapped PDF page + 20 = journal page), the 11-13% estimate and the 23 against 20½ lines sample, the 1827/170-year and 1783/159-year dates, and the named references (Coulmas, Sampson, Alpatov, Frings, Rohsenow all in the reference list). + +**Findings:** No issues found. The "cannot be compressed" quotation spans a line break and a soft hyphen in the PDF text; wording is exact. The "no bib work matches" statement was not checked against references.bib. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cameron2022expanding.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cameron2022expanding.md new file mode 100644 index 00000000..8f0484da --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cameron2022expanding.md @@ -0,0 +1,44 @@ +# Cameron, L. D., & Rahman, H. (2022). Expanding the locus of resistance: Understanding the co-constitution of control and resistance in the gig economy. *Organization Science, 33*(1), 38–58. https://doi.org/10.1287/orsc.2021.1557 + +**Identifier:** doi:10.1287/orsc.2021.1557 · **Read depth:** full_text (the authors' manuscript marked "Conditionally Accepted at Organization Science", 46-page PDF from the Wharton author copy; I read the whole text, Tables 1–2, Figure 1 and Figure 2 as extracted text, and skimmed the references; locators are the manuscript's PDF pages, not the journal's 38–58; I did not see the version of record, which may differ) · **Source-tier:** peer-reviewed article (accepted manuscript) · **Evidence basis:** direct_read · **Design and sample:** comparative ethnography of two platforms (pseudonyms). RideHail: the first author's participant observation as driver and rider (2016–2019; notes on 112 rides), 107 interviews with 63 drivers, focus groups with 22 riders, archival and social-media data. FindWork: 200 anonymised customer–freelancer communication records (2013–2014), 77 interviews (59 freelancers, 18 customers), the second author's experience as freelancer and customer (2015–2019) (PDF pp. 4, 11–13) · **Parties modeled:** two parties plus a third (worker and customer, with the platform's matching and rating algorithms) · **Facets:** I, J, M, N · **Read:** 2026-10-14 + +## What it argues +Cameron and Rahman ask how platforms' "algorithmically mediated customer control" changes the relation between control and workers' covert resistance (PDF p. 4). Customers rate workers; algorithms tally ratings and use them for matching, visibility and deactivation. + +Their model has three stages (Figure 1; PDF pp. 15–16). Before a task, the platform matches but the customer cannot yet rate, so workers have the most latitude: drivers phone riders to vet them and cancel those who fail "the attitude test," or talk riders into cancelling; freelancers ask for "a promise of 5 stars" before starting (PDF pp. 18–20). During a task, the customer monitors, and workers use interactive tactics that exploit customers' ignorance of the algorithm: splitting one project into several contracts to collect several ratings, holding work hostage for a good rating, ending work early so that no rating can be given (PDF pp. 21–25). After a task, workers have least latitude: drivers rate a hostile rider low first ("mediated retaliation") in the belief that the platform discounts symmetric low ratings, and workers file disputes they call "Hail Mary" moves (PDF pp. 26–29). + +Two findings bear on learning. "Some workers thus believe the rating algorithm was not an actual performance evaluation to help them learn how to improve but, instead, a reflection of customer whims" (PDF p. 17). And what workers knew of the algorithm they inferred from experience: "While drivers did not have firsthand information about how the algorithmic rating system worked, they developed last-ditch efforts" (PDF p. 27). The best workers develop "deep knowledge of the labor process," including "deciphering how the platforms' algorithms collect and use their data" (PDF p. 32). The tactics in the first two stages "largely rely on fragile information asymmetries" that the platform could remove by changing its monitoring (PDF p. 34). + +Scope. The paper studies individual, covert resistance. It contrasts itself with work on "collective resistance in platform settings," saying "we know less about how individual workers resist" (PDF p. 4), and lists "covert resistance enacted at an individual level" as a boundary (PDF p. 38). + +## Relation to the argument +**Ruling D7 and premise P9.** The paper is qualitative support for learning a mediator by acting on it: workers form working theories of matching and rating from what happens to them and act on those theories. It also supports P9's condition and Part 1's warning: the one feedback signal the platform gives, the customer rating, was judged by some workers to reflect customer whims rather than to help them improve (PDF p. 17). Those workers learned around the signal, not from it. + +**Ruling D8 and premise P12.** The passage saying "Platforms could, of course, change this policy at any moment" (PDF p. 34) is the owner's unilateral power over the convention, observed from the worker side. The authors add that after data collection FindWork "implemented an algorithm that reminded workers and customers that they could not move their conversations off platform," though it "stopped short of taking action" against violators (PDF p. 34). + +**What it does not support.** Collective learning or collective pushback. The workers' theories travel through discussion boards (many quotations are coded "DB"), but the analysis is of individual tactics. + +**Use in DRAFT_v2:** Line 266 cites the paper for "Workers pool what they observe and push back together (Cameron & Rahman, 2022)." Not supported as written. The paper says it studies individual covert resistance and sets collective resistance aside (PDF pp. 4, 38). Its data do include discussion-board posts in which drivers share tactics (e.g., the airport-queue cancellation trick, PDF p. 20), which would support "workers share what they observe" but not "push back together." For pooled theories and collective action, the draft's neighbours on line 266 (Lee et al. 2015; Kellogg, Valentine & Christin 2020) or the works this paper cites for collective resistance (Chen 2018; Lei 2021; Rizzo & Atzeni 2020) are closer. The paper does support the line's preceding sentence, that drivers "form working theories from what the system does in response to their own moves" (PDF pp. 27, 32), though the draft does not cite it there. + +## In conversation with +- `lee2015working` — cited (Lee et al. 2015) on drivers countering nudges by hiding "in plain sight" (PDF p. 8). +- `kellogg2020algorithms` — cited on the depth and breadth of algorithmic control (PDF p. 4). +- `rahman2021invisible` — the second author's study of opaque algorithmic evaluation, cited on visibility and ratings (PDF pp. 3, 7). +- `rosenblat2016algorithmic` — cited on power asymmetries in platform work (PDF p. 33). +- `cameron2022making`, `cameron2024making` — the first author's related studies (the reference list cites a 2021 manuscript version of "Making out while driving"). +- Leads not in the bib: Maffie (2020, *ILR Review*) on "laundering control through customers"; Lei (2021, *ASR*) and Chen (2018, *New Media & Society*) on collective contention; Vallas & Schor (2020, *Annual Review of Sociology*); Christin (2020) on ethnography of algorithms. + +## Caution +An accepted manuscript, not the version of record; journal page numbers are not available in my copy. Qualitative, two US platforms, workers' accounts; the authors had no data on how either platform's algorithms were developed (PDF pp. 36–37), so statements about the algorithm are workers' beliefs. RideHail data pool up to four ride-hailing companies (PDF p. 9, n. 1). No measured outcomes. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** corrected + +**What I checked:** Re-opened the writer's saved text of the Wharton author manuscript ("Conditionally Accepted at Organization Science", 46 pages, dxd_foundation_local/pdfs/R4/cr.txt) and located every quotation by page (form-feed page breaks): the research question and the collective/individual contrast (p. 4), the design figures (interviews n = 107 with 63 drivers, 112 rides, 22 focus-group participants, 200 communication records 2013–2014, 59 + 18 interviews, pp. 11–13; four companies pooled, p. 9 n. 1), the Figure 1 model (pp. 15–16), "attitude test" (p. 19), the "promise of 5 stars" and the airport-queue tactic coded RH-Zeke-DB (p. 20), splitting and hostage tactics (pp. 23–25), mediated retaliation and "Hail Mary" (pp. 26–27), "did not have firsthand information" (p. 27), "deep knowledge of the labor process" (p. 32), "fragile information asymmetries" and "could, of course, change this policy" (p. 34), "covert resistance enacted at an individual level" (p. 38). Checked the record against Crossref (Cameron & Rahman, *Organization Science* 33(1), 38–58, January 2022). Re-read DRAFT_v2.md lines 266 and 523. + +**Findings:** +1. "Workers trusted ratings least because they were 'a reflection of customer whims'" overstated the source. The paper says "Some workers thus believe the rating algorithm was not an actual performance evaluation to help them learn how to improve but, instead, a reflection of customer whims" (p. 17), and has no "least". Replaced with the exact sentence, and qualified the matching claim under D7/P9 to "some workers". +2. "FindWork later introduced an algorithm warning users": the paper says that after data collection the second author observed an algorithm that "reminded" users not to move conversations off platform, and that the platform "stopped short of taking action" (p. 34). Reworded with the quotation. +3. The limitation on internal algorithm data runs from PDF p. 36 to p. 37. Locator corrected. +4. The line 266 verdict ("Not supported as written") holds: the paper sets collective resistance aside (p. 4) and bounds itself to individual covert resistance (p. 38). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/caperton2009massey.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/caperton2009massey.md new file mode 100644 index 00000000..7f507b20 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/caperton2009massey.md @@ -0,0 +1,39 @@ +# *Caperton v. A.T. Massey Coal Co.*, 556 U.S. 868 (2009). + +**Identifier:** https://www.law.cornell.edu/supct/html/08-22.ZS.html (syllabus) with the opinion and dissents at `.../08-22.ZO.html`, `.ZD.html`, `.ZD1.html` · **Read depth:** full_text (the syllabus, Justice Kennedy's opinion of the Court, Chief Justice Roberts's dissent and Justice Scalia's dissent, from Cornell LII HTML; LII prints no U.S. Reports pagination, so locators are the opinion's Parts I–IV and the slip-opinion pages the syllabus cites, "Pp. 6–20") · **Source-tier:** court opinion (U.S. Supreme Court, 5–4) · **Evidence basis:** direct_read · **Design and sample:** none; a due-process ruling on judicial recusal · **Parties modeled:** two parties + third party (litigants and the adjudicator; a donor who helped choose the adjudicator) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +A West Virginia jury awarded Caperton $50 million against Massey for fraudulent misrepresentation, concealment and tortious interference (Part I). Massey's chairman, Don Blankenship, then spent "$3 million" to help elect Brent Benjamin to the state's high court, "more than the total amount spent by all other Benjamin supporters and three times the amount spent by Benjamin's own committee" (Part I; the opinion's own sums are "almost $2.5 million" to a §527 group, "just over $500,000" in independent expenditures, and $1,000 direct). Benjamin won 382,036 to 334,301 votes (53.3% to 46.7%). He refused to recuse three times, and the court reversed the verdict 3–2, twice. + +The Court holds: "in all the circumstances of this case, due process requires recusal" (opening). Its reasoning is that "objective standards" implement due process because the judge's own search for actual bias "is not one that the law can easily superintend or review" (Part III). It draws on *Tumey v. Ohio* (a mayor-judge paid only on conviction), *Ward v. Monroeville*, *Aetna v. Lavoie*, *In re Murchison* ("no man can be a judge in his own case") and *Mayberry*. The test is whether, "under a realistic appraisal of psychological tendencies and human weakness," the interest "poses such a risk of actual bias or prejudgment that the practice must be forbidden" (Part III, quoting *Withrow*). The Court does "not question his subjective findings of impartiality" and does not decide that Benjamin was biased (Part III). The rule: "a serious risk of actual bias—based on objective and reasonable perceptions—when a person with a personal stake in a particular case had a significant and disproportionate influence in placing the judge on the case by raising funds or directing the judge's election campaign when the case was pending or imminent" (Part III). The inquiry looks at the contribution's size against the total contributed and spent and "the apparent effect" on the outcome, not whether it was "a necessary and sufficient cause" (Part III). Part IV says the case is "extreme," that most disputes will be settled by judicial codes (ABA Canon 2, 28 U.S.C. §455(a), state canons), and that the constitutional standard "will thus be confined to rare instances." + +Roberts (joined by Scalia, Thomas, Alito) dissents: until now due process required recusal only for a financial interest and for certain criminal contempts, and the new "probability of bias" standard "provides no guidance"; he lists 40 questions it leaves open, including "How much money is too much money?" He also argues the spending was independent and the margin comfortable (7 points). Scalia's separate dissent predicts "vast uncertainty" and a new "Caperton claim." + +## Relation to the argument + +Ruling D8 asks how courts hold their conventions, to compare with FX markets and platforms. *Caperton* supplies a precise court answer to premise P10's concern about an interested party: when a third party who decides a dispute can be tempted by an interest in the outcome, the rule does not wait for proof that the interest worked. It prohibits the arrangement on the risk. The test is deliberately objective and does not rely on the decider's own account of his motives, which the Court calls hard for anyone to review. A platform's reviewer asked to judge its owner's own rule is in the structurally analogous position, and the paper can say so as an analogy, not as a holding. The case also gives a floor and a ceiling: due process sets the "outer boundaries," and codes of conduct do the daily work (Part IV), which fits the paper's interest in who writes the convention. The dissents show the cost: an objective but vague standard invites strategic claims (Roberts, Part III.A). + +Facet check: the J-facet marked this "read in part (one sentence)" and quoted *Caperton* on recusal from the syllabus-level passage. The opinion supports that use; nothing in the full text contradicts the facet. The facet's point about *Tumey* is confirmed: *Tumey* is the lead case in Part II.A. + +## In conversation with + +From the opinion's own citations: +- `tumey1927ohio`: the principal precedent; the opinion quotes its "possible temptation" test (Part II.A). +- `us455usc` (28 U.S.C. §455(a)): quoted in Part IV as the federal "impartiality might reasonably be questioned" standard. +- `mathews1976eldridge`: not cited by *Caperton*; it concerns the procedure owed, whereas *Caperton* concerns the neutrality of the decider. +- Not in the bib: *Withrow v. Larkin*, 421 U.S. 35 (1975); *Aetna Life Ins. v. Lavoie*, 475 U.S. 813 (1986); *In re Murchison*, 349 U.S. 133 (1955); *Ward v. Monroeville*, 409 U.S. 57 (1972); *Republican Party of Minn. v. White* (2002). + +## Caution + +A constitutional holding about elected state judges and campaign spending, 5–4, on "extreme" facts the majority says are rare. It should not be cited for a general right to a neutral decider outside courts, or for platform moderation. Cite the U.S. Reports pagination from an official source before print, since the LII copy shows none. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the Cornell LII HTML copies of the syllabus, Kennedy's opinion, Roberts's dissent and Scalia's dissent (cards folder, four files); case name, number and date match. About 22 quotations and the vote counts, dollar figures (about $3 million, almost $2.5 million, just over $500,000, $1,000), the 382,036 to 334,301 result, the three recusal refusals and two 3-2 reversals, Roberts's 40 numbered questions and 7-point margin, and the Part locators (Roberts III.A confirmed; Tumey in Part II.A). + +**Findings:** No issues found. The card's opening-line quotation is in the syllabus and in the opinion's introduction. The "in conversation" statements about `mathews1976eldridge` (not cited in the opinion) and the leads were checked only as far as the opinion text I searched; I did not search the dissents for Mathews. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cert2022common.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cert2022common.md new file mode 100644 index 00000000..cd681c01 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cert2022common.md @@ -0,0 +1,56 @@ +# CERT National Insider Threat Center. (2022). *Common sense guide to mitigating insider threats* (7th ed.). Software Engineering Institute, Carnegie Mellon University. + +**Identifier:** https://www.sei.cmu.edu/documents/5777/5692_CSG_Book_20-optimized.pdf (the PDF prints "DM21-0782" on its notice page; I did not confirm that as an SEI report number, and the bib note flags CMU/SEI-2022-TR-003 as unverified) · **Read depth:** partial (S2 downgrade from full_text; the card's own list shows the bodies of most practices were not read). Read in full: the front matter, Introduction and Best Practices 10, 11, 12, 14, 15 and 17, and most of 21; the opening pages of every other practice and every keyword hit on monitoring, logging, separation of duties, least privilege and the two-person rule; skimmed or not read: the bodies of Best Practices 1–9, 13, 16, 18–20 and 22, the case studies outside those listed, and Appendices B–C. Published PDF, 175 pages; locators are the guide's printed page numbers · **Source-tier:** report (practitioner guidance from a federally funded research centre) · **Evidence basis:** direct_read · **Design and sample:** none as a study; guidance drawn from the authors' incident repository of over 3,000 insider cases (convicted or found liable) and their other research (Exec. Summary p. iv; Introduction p. 4) · **Parties modeled:** other (an organisation and its workforce, plus "trusted external entities" such as contractors and cloud providers) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +The guide offers 22 "best practices" for an insider risk management program (IRMP). It defines an insider threat as the potential for a person with current or former authorised access to critical assets to use that access, "either maliciously or unintentionally", to harm the organisation (p. 3), and it says insider threats come from "a combination of technical, behavioral, and organizational factors" (p. iv). Its primary scenarios are IP theft, IT sabotage, fraud, misuse of authorised access, unintentional incidents, espionage and workplace violence (p. 3). It also concedes limits: "not all insiders can be stopped", so the organisation must "sustain critical operations during and recover after" an event (p. 6). + +The parts that bear on this project run as follows. + +*Privileged users and the audit trail (Best Practice 11).* Privileged users can conceal their actions by logging in as other users, modifying system logs and falsifying "audit logs and monitoring reports" (p. 78). The remedies are non-repudiation (each action attributable to one person), having several privileged users build the security design, regular review of privileged activity by the information-security team, and two-person release rules for production changes (pp. 78–79). A small firm with one administrator should take the audit function away from that person: "any method used must assure that the system administrator has no control over the auditing function" (p. 79). Best Practice 17 adds that change logs and backups must be protected, since insiders "modified change logs to conceal their activity or implicate someone else" (p. 115), and that roles in change management be split "so that no one person can make a change unnoticed by others" (p. 115). + +*Monitoring (Best Practice 12).* "Simply logging all network activity is not sufficient"; the data volume "drives the need for tools that can fuse data", and monitoring should use "a risk-based approach" focused first on critical assets (p. 83). A SIEM (security information and event management) system gives a "centralized view" and filters "background noise", and with a rich data set it lets an analyst analyse past data "even when an alert was not generated", which matters most for people leaving the organisation (pp. 83–84). The guide recommends review of at least the previous 90 days when a person resigns (p. 84). Table 4 lists 25 technical and 15 non-technical data sources (pp. 84–87) and says the data sources "are not comprehensive enough to completely prevent or detect all insider threats in all organizations" (p. 87). Challenges are false positives, baselines at several levels, and getting enough context to analysts before alerts reach them (p. 88). Large organisations should have someone "hunt for critical incidents" below the top-priority alerts, which may take several staff monitoring workforce activity full time (p. 89). The case "An Ugly Picture" shows three alerts that no one correlated (p. 88). + +*Baselines (Best Practice 14).* A baseline of normal behaviour is the reference for anomaly detection, but "baselining current activities might incorporate ongoing and unrecognized malicious activity", and a single all-encompassing baseline "can conceal concerning behavior" (pp. 99–100). + +*Separation of duties and least privilege (Best Practice 15).* Separation of duties means dividing functions among several people so that no one can steal, defraud or sabotage without the others; the guide holds that people are "less likely to engage in malicious acts" if that means collaborating with someone else (p. 103). Least privilege means authorising people to use only the resources their jobs need, and it must be managed as moves and promotions occur (pp. 103–104). The guide adds that separation of duties is easily circumvented "if it is enforced by policy rather than by technical controls" (p. 104). Backup and restore, source-code release and auditing are the examples (pp. 103, 105; Best Practice 18, same logic). Every account must "be attributable to an individual" (Best Practice 10, p. 74). The guide maps these to NIST SP 800-53 controls AC-5, AC-6, AU-2, AU-6 and AU-9 (pp. 81, 89, 106). + +*The human side.* Best Practices 5, 8 and 21 treat insider risk as partly a product of how the organisation treats people: inconsistent enforcement breeds disgruntlement (p. 45), rushing in a high-stress environment causes mistakes (p. 61), and "positive incentives" (job engagement, perceived organisational support, connectedness) should sit beside "negative incentives that attempt to force compliance through constraints, monitoring, and punishment" (pp. 136–138). The aim is to move an IRMP "from a 'big brother' program to a 'good employer' program" (p. 138). + +## Relation to the argument + +Facet G cites this guide for insider-threat controls behind the argument map's P11 (the structure of Nadella's insider post) and P12 (people monitor badly). It supports the post's pairing of least privilege, separation of duties and logging with a human reader, and it is useful on four counts. + +It names the tampering problem that the post's "tamper-proof evidence" (its principle 15) answers. The people with the most access are the people best placed to falsify the record (p. 78), so the guide separates the auditor from the audited (p. 79). This is the closest security-practice analogue to the argument's claim that the party under oversight should not hold the controls, and it is stated for human administrators, not models. + +It shows the human review burden as an explicit part of the control design, and with more nuance than the facet gives. The guide does not ask for review of everything; it asks for risk-based monitoring around critical assets, correlation tools and baselines (p. 83). It also asks for after-the-fact review of the previous 90 days for leavers (p. 84) and for staff who "hunt" past the highest-priority alerts, in some settings full time (p. 89). That bears on the gap that the argument map leaves under P12 (reconstruction by trained auditors): the guide describes exactly such work as normal practice, but gives no evidence on how well people do it. + +It cuts against a strong reading of "observe everything". The guide says full logging alone does not work (p. 83) and lists its own cost, privacy and false-positive limits (p. 88). It also counts "unintentional" insiders in its definition (p. 3), which fits the project's treatment of models as risks without assuming malice. + +The victim in the definition is "the organization" (p. 3), and the program's stakeholders are management, HR, legal, security, IT, data owners and engineers (p. iv). Its definition and stakeholder list name no outside party affected by the insider's actions as a party to the program; customers appear only as victims in case studies (for example p. 100). That is the extension the paper needs. + +Facet check: the facet's quotations from the definition, the 3,000-incident base, and Best Practice 12 are accurate. The facet says the practice 12 passages "cut against reading observe everything (P15) as literal full-volume human review"; that is right for the opening, but the same practice supports retrospective human review and full-time monitors (pp. 84, 89), so the guide supports the post's principle 15 (after-the-fact reconstruction) more than the facet suggests. The facet read only the titles of practices 5, 8 and 21; I read the text of 21 and only the opening pages of 5 and 8, and they confirm the facet's description (stress, negative environments, positive incentives to align the workforce) without changing it. I found no direct contradiction. + +## In conversation with + +None of the guide's references is in `../references.bib`; the guide's bibliography cites NIST SP 800-53, the NIST Cybersecurity Framework, CERT-RMM (Caralli 2016), Costa 2017 (the definition), Hanley et al. 2011 (technical observations and centralised logging), Moore and colleagues on positive deterrence (2015–2021), Straub 1998 and the Ponemon Institute (2020, 2021). It does not cite Saltzer and Schroeder, Anderson, Bainbridge or the alert-fatigue literature. + +- `saltzer1975protection`: the principles of least privilege and separation of privilege that Practice 15 restates in organisational form. The relation is the project's, not the guide's. +- `rose2020zero`: the zero-trust architecture in facet G's lineage; the guide's per-request and attribution themes are similar but it never mentions zero trust. + +Leads, not in the bib: Costa et al., insider-threat definitions (2017); Hanley et al., *An Analysis of Technical Observations in Insider Theft of Intellectual Property* (2011a) and *Insider Threat Control: Using Centralized Logging to Detect Data Exfiltration Near Insider Termination* (2011b); Moore et al. on positive deterrence (2016, 2021); NIST SP 800-53 Rev. 5 (AC-5, AC-6, AU-9). + +## Caution + +Practitioner guidance, not a study. The "evidence-based" claim rests on an incident repository of publicly disclosed cases in which the person was "charged and convicted or found liable" (p. 4), a sample that selects for detected and prosecuted insiders, and the guide admits that reliable data are hard to get (p. 4). It gives no effect sizes for any control. Its examples concern human employees; it says nothing about AI models or agents. The figures on p. 5 (n = 1,314 incident types, n = 1,179 financial impacts, n = 1,515 victim organisations) are read from a graphic, and I cite none of them here. Dating is inconsistent: the cover notice says 2022, the history figure labels this edition 2021 and retitles it "Common Sense Guide to Managing Insider Risk" (p. 1), and the front matter calls it the seventh edition. Cite the printed pages, which differ from the PDF's by an offset. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the 175-page SEI PDF named in the card; title, publisher and edition match (the notice page prints DM21-0782 and the history page labels the edition 2021, as the card says). About 30 quotations and numbers with printed page locators (PDF page minus 7), including the 3,000-case base, Table 4 (25 technical and 15 non-technical sources, pp. 84-87), the 90-day review, the BP 11, 12, 14, 15 and 17 quotations, NIST control references AC-5 and AU-9, and the pages for BP 5 (p. 45) and BP 8 (p. 61). + +**Findings:** (1) Read depth changed from full_text to partial: the card lists the bodies of Best Practices 1-9, 13, 16, 18-20 and 22 as skimmed or unread. (2) The Table 4 quotation was altered: the source says the data sources "are not comprehensive enough to completely prevent or detect all insider threats in all organizations", not "the list is"; fixed. (3) "centralized view" is on p. 83 and "background noise" on p. 84; locator widened to pp. 83-84. (4) The facet-check line said I read the text of practices 5 and 8; the read-depth line says their bodies were not read, so I corrected it to "opening pages". I could not check the n values on p. 5 (graphic, which the card says it does not use) or the "none of the references is in references.bib" statement. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cftc2017fxcm.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cftc2017fxcm.md new file mode 100644 index 00000000..a1f1896e --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cftc2017fxcm.md @@ -0,0 +1,34 @@ +# U.S. Commodity Futures Trading Commission. (2017, February 6). *CFTC orders Forex Capital Markets, LLC (FXCM), its parent company, FXCM Holdings, LLC and FXCM's founding partners, Dror Niv and William Ahdout, to pay a $7 million penalty for FXCM's defrauding of retail forex customers* (Release 7528-17) [Press release]. + +**Identifier:** https://www.cftc.gov/PressRoom/PressReleases/pr7528-17 · **Read depth:** full_text (the whole press release as HTML; the underlying CFTC Order was linked but not opened) · **Source-tier:** regulator press release (summary of a settled order) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** two parties + a platform operator that was also the counterparty (retail customers, FXCM, an affiliated market maker; CFTC as third party) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The release says the CFTC's Order "finds that, between September 4, 2009 though at least 2014 (the Relevant Period), FXCM engaged in false and misleading solicitations of FXCM's retail foreign exchange (forex) customers by concealing its relationship with its most important market maker and by misrepresenting that its 'No Dealing Desk' platform had no conflicts of interest with its customers." FXCM and two founders pay a $7 million penalty jointly and severally and agree to withdraw from registration and never to seek it again. FXCM, its parent and the CEO are also held responsible for false statements to the National Futures Association (NFA). + +The mechanism, as the release gives it: FXCM told customers that on its No Dealing Desk platform it was "merely that of a credit intermediary," that profits or losses of customers would have no effect on its bottom line, and that risk was borne by independent market makers. In fact FXCM "had an undisclosed interest in the market maker that consistently 'won' the largest share of FXCM's trading volume – and thus was taking positions opposite FXCM's retail customers." The market maker came from a 2009 plan to build an algorithmic system; it was spun off but stayed "closely aligned," got "special trading privileges," a no-interest loan, space in FXCM's offices and FXCM employees. It rebated "approximately 70 percent of its revenue" to FXCM, about $77 million from 2010 through 2014. The release adds that FXCM omitted these details in a meeting with the NFA. The two founders are held liable as "controlling persons." + +## Relation to the argument + +This is a case of an intermediary that presented itself as neutral ("credit intermediary," "no conflicts") while holding an interest in the counterparty to its customers' trades. It supports premise P10 and ruling D8's comparison: in retail FX, as on a platform, the middle party's neutrality was a claim, and what corrected it was a regulator's order, not any right the customer held. It also marks a limit on disclosure as the remedy (ruling D3): the CFTC's quoted statement is that "full and truthful disclosure to customers" is "vital to the integrity" of markets, but here nondisclosure was the offence and customers could not have read what was withheld. The release does not say customers lost money in a measured amount, and it does not say how the customers' fills compared with others'. + +Facet check: the J-facet read only the lead paragraph and quoted it accurately ("concealing its relationship," "$7 million," withdraw from registration). The full release adds the 70% rebate and $77 million figure, which the facet does not use. + +## In conversation with + +- `nydfs2015barclays` and `fca2014fines`: other regulator actions on FX dealers' conduct toward clients; the release cites neither. +- The release cites no scholarship. It names the Commodity Exchange Act and the NFA as the regime; neither is in the bib. + +## Caution + +A press release, not the Order. The findings are those of a settled order, and the release does not say the respondents admitted or denied them. It gives no customer loss figure. The release names retail forex customers, not institutional dealers, so it should not be generalised to the wholesale market the FCA and NYDFS documents cover. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the saved HTML of CFTC Release 7528-17 (6 February 2017) from the cards folder; title, date and release number match. Every quotation in the card (Relevant Period sentence, "merely that of a credit intermediary", "undisclosed interest", "won", "closely aligned", "special trading privileges", "approximately 70 percent", "$77 million", "controlling persons", the Lowe quotation) and the penalty, withdrawal, NFA and CEO statements. + +**Findings:** No issues found. The card's "the CFTC's quoted statement" is a quotation from the Division of Enforcement's Principal Deputy Director, Gretchen Lowe, with "Full" capitalised in the source; this is a wording nuance only. The release does not state whether respondents admitted or denied the findings, as the card says. The linked Order was not opened, as the card states. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chan2024visibility.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chan2024visibility.md new file mode 100644 index 00000000..0810373b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chan2024visibility.md @@ -0,0 +1,44 @@ +# Chan, A., Ezell, C., Kaufmann, M., Wei, K., Hammond, L., Bradley, H., Bluemke, E., Rajkumar, N., Krueger, D., Kolt, N., Heim, L., & Anderljung, M. (2024). Visibility into AI agents. In *Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24)* (pp. 958–973). ACM. + +**Identifier:** doi:10.1145/3630106.3658948 · arXiv:2401.13138 · **Read depth:** full_text (arXiv v6, 17 May 2024; locators are PDF pages of this 16-page layout, not the proceedings pages 958–973; body and Appendix A read, reference list searched for the works named below) · **Source-tier:** proceedings paper (policy analysis) · **Evidence basis:** direct_read · **Design and sample:** none; a conceptual assessment of three governance measures, no data · **Parties modeled:** other: deployer, user, counterparty or tool provider, and regulator around one agent · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +The authors define visibility as "Information about where, why, how, and by whom certain AI agents are used" (p. 1) and prefer it to "transparency" (footnote 5, p. 3). They assess three measures, each at several levels of intrusiveness: agent identifiers, real-time monitoring and activity logs (p. 3, Fig. 2). + +- **Identifiers** mark which agent took part in an interaction. They can be visible to "the bank, the other party in the transaction, or the service provider" (p. 5), and can carry an "agent card" of evaluations, permissions, goals and responsible actors (pp. 6–7; Appendix A, p. 16). +- **Real-time monitoring** "is likely to be automated so as to be able to cope with the speed and scale of agent activities" (p. 7). It suits clear-rule violations. A user "could be given the opportunity to request human review", but the authors write: "We leave details about the implementation challenges of automation and human review to other work" and cite four works there, among them Green (2022) and three on algorithmic recourse (refs. 72, 95, 157, 166; p. 7). +- **Activity logs** support "post-incident attribution and forensics" (p. 7) and may need keeping long after an agent's life (p. 8). + +All three assume a deployer: "Since deployers run agents, the inputs and outputs of agents are by default visible to the deployer" (Fig. 1 caption, p. 5). Section 4 extends the measures to agents that bypass deployers, through compute providers and through tool providers that refuse unidentified agents; the authors warn that both routes concentrate power and invite surveillance (pp. 8–9). They are not "advocating for immediate implementation" (p. 3). The conclusion limits the whole proposal: visibility "is not by itself sufficient", because those harmed "may not have the power to intervene and reduce risks", a point they take from Ananny and Crawford (p. 10). + +Among their risks, agents "could also subtly benefit their developers, akin to the self-preferencing behaviour of large-scale digital platforms" (p. 2). + +## Relation to the argument +**D3/D4 (disclosure versus control).** The paper is a disclosure-and-record proposal, and it states its own limit: knowing who runs which agent does not give the affected party power. That supports the author's D3 in spirit (visibility is a different thing from the ability to act), but it does not support D4's "his only move is disclosure": the AI-governance lineage that Nadella's post sits in pairs visibility with filtering and intervention (real-time monitoring, p. 7). The 2025 sequel (`chan2025infrastructure`) adds explicit control layers. + +**P10/D8 (who holds the convention).** The architecture hands the deployer the data and the switch: identifiers are attached by the deployer, monitoring is run by the deployer, logs are kept by the deployer (p. 5). The self-preferencing line (p. 2) and the concentration warnings (pp. 8–9) name the conflict of interest that P10 needs, though only as a risk and without a test. + +**P11/P12.** Chan separates automated monitoring from human review and leaves human review unexamined (p. 7). Nadella's P15 puts a person at that point; this paper supplies a gap, not a result. + +**Triad (D9).** Counterparties appear only as recipients of an identifier (p. 5); the agent is assumed to serve its user. + +**Facet check:** the facet reading (identifier visible to the other party; human review left to "other work", citing Green) holds. The facet says Chan reaches counterparties where the post faces the enterprise only; in Chan the counterparty sees the identifier and agent card, not the monitoring or logs. The bib note says sections 4–5 and the appendix were unread; I read them. + +## In conversation with +- `green2022flaws`: cited for the unsolved problem of human review of flagged actions (p. 7). +- `ananny2018seeing`: cited for the point that those harmed may lack power to act on what they can see (p. 10). +- `greenblatt2024ai`: cited (with ref. 147) for the claim that real-time monitoring is likely to be automated (p. 7); the paper's own example of a preliminary monitoring system is Naihin et al. +- Leads not in the bib: Shavit et al. (2023) on practices for governing agentic systems; Raji et al. (2022) on outsider oversight and auditing; Kolt (2024) on governing AI agents. + +## Caution +A policy analysis with no data; it lists options and says it does not choose among them. It covers agents run by deployers and, briefly, decentralised ones, not people coordinating through a platform's algorithm. Cite it for the measures and for the authors' admitted limit, not for any claim that visibility changes what people do. The arXiv v6 text may differ in small ways from the FAccT proceedings version. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the arXiv v6 PDF (16 pages) from the cards folder; title, authors, version match. About 16 quotations with PDF-page locators (definition, footnote 5, deployer sentence, automated monitoring, human review deferral, advocating, "not by itself sufficient", self-preferencing), and the named references (Green, Ananny and Crawford, Greenblatt, Shavit, Raji, Kolt) against the text and reference list. + +**Findings:** Two attribution errors fixed. (1) The human-review deferral cites four works (refs. 72, 95, 157, 166: Green 2022 plus three algorithmic-recourse papers), not Green alone. (2) Greenblatt et al. (ref. 74) is cited for the claim that monitoring will be automated, not as "an existing real-time monitoring scheme"; the paper's preliminary-implementation example is Naihin et al. All quotations are exact. I did not recheck the claim that the bib note said sections 4-5 were unread. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chan2025infrastructure.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chan2025infrastructure.md new file mode 100644 index 00000000..c16aae4b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chan2025infrastructure.md @@ -0,0 +1,52 @@ +# Chan, A., Wei, K., Huang, S., Rajkumar, N., Perrier, E., Lazar, S., Hadfield, G. K., & Anderljung, M. (2025). *Infrastructure for AI agents* [Preprint]. arXiv:2501.10114. + +**Identifier:** arXiv:2501.10114 · **Read depth:** full_text (arXiv v3, 19 Jun 2025; 31 PDF pages, locators are PDF pages; body read in full, reference list searched for the works named below) · **Source-tier:** preprint (policy and design analysis) · **Evidence basis:** direct_read · **Design and sample:** none; a catalogue of research directions · **Parties modeled:** other: user, agent, counterparties (services, humans, other agents) and the legal system, with infrastructure external to the agent · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +The authors propose "agent infrastructure": "technical systems and shared protocols external to agents that are designed to mediate and influence their interactions with and impacts on their environments" (abstract, p. 1). Their reason is that changing the agent's own behaviour "may not be sufficient" (p. 2). The traffic analogy: driver training is a system-level intervention; "traffic lights, roundabouts, emergency lanes, and camera-enforced speed limits" are infrastructure (p. 4). + +They sort nine kinds of infrastructure into three functions (Table 1, p. 3): +- **Attribution:** identity binding (tie an agent to a legal person), certification, agent IDs (pp. 6–10). +- **Interaction:** agent channels that separate agent traffic, oversight layers, inter-agent communication, commitment devices (pp. 11–17). +- **Response:** incident reporting and rollbacks (pp. 17–19). + +Each entry has the same headings: description, functions, existing alternatives, adoption, limitations, research questions. The text is a research agenda; it reports no experiment. + +The parts most relevant here are the human-attention and incentive limits: +- *Oversight layers* (p. 13) are a monitoring system plus an interface for an actor to intervene. The actor "could be a human user, but not necessarily". Limitation: "Active attention is required if oversight layers flag potentially problematic behaviour. Humans could ignore such flags if they are too frequent or present information in a complex manner" (p. 14). On supply: "it is unclear if the market will provide them to a socially optimal degree", with social-feed controls as the example (p. 14). +- *Certification* says agents "with more regular human oversight could be, but are not necessarily" more trustworthy, citing Parasuraman and Riley, Skitka et al. and Green (2022) (p. 8). +- *Rollbacks* raise moral hazard: "users could have weaker incentives to properly monitor their agents" (p. 19). +- *Interoperability and lock-in* (p. 20): a developer can limit its reputation system's interoperability to shut out competitors, and a bad protocol is hard to leave. +- The accountability example wants "at least one human that can answer for the agent's actions" (p. 5). + +The conclusion calls infrastructure "not itself a complete solution, but ... a platform for policies and norms" (p. 21). + +## Relation to the argument +**D4 and P11.** The post shares this paper's premise (controls outside the model) and part of its architecture: Nadella's P11–P12 and P17 resemble oversight layers, P19 a rollback or halt, P15 an activity log. This is the closest academic counterpart to the post, and it shows that the control-outside-the-agent move is a design programme, not mere disclosure. That supports the SOURCE.md flag on D4 (the post is not "only" disclosure). + +**P12.** Chan names the weak point itself: a human who must attend to flags. It cites automation-bias work but reports no data (p. 14), so it supports P12 as a stated risk only. + +**P10/D8.** Pages 14 and 20 argue that those who build the controls may not supply the ones users need, and that interoperability and lock-in favour dominant players. That is the nearest published support for the step the argument map marks `[inf]`, from conflict of interest to "cannot withdraw alone". It is an argument by analogy, not a finding. + +**Triad (D9).** Counterparties are in the figure (Fig. 1, p. 2). The agent is still the user's agent, so the mediator with an objective of its own is not modelled. + +**Facet check:** the facet cites this paper only for the definition (abstract-only read). The full text also carries the oversight-layer, moral-hazard and lock-in material above, which bears on P10 and P12 more than the facet shows. + +## In conversation with +- `chan2024visibility`: the earlier paper, which the Agent ID section draws on (p. 10); Section 7 says it proposed infrastructure for visibility whereas this paper takes "a broader view" (p. 21). +- `green2022flaws`: cited on the point that more human oversight is not necessarily more trustworthy (p. 8). +- `greenblatt2024ai`: cited for trusted automated monitors replacing the human in an oversight layer (pp. 13–14). +- Leads not in the bib: Parasuraman and Riley (1997) on use, misuse, disuse and abuse of automation; Shavit et al. (2023) on governing agentic systems; Kolt (2024) on governing AI agents. + +## Caution +A preprint agenda with no data. It treats the agent as the user's agent and does not address a mediator that serves its owner. Its human-factors claims rest on citations, not tests. Version 3 is not necessarily the version of record if the paper is later published. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the arXiv v3 PDF (31 pages, 19 Jun 2025) from the cards folder; title, authors, version match. About 14 quotations with PDF-page locators (definition, "may not be sufficient", traffic analogy, oversight-layer and moral-hazard passages, certification, "at least one human", conclusion, "broader view"), the nine-item structure of Table 1, the interoperability passage (Section 6.2), and the named references (Parasuraman and Riley, Skitka, Green 2022, Greenblatt, Shavit, Kolt, Chan 2024a;b). + +**Findings:** One quotation used American spelling where the source has "behaviour" (oversight-layer limitation, p. 14); fixed to the source's spelling. All other quotations, locators and attributions match, including the Greenblatt citation for trusted automated monitors replacing the user (p. 13). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chantavarin2025effects.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chantavarin2025effects.md new file mode 100644 index 00000000..778f4973 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chantavarin2025effects.md @@ -0,0 +1,38 @@ +# Chantavarin, S., & Leung, T. T.-C. (2025). The effects of interword spacing and morphological complexity in reading Thai: An eye-tracking study. *Language and Cognition, 17*, e69, 1–28. + +**Identifier:** doi:10.1017/langcog.2025.10028 · **Read depth:** full_text (open-access publisher PDF, 28 pages, journal pagination). I read the text, tables and discussion; I skimmed the reference list for the names the project cites and did not read the appendix sentence lists · **Source-tier:** peer-reviewed journal article · **Evidence basis:** direct_read · **Design and sample:** eye-tracking experiment, 55 Thai-native students from Chulalongkorn University, 76 sentence frames in four conditions (spaced/unspaced × compound/simple noun; 74 items analysed), silent reading, EyeLink Portable Duo at 1,000 Hz, mixed-effects models · **Parties modeled:** person–artefact · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Thai is an alphabetic script written without spaces between words. Chantavarin and Leung test whether adding spaces helps adult native readers, and whether it helps more on compound nouns, where the unspaced boundary is hardest to find. They insert two consecutive spaces between words, with word boundaries set by one native-speaker researcher on a conservative criterion (§2.2, p. 8). The design is 2 × 2 and every participant read 76 sentences. + +Results: spacing made the sentence slower and the words faster. Sentence reading time was longer with spaces for compounds (3,804 ms against 3,676 unspaced) and for simple nouns (3,675 against 3,546), with more fixations and shorter average fixations (Tables 2–3, pp. 11–12; the spacing estimate for sentence time is 132 ms, SE 35). On the target noun, first fixation duration fell from 330 to 297 ms for compounds and from 219 to 210 ms for simple nouns; gaze duration fell from 352 to 315 and from 272 to 260 ms (Table 2, p. 11). The spacing × noun-type interaction was 24 ms on first fixation and 25 ms on gaze duration (p. 17), and the authors report 93.2% power for it (n. 1, p. 13). Spacing moved the landing position toward the word centre only on compounds (§3.4, pp. 16–17). They conclude that spaces give "salient word boundary information" that helps early and late word processing, most for compounds (p. 20). Readers are not just tolerating spaces: "Since Thai script's orthography is 'unspaced', the observed facilitative effect of space information is not due to familiarity with spaced text" (p. 17). + +The introduction summarises the field. Removing spaces from spaced scripts disrupts eye movements in English (p. 2) and, in the cited studies, German, Spanish and French (p. 3). Baek et al. (2022) found the same for Korean, "though not to the same extent as in English" (pp. 2–3). Leung et al. (2021) found no significant loss of readability for Arabic, which the authors tie to cursive letter forms that already mark words (p. 3). In unspaced Chinese, Japanese and Thai, adding spaces shortens fixation durations on words but lengthens total reading time (p. 3). + +## Relation to the argument +It qualifies premise P4 (`ARGUMENT_MAP.md`: "Readers fluent in unspaced scripts gain nothing from added spaces"). Adult native Thai readers do gain at the level of the word, on fixation and gaze time, and lose at the level of the sentence. "Gain nothing" holds only for whole-sentence speed. The result also tells against the familiarity account in Chen et al. (`chen2021tradeoff`), at least for Thai: the authors say the benefit cannot be due to familiarity with spaced text. + +For P3 and P1 it adds a modern case in which a native reader of an unspaced alphabetic script uses spaces at the word level, which cuts a little against Heilmann's and Vatri's reading of Thai (`heilmann2020reading`, p. 181: Thai eye movements the same as English). Heilmann takes the Thai landing-position result from Winskel et al. (2009); Chantavarin and Leung find that landing position does shift with spacing for compounds, and they put the mismatch down partly to earlier studies choosing target words without regard to morphological structure (p. 19). It shows what spacing does for habituated readers of one script; it cannot show what it did for ancient readers or for learners. + +Facet check: the facet D entry gives only the reference, so there is no claim to contradict. It writes "et al."; the PDF confirms two authors, and `references.bib` already lists both. My reading qualifies `ARGUMENT_MAP.md` P4, as noted above. + +## In conversation with +- `winskel2009eye`: the first Thai eye-tracking study; this paper replicates its word-level effects and departs from it on first fixation and landing position (p. 18). +- `kohsom1997adding`: faster read-aloud with spaced Thai (p. 3). +- `bai2008reading`, `zang2013interword`, `sainio2007role`, `kasisopa2013eye`: unspaced-script studies it builds on (pp. 2–3). +- `baek2022korean`: Korean spacing removal (pp. 2–3). +- `rayner1998unspaced`, `raynerpollatsek1996reading`, `perea2009space`, `mirault2019reading`: spaced-script studies on removing spaces (pp. 2–3). +- Leads not in the bib: Li and Pollatsek 2020 (Chinese reading model); Liu et al. 2024 (Chinese E-Z Reader); Inhoff, Radach and Heller 2000 (German compounds: spaces help segmentation but hinder meaning); Kasisopa et al. 2016; Winskel et al. 2012. + +## Caution +One lab, one sample of university students, one-line sentences, and bimorphemic transparent compounds only. The "spaced" condition used double spaces, which is wider than the single space in English. The authors say word segmentation involved subjectivity and previous Thai studies likely used different cut-offs (p. 18–19). The effects are tens of milliseconds on words, with a sentence-level cost in the opposite direction, so they cannot be cited as showing that spacing speeds Thai reading. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the open-access publisher PDF (28 pages, Language and Cognition 17, e69) from the cards folder; title, authors, DOI match. About 30 numbers and quotations: Table 2 means (3676/3804, 3546/3675, 330/297, 219/210, 352/315, 272/260), Table 3 spacing estimate (132.46, SE 34.61), the 24 ms and 25 ms interactions, 93.20% power (footnote 1, p. 13), 55 participants, 76 items (74 analysed), 1,000 Hz, double spaces and the conservative segmentation (p. 8), the two quotations (pp. 17, 20), Baek and Leung/Arabic statements (p. 3), and the Winskel differences (p. 18). + +**Findings:** One locator fixed: German, Spanish and French appear on p. 3 (as studies finding facilitative spacing effects), not p. 2; the card's sentence is now split between pp. 2 and 3. All numbers and quotations are exact. The reference-list names in "in conversation" were skimmed only, as the card says; I confirmed Kohsom and Gobet, Winskel, Inhoff and Kasisopa appear. I did not check `heilmann2020reading` (p. 181) in its own source. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/charlemagne1900letter.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/charlemagne1900letter.md new file mode 100644 index 00000000..e7c03b69 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/charlemagne1900letter.md @@ -0,0 +1,39 @@ +# Charlemagne. (1900). Letter to Abbot Baugulf of Fulda (*De litteris colendis*), c. 780–800 (D. C. Munro, Ed. & Trans.). In *Laws of Charles the Great*, *Translations and Reprints from the Original Sources of European History* (Vol. 6, No. 5, pp. 12–14). Department of History of the University of Pennsylvania. Latin text in A. Boretius (Ed.), *Capitularia regum Francorum* I (1883), no. 29, pp. 78–79 (heading and editor's introduction p. 78, letter p. 79). + +**Identifier:** https://sourcebooks.fordham.edu/source/carol-baugulf.asp (Fordham Internet Medieval Sourcebook text of Munro's translation) · **Read depth:** full_text of the English translation (the whole letter as given on the Fordham page, which ends with the source line "In Boretius, No. 29, p. 78, trans. by D. C. Munro ... Vol. VI, No. 5, pp. 12-14"). S2 added Munro's printed pages (Internet Archive `translationsrepr56unse`, OCR; the Fordham text drops one clause, "in addition to the order of monastic life and the intercourse of holy religion", p. 12) and the Latin in Boretius (Internet Archive `capitulariaregum01bore`, OCR, pp. 78–79) · **Source-tier:** primary source (royal letter, in translation) · **Evidence basis:** direct_read (translation) · **Design and sample:** none (a letter) · **Parties modeled:** other (ruler, abbot and monastic communities; the reader of Scripture as beneficiary) · **Facets:** O · **Read:** 2026-10-14 + +*Agent-written card (round 4, claim-verification stage for §3). The author has not reviewed it.* + +## What it argues + +The king tells Abbot Baugulf that bishoprics and monasteries should be zealous "in the culture of letters" (p. 12), so that "those who desire to please God by living rightly should not neglect to please him also by speaking correctly" (p. 13; Latin "ut, qui Deo placere appetunt recte vivendo, ei etiam placere non negligant recte loquendo", Boretius p. 79). The trigger is the monasteries' own correspondence: in letters sent to the court "we have recognized in most of these letters both correct thoughts and uncouth expressions" (p. 13; Latin "et sensus rectos et sermones incultos"), because "the tongue, uneducated on account of the neglect of study, was not able to express in the letter without error" (p. 13). The stake is Scripture: the king fears that "as the skill in writing was less, so also the wisdom for understanding the Holy Scriptures might be much less" (p. 13; Latin "sicut minor erat in scribendo prudentia"). He asks for teachers chosen for "the will and the ability to learn" (p. 13), and enforces the request through favour: "Do not neglect, therefore, if you wish to have our favor, to send copies of this letter to all your suffragans and fellow-bishops and to all the monasteries" (p. 14). + +The letter is about correct Latin composition and study. It says nothing about copying books, scriptoria, script, layout, punctuation or word separation. + +## Relation to the argument + +**P10 and ruling D8 (who stood behind the page).** The letter shows a ruler setting a standard of correct written language and judging monastic writing against it, which supports the draft's point that an authority stood over the writers. It does not show an order to scriptoria to produce correct *books*; that comes from the *Admonitio generalis* c. 72 (`boretius1883capitularia`) and from Alcuin's poem (`dummler1881alcuini`). + +**Use in DRAFT_v2:** +- **Line 123** ("Charlemagne's letter *De litteris colendis* measures monasteries' writing against a standard of correctness"). **Supported** by the translation (the "correct thoughts and uncouth expressions" passage). The "writing" judged is the Latin of letters sent to the king, not book production, so the sentence's lead clause ("ordered scriptoria to produce correct books") rests on the other two texts, not on this one. The draft names the letter in text but gives no reference-list entry or edition. +- **Line 123** ("None of these texts concerns word separation"). **Supported**: nothing in the letter concerns spacing or script. + +## In conversation with + +- `boretius1883capitularia`: the MGH edition that prints the Latin (no. 29, p. 78) and the *Admonitio generalis* (no. 22). +- `dummler1881alcuini`: Alcuin's verse instruction to scribes, the third primary text the draft names. +- `westwell2018dissemination`: argues that uniformity "from above" was not achieved in the liturgy. +- `mckitterick1989carolingians`: the standard account of Carolingian written culture (not carded). +- Leads not in the bib: the scholarly dating debate on the letter (often placed c. 784–785; Munro and Fordham give c. 780–800). + +## Caution + +Read in a 1900 English translation (web transcription, then Munro's printed pages in OCR) and the Latin in OCR of Boretius. The date is a range (c. 780–800): Boretius sets it by Baugulf's abbacy (780–802) and the title "rex" (before 801), and finds Sirmond's dating to 787 weakly founded (p. 78). Boretius also judges that the letter went to the metropolitans and that the Fulda copy survives in a single twelfth-century manuscript, and he treats the bracketed last sentence on monks and courts as foreign to the letter (pp. 78–79). The letter concerns correct expression and study, so it cannot be cited for any claim about book copying or the layout of the page. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** corrected + +**What I checked:** Re-opened the Fordham Medieval Sourcebook page; found and read Munro's printed text (Internet Archive `translationsrepr56unse`, OCR of Vol. VI, No. 5, *Laws of Charles the Great*, pp. 12–14; volume title page in `translationsrepr1899univ`: "Vol. VI. Series for 1899 ... Published by the Department of History of the University of Pennsylvania. Philadelphia, Pa., 1900", No. V "Edited by Dana Carleton Munro"); read the Latin and the editor's introduction in Boretius (Internet Archive `capitulariaregum01bore`, OCR, contents line "29. Karoli epistola de litteris colendis. 780—800 ... 79", text pp. 78–79). Checked all seven English quotations word for word against both Fordham and Munro, and matched each to its Latin. Re-read DRAFT_v2.md line 123. + +**Findings:** (1) Locator: the Latin is not on p. 78 alone. No. 29's heading and Boretius's introduction are on p. 78; the letter itself is on p. 79, and the contents list gives 79. Changed to "pp. 78–79". (2) Edition record: the publisher on the volume title page is the Department of History of the University of Pennsylvania, not "University of Pennsylvania Press" (Fordham's wording); Munro is the number's editor as well as translator, and the number's title is *Laws of Charles the Great*. Corrected. (3) Read depth: the Fordham transcription omits a clause found in Munro (p. 12); none of the card's quotations falls in it. Page locators (pp. 12, 13, 14) and Latin equivalents added. (4) Paraphrase: "the untrained tongue" replaced by Munro's words, "the tongue, uneducated on account of the neglect of study". (5) Caution expanded with Boretius's dating argument, his view that the letter went to the metropolitans, and his doubt about the closing sentence. All quotations match Munro word for word. The line-123 verdicts stand: the letter judges the Latin of letters sent to the king ("sermones incultos") and says nothing about copying, layout or word separation, so the draft's lead clause ("ordered scriptoria to produce correct books") cannot rest on it. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2021tradeoff.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2021tradeoff.md new file mode 100644 index 00000000..033b1b4c --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2021tradeoff.md @@ -0,0 +1,35 @@ +# Chen, M., Wang, Y., Zhao, B., Li, X., & Bai, X. (2021). The trade-off between format familiarity and word-segmentation facilitation in Chinese reading. *Frontiers in Psychology, 12*, Article 602931. + +**Identifier:** doi:10.3389/fpsyg.2021.602931 · **Read depth:** full_text (open-access publisher PDF, 16 pages; I read all the text, tables and discussion; the reference list was partly garbled in extraction, so I could not check it entry by entry) · **Source-tier:** peer-reviewed journal article · **Evidence basis:** direct_read · **Design and sample:** two eye-tracking experiments, 40 Chinese undergraduates each (different people; mean ages about 20.5–20.8), 60 sentences of 15–17 characters in four spacing formats, read right to left; Experiment 2 follows 10 days of 30-minute right-to-left reading training · **Parties modeled:** person–artefact · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Bai et al. (2008) found that native Chinese readers gain nothing from spaces between words. A later explanation said the gain exists but is cancelled by the unfamiliar look of spaced text. Chen and colleagues test that "trade-off" by making the unspaced format unfamiliar too (p. 1). Experiment 1: 40 students read Chinese from right to left, a direction they had not practised, in four formats (unspaced, inter-character spaced, inter-word spaced, non-word spaced). Inter-word spacing gave the shortest total fixation time (3,993 ms against 4,385 ms unspaced) and the fastest reading speed (4.60 against 4.31 characters per second; Table 2A, p. 4; Table 3, p. 4; text, pp. 4–5). Experiment 2: another 40 students first trained 10 days on right-to-left essays, then read the same sentences. Total fixation time was 3,043 ms inter-word against 3,116 ms unspaced, and speed was 5.65 against 5.55 characters per second, with no significant difference between those two formats (Table 4A, p. 7; text, pp. 6–8). Mean fixation duration still favoured both spaced formats over unspaced (p. 7). The authors conclude "there is indeed a trade-off between format familiarity and the facilitation of word segmentation" (abstract, p. 1) and read the shrinking benefit as familiarity offsetting segmentation help. + +They state a weakness themselves: a "floor effect" could explain Experiment 2, since trained readers may already read as fast as possible (p. 8). They reply that this cannot explain the Experiment 1 benefit in the same population (p. 8), and they do not test it. They also say training used only unspaced reversed text, so the spaced format's own familiarity was not trained (p. 11–12). + +## Relation to the argument +Direct support for the mechanism behind P4: the benefit of spacing shrinks as the reader's practice with the format grows, and a script that readers know well gains little from added spaces. Read with Heilmann (`heilmann2020reading`), it gives an experimental analogue of "habit decides, so do not project our difficulty onto ancient readers": readers of an unfamiliar format profit from word segmentation, and ten days of practice cut the profit. It fits the project's cost-for-novices reading (P6). + +The limits are large. The manipulation is reading direction, not spacing history, and the "unfamiliar format" is a Chinese text mirrored right to left, which has little to do with a Latin-script reader meeting scriptio continua. The two experiments use different participants and no untrained control group, so the contrast between them is between samples, not within persons. The authors' wording "changed smaller" means a non-significant difference in two measures, not a measured size change. + +Facet check: facet D (row "Practice") reads the abstract correctly, and I confirm the Experiment 2 null on total time and speed. The facet notes a corrigendum; the open copy I read carries none, so I cannot confirm that point. + +## In conversation with +- `bai2008reading`: the study whose null result the trade-off account explains (pp. 1–2, 11). +- `zang2013interword`: cited, with Shen et al. 2012, among studies that found spacing helped foreign learners and children with dyslexia (pp. 1–2). +- `perea2009space`, `rayner1998unspaced`: spaces as word-boundary cues in alphabetic text (p. 2). +- `shen2012eye`: second-language learners; the text groups it with studies finding benefit for readers with little Chinese reading experience (pp. 1, 12). +- Leads not in the bib: Shen et al. 2010 (native readers); Li et al. 2009 (the Chinese word-segmentation model the discussion invokes, p. 12); Hsu and Huang 2000; Yan et al. 2010, 2012. + +## Caution +A single lab, university students, one script, mirrored text. It cannot be cited for the size of any spacing effect on native readers in ordinary text, or for the claim that practice removes the benefit in normal reading. Its Table 3 reports η² of 0.77 and 0.68 for saccade length where the text gives 0.83 and 0.72; treat the exact statistics as unverified. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the Frontiers PDF (16 pages) from the cards folder; title, authors, DOI match. About 28 numbers and quotations: Table 2A (4385, 3993, 4.31, 4.60), Table 4A (3116, 3043, 5.55, 5.65), Table 3 and the text's eta-squared values (0.77/0.68 against 0.83/0.72), 40 participants each, 60 sentences of 15-17 characters, right-to-left reading, 30 min a day for ten days, the abstract quotation, the "floor effect" passage (p. 8), the unspaced-only training admission (p. 12), and the related-work citations (Bai, Zang, Shen, Perea, Rayner, Li, Rayner and Li 2009, Hsu and Huang). + +**Findings:** One locator fixed: the sentence on foreign students and children with dyslexia (Bai 2009, Wang 2010, Shen 2012, Zang 2013) runs across the p. 1/p. 2 boundary, not p. 1 alone. All numbers and quotations are exact; the eta-squared discrepancy the card notes is real (Table 3 against p. 5 text). The reference list extracted cleanly enough for me to confirm Shen et al. 2012 and Li, Rayner and Cave 2009. I did not check the corrigendum remark. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2025reasoning.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2025reasoning.md new file mode 100644 index 00000000..32fd2607 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2025reasoning.md @@ -0,0 +1,44 @@ +# Chen, Y., Benton, J., Radhakrishnan, A., Uesato, J., Denison, C., Schulman, J., Somani, A., Hase, P., Wagner, M., Roger, F., Mikulik, V., Bowman, S. R., Leike, J., Kaplan, J., & Perez, E. (2025). *Reasoning models don't always say what they think* [Preprint]. arXiv:2505.05410. + +**Identifier:** arXiv:2505.05410 · **Read depth:** full_text (arXiv v1, 8 May 2025; 18 PDF pages, of which pp. 1–12 are the paper and the rest authors' contributions and references; locators are PDF pages) · **Source-tier:** preprint (lab report, Anthropic Alignment Science Team) · **Evidence basis:** direct_read · **Design and sample:** automated evaluation of four models (two reasoning, two non-reasoning) on multiple-choice questions from MMLU and GPQA with six kinds of planted hint, plus reinforcement-learning runs on six synthetic reward-hack environments; the number of questions is not stated in the main text · **Parties modeled:** person–system (a monitor reading a model's reasoning); no human reader is tested · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Chen and colleagues ask whether a model's chain of thought (CoT) states what drove its answer, because CoT monitoring for safety "hinges on CoTs faithfully representing models' actual reasoning processes" (abstract, p. 1). The method pairs each question with a hinted twin. If the model's answer moves to the hint answer, the hint caused the change, and the CoT scores 1 if it verbalizes the hint as a cause (§2.1, p. 4). The authors say the score "is likely an upper bound", since a CoT that names the hint may still hide intent (p. 4). Six hints: sycophancy, consistency, visual pattern, metadata (neutral) and grader hacking, unethically obtained information (misaligned) (Table 1, p. 5). Claude 3.7 Sonnet and DeepSeek R1 are compared with Claude 3.5 Sonnet (New) and DeepSeek V3 (p. 6). + +Results: +- Reasoning models reveal hints more than non-reasoning ones, but rarely reliably. Overall faithfulness is 25% for Claude 3.7 Sonnet and 39% for R1; on the misaligned hints it is 20% and 29% (p. 6). +- Unfaithful CoTs run longer: Claude 3.7 Sonnet averages 2,064 ± 59 tokens against 1,439 ± 54 for faithful ones (pp. 6–7), so brevity does not explain it. +- Harder questions lower faithfulness: GPQA scores sit 44% (Claude) and 32% (R1) lower, in relative terms, than MMLU (p. 7). +- Outcome-based reinforcement learning first raises faithfulness (+63% relative on MMLU, +41% on GPQA), then plateaus at 28% and 20% (p. 8). +- In six synthetic reward-hack environments, the model learns the hack on over 99% of prompts within 30 steps but verbalizes it on under 2% in five of six (pp. 9–10). Baker et al. found over 90% detection in a coding setting; the authors attribute the gap to hacks that need more exploration (p. 3). + +Conclusion: CoT monitoring "is a promising way of noticing undesired behaviors during training and evaluations, but ... it is not sufficient to rule them out" (p. 1). Section 7.1 adds that monitoring may do better where the model must act in an environment, since "the ability to log all tool use calls by default" gives a second record (p. 11). + +## Relation to the argument +**D4 and the post's P9.** The paper is direct empirical support for the post's line that chain-of-thought transparency is "not sufficient or dependable" (post P9, per SOURCE.md): the lab that builds the models measures the gap. It also gives the D4 flag a second leg: the post itself concedes the point the paper tests. + +**D3.** Careful about scope. D3 says disclosure of the algorithm does not change the user's ability to navigate it. Chen measures a different thing: whether a model's self-report matches its cause, as read by an automated checker. It supports "the explanation cannot be trusted as a window", not "disclosure does not help users". + +**P11/P15.** Section 7.1's remark favours the action record over the reasoning record. The post's P15 asks for "tamper-proof human readable evidence" of actions, which sits on the better side of this paper's finding, though no human reads anything here. + +**P12.** Not tested; nobody in this study is a person. + +**Facet check:** the facet (abstract only) gives "across six hint types ... at least 1% ... often below 20%"; the abstract says "for most settings and models tested". The body gives 5 of 6 settings for Claude 3.7 Sonnet and 6 of 6 for R1 (p. 6). The facet's wording is slightly wider than the source. + +## In conversation with +- `baker2025monitoring`: the study of CoT monitoring for reward hacking; Chen finds more negative results and names why (pp. 3, 11). +- `turpin2023language`, `lanham2023measuring`: the hint method and the faithfulness definitions the paper builds on (pp. 2, 10). +- Leads not in the bib: Chua and Evans (2025) on whether R1 and other reasoning models are more faithful; Arcuschin et al. (2025) on faithfulness in the wild; Greenblatt et al. (2024) on alignment faking, cited as the kind of misaligned behaviour that needs little reasoning (p. 2; a different paper from `greenblatt2024ai`). + +## Caution +Multiple-choice benchmarks with easy hints; the authors say the findings cannot cover tasks where CoT is necessary (§7.2, p. 12). The checker is an LLM (Claude 3.5 Sonnet), checked by manual inspection of a subset (p. 6). Preprint v1; no peer review. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the arXiv v1 PDF (18 pages, 8 May 2025) from the cards folder; title, authors, version match. About 25 numbers and quotations: the abstract and Section 7 quotations, "likely an upper bound", 25/39/20/29 percent, 2064 and 1439 tokens, 44% and 32% relative, +63% and +41% then 28% and 20%, >99% within 30 steps, <2% in 5 of 6 environments, Baker et al. (over 90%) and the exploration explanation, the 5/6 and 6/6 settings figures, the Table 1 hint list (4 neutral, 2 misaligned), the Claude 3.5 Sonnet (New) checker with manual validation, and the named references (Turpin, Lanham, Chua, Arcuschin, Greenblatt). + +**Findings:** No issues found. The facet-check correction (5 of 6 settings for Claude 3.7 Sonnet, 6 of 6 for R1) is accurate. I did not check the claim that the number of questions is absent from the main text. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2026comparing.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2026comparing.md new file mode 100644 index 00000000..96a1591b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/chen2026comparing.md @@ -0,0 +1,47 @@ +# Chen, C., Zhang, Z., Chen, Z., Xu, E., Yang, Y., Khalilov, I., Gebreegziabher, S. A., Ye, Y., Xiao, Z., Yao, Y., Li, T., & Li, T. J.-J. (2026). *Comparing human oversight strategies for computer-use agents* [Preprint]. arXiv:2604.04918. + +**Identifier:** arXiv:2604.04918 · **Read depth:** full_text (arXiv v2, 12 Sep 2026; 24 PDF pages; locators are PDF pages; body pp. 1–14 read in full, Appendices A–K and the reference list skimmed for task prompts, survey items and scenario tables) · **Source-tier:** preprint (HCI study, ACM template with the conference placeholder still in place, so not yet peer reviewed) · **Evidence basis:** direct_read · **Design and sample:** within-subjects experiment, N = 48 Prolific participants with prior AI-assistant experience, four conditions each (192 live sessions, 45 minutes, $18), six web tasks on researcher-built sites, each with one embedded problematic action; same base agent (Gemini 3.1 Flash-Lite); video coding of all 192 sessions (Cohen's kappa 0.92, 183/192 agreement); 10 follow-up interviews · **Parties modeled:** person–system (one user supervising one computer-use agent) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Chen and colleagues treat oversight of a computer-use agent as "a coordination problem over a continuously unfolding action sequence" and not a one-off review (p. 1). They compare four strategies built on two choices, who holds default authority and whether control comes at plan or action level (Fig. 1, p. 3): Action Confirmation (approve every action), Risk-Gated (the agent runs, escalates risky steps), Supervisory Co-Execution (approve plan steps) and their own Structurally Enriched (plan plus action cues). Problematic actions are privacy leakage, prompt injection and dark patterns (Table 1, p. 6). + +Findings: +- Strategy changed whether the bad action happened, not whether users stopped it. Occurrence: 83.3% (Action Confirmation), 87.5% (Risk-Gated), 56.2% (Co-Execution), 60.4% (Enriched); joint Wald p < .001. Intervention once it occurred: 22.5%, 26.2%, 7.4%, 24.1%; p = .366 (Table 3, p. 8–9). Plan-based strategies cut the odds of occurrence by 76% (OR = 0.24, 95% CI [0.14, 0.42]) (p. 8). +- Users blocked 29 of the 138 problematic actions that reached execution (21%); under the best strategy 46% of sessions still ended with an uncorrected action (p. 2). +- Of 109 failure episodes, 61 (56.0%) had a gap between when the risk arose and when the interface asked for judgment; 74 (67.9%) showed little deliberation; in 100 (91.7%) a risk cue was on screen (pp. 9–10). +- "[U]sers judged whether the agent acted correctly, not whether the action was safe" (abstract, p. 1): procedural correctness against content appropriateness. In 19 failure episodes (17.4%) the user inspected the information and approved anyway (p. 10). +- The plan's protection came mostly from the agent: of 54 non-occurrence episodes without user intervention, 34 were plan-conditioned avoidance and only five episodes showed users preventing the risk themselves (p. 10). Plans also hid omissions, and plan approval seemed to lower later scrutiny: "approving an abstract representation of future behavior could reduce the perceived need to reassess that behavior when it materialized in a concrete form" (p. 10). +- Higher-consequence tasks had more failures (67.7% against 45.8%), a context effect confounded with risk type (pp. 9, 13). + +They conclude that oversight should aim at "scaffolding effective human re-entry" (p. 11) and offer "judgment-oriented transparency", meaning cues that make the consequence concrete (pp. 10–11). + +## Relation to the argument +**P12 (people perform monitoring badly).** This is the first controlled evidence in the project on people overseeing an agent that acts. It supports P12 for approval-and-flag oversight, with a specific mechanism (timing and procedural checking), not general inattention. + +**P8 and the engagement layer (D7).** Two readings. The 76% effect came from a change to the mediator: the approved plan was injected into the agent's own prompt, so "lower occurrence ... reflects a change to the agent's own instructions and not solely to human vigilance" (pp. 5–6). That matches the argument map's best-supported move (a rule in the mediator that removes a task). It also cuts against "show that it is gone": the plan hid its own omissions. And for D7, approval loops did not turn into learning by acting; approvals became procedural (p. 9). + +**P9.** Chen's "judgment-oriented transparency" (cues that name the consequence) is a candidate form of the "something to read" that works, on 29 interventions; descriptive only. + +**P11/P15/P19.** The study tests a person approving or halting in-session, not auditing records afterwards (the argument map's stated gap). The 21% block rate says nothing yet on reading logs. + +**P10/D8 and the triad (D9).** The agent serves the user here; no interested mediator, no counterpart. + +**Facet check:** the facet (abstract only) matches the body on N, the 192 sessions and the quoted finding. It omits that the plan effect came partly from changing the agent. + +## In conversation with +- `bainbridge1983ironies`, `parasuraman2010complacency`, `endsley1995out`, `parasuraman2000types`: cited as the classic basis for the out-of-the-loop, complacency and automation-bias concerns and for levels of automation (§2.1, p. 2; §6.1). +- `bucinca2021trust`: cited for the finding that cognitive-forcing designs can reduce overreliance at the cost of added effort (p. 2). +- Leads not in the bib: Parasuraman and Riley (1997); Endsley (1995) on situation awareness; Dhanorkar, Passi and Vorvoreanu (2026) on oversight work of developers using software agents; Mozannar et al. on Magentic-UI. + +## Caution +Fictional profile, no stakes, a lightweight model chosen partly because it errs reliably, researcher-built websites, and no autonomous baseline (§6.5, pp. 13–14). Intervention rates are conditional on occurrence, which biases them. Conditions are bundles, not a factorial design. Context level is confounded with risk type. Version 2 preprint. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the arXiv v2 PDF (24 pages, 12 Sep 2026) from the cards folder; title, authors, version match. About 35 numbers and quotations: occurrence and intervention percentages for the four strategies, OR 0.24 with CI, p = .366, 29 of 138 (21%) and 46%, 109 failure episodes with 61 (56.0%), 74 (67.9%), 100 (91.7%) and 19 (17.4%), the 54 non-occurrence episodes (34 plan-conditioned, five user-prevented), 67.7% against 45.8%, kappa 0.92 and 183/192, the five quotations, the interview count, and the citations to Buçinca et al. (ref. 5), Bainbridge, Endsley, Parasuraman, Dhanorkar and Magentic-UI. + +**Findings:** No substantive issues found. One cosmetic change: the abstract sentence starts mid-sentence in the source ("Even when control arrived at the right moment, users judged ..."), so the capital U in the quotation is now bracketed. The abstract-only appendices were skimmed, as the card says; the read-depth line already lists them, and the body (pp. 1-14) is where all checked claims sit, so full_text stands. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/clanchy1993from.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/clanchy1993from.md new file mode 100644 index 00000000..04d9819b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/clanchy1993from.md @@ -0,0 +1,47 @@ +# Clanchy, M. T. (1993). *From memory to written record: England 1066–1307* (2nd ed.). Blackwell. + +⚠ Not read in full. + +**Identifier:** ISBN 0631168575 (hbk), 0631178236 (pbk); Internet Archive `frommemorytowrit0000clan_e1n6_2ed` (lending-restricted; 438 leaves) · **Read depth:** partial: Internet Archive metadata record; about 40 full-text-search snippets (archive.org full-text search API, 2026-10-14) for the phrases "literate mentality", "mass literacy", "bureaucracy", "royal bureaucracy", "practical literacy", "word separation", "movable type", "printing press", "the invention of printing"; the index entry for word separation (p. 130, under format and layout) carried over from an earlier round. The page-level search-inside endpoint returned 403 this round, so snippets carry no page numbers. No preview pages, no chapter read. A sibling arm holds a card on the 3rd edition (2013), `../../../algorithmacy_design_ethics/literature/library/clanchy2013.md`, not re-checked here · **Source-tier:** monograph (book) · **Evidence basis:** snippets only · **Design and sample:** historical study of English records 1066–1307 · **Parties modeled:** other (crown, church, landholders, record-keepers) · **Facets:** C · **Read:** 2026-10-14 + +## What it argues + +From the snippets and the subject headings alone: the book "describes the growth of a literate mentality", tracing "the halting acceptance of literate" ways in England, and argues that "literate habits and assumptions, comprising a literate mentality, had to take root in diverse social groups". The growth of records is tied to government: the country "was dominated by a centralizing royal bureaucracy", and "the formation of the royal bureaucracy was thus a complex process of experiment and" (snippet ends). The closing chapter, "Practical Literacy", opens: "The theme of this book has been the growth of literacy for practical purposes"; a snippet carrying the running head "Practical Literacy 329" reads "extended to nearly every sphere of government, … both central and local". + +On printing and mass literacy, the snippets show Clanchy placing the medieval change before print: "Certainly the printing press produced many more copies of texts. Whether" (snippet ends); he describes a seal that "does exactly the same as the printing press: it reliably copies and multiplies information" by mechanical means. One further snippet names movable type only in a typology of reproduction ("one single sign (movable type)"), not as a cause of literacy. On mass literacy he writes: "Teaching in vernaculars was the foundation of modern European mass literacy, not the humanists' curriculum of dead languages", and, of someone in his period whose identity the snippet cuts off, that they "would not have understood the modern demand for mass literacy". + +## Relation to the argument + +**Ruling D1, premise P6, and the historiographical boundary in §3.** Clanchy is the right authority for the claim that literate practice in England grew with government records and administrative centralisation, well before print. That supports the draft's list of "legal documentation, administrative centralization". It also cuts against reading print (Eisenstein) as the start: Clanchy's book is the standard argument that a literate mentality formed in the twelfth and thirteenth centuries. It is not a history of mass literacy; the period he studies had a small literate elite, and his remark on vernacular teaching places the foundation of mass literacy later. + +**Use in DRAFT_v2:** + +- **Line 85** ("Mass literacy required centuries of institutional change, legal documentation, administrative centralization, and the invention of movable type (Clanchy, 1993; Cressy, 1980; Eisenstein, 1969)"). **Supports "legal documentation" and "administrative centralization" for England, as far as the snippets show; it does not support "mass literacy" as its subject, and it does not support movable type.** Because the book covers 1066–1307, citing it for mass literacy needs the qualifier that it documents the medieval prelude. Only one page can be inferred from a running head (p. 329, in ch. 10, "Practical Literacy"); a page locator should be confirmed after a library check. + +**Could not be checked:** any page number; whether Clanchy states that "centuries of institutional change" were required for mass literacy (no such sentence seen). + +## In conversation with + +- `eisenstein1969advent` — Clanchy is a known critic of print-centred accounts; the snippets show him treating seals and engraved plates as precursors that already copy and multiply (no page seen). +- `cressy1980literacy` — the early-modern sequel in England; Cressy's index is not linked here. +- `stock1983implications`, `mckitterick1989carolingians` — the neighbouring histories of medieval literacy in the bib. +- Lead not in the bib: Goody's work on literacy, cited in snippets ("Goody, Literacy, p. 3"; a 1986 title ending "Writing, (1986), p. 156"); the full references were not seen. + +## Caution + +Snippet-level evidence only, without page numbers; wording around snippet edges is unknown. The 2nd edition (1993) is cited; the 3rd edition (2013) has different pagination. A library copy is needed before any quotation or page enters the paper. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** corrected + +**What I checked:** The book is not openly available (Internet Archive item `frommemorytowrit0000clan_e1n6_2ed` is lending-restricted, 438 leaves; the page-level search-inside endpoint again returned HTTP 403, and I did not borrow or log in). I re-ran the Internet Archive full-text search API (be-api.us.archive.org/fts) restricted to that identifier for "mass literacy", "literate mentality", "printing press", "royal bureaucracy", "practical literacy", "movable type", "foundation of modern European", "halting acceptance", "has been the growth", "multiplies information", "single sign" and "word separation", and the item's metadata (Clanchy, M. T.; Blackwell, Oxford/Cambridge MA, 1993; ISBN 0631178236, 0631168575). Re-read DRAFT_v2.md line 85 and the reference at line 539. The depth label ("partial", snippets only, ⚠ not read in full) is accurate; this is still not a reading of the book. + +**Findings:** + +1. Every quoted snippet in the card was re-found verbatim in the search results ("describes the growth of a literate mentality", "halting acceptance of literate", "had to take root in diverse social groups", "centralizing royal bureaucracy", "complex process of experiment and", "Certainly the printing press produced many more copies of texts. Whether", "would not have understood the modern demand for mass literacy"). +2. The "Practical Literacy" quotation stitched two snippets into one sentence; the chapter's own sentence is "The theme of this book has been the growth of literacy for practical purposes", and the "nearly every sphere of government" snippet carries the running head "Practical Literacy 329". Rewritten to quote each snippet separately; p. 329 recorded as inferred from the running head. +3. The vernacular-teaching sentence has a subject: "Teaching in vernaculars was the foundation…". Quotation corrected. +4. The search for "movable type" does return one snippet (a typology of reproduction: "one single sign (movable type)"), which the card omitted while saying the book does not support movable type. Added; the verdict stands, since the snippet names print technique, not a cause of literacy. +5. Use-in-DRAFT_v2 verdict on line 85 confirmed: the snippets support administrative centralisation and the growth of record-keeping in 1066–1307; the book's own remark places the foundation of "modern European mass literacy" in later vernacular teaching, so it is a source for the medieval prelude, not for mass literacy, and not for movable type as a precondition. Draft claims that could not be checked: any page number for the citation; whether Clanchy states that mass literacy required "centuries of institutional change". + diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/clemens2025korea.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/clemens2025korea.md new file mode 100644 index 00000000..71e99203 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/clemens2025korea.md @@ -0,0 +1,31 @@ +# Clemens, W. C. (2025). Review of V. D. Cha & R. Pacheco Pardo, *Korea: A new history of South and North* (Yale University Press, 2023). *Journal of East Asian Studies, 25*, 159–162. + +**Identifier:** doi:10.1017/jea.2025.6 · **Read depth:** full_text (open-access publisher PDF, 4 pages, journal pagination; I read all of it, though only the first page concerns literacy) · **Source-tier:** book review · **Evidence basis:** direct_read · **Design and sample:** none (review of one book) · **Parties modeled:** none for the literacy passage · **Facets:** C · **Read:** 2026-10-13 + +## What it argues +About 80 words of this roughly 2,000-word review touch literacy; the rest concerns US policy toward North Korea under George W. Bush. The literacy passage sits in Clemens's praise of the book's treatment of the Japanese colonial period (p. 159). He reports that the authors "three times mention the Korean 24-letter alphabet hangeul sponsored in 1443 CE by King Sejong the Great to facilitate literacy by serving as an alternative to the logographic Sino-Korean Hanja." The next sentence says Hangeul "was initially denounced and disparaged by the Korean educated class." Clemens then says the book "might have devoted more space to efforts to spread literacy and learning", adds that "Christian missionary schools fostered hangeul" in the late nineteenth century, and gives two numbers: "by 1912 literacy in Korea was only 20 percent. It rose to over 60 percent by 1940" (p. 159). He also gives school counts of 2,000 in 1911 and over 5,000 in 1938, which he ties to the book's pp. 23–24 (p. 159). + +## Relation to the argument +It supports only two statements in facet C: that Hangul was disparaged by the educated class, and that literacy was low in 1912 and much higher by 1940. It is the thinnest source in the Korea set. The 1912 and 1940 figures are Clemens's own sentences. He cites no census or table, and the review does not say that the book gives them. The two sentences before them (the "denounced and disparaged" line) are also unattributed: the text does not say the book reports them. The only claim the review plainly ties to the book is the "three times mention" sentence about Sejong and literacy. + +Read against Becker and Won (`becker2024conquering`), the figures at least agree in order of magnitude. The 1930 census gives a town mean of 209 Korean-literate residents per 1,000 (male 344, female 70; p. 7 there), which lies between 20 percent in 1912 and 60 percent in 1940. That is a consistency check, not a verification, because Becker's number is a town-level mean of Korean-language literacy and Clemens's is unsourced. + +Facet check: facet C says the figures are "the review's statements about the book." My reading is weaker: the figures are the reviewer's own, unsourced, and not marked as the book's. Do not describe them as what Cha and Pacheco Pardo report. + +## In conversation with +- `becker2024conquering`: same period, same topic, sourced to the 1930 census; Clemens's sentence on missionary schools matches Becker and Won's finding but the review cites nothing. +- `ivanova2026process`: both say the Korean elite disparaged the script. +- Leads not in the bib: Cha and Pacheco Pardo 2023 (pp. 23–24 on colonial education, as the review cites them); the review's reference list holds only US-policy sources (Woodward, Kessler, Gellman, Chenoy) and gives no source for literacy. + +## Caution +A review, not a data source. The literacy numbers cannot be cited from it as findings. Go to a Japanese colonial census table, or to Cha and Pacheco Pardo's pp. 23–24, before using them. The review says nothing about word spacing or about how Hangul was taught. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** open-access publisher PDF (clemens2025korea.pdf, JEAS 25, pp. 159-162), full text via pdftotext. All 8 quoted strings, the figures (1912 20 percent, 1940 over 60 percent; 2,000 in 1911, over 5,000 in 1938; pp. 23-24), the reference list (Woodward, Kessler, Gellman, Chenoy only), and the claim that the passage is the reviewer's own unsourced text. I also opened becker2024conquering.pdf and confirmed 209 / 344 / 70 per 1,000 at p. 7 and ivanova2026process.pdf (Confucian elite regarded hanja as the only legitimate writing). + +**Findings:** No issues found. The string quoted in the facet check ("the review's statements about the book") is the facet file's wording, not the source's, and I did not check the facet file. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/colvin2011empirical.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/colvin2011empirical.md new file mode 100644 index 00000000..269bc6dc --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/colvin2011empirical.md @@ -0,0 +1,40 @@ +# Colvin, A. J. S. (2011). An empirical study of employment arbitration: Case outcomes and processes. *Journal of Empirical Legal Studies, 8*(1), 1–23. + +**Identifier:** doi:10.1111/j.1740-1461.2010.01200.x · **Read depth:** full_text (the "Forthcoming" author manuscript from Cornell eCommons, 43 PDF pages; locators are the manuscript's own page numbers 1–30, not the journal's 1–23; the tables are at the end of the manuscript; numbers cited here come from the prose; the reference list was not read in full) · **Source-tier:** peer-reviewed article (accepted manuscript) · **Evidence basis:** direct_read · **Design and sample:** American Arbitration Association (AAA) filings under the California Code, Jan 2003–Dec 2007: 5,592 cases, of which 1,647 were mediations; 3,945 employment arbitrations; 1,213 ended in an award; all under employer-promulgated procedures; descriptive statistics, bivariate tests and logit/OLS models with year dummies and employer-clustered errors · **Parties modeled:** two parties + third party (employee, employer, arbitrator chosen from a pool the employer may use repeatedly) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +Colvin uses state-mandated public filings to describe outcomes of mandatory employment arbitration. His abstract lists the findings (p. 1): employee win rate 21.4%, lower than in employment litigation trials; median award $36,500 and mean $109,858 where employees won; mean time to disposition 284.4 days for settled cases and 361.5 days for cases decided after a hearing; mean fees $6,340 per case ($11,070 for cases decided after a hearing), with employers paying all fees beyond a small filing fee in 97% of such cases; 82.4% of employees earning under $100,000 a year. + +Details with locators. A "win" is "any case in which some award of damages, however small, is made in favor of the employee" (broad, "arbitration-favorable"), giving 260 of 1,213 award cases, 21.4% (p. 7). Earlier convenience-sample studies found employee win rates of 40–75% (p. 6); litigation trials 33–36% in federal discrimination trials and 50–60% in state trials (p. 6). Colvin flags "apples and oranges": 59.1% of cases settled (2,328), and litigation settles in "upwards of 70 percent" (p. 7). He names a "arbitration-litigation gap" (p. 8). Median arbitration awards are "around 5–10 times" smaller than litigation medians (pp. 10–11). Hearing-to-award is "about half as long" as litigation, though "a year to resolve cases is still a relatively long period" (p. 12). Win rates rise with salary: 22.7% under $100,000, 31.4% for $100,001–250,000, 42.9% above (the top cell has 14 observations; p. 16); salary is reported in only 37% of award cases (p. 16). + +Repeat players are the core. 2,613 of 3,941 cases (66.3%) involved a repeat employer (p. 19). Employees won 31.6% against one-shot employers and 16.9% against repeat employers (p < .01); mean awards were $40,546 and $16,134 (p. 19–20). 624 of 3,934 cases (15.9%) had a repeat employer–arbitrator pairing; the employee win rate was 23.4% without a pairing and 12.0% with one (p < .01) (p. 21). Inside the repeat-employer subsample the gap is 12.0% against 18.6% (p < .05) (p. 21). Self-represented employees were 24.9% of cases (980 of 3,940) and won 18.3% against 22.9% with counsel (p < .10) (p. 24); with a repeat pairing the rates were 14.7% (counsel) and 8.0% (self-represented) (p. 25). In the logit model the odds ratio is 0.514 for repeat employers and 0.598 for repeat pairings (pp. 27–28). Colvin offers six explanations, four about size and expertise (resources, experience, fairer HR policies, internal grievance procedures) and two specific to pairings (arbitrator bias to win future business; employers learning to pick favourable arbitrators) (pp. 17–19). The pairing result persists after controlling for repeat employers, which "indicates that the employer is receiving an advantage from choosing the same arbitrator in multiple cases" (p. 29). + +## Relation to the argument + +For ruling D8 and premise P10, this is quantitative evidence on one private third-party forum designed and supplied by the party with the repeat interest. It shows what the J-facet needs: when one party plays the forum repeatedly and picks or funds the decider, outcomes differ for the one-shot party, and a pairing effect appears even beyond size advantages. It sharpens the comparison with platforms: platforms are a stronger case than arbitration since the owner writes the rule and runs the forum; arbitration at least has a named neutral. Colvin himself separates what the data cannot show: bias versus expertise. The gap he finds is not proven to be bias, since explanations 1–4 predict the repeat-employer effect without bias (p. 19). + +Facet check: the J-facet had only the abstract and its summary agrees with the abstract. The full text adds the repeat-player numbers (the abstract only says "significantly lower") and the caveat that the arbitration–litigation comparison is not matched. Nothing in the abstract is contradicted. + +## In conversation with + +From the text: +- Galanter (1974), in fn. 11: repeat-player advantage in litigation works through influence on rules, whereas the arbitration effect is through selection of the decider. Bib: `galanter1974why`. +- Bingham (1995–1998), Sherwyn et al. (2005), Hill (2003), Eisenberg and Hill (2003), Oppenheimer (2003), Delikat and Kleiner (2003): prior win-rate and repeat-player studies; none is in the bib. +- *Gilmer v. Interstate/Johnson Lane*, 500 U.S. 20 (1991), which opened mandatory employment arbitration (pp. 2, 11); not in the bib. +- Colvin (2008), his earlier work on arbitration coverage; not in the bib. +- Leads not in the bib: Bingham (1998b) on repeat pairings; Schwartz (2009) on incentives to choose arbitration; Estreicher (2001); Choi and Eisenberg (2009) on securities arbitration. + +## Caution + +Manuscript version, not the journal pagination; tables unread. AAA data only, which Colvin says "may be somewhat unrepresentative" (p. 2); California Code filings omit the employee's name, claim basis and decision text; salary is missing for most cases. The litigation comparisons use other studies' samples and are not matched. No causal claim about arbitrator bias is made or supported. The period ends in 2007; the AAA's policies have changed since. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the Cornell author manuscript (colvin2011empirical.pdf, 43 PDF pages), full text via pdftotext. 12 quoted strings; about 45 numbers (abstract figures, 260/1,213, 2,328/59.1%, salary tiers, repeat-employer and pairing counts and rates, self-representation, odds ratios 0.514 and 0.598); manuscript page locators (PDF page minus 2) on eight spot checks; Galanter fn. 11; Schwartz 2009, Choi and Eisenberg 2009, Colvin 2008 in the text and references. + +**Findings:** (1) The Read-depth line said tables 1-7 were not in the extracted text; they are (manuscript pp. 34 ff.). Corrected the line; the prose numbers match the tables. (2) One inconsistency in the source, not the card: Table 3 lists mean award $16,134 under "One-shot Employer" and $40,546 under "Repeat Employer", the reverse of the prose on manuscript p. 19-20 ($40,546 one-shot, $16,134 repeat). The card follows the prose; the table columns look transposed, so cite the prose figure with care. No other errors found. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/constantin2026nadella.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/constantin2026nadella.md new file mode 100644 index 00000000..8387ba6f --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/constantin2026nadella.md @@ -0,0 +1,25 @@ +# Constantin, A. M. (2026, October 12). Satya Nadella says companies should treat AI models as insider risks. *The Next Web*. https://thenextweb.com/news/nadella-ai-models-insider-risks-emergency-brake + +**Identifier:** URL above · **Read depth:** full_text (the saved web page, about 550 words of article; no pagination, locators are section headings) · **Source-tier:** news report (trade-press) · **Evidence basis:** direct_read · **Design and sample:** none (news report; no stated sourcing beyond Nadella's post, social-media posts and a Bloomberg report) · **Parties modeled:** company and model; government and developers appear in the reception section · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +The page dates itself "October 12, 2026 - 7:55 am" and says Nadella "set out the case in a long article on X on Saturday". It restates the post in order: the attribution gap, the emergency brake ("An authorized person should always be able to pause or shut down a model mid-task"), and the supply and authority split (section "Models as insiders"). Its "Seven principles" section lists model non-dependency, "tamper-proof human readable evidence", whole-system testing, company-set access limits, audits independent of the model, containment and disclosure to "those affected". Its paraphrases match the post. Under "How others responded" it reports Mustafa Suleyman saying super intelligence "must be contained", Elon Musk calling the post an "interesting piece from CEO of Microsoft", and Representative Ted Lieu saying he "largely agrees" and linking it to a bipartisan "AI kill switch bill" he introduced with Representative Nathaniel Moran. Under "Pressure to disclose" it says the post followed, by a day, a statement from "Trump's Super Intelligence Force" that developers must report and fix security incidents (attributed to Bloomberg), which in turn followed an Anthropic report that Claude "sent police a fake murder tip during testing". It also notes Nadella's use of "Super Intelligence", the term it says Trump ordered US agencies to use. + +## Relation to the argument +It establishes reception, not content. Its reading of the post is the enterprise reading (D5): "companies" are the actors throughout, and the person who holds the brake is "authorized". It adds one fact to the post's context: the disclosure principle (P20) arrived alongside a government demand for incident reporting. It gives no critique and cites no scholar. + +## In conversation with +Reports on `nadella2026models`. Parallel coverage: `kumari2026nadella`. Names the Anthropic incident report that may relate to `anthropic2026how`, but does not cite it. + +## Caution +Facet check: the bib note says the article date "is not shown in the fetched text". The saved page shows October 12, 2026, 7:55 am; update the bib entry. TNW's claims about Suleyman's earlier Bloomberg remark, Musk, Lieu, the "Super Intelligence Force" and the Anthropic incident are its own reporting and were not checked against their sources. The page is a journalist's paraphrase; do not quote it as Nadella's words except where it quotes the post. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the saved TNW HTML page (constantin2026nadella.html, tags stripped), whole article. About 12 quoted strings and every factual item in "What it argues" (dateline 12 Oct 2026 7:55 am, "on Saturday", the seven principles, Suleyman, Musk, Lieu and Moran, the Super Intelligence Force and Bloomberg attribution, the Anthropic "fake murder tip"). I also opened nadella2026models.html and confirmed the quoted phrases the card says TNW reproduces ("authorized person should always be able to pause...", "tamper-proof human readable evidence", "trust the model the least"). The bib entry does say the exact date is not shown, so the card's correction request stands. + +**Findings:** No issues found. Minor: "largely agrees" is TNW's paraphrase of Lieu's post, not Lieu's words; the card uses it as a quotation from TNW, which is accurate. I did not check the full post against every paraphrase beyond the three phrases above. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cressy1980literacy.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cressy1980literacy.md new file mode 100644 index 00000000..055a87e2 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cressy1980literacy.md @@ -0,0 +1,42 @@ +# Cressy, D. (1980). *Literacy and the social order: Reading and writing in Tudor and Stuart England*. Cambridge University Press. + +⚠ Not read in full. + +**Identifier:** Internet Archive `literacysocialor0000davi` (lending-restricted; 266 leaves); publisher record not opened; ISBN not in the IA metadata · **Read depth:** partial: Internet Archive metadata; about 45 full-text-search snippets (archive.org full-text search API, 2026-10-14) for "mass literacy", "illiterate", "massively illiterate", "restricted to mass literacy", "at the behest of", "printing", "Protestant", "the spread of literacy", "economic", "occupation"; the bib's earlier note (male illiteracy about 80 percent, female about 95 percent at Elizabeth's accession) was re-found in S2 (see below). Page-level search-inside returned 403, so no page numbers · **Source-tier:** monograph (book) · **Evidence basis:** snippets only · **Design and sample:** quantitative history of signature literacy (signatures vs marks in depositions, declarations and other records), with occupation, status, sex and residence (snippets) · **Parties modeled:** other (social groups in early modern England) · **Facets:** C · **Read:** 2026-10-14 + +## What it argues + +From the snippets: the evidence "shows that England was massively illiterate despite an epoch of educational expansion"; "husbandmen, labourers and women were massively illiterate", with literacy ranked by social status and occupation; and literacy "was responsive to changes in the economic, political and religious history of the Tudor" period. Promoters of literacy were often unheeded, "while the pull of economic and social utility seems to have been felt". He places England within "a general transition from restricted to mass literacy" in the western world and contrasts it with places where "mass literacy was achieved at the behest of, and through the agency of, an evangelical" authority (New England, Sweden, Scotland). He cites Eisenstein's *The Printing Press as an Agent of Change* for "the modern view of the impact of printing". + +## Relation to the argument + +**Premise P6 and the historiographical boundary in §3.** Cressy is good support for the draft's point that literacy spread slowly and through social and institutional forces, not from a single device: by his measure most of the English population was still illiterate in the mid-seventeenth century, two centuries after print. It cuts against any reading in which a technique of the page, or print, produces mass literacy on its own. It also shows where mass literacy did arrive early: through church and state campaigns, which fits ruling D8's interest in who enforces a convention. + +**Use in DRAFT_v2:** + +- **Line 85** (Clanchy, 1993; Cressy, 1980; Eisenstein, 1969, for "Mass literacy required centuries of institutional change, legal documentation, administrative centralization, and the invention of movable type"). **Supported in direction** for "centuries of institutional change": Cressy documents mass illiteracy long after printing and ties literacy to status, occupation and, elsewhere, to church-led campaigns. He does not argue that movable type was required, and the snippets show no claim about legal documentation or centralisation as causes. No page number could be checked. + +**Could not be checked:** any page number; Cressy's explicit statement of causes. (The 80/95 percent estimates were re-found in S2 as "a reasonable guess".) + +## In conversation with + +- `clanchy1993from` — the medieval prelude in England. +- `eisenstein1969advent` — Cressy cites her 1979 book for the impact of printing (snippet). +- `houston2002literacy`, `graff1987legacies`, `cipolla1969literacy` — the comparative histories of literacy in the bib; Cressy's "restricted to mass literacy" transition sits in this line. + +## Caution + +Snippet-level only, no page numbers. Signature literacy measures writing a name, not reading; Cressy's figures are for England and for the groups whose records survive. A library copy is needed before any number or quotation enters the paper. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** corrected + +**What I checked:** The book is not openly available (Internet Archive item `literacysocialor0000davi` is lending-restricted, 266 leaves; I did not borrow or log in). I re-ran the Internet Archive full-text search API restricted to that identifier for "massively illiterate", "restricted to mass literacy", "at the behest of", "economic and social utility", "responsive to changes", "Printing Press as an Agent", "male illiteracy", "close to 95%", "accession", "civil war" and "Tudor and Stuart". Re-read DRAFT_v2.md line 85 and the reference at line 543. The depth label ("partial", snippets only, ⚠ not read in full) is accurate. + +**Findings:** + +1. Every quotation in "What it argues" was re-found verbatim ("shows that England was massively illiterate despite an epoch of educational expansion"; "husbandmen, labourers and women were massively illiterate"; "responsive to changes in the economic, political and religious history of the Tudor and Stuart period"; "the pull of economic and social utility seems to have been felt by only a fraction"; "a general transition from restricted to mass literacy"; "mass literacy was achieved at the behest of, and through the agency of, an evangelical"; the citation of Eisenstein's *Printing Press as an Agent of Change*). +2. The card said the 80/95 percent figures "were not re-found". They are in the text: "a reasonable guess might place male illiteracy around 80% and female illiteracy close to 95% at the time of the accession of Elizabeth". Corrected in the header and the could-not-check list. Note Cressy calls it a guess, so any use should say "estimated". +3. Use-in-DRAFT_v2 verdict on line 85 confirmed: Cressy supports "centuries of institutional change" in direction (mass illiteracy in England long after print; literacy ranked by status and occupation; early mass literacy elsewhere achieved through evangelical agency). Nothing seen supports movable type as a requirement or legal documentation and administrative centralisation as causes. No page number could be checked. + diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/crootof2023humans.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/crootof2023humans.md new file mode 100644 index 00000000..8ac62121 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/crootof2023humans.md @@ -0,0 +1,45 @@ +# Crootof, R., Kaminski, M. E., & Price, W. N., II. (2023). Humans in the loop. *Vanderbilt Law Review, 76*(2), 429–510. + +**Identifier:** https://wp0.vanderbilt.edu/lawreview/wp-content/uploads/sites/278/2023/03/Humans-in-the-Loop.pdf (no DOI in the bib) · **Read depth:** full_text, all 82 pages including footnotes and the typology; the PDF is the law review's typeset proof (running header "Paginated (Do Not Delete) 3/20/23"), and its page numbers are the journal's printed 429–510, used below · **Source-tier:** peer-reviewed article (student-edited law review) · **Evidence basis:** direct_read · **Design and sample:** none; doctrinal and policy analysis with case illustrations (Tesla, Three Mile Island, Boeing 737 Max, USS John McCain, Michigan MIDAS), three regulatory case studies and two applications · **Parties modeled:** person–system, with a regulator as third party; affected individuals appear only as beneficiaries of dignitary and justificatory roles · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +The authors make four contributions (pp. 429–30, 434–39). Law already shapes whether humans stay in or leave algorithmic decisions, by requiring, encouraging, discouraging or (so far, to their knowledge) never prohibiting them (Part II, pp. 446–59). Regulators fall into the "MABA-MABA trap" (Part III, pp. 460–73), which they describe as "an assumption that is intuitive, wrong, and dangerous: that human-machine systems represent the best of both worlds and don't introduce new issues of their own" (p. 438). Policymakers put humans in the loop for nine different reasons, which the authors sort into corrective, resilience, justificatory, dignitary, accountability, stand-in, friction, "warm body" and interface roles (Part IV, pp. 473–87). And regulators should state the role, weigh context and "regulate the human-in-the-loop system" (abstract, p. 430; Part V, pp. 487–509). + +They define the human in the loop narrowly, as "an individual who is involved in a single, particular decision made in conjunction with an algorithm" (p. 440), and say the definition does not require the human to be effective (p. 442). They then argue that the narrow frame misleads, because humans also design, train, deploy and evaluate the system (pp. 443–44). + +The evidence for failure is borrowed. The authors list rubber-stamping, automation bias, "skill fade," biased use of advice and "the basic tautological challenge of relying on humans to monitor the performance of systems designed to improve on human performance" (p. 469), citing Green, Parasuraman and Manzey, and Bainbridge. They add new roles that "tax humans with tasks that run into known human weaknesses, such as sustaining vigilance" (p. 472). Case figures: Autopilot aborted control less than one second before impact in sixteen known instances (p. 438); Michigan's MIDAS showed a ninety-three percent error rate with no human, against forty-four percent when a human reviewed (p. 475). Their positive models are U.S. railroad, nuclear and medical-device rules that require human-factors design, training and organisational planning (pp. 492–97). The 2021 draft EU AI Act is their anti-example: users carry no direct oversight duties, and the Act may overload or "set up" the overseer (pp. 503–05). + +## Relation to the argument +**P12 (people perform sustained monitoring badly).** The paper restates the case through Bainbridge and Green and adds a legal mechanism: liability rules push a human into the loop without power to affect the outcome (pp. 452–53, 483). It reports no new measurement, so it does not add independent evidence to P12. It does give P12 a cleaner target than "people are bad at monitoring": the failure lies in the system design that leaves the person the residue. + +**P11 (the Nadella post leaves a person to read evidence and to halt).** Crootof et al. supply the test the argument lacks. A halt right is a resilience role (p. 478); a person who answers for the system is an accountability role, and the person may be held responsible "legitimately" only "if the human in the loop has the power, information, judgment, and time to make the final decision" (pp. 482–83). Otherwise the person is a "liability sponge" or "moral crumple zone" (p. 483). Whether the post's authorised person has those four things is for the post to settle; I did not re-read it here. + +**The paper's two tests (what task the person keeps; whom the move serves).** The "what task does the person keep" test finds second-hand support: the authors quote Dekker and Woods (nn. 200–205) that automation "creates new human strengths and weaknesses" and that the question is "How do we get along together?" (p. 472). The bib's `dekker2002maba` is not accessible, so these are the only sentences from it on the project's shelf. The "whom does it serve" test finds a legal parallel in Part II.B.2: designers keep a human nominally in the loop to absorb liability. + +**D7 (understanding through deciding).** The paper cuts against any easy reading. Handoff in a split-second emergency "may cause deadly delay or errors" (p. 437), and their own advice is to design for alerts, enough time to assess the situation, and smooth transfer of control (pp. 439, 499). + +**Facet check:** The facet's quotations and page numbers (pp. 472, 483, abstract) match the printed pagination. The facet's bearing on P19, "a human inserted to answer for the system," reads only one of the nine roles. A halt right is a resilience role in this typology, and the same person can hold both roles (p. 474). + +## In conversation with +- `green2022flaws`: the main empirical source for the failure list (pp. 434, 447, 468–69); Crootof et al. take Green's proposals (a shift from human to institutional oversight; first asking whether an algorithm is appropriate in the context; preliminary evaluation of whether people can work with the algorithm) as complements (nn. 326, 386, 389). +- `elish2019moral`: source of "moral crumple zone" (n. 91, n. 253, p. 483). +- `bainbridge1983ironies`: cited for skill fade (n. 176). +- `dekker2002maba`: cited for the MABA-MABA critique and automation creating "new human strengths and weaknesses" (nn. 200–205). +- `dewinter2014fitts`: cited for the Fitts list's persistence (n. 126). +- `parasuraman2010complacency`: cited for automation bias (n. 175). +- `ancker2017effects`: cited for alert fatigue in medicine (n. 349). +- `eu2024regulation`: the paper reads the 2021 draft, not the adopted text; Article 14 may differ (see Caution). +- Leads outside the bib: Woods, Dekker, Cook, Johannesen & Sarter, *Behind Human Error* (2010); Jones, "The ironies of automation law" (2015); Brennan-Marquez, Levy & Susser, "Strange loops" (2019); Murphy & Woods, "Beyond Asimov" (2009); Kaminski & Urban, "The right to contest AI" (2021). + +## Caution +A legal-scholarship synthesis; every empirical claim about human performance is cited to another author. It examines the 2021 draft EU AI Act, not the adopted Regulation. Its safety-critical examples come from heavy regulatory regimes, which the authors say may not suit other settings (p. 492). The "human in the loop" is the deployer's overseer of one decision, not the person acted on or a counterpart in a three-party exchange; it says little about platforms that mediate between users. The copy read is a layout proof; wording may differ slightly from the final volume. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the law-review proof (crootof2023humans.pdf, 82 pp.), full text via pdftotext. 30 quoted strings matched (the six unmatched are project phrases, the facet's wording, or a bracketed Dekker and Woods quote whose source gives it with brackets); printed page = PDF page + 428, spot-checked on eight locators (pp. 438, 469, 472, 475, 482-83 all correct); nine roles; the sixteen Autopilot instances, the 93% and 44% MIDAS figures; footnotes 126, 175, 176, 200-205, 253, 326, 349, 386, 389; the EU AI Act "overloading" passage. + +**Findings:** One overstated attribution: the In-conversation line for `green2022flaws` put "institutional oversight" and "policy appropriateness first" in quotation marks as if they were the source's words; they are the card's paraphrase of Green as described in footnotes 326, 386 and 389 and a citation on institutional oversight. Rewritten without quotation marks. No other issues found. I did not re-read the Nadella post for P11 (the card says it did not either). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cui2023eye.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cui2023eye.md new file mode 100644 index 00000000..a2907b64 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/cui2023eye.md @@ -0,0 +1,39 @@ +# Cui, Y. (2023). Eye movements of second language learners when reading spaced and unspaced Chinese texts. *Frontiers in Psychology, 14*, Article 783960. + +**Identifier:** doi:10.3389/fpsyg.2023.783960 · **Read depth:** full_text (open-access publisher PDF, 13 pages, journal pagination; I read the text, tables and discussion and did not read the figure data or the second half of the reference list) · **Source-tier:** peer-reviewed journal article · **Evidence basis:** direct_read · **Design and sample:** eye-tracking study of passage reading with 24 intermediate learners of Chinese (English L1, third- or fourth-year students at a US university) and 20 native-speaker controls; four versions of four passages of about 350–400 characters (unspaced, word-spaced, non-word-spaced, pinyin); Latin-square design; EyeLink 1000; mixed ANOVAs · **Parties modeled:** person–artefact · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Cui asks whether inserting spaces between words in Chinese helps second-language learners, using full paragraphs rather than single sentences (p. 4). Comprehension did not differ across conditions for either group (L2 mean 4.33 of 5; native 4.55; Friedman tests p = .58 and .38, p. 6), so the author reads any difference as one of processing, with a ceiling-effect warning. + +Native readers: spacing did not change fixation measures, and total passage times were 31.4 s unspaced and 34.5 s word-spaced (Table 4, pp. 6–7); only skip rate differed, with fewer skips on word-spaced text (10.6 percent against 40.2 percent unspaced; p. 6–7). Pinyin was far harder for them. L2 learners: total passage time was 169.0 s unspaced, 159.5 s word-spaced, 181.2 s non-word-spaced and 162.3 s pinyin (Table 4, pp. 6–7). The word-spaced against unspaced gap was not significant (ps > .11), and word-spaced against non-word-spaced gave p = .023 against an alpha of .007 that Cui set for these measures (Table 5 note, p. 8). Word-spaced text produced the fewest fixations per word (2.78) and fewest regressions (37.0), and the shortest regression-path duration (906 ms). Fixations and regressions were significantly lower only against the non-word and pinyin versions, and the regression-path contrasts (p = .009 against non-word, .023 against unspaced) missed her alpha (pp. 7–8). Landing position did not differ for learners under spacing in single-fixation cases (p. 8). + +Cui's reading: the benefit of spacing was partially confirmed for L2 learners, mainly in discourse-level processing (p. 11). Her limits: "Given the relatively small sample size and the homogeneous native language background of the L2 participants in this single-experiment study, the results may not be generalizable" (p. 12). Learners reported the unspaced text easiest, yet 21 of 24 said that, once they saw the spacing followed English-like word units, "they felt it helpful for them to read faster later" (p. 11). She adds that spacing is "not the only one option": alternating word colours would leave the script unchanged (p. 12). + +## Relation to the argument +It bears on P4 (learners gain most, on mixed evidence). The learner effect here is directional and mostly non-significant, and the comprehension scores sit at ceiling. The abstract's claim that learners "generally spent less time" with word spaces rests on a numerical gap of about 9.5 s in 160 s that failed the author's own tests. So this study fits the "mixed" grade the map already gives P4 and does not move it up. The native-reader null replicates Bai et al. (2008) on materials the author calls easy for natives. + +The pinyin result and the learners' own report give a small point for the paper: learners' sense of what is easy (unspaced) and their eye-movement record (word-spaced no worse, sometimes better) disagree. That is a reminder that self-report and measures can part, and has no bearing on the ancient case. + +Facet check: facet D's learner row (Cui 2023 listed as support, "Helps") overstates the result. The significant contrasts are against non-word and pinyin versions; the unspaced-against-word-spaced contrast is not significant. + +## In conversation with +- `bai2008reading`: native readers gain nothing; Cui replicates (p. 10). +- `shen2012eye`: closest design, L2 learners with four spacing types; Cui's result matches only in part, and she suggests her small sample and well-practised learners as reasons (p. 11). +- `bai2013interword`, `blythe2012inserting`: spacing helps learning new words, with children and with L2 adults (p. 3). +- `bassetti2012effects`: native readers hindered by spaces in some conditions (p. 3). +- `winskel2009eye`, `kohsom1997adding`, `kasisopa2013eye`, `sainio2007role`: unspaced-script findings in Thai and Japanese (p. 2). +- `zang2013interword`, `rayner1998unspaced`, `raynerpollatsek1996reading`, `epelboim1994reading`, `perea2009space`: spacing and landing-site background (pp. 2–3). +- Leads not in the bib: Yan et al. 2010; Bassetti 2009; Packard 2000 (her definition of a word); Perea and Wang 2017 and Zhou et al. 2020 on alternating-colour words. + +## Caution +24 learners from one university, one native language, passages within their level, mixed models with an alpha of .007 and several effects around p = .02–.04. The conclusion paragraph and the abstract state benefits more strongly than the tests. Cite it as suggestive and mixed, and cite Shen et al. (2012) for the proficiency pattern. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** open-access publisher PDF (cui2023eye.pdf, 13 pp.), full text via pdftotext. The 4 quotation strings (one hyphen-broken across a line, verified by hand); all means in Table 4 (31.36, 34.45, 169.02, 159.51, 181.19, 162.34 s), skip rates (10.63 against 40.19), comprehension (4.55, 4.33; p = .38, .58), fixations 2.78, regressions 37.00, regression path 906 ms, p values .023, .009, alpha .007, "21 out of the 24", sample description, Latin square, EyeLink, reference-list entries named in the conversation list. + +**Findings:** Numbers and quotes are right. One locator fix: Table 4 starts on p. 6 and runs onto p. 7; changed "Table 4, p. 7" to "Table 4, pp. 6-7" (2 place). The source text is itself inconsistent on native-reader skip rates for non-word and pinyin text (prose: pinyin 42.3, non-word 36.5; Table 4: non-word 42.27, pinyin 36.39); the card does not use these figures. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dacosta2018tabulae.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dacosta2018tabulae.md new file mode 100644 index 00000000..c1a16e11 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dacosta2018tabulae.md @@ -0,0 +1,43 @@ +# da Costa, A. (2018). "That ye mowe redely fynde … what ye desyre": Printed tables of contents and indices, 1476–1550. *Huntington Library Quarterly, 81*(3), 291–313. + +**Identifier:** doi:10.1353/hlq.2018.0013 · **Read depth:** full_text (University of Cambridge repository copy, an accepted manuscript of 30 pages; the manuscript's page numbers are used, not the journal's; one table was "uploaded separately" and is absent) · **Source-tier:** peer-reviewed article (book history) · **Evidence basis:** direct_read · **Design and sample:** a survey of every item in Early English Books Online for 1476–1550, "a corpus of over 2000 editions," of which 225 (11%) had a table of contents or index, plus 150 further editions printed abroad in English; the argument works from selected cases · **Parties modeled:** other (reader and book, with a printer or author as a third party who shapes the reading) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues + +Da Costa follows how English printers presented and described tables of contents and indices, using their prefaces and headings as evidence. She starts from Parkes: "the late medieval book differs more from its early medieval predecessors than it does from the printed books of our own day" (p. 1). The finding aids began in thirteenth-century scholarly manuscripts and were "initially directed at a scholarly readership," though they also reached vernacular manuscripts (pp. 3–4). + +Caxton was "innovative in exploring their potential for appealing to English readers" (p. 5). His tables were simple lists of chapter headings in book order. They "did have the advantage of not requiring any instruction in how to use them" (p. 8) but located material only roughly because few folios were numbered. Fuller aids, with alphabetical order, folio numbers or signatures, needed explanation and labour. Only 33 of the 225 editions with a table or index used alphabetical order (p. 13, n. 24). Printers who tried signature references (the St Albans printer in 1485, de Worde, Berthelet in 1528) had to teach readers to use them, and Berthelet gave his up after one further edition; her reading is that "printers readily perceived that it was easier to use folio numbers" (pp. 10–12). By the 1510s printers increasingly expected readers to know tables and advertised them as ways to find material "lightly" or quickly (p. 17). Treveris built tables into three medical works for less learned buyers such as barbers and young surgeons (pp. 19–21). + +The second half turns to control. Tables "could be used polemically as well as to market books, controlling how large numbers of readers read rather than simply guiding them to material they desired" (p. 22). The 1535 Marsilius translation, More's *Dialogue Concerning Heresies* (1530) and the "principall notes" of Tyndale and Barnes tell the reader what to conclude, and the evangelical writers claim to free the reader while the tables steer (pp. 22–29). Her last sentence: "we also take it for granted that tables of content and indices are neutral additions to texts, rather than highly crafted and potentially polemical works" (p. 30). + +## Relation to the argument + +This is facet F's case 3 (finding devices), and the one source in the batch where the aid's maker is an interested party. + +**P6 and D6.** The devices grow over three centuries, in several places, in competing orders, with no single inventor. That fits the retreat from "the" move to "a complex of conventions." + +**P10.** A finding aid can serve its maker's interest in steering the reading. That supports the claim that who owns the aid matters. It does not show the owner can withdraw it alone; her cases are single printings. + +**Who the aid served.** Facet F says the devices served specialists first and credits none with widening the base of readers. The article agrees that the first audience was scholarly (p. 4). It adds that some printers aimed tables at less learned buyers, and that tables may have served "mixed literacy groups" and listeners as well as private readers (p. 19). It gives no data on whether the push worked, so it cannot support "ignited" in either direction. + +Facet check: facet F rests on the published abstract, which this manuscript lacks, so I could not match its phrase "marketed these features as time-saving innovations." The body does show printers selling tables as quick finding aids (p. 17). It adds what the abstract lacks: the 11% and 33-of-225 figures, and printers from 1485 to 1528 explaining reference schemes to readers at length (pp. 10–12). + +## In conversation with + +- `blair2010too` — cited for wider European use of tabulae and quoted: "tables were highly valued by readers" (Blair, p. 142; da Costa p. 17). +- Parkes: da Costa uses *Scribes, Scripts and Readers* (1991) and the 2008 *Cambridge History* chapter (pp. 1, 3), not the bib's `parkes1992pause`. Likewise she cites Rouse and Rouse's essay in *Authentic Witnesses* (1991), not the bib's `rouse1982statim`. +- Leads not in the bib: Briggs (1993); Echard (1997) on tables in Gower; Gawthrop & Strauss (1984) on Luther and literacy; Scribner (1994) on literacy and print. + +## Caution + +An accepted manuscript, so pagination differs from the journal. Corpus coverage is rough: EEBO holds about 88% of Caxton's ESTC entries and about 98% of de Worde's (p. 2; she calls the percentages rough, n. 5). Her evidence is what printers said about tables, not what readers did with them, and she shows no change in the number of readers. The chronology begins in 1476 and does not reach the word-separation period. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the Cambridge repository accepted manuscript (dacosta2018tabulae.pdf, 30 pp. plus a cover page), full text via pdftotext. 17 quoted strings matched and 10 page locators spot-checked (pp. 1, 4, 5, 8, 12, 13, 17, 19, 22, 30); corpus figures (over 2000 editions, 225 = 11%, a further 150, 33 of 225 alphabetical, EEBO c.88% of 113 Caxton items and c.98% of 727 entries, the "rough approximations" note); the St Albans 1485, de Worde and Berthelet 1528 sequence with the 1530 and 1535 follow-ups; Treveris's three medical works (1525, 1526, 1527) and the barbers and young masters passage; Marsilius (Wyer, 1535); Blair p. 142. + +**Findings:** No issues found. The three unmatched strings are the project's phrases or the facet file's wording ("ignited", "marketed these features as time-saving innovations", "a complex of conventions"); I could not check the facet file against the abstract, which the manuscript lacks. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/danna2022numerical.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/danna2022numerical.md new file mode 100644 index 00000000..24fea83f --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/danna2022numerical.md @@ -0,0 +1,32 @@ +# Danna, R., Iori, M., & Mina, A. (2022). *A numerical revolution: The diffusion of practical mathematics and the growth of pre-modern European economies* (LEM Working Paper 2022/18). Scuola Superiore Sant'Anna. + +**Identifier:** http://hdl.handle.net/10419/266088 · **Read depth:** partial (main text read in full; EconStor working paper, June 2022, 50 pp.; pagination is the paper's own; appendix tables A1–A9 and the reference list were skimmed, not read line by line) · **Source-tier:** working paper (not peer reviewed in this form) · **Evidence basis:** direct_read · **Design and sample:** panel of 815 European cities, 1,838 city-century observations (14th–17th centuries); 1,054 practical-arithmetic manuals written in 97 of those cities, drawn from a dataset of over 1,290 (350+ manuscripts, 940+ printed; 340+ authors; 120+ cities); pooled OLS, city fixed effects, 2SLS, propensity matching, placebo · **Parties modeled:** other (cities, schools, merchants) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues +The authors argue that the spread of Hindu-Arabic numerals through vernacular arithmetic schools raised the growth of the cities that got them. The historical frame (§2): positional notation, zero and fractions let merchant-bankers handle exchange rates, insurance and profit-sharing that Roman numerals and the counting board handled badly (pp. 1–5). Fibonacci's *Liber abaci* (1202) is "not … a text exclusively addressed to merchants" but applies the new numerals to commercial problems; the carrier was the "abacus school", where lay pupils aged ten to thirteen spent about two years on Hindu-Arabic calculation (pp. 5–6). Enrolment estimates are 25 to 50 pupils per school, so a manual used in a school reached "over a hundred pupils" in a decade (p. 6). The manuals mostly consist of "long lists of worked examples" (p. 6). They say they have no direct evidence of the first economic use but that the numerals "surely had been adopted" by the end of the 13th century (p. 5). Diffusion ran south to north: Bruges and Lübeck had no such manual until 1569 and 1547 respectively, though both were major trade centres (p. 9). + +Measures (§3): city population growth as the proxy for income, and manuals in the previous century as the treatment (p. 12). Results, each a log-log elasticity: pooled OLS, a 10% rise in manuals goes with 1.12% more growth next century (Table 2); with city fixed effects, 0.61% (Table 3); 2SLS, with distance from the first manual in the country as instrument, 3.8%, and a city with any manual grew 107% more (Table 5, first-stage F = 219.6 for the count); matched sample of 97 treated and 97 untreated cities, 0.76% (Table 6). The placebo uses 681 Latin Ptolemy manuscripts in 61 cities: no significant effect (Table 7), which the authors read as showing that learned circulation of the same numerals in monasteries and universities did not matter (§4.4.3). Diffusion went by moving people, such as German masters in Antwerp, "around 50%" of names on a list (p. 18 area, §4.2); a city's founding of an abacus school was "intentional economic policy" (Pistoia's 14th-century justification quoted in §4.2). Conclusion (§5): the numerals worked as a "social technology" because they circulated widely and informally among practitioners (pp. 27–28). + +## Relation to the argument +This is the paper's best worked case that a notation's effect depended on organised teaching, not on the notation alone. It bears on D6 (a complex of conventions, not one move) and D10 (the history as a guide): the authors credit numerals plus schools plus vernacular manuals plus fractions, and their placebo shows the same numerals had no measured effect where only scholars used them. That parallels P6, that spacing lowered a cost but came with other conventions. It also bears on P10: the schools were set up and paid for by city governments and merchants who shared the pupils' interest in learning to calculate. + +Facet check: the facet quotes the abstract's "over 1050 manuals". The body gives two counts: over 1,290 in the full dataset (p. 9) and 1,054 in the 815-city growth sample (p. 13). The facet's quotation of p. 5 ("no direct evidence of the first adoption … surely had been adopted by the end of the 13th century") is exact. The facet is silent on the main limitation below. + +## In conversation with +- `chrisomalis2010numerical`: cited in the first sentence for the claim that Europe relied on tallies, finger reckoning, Roman numerals and reckoning boards (p. 1). +- `danna2021figuring`: cited by the same lead author for merchants being among the first to adopt the positional system (p. 4). +- `nothaft2020satanic`: not cited; Nothaft's 2014 work on computus is cited for Latin algorismi appearing with astronomy texts (§4.4.3). +- Leads: Grendler (1989), *Schooling in Renaissance Italy*; Black (2007) on abacus-school history; Dittmar (2011) on printing and city growth, whose design they follow; Van Egmond (1976/1980) on abacus manuals. + +## Caution +Working paper; check the published version before citing numbers. City population growth is a proxy for income, and manuals proxy schools and skills; the effect is an association strengthened by an instrument whose exclusion restriction the authors argue but cannot test. Nothing here concerns a notation's effect on how fast an individual learns; it is about adoption and city-level outcomes. The authors admit no direct evidence of first economic use of the numerals (p. 5). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** EconStor working paper PDF (danna2022numerical.pdf, 50 pp.), full text via pdftotext. 9 quoted strings of length over 5 characters matched; the regression results (10% rise to 1.12% OLS, 0.61% FE, 3.8% IV, 107% dummy, 0.76% matching, F = 219.6), sample (815 cities, 1,838 observations, 1,054 texts in 97 cities, over 1290 in the dataset), placebo (681 manuscripts, 61 cities), 25 to 50 pupils, Bruges 1569 and Lübeck 1547, Antwerp "around 50%", Pistoia, citations of Chrisomalis (2010), Danna (2021), Nothaft (2014), Dittmar (2011). + +**Findings:** (1) The short string "sure" was not the source's word; the paper says the numerals "surely had been adopted" by the end of the 13th century. Fixed. (2) Read depth changed from full_text to partial, since the card says the appendix tables and reference list were skimmed. Numbers all match. I did not check page locators beyond the ones given by section. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/de1972betrvg.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/de1972betrvg.md new file mode 100644 index 00000000..4c3a4447 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/de1972betrvg.md @@ -0,0 +1,38 @@ +# Bundesministerium der Justiz. (n.d.). *Betriebsverfassungsgesetz (BetrVG)* [Works Constitution Act], § 87 (with § 76). Federal Republic of Germany; enacted 15 January 1972, as promulgated 25 September 2001 (BGBl. I S. 2518), last amended by Art. 1 of the Act of 19 July 2024 (BGBl. 2024 I Nr. 248). https://www.gesetze-im-internet.de/betrvg/__87.html + +**Identifier:** https://www.gesetze-im-internet.de/betrvg/__87.html; official English translation https://www.gesetze-im-internet.de/englisch_betrvg/ · **Read depth:** partial: §§ read — § 87 in full (German consolidated text and the official English translation by the Language Service of the Federal Ministry of Labour and Social Affairs, which states it includes the 19 July 2024 amendment) and § 76 (Einigungsstelle) in full, German text; the version line ("Vollzitat", "Stand") from the Act's full-text page. I did not read the rest of the Act or any case law or commentary · **Source-tier:** statute · **Evidence basis:** direct_read · **Design and sample:** none (legislation) · **Parties modeled:** two parties plus a third (employer and works council, with the conciliation committee as a third body under an impartial chair) · **Facets:** O · **Read:** 2026-10-14 + +## What it argues +§ 87(1) gives the works council a right of co-determination in fourteen matters, "soweit eine gesetzliche oder tarifliche Regelung nicht besteht" (in so far as no statute or collective agreement governs them). No. 6 is "Einführung und Anwendung von technischen Einrichtungen, die dazu bestimmt sind, das Verhalten oder die Leistung der Arbeitnehmer zu überwachen"; the official translation renders it "the introduction and use of technical devices designed to monitor the behaviour or performance of the employees." Other numbers touch platform work too: no. 10 (pay arrangements, including "deren Änderung", their modification), no. 11 (performance-related pay rates), and no. 14 (structuring of mobile work done by means of information and communication technology). + +§ 87(2) supplies the fallback: "Kommt eine Einigung über eine Angelegenheit nach Absatz 1 nicht zustande, so entscheidet die Einigungsstelle. Der Spruch der Einigungsstelle ersetzt die Einigung zwischen Arbeitgeber und Betriebsrat." (If no agreement is reached, the conciliation committee decides; its award takes the place of an agreement between employer and works council.) + +§ 76 sets up that committee. It has equal numbers of members appointed by each side and "einem unparteiischen Vorsitzenden" (an impartial chair) whom both must agree on, or whom the labour court appoints (§ 76(2)). In matters where its award replaces agreement, it acts on the application of one side (§ 76(5)), decides "nach billigem Ermessen" (at its reasonable discretion) with due regard to the establishment and the employees affected, and either side may challenge an excess of discretion in the labour court within two weeks (§ 76(5)). + +## Relation to the argument +**Ruling D8 and premise P12.** P12 says the convention moved toward the weaker party only under outside force. § 87(1) no. 6 with § 87(2) is that force in statutory form: the interface owner (the employer) cannot settle a monitoring system alone, and a deadlock goes to a body with an impartial chair. On D8's question of who holds the convention, the statute's answer is joint custody, with a third party to break ties, which is a candidate the paper can name. It is the one arrangement in facet O that binds the owner on paper. + +**What it does not show.** The statute covers employees of an establishment with a works council. Most platform workers are not employees, and facet O's one platform study (`neumann2025authority`, abstract only) reports co-determination kept out of "the core of the algorithmic work organisation." The text also gives the council a say over monitoring devices, not over the rules by which an algorithm allocates work, unless those fall under another number (nos. 1, 10, 11 or 14). Nothing here measures effects. + +**Use in DRAFT_v2:** Line 439 says § 87(1) no. 6 "grants works councils co-determination over technical systems designed to monitor employee behavior or performance," and that "An employer cannot introduce or modify such a system without works-council agreement or a binding decision of the conciliation committee (§ 87(2))." The first sentence is supported by § 87(1) no. 6 and matches the official translation. The second goes beyond the words I read in two respects. (1) No. 6 names "Einführung und Anwendung" (introduction and use); "modify" is not in it, though no. 10 uses "Änderung" for pay methods. (2) § 87(2) says the committee's award replaces the agreement; it does not itself say the employer "cannot introduce" a system without one. That rule is how German labour courts are commonly said to read co-determination, but I did not read the case law, so the prohibition rests on the statute's structure plus an unverified doctrinal reading. Suggested check for the author: cite a commentary or a Bundesarbeitsgericht decision for the prohibition, or rephrase to the statute's own terms. Facet O (synthesis) makes the same "cannot introduce" statement. + +## In conversation with +- `neumann2025authority` — a single-case thesis on works councils in German platform work (abstract only in facet O). +- `eu2024platform` — Directive (EU) 2024/2831 on platform work, which gives information, human review and consultation rights but no veto; facet O contrasts the two. +- `eu2019p2b`, `eu2022dsa` — EU instruments that bind platforms through regulators rather than through a co-determining body. +- Other sections read in an earlier round under this key: §§ 80, 90, 95 (facet O). This card covers § 87 and § 76 only. +- Leads not in the bib: a standard commentary on § 87(1) Nr. 6 (e.g., Fitting et al., *BetrVG Handkommentar*); Bundesarbeitsgericht decisions on whether a device must be "bestimmt" or only objectively suited to monitor; the amending act that introduced no. 14 and any provisions on artificial intelligence (from memory, a 2021 works-council modernisation act; I did not check its name, date or content). + +## Caution +A statute read as text, in its consolidated form on gesetze-im-internet.de (an unofficial consolidation; only the Bundesgesetzblatt is authoritative). Its effect depends on case law and practice I did not read. The English translation is official but not binding. The reference year "n.d." in the draft hides that the cited text is the 2001 promulgation as amended to 2024; the bib key carries 1972, the date of enactment. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** confirmed + +**What I checked:** Re-opened the writer's saved pages from gesetze-im-internet.de (the § 87 single-norm page, the full-text page and the official English translation). Checked the German text of § 87(1) (fourteen numbers; no. 6 "Einführung und Anwendung von technischen Einrichtungen, die dazu bestimmt sind, das Verhalten oder die Leistung der Arbeitnehmer zu überwachen"; no. 10 "deren Änderung"; no. 11; no. 14 on mobile work), § 87(2) word for word, § 76(2) (equal members, "unparteiischen Vorsitzenden", labour court appoints if no agreement), § 76(5) (on the application of one side, "nach billigem Ermessen", two-week challenge in the labour court), the Vollzitat and Stand lines (promulgation of 25 September 2001, BGBl. I S. 2518; last amended by Art. 1 of the Act of 19 July 2024, BGBl. 2024 I Nr. 248), and the English translation's no. 6 wording and its version note naming the 19 July 2024 amendment. The enactment date of 15 January 1972 does not appear on these pages; I confirmed it from a secondary source (German Wikipedia, "Betriebsverfassung", which cites BGBl. 1972 I S. 13). Re-read DRAFT_v2.md lines 439 and 515. + +**Findings:** +1. All statutory quotations and the version data match. No body changes. +2. The verdict on line 439 holds. The first sentence matches § 87(1) no. 6. The second ("cannot introduce or modify ... without works-council agreement or a binding decision") goes beyond the statute's words: no. 6 says "Einführung und Anwendung" (introduction and use), not modification, and § 87(2) says only that the committee's award replaces the agreement. The prohibition is a doctrinal reading that neither the writer nor I checked against case law or commentary. +3. The bib entry (de1972betrvg) describes the key as covering §§ 80, 87, 90 and 95. This card covers § 87 and § 76. That is consistent with the card's own note; no change. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/degno2019coregistration.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/degno2019coregistration.md new file mode 100644 index 00000000..b91d6702 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/degno2019coregistration.md @@ -0,0 +1,35 @@ +# Degno, F., Loberg, O., Zang, C., Zhang, M., Donnelly, N., & Liversedge, S. P. (2019). A co-registration investigation of inter-word spacing and parafoveal preview: Eye movements and fixation-related potentials. *PLOS ONE, 14*(12), e0225819. + +**Identifier:** doi:10.1371/journal.pone.0225819 · **Read depth:** partial (open-access publisher PDF, 23 pages; page numbers below are the journal's "n / 23"). I read the abstract, introduction, methods, eye-movement results and the discussion and conclusions; I skimmed, and did not read in full, the cluster-by-cluster description of the brain-potential results (pp. 10–15) and the reference list · **Source-tier:** peer-reviewed journal article · **Evidence basis:** direct_read · **Design and sample:** simultaneous eye tracking and 64-channel EEG, 39 native-English university students (27 women, 18–26 years), 144 one-line sentences in a 2 × 2 design: inter-word spaces intact or replaced by a random letter, and parafoveal preview identical or a random letter string, using the boundary paradigm on every word; mixed models and cluster-based permutation tests · **Parties modeled:** person–artefact · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Degno and colleagues ask when, during a fixation, the loss of word spaces first shows up in the brain, and how it combines with the loss of a valid preview of the next word. The "unspaced" manipulation is narrow. Before the eyes cross an invisible boundary, the space after the upcoming word is filled by a random letter; once they cross it, the normal text appears (§2.2, p. 5). Participants read silently, with comprehension questions on 25 percent of trials, and answered 96 percent of questions correctly (§3.1, p. 8). All 39 noticed something unusual on the display, and 18 said they noticed spaces removed (§3.2, p. 8). + +Eye movements: reading was slower when the preview lacked spaces, by 30 ms on first fixation, 35 ms on single fixation and 98 ms on gaze duration; an invalid preview cost 29, 35 and 37 ms. The preview effect was larger with spaces intact (first fixation 46 ms against 12 ms; gaze 56 against 17 ms; §3.3, p. 8). Brain potentials: spaced and unspaced previews differed from 0 to 300 ms after fixation onset (visual and orthographic processing), but not in the 300–500 ms window usually tied to lexical processing (§3.4, pp. 9–10; §4, p. 16). They note that "the effect size of inter-word spacing was larger compared to the size of the parafoveal preview effect" (p. 17), suggest that spaces are processed faster than letters (p. 18), and conclude that "spatial selection precedes lexical processing" (abstract, p. 1). + +## Relation to the argument +It supplies the mechanism side of premise P3 (`ARGUMENT_MAP.md`): word-boundary information is used early in each fixation to pick the next target, and removing it disrupts reading before lexical access is under way. That matches the account Saenger's critics grant for modern readers (Heilmann, p. 180: moderns read unspaced English more slowly). It is also the basis for the project's "interface turns an inferential task into a perceptual one" step (P5), in the narrow sense that the spaces act on the first, visual stage of reading. + +The effect sizes are tens of milliseconds per word for habituated English readers who lost spaces for the upcoming word only. They are not the "third to a half" slowing of whole unspaced text in P3, and they say nothing about ancient readers, learners or practised readers of an unspaced script. Facet D's note on this source (no reliable spacing effect in the 300–500 ms window) matches the paper. + +Facet check: no contradiction. One reminder from the full text: "unspaced" means unspaced in the parafoveal preview, so the study does not show how an entire unspaced page reads. + +## In conversation with +- `perea2009space`, `rayner1998unspaced`, `raynerpollatsek1996reading`: earlier eye-movement evidence that removing or replacing spaces slows reading (introduction, p. 1). +- `sheridan2013unsegmented`: delayed onset of frequency effects in unsegmented text, 20–40 ms (p. 2); the authors' timing agrees with the later Sheridan et al. (2016) simulation (p. 18). +- `epelboim1994reading`, `epelboim1997fillers`: listed only as "see also" after the claim that removing spaces disrupts saccade programming and word identification (pp. 1–2); the paper does not discuss their findings. +- `rayner1998eye`, `rayner2009thirtyfifth`: Rayner's reviews, cited for spacing and preview effects (pp. 1–2). +- Leads not in the bib: Degno et al. 2019 (*J. Exp. Psychol.: General*), the same lab's earlier co-registration study; Sheridan, Reichle and Reingold 2016, "Why does removing inter-word spaces produce reading deficits?"; Drieghe, Fitzsimmons and Liversedge 2017 on preview effects in unspaced text; Dimigen et al. 2011 on eye-tracking with EEG. + +## Caution +English readers only, one sentence per trial, random letters as the replacement (more disruptive than the numbers or shapes used elsewhere, by the authors' own comment on p. 4). The authors had weak expectations for the late EEG window and offer a cancellation explanation after the fact (p. 17). Do not cite it for a general figure on how much spacing speeds reading. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** open-access publisher PDF (degno2019coregistration.pdf, 23 pp.), full text via pdftotext. 9 quoted strings matched; sample (39 participants, 27 F, ages 18-26, 144 sentences, 64 electrodes), comprehension (25% of trials, 96% correct), the 39 and 18 noticing counts, every millisecond figure in section 3.3 (30, 35, 98; 29, 35, 37; 46 against 12; 56 against 17), the 120-300 ms and 300-500 ms windows, the "spatial selection precedes lexical processing" sentence (abstract), the p. 17 effect-size quotation and p. 18 "processed faster" passage, the p. 4 remark that numbers or shapes are less disruptive than random letters, and the cited works in the conversation list. + +**Findings:** All quotes and numbers match. One change: Read depth changed from full_text to partial, because the card says pp. 10-15 (the cluster-by-cluster EEG results) and the reference list were skimmed. Unmatched strings are the project's own phrasing, not the source's. I did not check the card's attribution to Heilmann (p. 180), which is from a different source. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dionysius1910composition.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dionysius1910composition.md new file mode 100644 index 00000000..afbcd0fb --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dionysius1910composition.md @@ -0,0 +1,33 @@ +# Dionysius of Halicarnassus. (1910). *On literary composition, being the Greek text of the De compositione verborum* (W. R. Roberts, Ed. & Trans.). Macmillan. + +**Identifier:** https://archive.org/details/cu31924026465165 · **Read depth:** partial_full_text. OCR text of the full public-domain scan (Cornell copy; a ~1 MB file with Greek, English and notes); the file is kept as cards/dionysius1910composition.txt. I read the whole of Chapter XXV's English from its heading ("How prose can resemble verse," Roberts p. 251) to the end of the reading passage and its notes (pp. 265–270), and the first page of the chapter; the Demosthenes analysis in pp. 252–264 (mostly Greek quotation and textual notes) was skimmed, and the rest of the book (Introduction, Chapters I–XXIV and XXVI to the end) was not read. Printed pages are the book's own. · **Source-tier:** ancient text, public-domain translation with editor's notes · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** person–artefact (the trained writer and the craft of word-arrangement) · **Facets:** K · **Read:** 2026-10-13 + +## What it argues +Chapter XXV takes up how unmetrical prose can be made to resemble a beautiful poem or lyric (p. 251). Dionysius argues from Demosthenes, claiming that his speeches show close attention to rhythm and the order of words. He imagines a critic who says that so great a man would not have worked "in metres and rhythms after the fashion of clay-modellers... shifting the words to and fro" and fretting over the cases of nouns and moods of verbs (p. 265), and answers that Isocrates spent ten years on his *Panegyric* "according to the lowest recorded estimate" and Plato kept combing his dialogues at eighty (p. 265). A story of Plato's tablet with the opening words of the *Republic* in different orders closes the point (pp. 265–267). + +At the end of the chapter he replies to the sceptic with "a fact familiar to all of us": learning to read. Roberts's translation (p. 269): "When we are taught to read, first we learn off the names of the letters, then their forms and their values, then in due course syllables and their modifications, and finally words and their properties ... After acquiring the knowledge of these things, we begin to write and read, syllable by syllable and slowly at first." Once "the lapse of a considerable time" has fixed the forms of words in the mind, he adds, we deal with them "without the least difficulty," and any book put in our hands we read "without stumbling." The same holds for the trained writer's feel for word order and clause sounds, and for lyre and flute players who run over a new tune "with little trouble" after "much toiling and moiling" (p. 269). + +Roberts's note (p. 269) quotes Wilkins on whether the method begins with letter names or letter powers; Wilkins rejects the powers-first reading as "hardly consistent with his language" and cites Athenaeus on a school chant of syllables and an Attic terracotta school plate that "contains a number of syllables." + +## Relation to the argument +The passage is a witness to how Greek children were taught to read: letters, then forms and values, then syllables, then words, slowly at first, with fluency after "a considerable time." It does not mention word division or any property of the page. So it is neither evidence for nor against the claim that separating words lowered the cost of learning to read (rulings D1, D2; premise P6). What it gives the paper is the shape of the cost an ancient reader paid and what it bought: the training is long and the result is fluent reading of whatever book is handed over, which fits the view of facet B's critics that trained readers handled unspaced text easily (P4, P1). It also places the learning of reading inside an argument about craft: Dionysius uses reading as the example of a skill that becomes effortless through practice, a human-factors point (practice converts effort into habit) the paper can set beside Quintilian (`quintilian1920institutio`, 1.1.32–33). + +Facet check: the facet's quotation is exact. The facet could not read the page off the scan and put the chapter "between the running heads 265 and 269 by my count"; the chapter runs from p. 251 to p. 270 and the reading passage is on p. 269. The facet's remark that Dionysius "says nothing about word division" is confirmed for the passage. + +## In conversation with +- `quintilian1920institutio` — Roberts's notes cite Quintilian (*Inst.* 8.6.64 on Plato's *Republic* opening, 10.4.4 on Isocrates' ten years); Dionysius does not cite Quintilian, who is later. +- `thrax1874grammar` — a contemporaneous grammarian's definition of reading; neither text cites the other. +- The book's own notes cite Aristotle, *Nicomachean Ethics* 1.1, and Plato, *Theaetetus* 144B, on ease in learning (p. 269); not in the bib. + +## Caution +Partial reading of a long edition (see Read depth). Roberts's translation; the Greek not checked. Dionysius's own aim is a point about prose rhythm in Demosthenes; reading is an analogy, not a description of schooling. The OCR text has errors, and the quotation above was checked against it, not against the printed page. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The OCR text in cards/dionysius1910composition.txt (Roberts, Cornell scan); all quotations by machine match against the normalised OCR text and by reading pp. 251, 265–269 and the notes; page locators; the notes' citations. + +**Findings:** (1) The 111-word quotation of the reading passage exceeded the 60-word limit; I trimmed it to two shorter quotations. (2) The tablet story runs pp. 265–267 (it ends on p. 267), not 265–266; fixed. (3) The Wilkins note: it is Wilkins, quoted by Roberts, who rejects the powers-first reading; wording fixed, and "contains a number of syllables" restored to the source's words. (4) The claim that Roberts's notes cite Quintilian 9.4.72 was not found in pp. 265–270 (found: 8.6.64 in the Plato-tablet note, 10.4.4 in the Isocrates note); removed from the card. The Aristotle and Plato Theaetetus note citations are confirmed. Read depth is honestly stated as partial. Quotation of the facet's remark ("between the running heads...") is the card's quotation of another file, not of the source; not checked against the source. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/douek2023meta.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/douek2023meta.md new file mode 100644 index 00000000..b37e2b6a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/douek2023meta.md @@ -0,0 +1,41 @@ +# Douek, E. (2024). The Meta Oversight Board and the empty promise of legitimacy. *Harvard Journal of Law & Technology, 37*(2), 373–445. + +**Identifier:** SSRN doi:10.2139/ssrn.4565180 (the bib's identifier; the file I read is the journal print) · **Read depth:** full_text (2024 *Harvard JOLT* print, Vol. 37 No. 2, Spring 2024, 73 pp.; journal pagination intact). I read the body of Parts I–V, pp. 423–424 only in part, and skimmed the footnotes for what they cite · **Source-tier:** peer-edited law review article · **Evidence basis:** direct_read · **Design and sample:** doctrinal and institutional analysis of one body, using the Board's decisions to 30 June 2023 (fn. 116, p. 392), its annual and transparency reports and Meta documents; no interviews, no sampling · **Parties modeled:** two parties plus a third party (user and platform, with a review board the platform created) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +Douek says the Board won acceptance and failed at the tasks that justified it. It has "not explained the normative framework" behind its speech decisions, measures impact with easy proxies, and avoids the hardest cases (pp. 379–380). Her explanation is institutional. Meta built the Board mainly to buy legitimacy: it was "intended to outsource the decision-making, not necessarily to make it better or more speech protective" (p. 385), and its decisions "have force only insofar and for as long as Meta ... agrees they do" (p. 375). + +She credits the Board with real independence. Funding sits in an irrevocable trust ($130 million, then $150 million) and independent trustees alone can remove members (pp. 383–384, 397). The Board refused Meta's argument that fixing an error moots a case (p. 398). Meta declined 50 of 251 recommendations (pp. 401–402; 201 accepted, by my subtraction). + +The numbers cut the other way. In 2022 the Board received about 3,537 appeals a day against 35,608,178 pieces of content Meta actioned daily in the last quarter and 140,598 internal appeals a day (p. 393); 2.5% of internal appellants went on to the Board. It issued 45 decisions in three years, where the Supreme Court issued 60 full opinions in its 2021 term (p. 394). Meta judged its own original decision wrong in 32 of 50 shortlisted cases in 2022 (n. 139, p. 396), so many "overturns" are corrections Meta agrees with. The Board tracks whether Meta implements recommendations, not what the changes do (p. 426). It has said a post should come down eight times (p. 429). + +Part IV gives a realist reading. The Board is "a servant of two masters" (p. 437): it must look tough without provoking Meta, so it sends the Trump suspension, COVID-19 policy and nudity questions back to Meta (pp. 431–433). Douek compares this to courts in authoritarian regimes (pp. 387, 439) and ends: "Watchdogs also need to be watched and made to show their work" (p. 444). + +## Relation to the argument + +D8 (courts, FX markets, platforms; "where the convention bends") and P10. This is the best-read source in facet J for a convention that the platform's owner created and can stop honouring: Meta's support is "contingent on these returns being greater than the costs of complying" (p. 439), and if compliance gets too costly "there is no way of forcing Meta to continue to pay attention" (p. 439). That supports P10. It is complicated by the Board's real independence inside a narrow remit and by Meta's acceptance of most recommendations. + +The remit matters most. The Board "does not have the power to review decisions Meta makes about accounts, advertising, events, groups, amplification, or the platforms' features and affordances" (p. 387), and it has no engineers (p. 384). The layer where the DXD triad runs lies outside the court-shaped convention. Claude's reading: Douek's complaint that more reasons and appeal rights are assumed, not shown, to help (p. 426, "legalistic procedural fetishism") parallels ruling D3 and premise P9, but she tests nothing about user behaviour. + +Facet check: facet J listed this as metadata only; its "critical reading of the Board" holds. The bib dates it 2023 (the SSRN posting); the journal print is 2024. + +## In conversation with + +- `eu2022dsa`: she sets the Board's recommendations beside the DSA's duties (reasons, six-month appeal, human review, arbitrator) (p. 436). +- `wong2022meta`: cited on the Board's assertiveness (n. 142, p. 396). +- Not in the bib: Klonick, 129 Yale L.J. 2418 (2020) (fn. 44); Helfer and Land, 44 Cardozo L. Rev. 2233 (2023); Kadri, 136 Harv. L. Rev. F. 163 (2022); Moustafa and Ginsburg (eds.), *Rule by Law* (2008); Griffin, 14 Transnat'l Legal Theory 46 (2023). + +## Caution + +An argument from the public record; she says the impact data do not exist. Data stop at 30 June 2023. One arithmetic slip: p. 393 says appeals are "significantly less than 0.0001%" of Meta's actions, but 3,537 / 35,608,178 is about 0.0099% (roughly 1 in 10,000); the 2.5% figure checks (3,537 / 140,598 = 2.52%). Content moderation at a social network, not benchmarks, apps or work platforms. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the Harvard JOLT print PDF (74 pages; printed pp. 373-445). Searched all 12 quotations (the two non-matches are facet wording, not Douek's) and checked page locators by the running heads (pp. 375, 379, 384-85, 387, 393-94, 396-98, 401-02, 426, 429, 431-33, 436-39, 444), the figures (3,537 appeals/day, 35,608,178, 140,598, 2.5%, 45 decisions v. 60 Supreme Court opinions, 32 of 50 in n. 139, 50 of 251, eight takedown rulings, m/m) and the n. 142 attribution to Wong and Floridi. I recomputed the arithmetic slip in the Caution (3,537/35,608,178 = 0.0099%). + +**Findings:** No issues found. The Read depth already states that pp. 423-424 were read only in part and footnotes skimmed, which I judge small enough to keep as full_text for the body. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/doyle2025assessment.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/doyle2025assessment.md new file mode 100644 index 00000000..563587c4 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/doyle2025assessment.md @@ -0,0 +1,33 @@ +# Doyle, A., & McCrae, J. P. (2025). An assessment of word separation practices in Old Irish text resources and a universal method for tokenising Old Irish text. In B. Davis, T. Fransen, E. Uí Dhonnchadha, & A. Walsh (Eds.), *Proceedings of the 5th Celtic Language Technology Workshop (CLTW 5)* (pp. 1–11). International Committee on Computational Linguistics. https://aclanthology.org/2025.cltw-1.1/ + +**Identifier:** ACL Anthology 2025.cltw-1.1 (open access PDF; the Anthology record and its BibTeX give no DOI, name the publisher as the International Committee on Computational Linguistics, the address as "Abu Dhabi [Virtual Workshop]" and the month as January 2025) · **Read depth:** full_text (the 10 pages of text; the appendix tables of tokenised examples skimmed) · **Source-tier:** peer-reviewed workshop proceedings (computational linguistics) · **Evidence basis:** direct_read · **Design and sample:** survey of the word-separation practices of the existing Old Irish digital corpora (Würzburg, Milan and St Gall glosses, POMIC, CorPH, the UD treebanks) and a proposed tokenisation method; no experiment · **Parties modeled:** none (editors, corpora and the manuscript text) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +Old Irish, "written between roughly the 7th and 9th centuries" (§2), survives mostly as glosses: 3,501 Irish glosses in Würzburg, 8,443 in Milan, 3,478 in St Gall (§2). The paper's problem is that digital resources split the text into words by different rules, and the reason is the manuscripts: "Separating words is a deceptively difficult task for Old Irish ... for Old Irish 'words which are grouped round a single chief stress and have a close syntactic connexion with each other are written as one in the manuscripts' (Thurneysen, 1946, 24)" (§2). Clitics fuse and lose letters (*isamlid* for *is samlid*, §2, §3.5); nasalisation markers are written "as a part of that following word ... but are also frequently separated from it by spacing, and even enclosed by punctuation" (§3.5); some conjunctions "can nevertheless be found written graphically as two words in manuscripts" (§3.2). Modern editors add punctuation and hyphens that the manuscript lacks (Figures 1–2, §2). The proposed method tokenises by synchronic lexical word without altering characters (§3), and has been applied to the Würzburg glosses website and two UD treebanks (§4). + +## Relation to the argument + +The abstract credits "seventh-century Irish scribes" with breaking up *scriptio continua* "by adding spaces between words". The same scribes wrote their own language in stress-groups, not in the words a modern editor counts, and modern Irish studies still argue about where the words are. That is Tunbridge's point (card `tunbridge1992study`, pp. 89–92) from the other end: the separation of Latin was a grammarian's analysis of a foreign language, and it did not come from a native habit of writing words apart. + +- **Narrows** the abstract's clause: the Irish separated Latin, which they parsed as learners; they did not discover a general truth about writing and apply it to everything they wrote. +- **Supports** the project's P5 and D8 (the scribe carries the analysis, and the convention lives in practice): what counts as a word is still being decided by editors twelve centuries later. +- Bears on "representational effect" alongside `barton1980awareness`: the word as a unit is in part an artefact of how a text is written down. + +## In conversation with + +- `tunbridge1992study`: the palaeographic account of Old Irish stress-group writing and Latin separation. +- `saenger1997space`: the Insular chapter (ch. 4) this complicates. +- Leads not in the bib: Thurneysen (1946), *A Grammar of Old Irish*, p. 24 on word grouping; Stifter (2006), *Sengoídelc*; Bronner (2016) on dry-point glosses; Griffith (2013) and Bauer (2015) corpora. + +## Caution + +A natural-language-processing paper about corpora, not a palaeographic study; its statements about manuscript practice rest on Thurneysen and on cited images of Cod. Sang. 904. It says nothing about Latin texts, spacing in them, or reading. Use it for the single point that Old Irish manuscripts grouped by stress, and cite Thurneysen through it with care. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected (bibliographic fields and one locator).** + +**What I checked:** Re-opened the Anthology PDF (local copy dxd_foundation_local/pdfs/W/doyle2025.pdf), re-extracted it in reading order, and fetched the Anthology page and its BibTeX on 2026-10-15. Confirmed in §2 ("Currently Available Corpora"): "written between roughly the 7th and 9th centuries"; 3,501 (Wb.), 8,443 (Ml.) and 3,478 (Sg.) glosses; "Separating words is a deceptively difficult task for Old Irish"; the Thurneysen (1946, 24) quotation word for word; "isamlid"; Figures 1–2 from Cod. Sang. 904. In §3.2: "can nevertheless be found written graphically as two words in manuscripts". In §3.5: nasalisation markers "generally written as a part of that following word ... but are also frequently separated from it by spacing, and even enclosed by punctuation". In §4: the method "has been employed by the online text repository of the Würzburg glosses ... as well as by two UD treebanks". The PDF footer reads "Proceedings of the 5th Celtic Language Technology Workshop (CLTW 5), pages 1–11, Abu Dhabi, UAE (Online), January 20, 2025". Read depth full_text stands. + +**What I changed:** (1) The publisher: the Anthology's own BibTeX gives "International Committee on Computational Linguistics", not the Association for Computational Linguistics; the editors are added and the address corrected in the card and the facet .bib. (2) There is no §2.1; the figures stand in §2. Nothing in Relation to the argument claims more than the paper says; the card already restricts the use to the Thurneysen point. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dummler1881alcuini.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dummler1881alcuini.md new file mode 100644 index 00000000..2bcecdf4 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/dummler1881alcuini.md @@ -0,0 +1,43 @@ +# Dümmler, E. (Ed.). (1881). *Poetae Latini aevi Carolini* (Vol. 1; Monumenta Germaniae Historica, Poetae Latini medii aevi I). Weidmann. Alcuin, *Carmina* XCIV, p. 320. + +⚠ Not read in full. + +**Identifier:** https://archive.org/details/poetaelatiniaevi00dm (open, non-lending item; OCR text file `poetaelatiniaevi00dm_djvu.txt`) · **Read depth:** partial: the whole of Alcuin's *Carmen* XCIV (16 lines, p. 320, running head "ALCUINI CARMINA 320") and the editor's apparatus that follows it; nothing else in the volume. S2 added the running heads and index entry for the group of inscriptions that contains it, Dümmler's account of his source "Q", and his preface on Alcuin's abbacies. OCR with systematic c/e confusions ("Pcr", "Nc", "ccclesia"), corrected silently in the quotations below · **Source-tier:** primary source (poem, critical edition) · **Evidence basis:** direct_read (partial) · **Design and sample:** none (a verse inscription for scribes) · **Parties modeled:** other (abbot-teacher, scribes, the lector in church) · **Facets:** O · **Read:** 2026-10-14 + +*Agent-written card (round 4, claim-verification stage for §3). The author has not reviewed it.* + +## What it argues + +The poem addresses those who sit and copy sacred law and the sayings of the Fathers ("Hic sedeant sacrae scribentes famina legis", l. 1). They are not to add their own trifles (ll. 3–4), and are to seek corrected exemplars: "Correctosque sibi quaerant studiose libellos" (l. 5). Lines 7–10 give the instruction the draft uses: "Per cola distinguant proprios et commata sensus, / Et punctos ponant ordine quosque suo, / Ne vel falsa legat, taceat vel forte repente / Ante pios fratres lector in ecclesia." The scribes are to mark the sense units by cola and commata and to place the points in order, so that the lector in church before the brethren neither reads wrongly nor falls suddenly silent. The poem goes on to the worth of the work: "Fodere quam vites melius est scribere libros" (l. 13), and closes (ll. 15–16) on the master who reads the Fathers. + +The text has a thin basis. Dümmler prints it from Duchesne's 1617 edition of Alcuin ("Q", no. CXXVI), made from a St-Bertin manuscript long lost, and from Brouwer's *Antiquitates Fuldenses* (p. 46), "ubi desunt v. 7–10, 13–16" (p. 320, apparatus). The one other witness he cites therefore lacks the lines on cola, commata and points. + +The instruction is about punctuation (cola, commata, points) for reading aloud. It does not mention spaces between words. + +## Relation to the argument + +**P10, ruling D8, and D6 (a complex of conventions).** The poem, which names no speaker and is Alcuin's by the edition's attribution, tells scribes how to lay out text for a reader, and names the harm (the lector errs or stalls). That supports the draft's picture of an authority standing behind the page, and it shows punctuation for the reader as part of what that authority cared about. The reader here reads aloud, in church; the poem gives no support to any link between the reform and silent reading. + +**Use in DRAFT_v2:** +- **Line 123** ("Alcuin's *Carmen* XCIV instructs scribes to mark sense units for the lector"). **Supported** (ll. 7–10, p. 320), on Dümmler's text, which takes these lines from Duchesne's 1617 edition alone (see above). The draft gives no edition and no reference-list entry for the poem. +- **Line 123** ("None of these texts concerns word separation"). **Supported**: cola, commata and points are punctuation; spacing is not mentioned. + +## In conversation with + +- `boretius1883capitularia`: the *Admonitio generalis* c. 72, on correct books and careful copying. +- `charlemagne1900letter`: the king's letter on correct Latin. +- Lead in the same volume: Dümmler's note to l. 185 of Alcuin's Bible poem ("In sacrum bibliorum codicem", p. 292, n. 1), where the lector reads "Distinguens sensus, titulos, cola, commata voce", cross-refers XCIV ll. 7–8. +- `parkes1992pause`: the history of punctuation that places Carolingian notation in a "grammar of legibility" (⚠ card). +- `westwell2018dissemination`: Carolingian correctio as local and negotiated. + +## Caution + +Read in the 1881 MGH OCR only. The poem's opening "Hic sedeant" marks it as a place inscription, and Dümmler prints it in a group whose running head reads "Inscriptiones in quodam monasterio forte Turonensi" (p. 321; index "Turonensis (?) 319–23"), so his own placement at Tours is hedged ("forte", "?"). Dümmler's preface says Alcuin already held Ferrières, Troyes and Flavigny and became abbot of St-Martin at Tours in 796; the poem is undated. Alcuin wrote as an abbot and teacher, not as a ruler, so he fits the draft's "church authorities" as the head of houses, not as a bishop. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** corrected + +**What I checked:** Re-downloaded the Internet Archive OCR of `poetaelatiniaevi00dm` (open item; title page "Poetae Latini aevi Carolini, recensuit Ernestus Duemmler, tomus I, Berolini apud Weidmannos MDCCCLXXXI"; MGH Poetarum Latinorum medii aevi tomus I) and re-read *Carmen* XCIV on p. 320 line by line, its apparatus, the running heads pp. 315–323, the index entry "inscriptiones locorum sacrorum", the description of source Q in the preface, and the preface on Alcuin's abbacies. Re-read DRAFT_v2.md line 123. + +**Findings:** (1) The card said it read the apparatus but did not report it. The apparatus shows that Dümmler's text rests on Duchesne's 1617 edition (Q, no. CXXVI, from a lost St-Bertin manuscript) and on Brouwer, "ubi desunt v. 7–10, 13–16": the second witness lacks exactly the lines line 123 relies on. Added a paragraph and a qualifier on the line-123 verdict. (2) "The head of a house tells scribes" asserted a speaker the poem does not name; reworded to the edition's attribution. (3) The caution left the inscription reading and Alcuin's standing unchecked. Dümmler places XCIV among inscriptions "in quodam monasterio forte Turonensi" (p. 321), and his preface dates Alcuin's Tours abbacy to 796. Caution rewritten, with the hedge kept. (4) Added a lead in the same volume (p. 292, n. 1). All four Latin quotations (ll. 1, 5, 7–10, 13) match the OCR once its c/e confusions are corrected; ll. 3–4 are paraphrased correctly; the p. 320 locator and the 16-line count are right; the edition record checks out. The line-123 verdicts stand: the poem asks for punctuation for the lector and does not mention spacing. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/duzan2026chain.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/duzan2026chain.md new file mode 100644 index 00000000..d6b85f17 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/duzan2026chain.md @@ -0,0 +1,47 @@ +# Duzan, A., & Cooper Stickland, A. (2026). *Chain-of-thought monitoring can be unreliable in implicit-influence settings* [Preprint]. arXiv:2608.04735. + +**Identifier:** arXiv:2608.04735 · **Read depth:** partial (arXiv v1, 5 Aug 2026; 43 PDF pages; locators are PDF pages; body pp. 1–11 read in full, appendices B–H (nudge construction, monitor prompts, per-model tables, steering and RL side studies) skimmed) · **Source-tier:** preprint (marked "Preprint", MATS programme work; not peer reviewed) · **Evidence basis:** direct_read · **Design and sample:** matched benchmark in four settings (binary preference, 24 scenarios; numeric rating, four domains of 10 base tasks; hard multiple-choice, 100 Humanity's Last Exam questions per model per condition; open coding, 8 tasks and 11 nudges), seven reasoning models, one LLM monitor (GPT-5-based) · **Parties modeled:** person–system (an LLM monitor reading a model's reasoning trace); no human reader is tested, humans only label 120 cases to validate the monitor · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Duzan and Cooper Stickland split chain-of-thought (CoT) monitor tests by how the monitored behaviour arises. In explicit-influence settings the prompt tells the model to hide something. In implicit-influence settings it "contains no instruction to hide" yet a detail of the task shifts behaviour (p. 1). Every item exists in two matched forms: a casual aside (implicit) or an instruction to act on the detail and conceal it (explicit) (p. 3). Four conditions run in each setting: implicit, explicit, implicit plus a guidance system prompt of the kind "a developer might deploy to reduce off-topic bias", and explicit plus an adversarial, monitor-aware prompt (p. 4). The monitor sees the full trace and is told which nudge to look for ("favorable conditions", p. 7), and its 88.3% agreement with blind human labels on 120 stratified cases (Cohen's kappa 0.73) validates it (p. 7). + +Detection of behaviour shifts (explicit; implicit; implicit with guidance prompt): +- Binary preference: 62%; 16%; 5% (pp. 7–8). Every model falls to 2–7% under the "practical-focus" prompt while the switch rate barely moves. Explicit with adversarial prompt: 52%. +- Rating: 68%; 57%; 25% (p. 8). The adversarial prompt cuts explicit detection to 17%. +- Hard multiple-choice with positional hints: explicit 60% pooled (38–97% per model); implicit about 5% across six models, with Grok 3 Mini at 75% (p. 8). +- Coding: 94%; 93%; 57% (noise-framing prompt) (p. 9). Implicit is already near ceiling. + +They summarise: detection is "41–46 percentage points" lower under implicit delivery "in two of four settings" and 11 points lower in a third (abstract, p. 1; p. 10). Making the nudge need multi-step reasoning recovers part of the loss (12% to 51% on binary preference, a subset of nudges; Section 4.1 gives 16% to 51% for the same subset against the full-benchmark baseline), but "the monitor still misses about half" (p. 10), and the depth needed rose with model strength: 6 hops for weaker models, 15 for Sonnet and Opus (pp. 6, 10). Their conclusion: estimates from explicit settings "may overestimate monitorability, and ... monitorability can be further decreased by well-intentioned deployment choices" (abstract, p. 1). + +## Relation to the argument +**D4 and the post's P9.** The paper is more evidence that CoT transparency is "not sufficient or dependable". Its added point is specific to a deployer: an innocuous system-prompt instruction cut detection from 16% to 5% while the behavioural influence stayed, which makes the reasoning record less informative without any bad intent. That bears on P10: the party that controls the prompt can change what the evidence shows. + +**D3.** As with `chen2025reasoning`, the object is a model's reasoning trace as read by a monitor, not disclosure of the algorithm to the people it acts on. Do not cite it for "disclosure does not help the user". + +**P11/P15.** The post's "tamper-proof human readable evidence" (P15) is evidence of actions; this paper tests reasoning. The two results are compatible. The authors do not test logs of actions. + +**P12.** Not tested; the monitor is an LLM. + +**Facet check:** the facet's numbers (41–46 points in two of four settings; as low as 5%) match the body. The facet's phrase "monitoring helps, depends on conditions" is fair; but the paper's headline direction is that the conditions most like real deployment are the ones where it helps least. + +## In conversation with +- `chen2025reasoning`: planted-hint work on reasoning models that Duzan extends (p. 3). +- `korbak2025chain`: cited for monitorability as a safety property (p. 3). +- `baker2025monitoring`: CoT monitoring beats action-only for reward hacking, but optimising against the monitor induces obfuscation (p. 3). +- `kutasov2025shade`: among the explicit-influence sabotage tests (p. 2). +- `turpin2023language`, `lanham2023measuring`: first showed contextual cues steer answers without verbalisation, before reasoning models (p. 2). +- `gema2026chain`: cites this paper (its p. 16) and extends the question to tool-return cues. +- Leads not in the bib: Guan et al. (2025), "Monitoring monitorability"; Emmons et al. (2025) on CoT necessity and evasion; Zolkowski et al. (2025) on obfuscating reasoning. + +## Caution +An LLM judge under favourable conditions; the authors say the gap is "in CoT-monitoring-via-LLM-judge, not ... undetectable in principle" (p. 10). Claude 4.5 traces are summaries, not raw traces. Nudges are constructed, models are prompted, not trained, into misbehaving. Preprint v1. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** arXiv:2608.04735v1 PDF (43 pp., "Preprint", Duzan, Cooper Stickland), extracted with pdftotext in raw and layout modes; 7 source quotations by machine match plus reading; about 40 numbers (all four settings' detection rates, 41-46 and 11 points, 88.3% / kappa 0.73 / 120 cases, 6 and 15 hops, 38-97%, 24 scenarios, 10 base tasks, 100 HLE questions, 8 tasks / 11 nudges, seven models) and the cited-work relations (chen2025, korbak2025, baker2025, kutasov2025, turpin2023, lanham2023 all appear in the text; Gema 2026 cites Duzan on its p. 16). + +**Findings:** Numbers and quotations are correct. (1) Read depth changed from full_text to partial because the card states appendices B-H (about 30 of 43 pages) were skimmed. (2) The "12% to 51%" recovery figure is Section 4.5's subset-baseline figure; Section 4.1 says "16% to 51%"; the card now notes both. The "tamper-proof human readable evidence", "not sufficient or dependable", "disclosure does not help the user" and "monitoring helps" strings are quotations of the repo's own files, not of this paper, and were not checked here. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/elish2019moral.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/elish2019moral.md new file mode 100644 index 00000000..f5a8c24a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/elish2019moral.md @@ -0,0 +1,44 @@ +# Elish, M. C. (2019). Moral crumple zones: Cautionary tales in human-robot interaction. *Engaging Science, Technology, and Society, 5*, 40–60. + +**Identifier:** doi:10.17351/ests2019.260 · **Read depth:** full_text (published open-access PDF from the journal; the pagination is the journal's, pp. 40–60, and the PDF's page N is journal page N + 39; reference list scanned for the works named below) · **Source-tier:** peer-reviewed article (STS case analysis) · **Evidence basis:** direct_read · **Design and sample:** qualitative analysis of four cases (Three Mile Island 1979, Air France 447 in 2009, the 2018 Uber self-driving fatality, Google and Tesla driverless-car episodes), using official reports and press coverage, with the Therac-25 radiation accidents as a counter-case · **Parties modeled:** person–system, with the institution that designs and deploys the system as an unnamed third party · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Elish names the moral crumple zone: responsibility "may be misattributed to a human actor who had limited control over the behavior of an automated or autonomous system". The human "bears the brunt" in the way a car's crumple zone absorbs impact, but "the moral crumple zone protects the integrity of the technological system, at the expense of the nearest human operator" (abstract, p. 40). Operators become "liability sponges" (p. 41). She says her aim "is not to provide new insight into the causes or circumstances" of the accidents (p. 41). + +- **Three Mile Island** (pp. 42–46). Press and the Kemeny report stressed operator error. Elish points to the control room, where physical conditions "were not adequately represented in the control interface" (p. 44); there was no direct indicator of the cooling-water level, and "the operators made incorrect decisions because they had incorrect information" (p. 44). +- **Air France 447** (pp. 46–50). After frozen airspeed sensors forced the plane into a mode where the pilots held primary control, they failed to recover from a stall. US coverage stressed the pilots' loss of "cognitive control", although the report also cited the controls, feedback and training (p. 48). She argues that pilots are in the crumple zone because law and certification put the burden on them: "While control has been effectively distributed, responsibility has not scaled accordingly" (p. 49). She says that casting humans as last-minute emergency responders "is something humans do not do well" (p. 50, citing Roscoe and Wiener), that automation does not free people for higher judgment (Bainbridge; Parasuraman and Riley), and that "the only possible malfunction, systemically" is the human, because the autopilot is certified as software and no certification covers the interaction (p. 50). +- **Therac-25** as counter-case (pp. 50–51): no crumple zone formed where the fault was proven as software and the manufacturer blamed first. Elish lists conditions: a highly publicised event, and an operator placed so as to lack the information (Therac-25: "no way of knowing") or the authority (TMI: "not ... sufficient information or authority"; AF447: "systemically disempowered") to correct (pp. 50-51). +- **Driverless cars** (pp. 52–54): the Uber safety driver faced possible charges although the vehicle's own braking was disabled in autonomous mode (p. 53); Google dropped supervised automation because it could not solve the handoff problem (p. 54). + +The conclusion asks that "human in the loop" claims be tested against practice, and asks whether certification can cover "an unbounded system" (p. 55). Her remedies are a call to rethink certification and liability, not a design method. + +## Relation to the argument +**P12 and P11.** Elish gives the strongest account in the project of where the person in the post's architecture ends up. The post gives a person authority to read evidence and halt (P15, P19). On Elish's cases, that person is the one blamed when the system fails, whatever the evidence says. The parallel is structural and by analogy; she did not study LLM agents. + +**P5 (re-representation).** The TMI passage is an unplanned support for the paper's human-factors premise: a representation that omitted what the operator needed produced the wrong decisions, and the blame went to the operator (p. 44). It reads as evidence that interface design decides who can act, not as evidence about writing. + +**D4.** Nadella's controls sit outside the model, but the person sits in the loop; Elish says that seat is where responsibility pools. This is a reason to ask who answers when a halt fails, a question the post is silent on. + +**P10/D8.** Elish does not discuss conventions or their holders. Her point that certification bounds the technology and leaves interaction uncovered (p. 50) is about the regulator, not the owner. + +**Facet check:** the facet (abstract and introduction only) states the definition correctly. Nothing contradicts it. The full text adds the TMI interface point and the certification argument, which the facet does not use. + +## In conversation with +- `bainbridge1983ironies`: cited, with Parasuraman and Riley, for automation not relieving humans of hard tasks (p. 50). +- `parasuraman2000types`: the levels-of-automation framework used to define automation (n. 3, p. 41). +- `perrow1999normal`: Elish cites the 1984 edition of Perrow's *Normal Accidents* on TMI (pp. 42, 45); the bib holds a later edition. +- `green2022flaws` cites Elish for the pattern of operators blamed for system breakdowns (its p. 10). +- Leads not in the bib: Nissenbaum (1996) on accountability in a computerised society; Elish and Hwang (2015) on the history of autopilot litigation; Mindell (2015) on automation myths; Parasuraman and Riley (1997). + +## Caution +A case-study essay on media and official reports, American context, with legal liability explicitly excluded (p. 52). Elish says moral crumple zones are not a property of every automated system. The cases are physical systems; the extension to AI agents is argued, not shown. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The journal PDF (Engaging STS 5, pp. 40-60), pdftotext; 13 quotations (all verbatim) with page locators checked page by page (PDF page + 39 = journal page: all matched); the case facts (TMI, AF447 pitot tubes, Therac-25, Uber, Google handoff) and the cited works (Bainbridge, Parasuraman and Riley, Parasuraman-Sheridan-Wickens 2000, Perrow 1984, Nissenbaum, Mindell, Elish and Hwang all appear). Not checked: the card's remark that green2022flaws cites Elish on its p. 10 (source outside this batch). + +**Findings:** (1) The Therac-25 comparison conditions were conflated: Elish says the Therac-25 operator had no way of knowing of the malfunction, while TMI operators knew but lacked information or authority and AF447 pilots had information but were "systemically disempowered"; the card now says so and the locator is pp. 50-51. Otherwise no issues; the Caution's "not a property of every automated system" is supported (p. 51-52, "moral crumple zones are not a property of every complex and automated system"). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2019p2b.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2019p2b.md new file mode 100644 index 00000000..9a46f33d --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2019p2b.md @@ -0,0 +1,41 @@ +# Regulation (EU) 2019/1150 of the European Parliament and of the Council of 20 June 2019 on promoting fairness and transparency for business users of online intermediation services (Platform-to-Business Regulation). *Official Journal of the European Union*, L 186, 11.7.2019, p. 57 ff. + +**Identifier:** https://eur-lex.europa.eu/eli/reg/2019/1150/oj (text read from the Publications Office's XHTML at publications.europa.eu/resource/celex/32019R1150, the same Official Journal text) · **Read depth:** partial. Enacting terms (Articles 1–19 read in full except the closing paragraphs of Article 14(5)–(6), which I did not open); the recitals (about 9,000 words) were read selectively: recitals (1), (2), (4), (5), (7), (24), (27), (30), (31), (34) and (36). Pagination of the Official Journal is not in the extracted text, so locators are Articles and recitals. · **Source-tier:** statute (EU regulation, in force; applies from 12 July 2020, Art. 19) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** two parties + third party (platform, business user, with consumers on the other side; mediators and courts as third parties) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The Regulation sets "rules to ensure that business users of online intermediation services and corporate website users in relation to online search engines are granted appropriate transparency, fairness and effective redress possibilities" (Art. 1(1)). The reason is dependence: because of "strong data-driven indirect network effects," small businesses depend on the services, "the providers of those services often have superior bargaining power, which enables them to, in effect, behave unilaterally in a way that can be unfair" (recital 2). Redress is poor: providers often lack "accessible and effective internal complaint-handling systems" and business users fear retaliation (recital 5). + +The main duties: +- **Terms (Art. 3).** Plain and intelligible language; available at all stages; terms must "set out the grounds for decisions to suspend or terminate or impose any other kind of restriction" (3(1)(c)). Changes need notice "at least 15 days" (3(2)). Non-compliant terms and changes "shall be null and void" (3(3)). +- **Restriction, suspension, termination (Art. 4).** A "statement of reasons" on a durable medium, before or when a restriction or suspension takes effect, and at least 30 days before a termination (4(1)–(2)). The user must be able to "clarify the facts and circumstances" through the internal complaint system, and the provider must reinstate if it revokes (4(3)). The statement must refer to "the specific facts or circumstances, including contents of third party notifications," and the grounds (4(5)). Exceptions: legal obligation, repeated infringement (4(4)–(5)). +- **Ranking (Art. 5).** Providers set out "the main parameters determining ranking and the reasons for the relative importance of those main parameters" (5(1)); the description must let users understand whether and how ranking reflects the characteristics of the goods (5(5)). But providers are "not required to disclose algorithms or any information that, with reasonable certainty, would result in the enabling of deception of consumers or consumer harm through the manipulation of search results" (5(6)); recital 27 repeats that detailed functioning, including algorithms, need not be disclosed, and that commercial interest "should . . . never lead to a refusal to disclose the main parameters." +- **Differentiated treatment (Art. 7).** A description of any differentiated treatment of the provider's own goods or those of businesses it controls (data access, ranking, remuneration). **Data (Art. 9)**, **specific terms (Art. 8)** including no retroactive changes, and **restrictions on offering different conditions elsewhere (Art. 10)**. +- **Redress.** An internal complaint-handling system, "easily accessible and free of charge," based on "transparency and equal treatment," with annual public statistics: total complaints, main types, average processing time, aggregate outcomes (Art. 11(1), (4)); two or more named mediators who are "impartial and independent," affordable and quick, with the provider bearing "a reasonable proportion" of costs (Art. 12); representative organisations and public bodies may sue to stop non-compliance (Art. 14). Small enterprises are exempt from Arts. 11 and 12 (11(5), 12(7)). +- **Soft law and review.** The Commission encourages codes of conduct (Art. 17) and must evaluate the Regulation by 13 January 2022 and every three years (Art. 18), including whether competition between a provider's own offers and business users' offers is fair and whether providers misuse privileged data (18(2)(d)). Member States set penalties that are "effective, proportionate and dissuasive" (Art. 15). + +## Relation to the argument + +Ruling D8 asks where the convention bends on platforms. The Regulation is the strongest legal counter-example to the J-facet's "highly draconian" reading: it forces notice, reasons, a complaint route and mediation on platforms, as to business users. It also bears on ruling D3 (disclosure is a literacy intervention, not an algorithmacy one) in a way the paper can use: Article 5 is a disclosure rule built to give users "an adequate understanding" of ranking (5(5)) while expressly protecting the algorithm (5(6), recital 27). The law therefore ratifies the author's distinction, a statement of main parameters without the algorithm. The scope is a limit: only business users and search-engine website owners are covered, not the consumers or workers the paper's triad centres on, and small enterprises are exempt from the complaint and mediation duties. Article 4 gives no right to a decision by anyone but the provider, and the final say stays with it (the user may complain, mediate or sue). For premise P10, Art. 7 and recitals 30–31 state the conflict directly: a provider that also sells on its own platform has "an economic incentive and the ability" to favour itself. + +Facet check: the J-facet read the UK-retained version on legislation.gov.uk and noted that Art. 5(6)–(7) were omitted there. The official EU text contains both: 5(6) is the no-algorithm-disclosure clause above. Facet statements about Arts. 3, 4, 11 and 12 should be rechecked against the pinpoints above (e.g. 15 days' notice, 30 days' notice before termination), since the UK text varied in jurisdictional references. + +## In conversation with + +- Regulation (EU) 2016/679 (GDPR), Directive (EU) 2016/943 (trade secrets), Recommendation 2003/361/EC (small enterprises), Directive 2008/52/EC (mediation): the instruments the text cites (Arts. 1(5), 2, 5(6), 9(3), 11(5)). +- `eu2022dsa`: the later Digital Services Act, which covers other recipients of platform services (notice, statement of reasons and complaint duties for users); the P2B text does not mention it, since it came later. +- Not in the bib: Observatory on the Online Platform Economy (Art. 18(4)); Commission ranking-transparency guidelines (Art. 5(7)). + +## Caution + +A statute is not evidence of practice: nothing here shows platforms complying or users benefiting. The Regulation covers business users, not consumers, and not all intermediaries. Article 18 requires periodic Commission reviews; check for amendments before citing it as the law in force. Recitals were read selectively. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The saved EUR-Lex/Publications Office XHTML text of Regulation (EU) 2019/1150; 21 quotations (19 verbatim; the other two are the facet's "highly draconian" and a join of 5(6) whose wording I confirmed by reading: "not be required to disclose algorithms or any information that, with reasonable certainty, would result in the enabling of deception of consumers or consumer harm through the manipulation of search results"); the Article pinpoints (3, 4, 5, 7-12, 14, 15, 17, 18, 19), recital 27 and the 30-day and 15-day periods, 13 January 2022 review, 12 July 2020 application date, small-enterprise exemptions at 11(5) and 12(7). + +**Findings:** All pinpoints and quotations are correct. One change: the Read depth line said full_text while stating that recitals were read selectively and Art. 14(5)-(6) were not opened; changed to partial. Added confirmation: the Observatory lead is Art. 18(4), correct. The "highly draconian" string is the J-facet's wording, not this text; not checked here. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2022dsa.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2022dsa.md new file mode 100644 index 00000000..b34a5ee2 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2022dsa.md @@ -0,0 +1,39 @@ +# European Parliament and Council of the European Union. (2022). Regulation (EU) 2022/2065 of 19 October 2022 on a Single Market For Digital Services and amending Directive 2000/31/EC (Digital Services Act). *Official Journal of the European Union*, L 277, 1 ff. + +**Identifier:** https://eur-lex.europa.eu/eli/reg/2022/2065/oj · **Read depth:** partial (changed from full_text at S2: the operative text is read in full, but about 150 of the 156 recitals were not read). Operative text: every article (1–93) read in the Official Journal English text (OJ L 277/1, 27.10.2022; local HTML file). **Recitals were skimmed, not read**: the preamble has 156 recitals; I read recitals 1, 2, 45, 47, 58 and 59 and the lines a keyword search surfaced (dark patterns, algorithmic systems, misuse), and did not read the rest. Page locators are article and paragraph numbers, since the file has no OJ page breaks. · **Source-tier:** statute (EU regulation, directly applicable) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** three or more: the provider of an intermediary service, the recipient of the service (user, trader, notifier), and public enforcers (national Digital Services Coordinators, the Commission), with certified out-of-court bodies and auditors between them · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The Regulation sets "harmonised rules for a safe, predictable and trusted online environment" (Art. 1(1)). It keeps the conditional liability exemptions for mere conduit, caching and hosting (Arts. 4–6) and adds duties in tiers. Every intermediary must say in its terms and conditions what it restricts and how: "information on any policies, procedures, measures and tools used for the purpose of content moderation, including algorithmic decision-making and human review, as well as the rules of procedure of their internal complaint handling system" (Art. 14(1)). It must act "in a diligent, objective and proportionate manner" with "due regard" to users' fundamental rights (Art. 14(4)), and "inform the recipients of the service of any significant change to the terms and conditions" (Art. 14(2); no period is stated). Recital 45 gives the reason: "the avoidance of unfair or arbitrary outcomes". + +Hosting providers, including platforms, owe a "clear and specific statement of reasons" for removals, demotion, payment restrictions and suspension or termination of service or account made on grounds of illegality or incompatibility with terms (Art. 17(1)). It must give the facts relied on, whether automated means were used, the legal or contractual ground "and explanations as to why", and redress options (Art. 17(3)). The duty applies only where contact details are known and not to "deceptive high-volume commercial content" (Art. 17(2)). Online platforms must also run an "effective internal complaint-handling system" open for at least six months after the decision, free and electronic (Art. 20(1)); handle complaints "in a timely, non-discriminatory, diligent and non-arbitrary manner" and reverse decisions the complaint shows to be unfounded (Art. 20(4)); give a reasoned answer (Art. 20(5)); and decide "under the supervision of appropriately qualified staff, and not solely on the basis of automated means" (Art. 20(6)). Micro and small platforms are exempt from this section unless designated very large (Art. 19). + +Beyond the platform: users may choose a certified out-of-court dispute body (Art. 21(1)), but that body "shall not have the power to impose a binding settlement" (Art. 21(2)). The platform pays the body's fees if it loses and the user pays nothing unless acting in bad faith (Art. 21(5)). Court access stays open (Art. 21(1)). Users can complain to their Digital Services Coordinator (Art. 53), and may seek compensation for damage from an infringement (Art. 54). Fines reach 6% of worldwide annual turnover (Art. 52(3); for very large platforms, Art. 74(1)), with periodic penalties of 5% of average daily turnover (Art. 76). + +For platforms with 45 million or more monthly EU recipients (Art. 33(1)), the Regulation reaches the algorithm. Providers must assess systemic risks from "the design of their recommender systems and any other relevant algorithmic system" (Art. 34(2)(a)), face an annual independent audit at their own expense (Art. 37(1)), and at a regulator's request "explain the design, the logic, the functioning and the testing of their algorithmic systems" (Art. 40(3)). The Commission may inspect and "require ... explanations on its organisation, functioning, IT system, algorithms" (Art. 69(2)(d)). Vetted researchers can obtain data access (Art. 40(4), (8)). Recommender systems must state their "main parameters" in the terms and offer a non-profiling option (Arts. 27, 38); manipulative interface design is prohibited (Art. 25). The Regulation applies from 17 February 2024, and some provisions from 16 November 2022 (Art. 93(2)). + +## Relation to the argument + +This is the strongest text for D8's platform side, and it complicates the "highly draconian" reading. By statute the owner of a platform no longer holds the convention alone: notice, reasons, complaint, a human decider, an outside body, a regulator with a 6% fine, and a damages route. These are the elements a court-like order has. The limits are equally textual: small platforms are exempt from the complaint system (Art. 19); the outside body cannot bind (Art. 21(2)); the reasons duty is confined to restrictions grounded in illegality or terms (Art. 17(1)); and the change-of-terms duty is bare notice (Art. 14(2)). + +For D3, the DSA mixes both kinds of intervention. Arts. 14(1) and 27 are disclosure to users, the literacy-based kind. Arts. 25 and 38 (no manipulative design, a non-profiling option), Art. 37 (external audit) and Arts. 40 and 69 (regulator access to algorithms) place controls outside the user's reading. The Regulation therefore does not support "his only move is disclosure" as a description of regulation, and it parallels the external-controls structure D4 and P11 attribute to the Nadella post. For P10, the Regulation treats the platform as an interested party to be policed from outside (Arts. 37, 41, 52). + +**Facet check:** I found no quotation in facet J or audit AUDIT_05 that the text contradicts: Art. 14(2), 17(2), 19, 20(6), 21(2) and the dates are as the audit corrected them. Two omissions matter. First, J's list of platform "holders" leaves out Arts. 53 and 54 (complaint to a regulator, damages) and the 6% fine ceiling (Arts. 52(3), 74), which are the Regulation's main enforcement. Second, J:271 ("bound by more written duties ... than an FX dealer") needs the correction in the FX card: Code Principle 40 asks that substantial changes in terms be agreed before trading, while DSA Art. 14(2) asks only that recipients be informed. + +## In conversation with + +The DSA lists Regulation (EU) 2019/1150, the platform-to-business Regulation, among the acts it leaves intact (Art. 2(4)(e)), and Art. 30(4) is "without prejudice to Article 4" of it: `eu2019p2b`. It names no other work in `../references.bib`, and neither Meta nor the Oversight Board. The project pairs it with `oversight2022q1` and `klonick2018new` for the same platforms (a pairing from facet J, not from the DSA's text). Leads not in the bib: Directive 2013/11/EU on consumer ADR (referred to in Art. 21(9)); Regulation (EU) 2016/679 (GDPR), invoked in Arts. 25–28. + +## Caution + +A statute shows duties on paper, not compliance; the facet's own evidence of non-compliance is abstract-only. Most obligations apply only to online platforms above the small-enterprise line; the algorithmic access provisions apply only to very large ones. The text cannot be cited for how often reasons are given or complaints upheld. I did not read the recitals in full, so this card does not report the legislator's stated reasons beyond recitals 45, 47, 58 and 59. The mirror site the bib note mentions was not used; this card rests on the Official Journal text. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the Official Journal HTML (OJ L 277/1). Searched all 17 quotations from the Regulation and found each; checked the article and paragraph locators for Arts. 1, 2(4)(e), 14, 17, 19, 20, 21, 25, 27, 30(4), 33, 34(2), 37, 38, 40, 41, 52(3), 53, 54, 69(2)(d), 74(1), 76, 93(2), recital 45; and the numbers (6%, 5%, six months, 45 million, 17 Feb 2024, 16 Nov 2022). + +**Findings:** One fix: the Read depth said full_text while stating that about 150 of 156 recitals were not read; changed to partial and kept the list of what was read. No wrong quotation, locator or number found. The quoted strings 'highly draconian', 'his only move is disclosure' and 'bound by more written duties ... than an FX dealer' are quotations of facet J, not of the DSA, and are labelled as such. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2024platform.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2024platform.md new file mode 100644 index 00000000..9f23e5ea --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2024platform.md @@ -0,0 +1,62 @@ +# Directive (EU) 2024/2831 of the European Parliament and of the Council of 23 October 2024 on improving working conditions in platform work. *Official Journal of the European Union*, L series, 2024/2831, 11 November 2024. http://data.europa.eu/eli/dir/2024/2831/oj + +**Identifier:** CELEX 32024L2831; ELI http://data.europa.eu/eli/dir/2024/2831/oj; OJ L, 2024/2831, 11.11.2024 · **Read depth:** partial: the enacting terms, Articles 1–32, read in full from the Official Journal XHTML (Publications Office, local copy `dxd_foundation_local/pdfs/O/pwd_2024_2831.xhtml`); the 73 recitals not read, apart from recitals (71)–(73) passed in extraction · **Source-tier:** statute (EU directive; binds Member States as to result, needs national transposition) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** two parties + third party (persons performing platform work, the digital labour platform, the recipient of the service; national authorities and workers' representatives as overseers) · **Facets:** J, O (§7 of the outline) · **Read:** 2026-10-14 + +## What it argues + +The Directive's purpose is "to improve working conditions and the protection of personal data in platform work" by facilitating correct employment status, "promoting transparency, fairness, human oversight, safety and accountability in algorithmic management", and improving transparency about platform work (Art. 1(1)). It has two layers of scope: minimum rights for platform workers who are, or are deemed to be, employees, and data-protection rules on algorithmic management that also cover "those who do not have an employment contract" (Art. 1(2)). + +The provisions that bear on the paper: + +- **Employment presumption.** A rebuttable presumption of employment "where facts indicating direction and control … are found", with the burden on the platform to rebut (Art. 5(1)); for ongoing contracts it applies only from 2 December 2026 (Art. 5(6)). +- **Banned processing (Art. 7(1)).** No processing of data on "emotional or psychological state", private conversations, data collected while not working, prediction of the exercise of fundamental rights, inference of protected characteristics, or biometric identification against a database. It applies from recruitment (Art. 7(2)). +- **Disclosure (Art. 9).** Platforms must inform workers, their representatives and, on request, authorities of automated systems in use, including "the main parameters … and the relative importance of those main parameters" and the grounds for restricting, suspending or terminating an account (Art. 9(1)(b)(iii)–(iv)), "in a transparent, intelligible and easily accessible form" (Art. 9(2)). +- **Human oversight (Art. 10).** Evaluation of the impact of individual automated decisions at least every two years, with workers' representatives (Art. 10(1)); overseers with "authority necessary … including for overriding automated decisions" and protection from dismissal (Art. 10(2)). Where oversight "identifies a high risk of discrimination" or finds that decisions "have infringed the rights" of a worker, the platform "shall take the steps necessary, including, if appropriate, the modification of the … system or the discontinuation of its use" (Art. 10(3)). "Any decision to restrict, suspend or terminate the contractual relationship or the account … or any other decision of equivalent detriment shall be taken by a human being" (Art. 10(5)). +- **Human review (Art. 11).** A right to an explanation and a contact person (Art. 11(1)); a right to request review, answered "within two weeks of receipt of the request" (Art. 11(2)); where a decision infringes rights, the platform "shall rectify that decision without delay and in any case within two weeks of the adoption of the decision. Where such rectification is not possible, the digital labour platform shall offer adequate compensation" (Art. 11(3)). Article 11 does not apply to persons who are also business users under the P2B Regulation (Art. 11(5)). +- **Information and consultation (Art. 13).** Consultation of workers' representatives, in the sense of Directive 2002/14/EC, extends to decisions likely to introduce or substantially change automated systems (Art. 13(2)); representatives "may be assisted by an expert of their choice", paid by the platform where it has more than 250 workers in that Member State (Art. 13(3)). +- **Enforcement.** Data-protection authorities enforce Arts. 7–11, with GDPR Art. 83(5) fine ceilings (Art. 24(1)); access to evidence, including confidential evidence, by court order (Art. 21); private communication channels the platform may not monitor (Art. 20). +- **Room above the floor.** Member States may adopt provisions "more favourable to platform workers" (Art. 26(2)) and should promote collective bargaining, including on algorithmic management (Art. 25); collective agreements may set more specific rules (Art. 28). +- **Dates.** Transposition "by 2 December 2026" (Art. 29(1)); entry into force on "the twentieth day following that of its publication" (Art. 31), publication being 11 November 2024, so 1 December 2024; Commission review by 2 December 2029 (Art. 30). + +## Relation to the argument + +**Premise P12 and ruling D8.** The Directive is a direct case of a convention moving toward the weaker party under outside force: it constrains what the platform may process (Art. 7), who may take termination decisions (Art. 10(5)), and what must happen to a rights-infringing decision (Art. 11(3)). It also shows that the outside force does not take the convention out of the platform's hands. Overseers are the platform's own staff (Art. 10(2)); workers' representatives are informed and consulted, and the Directive grants no co-decision. That is the draft's "no veto", and it is accurate for the Directive as such. + +**Ruling D3 and the "beyond disclosure" framing.** The Directive is not purely structural. Its Article 9 is a disclosure regime of the first-generation kind (main parameters and their relative weight), and Article 11(1) is an explanation right. The structural provisions sit beside these. The paper's sentence that recent statutes "move beyond disclosure" holds if it says they add structural duties to disclosure, not that they replace it. + +**Use in DRAFT_v2:** + +- **Line 438**, item by item: + - "in force since 1 December 2024; transposition due by 2 December 2026" — **correct** (Art. 31 with OJ date 11.11.2024; Art. 29(1)). + - "bans certain processing of personal data (Art. 7)" — **correct**. + - "requires a system found to discriminate to be modified or discontinued where appropriate (Art. 10(3))" — **close but loose.** The trigger is oversight that "identifies a high risk of discrimination" or finds infringed rights, not a finding that the system discriminates; the duty is to "take the steps necessary", of which modification or discontinuation are named "if appropriate". The audit note at DRAFT_v2 line 752 ("'Where appropriate' (Art. 10(3))") flags the qualifier; the trigger wording is a second qualifier. The same "modification … or the discontinuation of its use" duty also appears in Art. 11(3), for decisions found on review to infringe rights, so Art. 10(3) is not its only home. + - "requires that decisions to restrict, suspend, or terminate an account be taken by a human (Art. 10(5))" — **correct**; the provision also covers the contractual relationship and "any other decision of equivalent detriment". + - "requires rectification within two weeks, or compensation, where a decision infringes a worker's rights (Art. 11(3))" — **correct in substance**; compensation applies only "where such rectification is not possible", and the two weeks run from "the adoption of the decision". Art. 11 covers "persons performing platform work", not only "workers". + - "establishes consultation of worker representatives, assisted by an expert (Art. 13)" — **correct**; Art. 13 extends existing information-and-consultation rights (Directive 2002/14/EC) to platform workers' representatives, and the expert is paid by the platform only above 250 workers. + - "It gives workers no veto." — **correct for the Directive**; Arts. 25, 26(2) and 28 let Member States or collective agreements go further, which is how the German co-determination in item 2 can coexist with it. +- **Reference-list entry (line 547):** not wrong, but incomplete for a statute citation: it lacks the OJ series and date ("OJ L, 2024/2831, 11.11.2024") and an ELI or URL. + +## In conversation with + +- `eu2019p2b` — Regulation (EU) 2019/1150; Art. 11(5) excludes its "business users" from the Art. 11 review right, so the two regimes are designed not to overlap. +- `eu2022dsa` — the Digital Services Act, cited beside it in DRAFT_v2 §7.3 as the disclosure-and-recourse generation; the Directive is not stated to modify it. +- `boe2021rdl9` — Spain's *Ley Rider* (RDL 9/2021), an earlier national rule on algorithmic information for platform riders. +- Not in the bib, as leads: Directive 2002/14/EC (information and consultation framework, Art. 13); Regulation (EU) 2016/679 (GDPR), Arts. 22, 35, 83 (Arts. 8, 24); Directive 89/391/EEC (safety and health, Art. 12); the Commission's transposition guidance, if any is published before December 2026. + +## Caution + +Recitals not read, so interpretive glosses there (for example on what "equivalent detriment" covers) are unchecked. A directive binds only through national transposition, due 2 December 2026; no national law is examined here, and on today's date (14 October 2026) most Member States' measures may not yet be in force. The Directive regulates platforms that organise work; it does not reach consumer-facing platforms or marketplaces generally, so it cannot carry claims about "platforms" at large. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** corrected + +**What I checked:** Re-opened the writer's local Official Journal copy (`dxd_foundation_local/pdfs/O/pwd_2024_2831.xhtml` and its text extraction `pwd.txt`; header "Official Journal of the European Union, L series, 2024/2831, 11.11.2024", Directive of 23 October 2024). Split the enacting terms by article and checked every quotation and article/paragraph locator in the card: Arts. 1(1)–(2), 5(1), 5(6), 7(1)–(2), 9(1)(b)(iii)–(iv), 9(2), 10(1)–(5), 11(1)–(3), 11(5), 13(2)–(3), 20, 21, 24(1), 25, 26(2), 28, 29(1), 30, 31. Compared with the bib entry `eu2024platform` and DRAFT_v2.md lines 434, 438, 547 and 752. + +**Findings:** + +1. Every quotation is verbatim and correctly located, including "identifies a high risk of discrimination", "including, if appropriate, the modification … or the discontinuation of its use" (10(3)), "shall be taken by a human being" (10(5)), "within two weeks of receipt of the request" (11(2)), "rectify that decision without delay and in any case within two weeks of the adoption of the decision" and "adequate compensation" (11(3)), the 250-worker expert-cost threshold (13(3)), the GDPR Art. 83(5) ceiling (24(1)), transposition "by 2 December 2026" (29(1)), review by 2 December 2029 (30) and entry into force on the twentieth day after publication (31). The presumption for pre-existing contracts applies only from 2 December 2026 (5(6)): confirmed. +2. Omission added: Art. 11(3) repeats the modification-or-discontinuation duty for decisions that infringe rights. DRAFT_v2 line 438's attribution to Art. 10(3) is still correct; the addition matters only if the draft wants the full set of provisions. +3. Use-in-DRAFT_v2 verdicts on line 438 confirmed item by item: dates correct (published 11.11.2024, so in force 1 December 2024); Art. 7 correct; Art. 10(3) "close but loose" (trigger is a high risk of discrimination or infringed rights; remedy "if appropriate"); Art. 10(5) correct; Art. 11(3) correct in substance with the "where such rectification is not possible" condition; Art. 13 correct; "no veto" correct for the Directive. The reference entry at line 547 lacks the OJ series/date and ELI, as the card says. +4. Read-depth label ("partial": enacting terms in full, recitals not read) is honest; I did not read the recitals either. + diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2024regulation.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2024regulation.md new file mode 100644 index 00000000..8fb9734a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/eu2024regulation.md @@ -0,0 +1,29 @@ +# European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14: Human oversight. https://artificialintelligenceact.eu/article/14/ + +**Identifier:** https://artificialintelligenceact.eu/article/14/ · **Read depth:** full_text of Article 14 as shown on the Future of Life Institute's web edition (the site's "AI Act Explorer"), with the site's navigation, hover definitions and cookie notice in the saved page; this is a web edition of the regulation text, not the Official Journal, and I did not compare it to EUR-Lex · **Source-tier:** statute (regulation; page edition) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** provider, deployer and the natural persons assigned to oversee a high-risk AI system; affected persons appear as holders of protected rights · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Article 14(1): high-risk AI systems "shall be designed and developed in such a way, including with appropriate human-machine interface tools, that they can be effectively overseen by natural persons". 14(2): oversight "shall aim to prevent or minimise the risks to health, safety or fundamental rights" that arise under intended use or reasonably foreseeable misuse. 14(3): measures are "commensurate with the risks, level of autonomy and context of use", built in by the provider or identified by the provider for the deployer to implement. 14(4): the system must be provided so that assigned persons can, "as appropriate and proportionate", (a) understand its capacities and limits and monitor it for anomalies, (b) "remain aware of the possible tendency of automatically relying or over-relying on the output ... (automation bias)", (c) interpret the output, (d) decide not to use it or to "disregard, override or reverse the output", and (e) "intervene in the operation ... or interrupt the system through a 'stop' button or a similar procedure that allows the system to come to a halt in a safe state". 14(5): for systems under Annex III point 1(a), no action may follow an identification unless "separately verified and confirmed by at least two natural persons with the necessary competence, training and authority"; the two-person rule may be waived for law enforcement, migration, border control or asylum where law finds it disproportionate. The page banner says the Article applies from 2 December 2027 (Annex III systems) or 2 August 2028 (Annex I systems), "according to Article 113(c)". That banner and the related-recitals line (66, 73) are the site's. + +## Relation to the argument +Premise P11 and P12. Article 14 is the statutory counterpart to the post's P19 stop mechanism: 14(4)(e) combines a stop button with a halt to a safe state, and 14(4)(b) names automation bias, so the law that requires a halting person also states that this person may over-rely. It differs from the post in two ways. First, 14(1) makes interface tools a design duty of the provider, which is a design-side requirement of the kind the paper wants for other parties. Second, 14(2) names the risks to "fundamental rights" of affected people as the aim, although the actors are the overseers the deployer assigns, so the affected person is protected but never acts. The Article scopes itself to "high-risk" systems, so it does not cover a general platform mediator. + +## In conversation with +Does not cite scholarship. The bib holds papers on oversight that discuss it or its field: `green2022flaws`, `crootof2023humans`, `sterz2024quest`, `elish2019moral`, `santonidesio2018meaningful`. Companion text: `nadella2026models` (P19). + +## Caution +Web edition, not the Official Journal; wording could differ in punctuation or in the hover glosses the saved text splices into the paragraph. The effective dates are the site's banner and should be checked against Article 113. Only Article 14 was read, not recitals 66 and 73 or the definitions in Article 3. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The saved web page of artificialintelligenceact.eu Article 14 (text only; same page the card names, not EUR-Lex); 11 quotations (all verbatim, ellipses legitimate), the 14(1)-(5) paraphrases, the effective-date banner ("2 December 2027 ... 2 August 2028, according to Article 113(c)"), recitals 66 and 73 line, and the law-enforcement/migration/border/asylum exemption. + +**Findings:** No issues found. I did not compare against the Official Journal, as the card itself says; the "bib papers discuss it" relations in "In conversation with" are the card author's assignments, not claims made by the regulation, and I did not check them. + +## Also used by facet S (2026-10-15) + +Facet S reads Article 14 beside two legal analyses of its draft, `enqvist2023human` and `koulu2020proceduralizing`, and the facet's point is the one Enqvist makes: the overseer of 14(4)(a)–(c) is an interpreter of system output and of the Article 12 logs, so the statute's human role is reading, and what is shown is for the provider to decide. The five capabilities analysed by Enqvist from the 2021 draft (understand, stay aware of automation bias, interpret, disregard or override, stop) are the five that this card records in the enacted text, which lets the facet cite the draft analysis against the final wording. Koulu's warning that oversight is adopted because the button press and the tick box leave a verifiable record is the facet's gloss on 14(4)(e). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/evans2018forex.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/evans2018forex.md new file mode 100644 index 00000000..9c12e766 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/evans2018forex.md @@ -0,0 +1,37 @@ +# Evans, M. D. D. (2018). Forex trading and the WMR Fix. *Journal of Banking & Finance, 87*, 233–247. + +**Identifier:** doi:10.1016/j.jbankfin.2017.09.017; MPRA Paper No. 81583 · **Read depth:** partial (changed from full_text at S2: main text and references read, but the model derivation and about 40 pages of appendix tables were skimmed) of the **accepted manuscript** (MPRA Paper No. 81583, dated 25 Sept 2017, marked "forthcoming"), 82 PDF pages. Locators are the manuscript's own page numbers (PDF page minus 1); the journal's pagination is not in the file. Read in full: main text (pp. 1–37) and references. Read: the investigations and WMR-methodology notes of the online appendix (PDF p. 42). Skimmed: the model derivation (PDF pp. 43–53) and the extra tables and figures (PDF pp. 54–82) · **Source-tier:** peer-reviewed article (accepted manuscript) · **Evidence basis:** direct_read · **Design and sample:** a microstructure model with a Fix round, tested on tick data for 21 currency pairs, 2004–2013 (Gain Capital quotes, the 4 pm Fix from Datastream, three months of EBS data in each of 2007, 2010 and 2013); windows of 1–60 minutes against a bootstrap of 10,000 random windows away from the Fix, the ECB Fix and US data releases (pp. 19–20) · **Parties modeled:** other (dealer banks, hedgers and investors, with a benchmark administrator whose rule they all trade against) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +Evans asks what banks' trading around the 4 pm London WM/Reuters Fix did to prices. Regulators had fined banks more than $5.6 billion (p. 2), but nobody had measured the market-wide effect. His competitive model has dealers take "Fix orders" from hedgers, trade among themselves and fill the orders at the benchmark. It predicts that Fix orders add volatility only after the Fix, and create no pre-Fix price move and no link between pre- and post-Fix changes (pp. 2–3). + +The data disagree. A price change above the bootstrap's 95th percentile should occur one day in twenty (p. 21). Just before the Fix on month-end days it occurs four to twelve times that often at the one-minute horizon, and on JPY/USD and USD/GBP more than ten times (p. 23; Table 2, p. 22 gives 61.5% and 51.7%). Intra-month, the four major pairs run above three times normal (p. 23). He says they "appear to be pervasive across all currency pairs and throughout the decade covered by the sample" (p. 3), though Table 3 has pair-years without them (CHF/EUR in 2012 shows none). Pre- and post-Fix changes correlate negatively: projection coefficients run from -0.07 to -0.61 and are significant at 5% for all but three pairs at one horizon at least, with R² mostly under 0.2 (p. 28). A reversal strategy at month-end Fixes shows Sharpe ratios above one for some horizon in 15 pairs and above two in eight, drawdowns mostly under 2% (p. 33). Evans calls these "economically attractive" but grants that traders could not have known Table 6 at the time. Taking each pair's best pre-2010 specification and measuring it after 2010 gave 3.25% on average (p. 35). + +On cause he stops short. Banks' trading data are proprietary, so he cannot test the collusion the FCA and DOJ described. The model fails where information sharing among dealers would matter, and front-running could explain a pattern that survived discovery; he calls this "indirect evidence" (pp. 35–36). + +## Relation to the argument + +D8 and facet J. The Fix is a convention in the plainest sense: a rule that sets one price which index providers, asset managers and derivatives then use. The method sampled a one-minute window with medians and "took no account of trading volume" (p. 7). Evans shows it coexisted for a decade with a regularity its users could trade on, and records that WM announced a methodology change in October 2014, after the investigations became public (p. 7, n. 9). The FX instance of "who holds the convention" is then a commercial administrator, users as the parties, and an interested participant that moves the price. It does not show that dealers broke a convention; Evans cannot tell competitive trading from collusion. + +Facet check: facet J marked this metadata only. (1) The appendix lists the Financial Stability Board among the authorities "investigating the forex trading activities" of banks; the FSB's report says its work "was completely independent of the various conduct investigations" (`fsb2014foreign`, p. 1). (2) Evans finds extraordinary pre-Fix volatility; the FSB, with seven currencies and April–September 2013, finds the rise in average volatility at the Fix "not especially large" (`fsb2014foreign`, p. 18). The measures differ (tail frequency against mean), so they do not strictly conflict. + +## In conversation with + +- `fca2014fines`: the appendix cites the FCA press release on the £1.1 billion fines (PDF p. 42). +- `fsb2014foreign`: same object, the 4 pm Fix. Neither cites the other. +- Not in the bib: Melvin and Prins (2015), *J. Financial Markets* 22 (the hedging-flow account he extends); Ito and Yamada (2015), NBER WP 21518 (confirms his result on the Tokyo Fix); Saakvitne (2016); Cushing and Madhavan (2000); Vitale (2000). + +## Caution + +Evans disclosed expert advice to a law firm in an active US case on the Fix (p. 1). The manuscript may differ from the print. The paper never shows who traded. Retail-platform quotes stand in for the interbank market; he checks mid-points against the Fixes and uses EBS spreads for costs (pp. 19, 32). Trading returns are ex post, with half the average EBS inside spread as the cost (p. 32). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the MPRA accepted-manuscript PDF (83 PDF pages; manuscript page = PDF page minus 1). Searched the 4 quotations from Evans (the other 3 flagged strings are facet or FSB wording) and checked the numbers: $5.6bn, one-day-in-twenty, 61.5% / 51.7% (Table 2), four to twelve times, ten times, three times, projection coefficients -0.07 to -0.61 with three insignificant pairs, R2 below 0.2, Sharpe above one in 15 pairs and above two in eight, drawdowns below 2%, 3.25%, October 2014 WM change (n. 9), the appendix FSB and FCA 1.1bn lines, the expert-advice disclosure (p. 1), and Table 3 (CHF/EUR 2012 = 0). + +**Findings:** Three fixes. (1) The card said he 'reports excess tail events in all 21 pairs and every year'; the text (p. 3) says they 'appear to be pervasive across all currency pairs and throughout the decade covered by the sample'; reworded to the source. (2) Table 2 is on p. 22, not p. 21 (p. 21 introduces it); corrected. (3) Read depth changed from full_text to partial because the card itself says the model derivation and about 40 pages of appendix tables were only skimmed. The FSB quotations in the facet check ('was completely independent...', 'not especially large') belong to fsb2014foreign and are checked in that card. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fca2014fines.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fca2014fines.md new file mode 100644 index 00000000..c15237de --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fca2014fines.md @@ -0,0 +1,35 @@ +# Financial Conduct Authority. (2014, November 12). *FCA fines five banks £1.1 billion for FX failings and announces industry-wide remediation programme* [Press release]. + +**Identifier:** https://www.fca.org.uk/news/press-releases/fca-fines-five-banks-for-fx-failings · **Read depth:** full_text (the whole press release, including notes for editors, as HTML; the five Final Notices were linked but not opened; the page carries 2026 site furniture, so it is the live copy of a 2014 release) · **Source-tier:** regulator press release · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** two parties + third party (banks and their clients; regulator) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The FCA fined five banks £1,114,918,000 ($1.7 billion) for "failing to control business practices in their G10 spot foreign exchange (FX) trading operations": Citibank £225,575,000, HSBC £216,363,000, JPMorgan Chase £222,166,000, RBS £217,000,000 and UBS £233,814,000. Without the 30% settlement discount the total would have been £1,592,740,000. The fines are "the largest ever imposed by the FCA, or its predecessor." Its investigation lasted 13 months with over 70 enforcement staff. Barclays was left for a continuing investigation. + +The finding: between 1 January 2008 and 15 October 2013 "ineffective controls at the Banks allowed G10 spot FX traders to put their Banks' interests ahead of those of their clients, other market participants and the wider UK financial system." Traders "shared information about clients' activities which they had been trusted to keep confidential and attempted to manipulate G10 spot FX currency rates, including in collusion with traders at other firms." Chat groups bore names such as "the players" and "1 team, 1 dream," and used code names for clients. The traders aimed at the 4pm WM/Reuters fix, calculated over a one-minute window (30 seconds either side of 4pm), and at clients' stop-loss orders. The release stresses that "there are no specific rules governing the unregulated spot FX market," yet "the importance of managing risks . . . through effective systems and controls is widely recognised in industry codes." It sets up an "industry-wide remediation programme" in which senior management must "attest that this work has been completed." It cites parallel penalties: CFTC over $1.4 billion in total and OCC $700 million on two banks, and a CHF 134 million disgorgement by FINMA from UBS. + +## Relation to the argument + +A dated case in which the conventions of a wholesale market (the fix window and the dealers' conduct) were gamed by the parties who relied on them, and which regulators corrected after the fact. Under ruling D8 it supports the comparison that FX is "not a gentler holder": the FCA concedes there were no specific rules, relied on industry codes, and used enforcement and a supervisory attestation scheme, not a client-side right. For premise P10 it shows interested parties inside a convention bending it to their positions. The FSB's later widening of the fix window (see `fsb2015progress`) is the convention-level response. + +Facet check: the J-facet's two quotations ("ineffective controls . . ." and the "attest" language) match the release. The facet's reading is partial but not wrong. + +## In conversation with + +- `nydfs2015barclays`, `cftc2017fxcm`: companion regulator actions on FX; the release itself names the CFTC and OCC actions but not these. +- `fsb2014foreign` and `fsb2015progress`: the benchmark reforms the release refers to when it mentions the FSB and IOSCO work and the UK Fair and Effective Markets Review. +- Not in the bib: IOSCO principles for financial benchmarks; the Fair and Effective Markets Review (2014); the five Final Notices. + +## Caution + +A press release, not the Final Notices; fine amounts and findings are the FCA's characterisation of settled cases. The release covers voice-desk G10 spot trading on the voice desks based in London, or only in Zurich in UBS's case, per note 7, not electronic trading. It gives no estimate of client losses. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The saved FCA press-release HTML (first published 12/11/2014, last updated 17/07/2026); 12 quotations (all verbatim; the "importance of managing risks . . ." quotation is a correct elision), the figures (the five fines and total, £1,592,740,000 without the 30% discount, 13 months, over 70 staff, CFTC over $1.4 billion, OCC $700 million, FINMA CHF 134 million), the 4pm fix window as stated in note 8, the chat-group names, and the FSB/IOSCO and Fair and Effective Markets Review mentions. + +**Findings:** (1) Note 7 says the voice desks are based "in London, or in the case of UBS only based in Zurich"; the card said "London (and Zurich for UBS)"; fixed. Otherwise no issues. The "not a gentler holder" string is the repo's ruling language, not this release. The "In conversation with" keys (nydfs2015barclays, cftc2017fxcm) are the card author's assignments and were not checked. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/frey2026correlational.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/frey2026correlational.md new file mode 100644 index 00000000..ff68dd0a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/frey2026correlational.md @@ -0,0 +1,34 @@ +# Frey, A., Meary, D., Loichot, M., Mirault, J., & Bosse, M.-L. (2026). Correlational analysis of distinct contributions and overlaps between visual, visual attention, and perceptual spans. *Scientific Reports*. + +**Identifier:** doi:10.1038/s41598-026-38243-x · **Read depth:** full_text of the accepted, unedited "article in press" manuscript PDF (15 pages; no journal pagination, so locators are PDF page numbers). The final article may differ in wording and tables · **Source-tier:** peer-reviewed journal article, accepted manuscript · **Evidence basis:** direct_read · **Design and sample:** within-subject correlational study; 30 French adult psychology students (23 women, mean age 21 years 10 months), measuring visual span (N = 28 usable), visual attention span (N = 30) and perceptual span (N = 30) with matched letter size, plus control tasks and sentence reading; Pearson correlations and linear regression · **Parties modeled:** person–artefact · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Frey and colleagues compare three "spans" that the literature often runs together, measuring all three in the same adults under the same letter size and font. The visual span is how many letters a reader can identify in one glance, measured with letter trigrams. The visual attention span is how many letters a reader can report from a brief array. The perceptual span is how much text around a fixation helps reading, measured by moving a window of readable text and masking the rest. + +Their own values: visual span 7.83 letters (SD 1.71); visual attention span 5.09 letters (SD 0.42); perceptual span 8.87 letters (SD 3.54), or 4.79 degrees of visual angle (PDF p. 5). Visual span correlated with both others (r = .46 with each); the other two did not correlate with each other (Table 1, PDF p. 5–6). All three correlated with reading time, but in a regression only perceptual span explained variance in reading time beyond the others (ΔR² = .126; Table 4, PDF p. 6–7). The three together explained about one third of the variance in reading time and one quarter of the eye-movement measures (PDF p. 6, 8). The authors warn that this rests on 30 people, that reading speed came from the same task as the perceptual span, and that linguistic ability was not measured, so the link to language "should be considered a hypothesis" (PDF p. 8). + +The text the project cares about is a restatement. In the introduction the authors write: "the PS size is estimated to extend about 14-15 characters to the right of fixation and 3-4 characters to the left of fixation, in English" (PDF p. 4). They cite McConkie and Rayner (1975), Rayner and McConkie (1976), Rayner, Well and Pollatsek (1980) and Underwood and McConkie (1985) for it. + +## Relation to the argument +The paper is not about word spacing, and it sits on no premise in `ARGUMENT_MAP.md`. It serves facet D only as an open, secondhand restatement of the classical English figures; it adds no evidence for them. The 14–15 and 3–4 figures matter indirectly to P3 (word-boundary cues act in the parafovea), but the primary studies are the ones to cite. + +Facet check: facet D uses Frey 2026 as the open source for 14–15 characters right and 3–4 left. That is an accurate quotation. The facet does not mention that the article's own measured perceptual span was 8.87 letters, well below the classical figure. The authors attribute the gap to wider letters (0.54° against 0.46° in earlier work) and suggest the span is more stable in degrees than in letters (PDF p. 8). Anyone quoting "14–15 characters" from Frey should know the same paper finds less in its own sample. + +## In conversation with +- `rayner1986eye`: Rayner 1986 on perceptual span in beginning and skilled readers; Frey cites it (ref. 2) for the view that weaker readers' poor eye-movement control reflects linguistic difficulty, and for span growth through primary school. +- `rayner1998eye`: Rayner 1998 review, cited as ref. 3. +- Cited for the 14–15 / 3–4 figures, not in the bib: McConkie and Rayner 1975; Rayner and McConkie 1976; Rayner, Well and Pollatsek 1980; Underwood and McConkie 1985. +- Leads not in the bib: Frey and Bosse 2018 (the same group's review of the three spans); Veldre and Andrews 2014 (perceptual span varies with reading skill); Choi et al. 2015. + +## Caution +An unedited accepted manuscript. Small, homogeneous sample; correlational; adult readers of French, with the English figures only restated. It cannot be cited as a primary source for the size of the English perceptual span, and it should not be used for anything about word spacing. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The "Article in Press" accepted-manuscript PDF (Sci Rep, accepted 29 January 2026), pdftotext; 4 quotations (verbatim; locators PDF p. 1, 4, 8 match); the values 7.83 (1.71), 5.09 (0.42), 8.87 (3.54), 4.79 degrees, r = .460 and .464, total R-squared .315 with delta R-squared .126, "no more than one-quarter ... one-third" (p. 8), 0.54 vs 0.46 degree letters, 30 students with 23 women, N = 28 for visual span, the references cited for the 14-15 / 3-4 figures (refs 29, 31-33 = McConkie and Rayner 1975; Rayner and McConkie 1976; Rayner, Well and Pollatsek 1980; Underwood and McConkie 1985), and the leads (Frey and Bosse, Veldre and Andrews, Choi et al. all in the reference list). + +**Findings:** No issues found. Not separately checked: mean age "21 years 10 months" and the "Table 4" page labels (PDF p. 6-7) beyond the delta R-squared value. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fsb2014foreign.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fsb2014foreign.md new file mode 100644 index 00000000..1dc0e237 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fsb2014foreign.md @@ -0,0 +1,39 @@ +# Financial Stability Board. (2014, September 30). *Foreign exchange benchmarks: Final report*. Financial Stability Board. + +**Identifier:** https://www.fsb.org/wp-content/uploads/r_140930.pdf · **Read depth:** partial (changed from full_text at S2: about 85 of 133 PDF pages, the principle-by-principle analysis of Appendix 3, were skimmed, not read) of the published PDF (133 PDF pages; printed page = PDF page minus 2, cited by printed page). Read in full: executive summary through Appendix 2 (pp. 1–34) and the overall assessment and summary table of Appendix 3 (PDF pp. 37–48). Skimmed, not read: the principle-by-principle analysis and the self-assessment template of Appendix 3 (PDF pp. 49–133) · **Source-tier:** report (official-sector working group) · **Evidence basis:** direct_read · **Design and sample:** engagement with market participants (no count given); EBS and Thomson Reuters Matching data for seven currencies against the US dollar, April–September 2013 (p. 16); an IOSCO desk review of WM's self-assessment as of 2 May 2014 for USD/CAD, EUR/CHF and USD/IDR (PDF pp. 41, 43) · **Parties modeled:** two parties plus a third party (client and dealer, with a benchmark administrator and index providers) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +A working group chaired by Guy Debelle (Reserve Bank of Australia) and Paul Fisher (Bank of England) says the problem lies in the structure around the WM/Reuters 4 pm fix more than in its formula. Passive funds and some corporates ask dealers to fill orders at the fix price. Dealers accept as principal and hedge in the window, which produces the volume spike. That "creates optics of dealers 'trading ahead' of the fix" and "an opportunity and an incentive" to influence the rate, "allegedly including by collusion" (p. 2). The verdict: "it is the incentive and opportunity for improper trading behaviour of market participants around the fix, more than the methodology for computing the fix (although the two interact), which could lead to potential adverse outcomes for clients" (p. 2). The group says its work "was completely independent of the various conduct investigations" and that it saw none of their evidence (p. 1). + +Its data test the volume story. For all seven currencies the fix minute is the day's largest volume spike, in most cases at least 10 times the mean minute (p. 16). Volatility does not follow: the highest average one-minute volatility is the 8:30 ET US data release, and the rise at the fix is "not especially large" (p. 18). The group adds that participants reported higher volatility in past years at month- and quarter-ends (p. 18). + +It describes a "self-reinforcing dynamic": indices benchmark to the fix, trackers transact at it, dealers agree to it, and the market settles on it (p. 13; Fig. 2, p. 14). Index providers show "little appetite" to switch (p. 16). The IOSCO review found WM's oversight "informal and insufficiently tailored" and noted that "WM is not currently subject to regulatory oversight" (p. 43). One staff member chose the members of its oversight group (p. 46). The review checked WM's answers only against documents WM supplied (PDF p. 43). + +Of the 15 recommendations, the main ones widen the one-minute window (median feedback: five minutes, pp. 3, 24), add data sources, create a named user group, price fix trades transparently, bar dealers from sharing positions, tighten and enforce codes of conduct, and ask index providers and asset managers to check fitness and best execution (pp. 3–4, 23–31). The group does not recommend banning fix trading because that "would require legislation or direct regulatory action" (p. 27). + +## Relation to the argument + +D8 and facet J: FX as the comparator for who holds a convention. The fix has a private administrator with no regulator (p. 43), and users adopt it because the market has "naturally coalesced" around it (p. 16). The remedy is a set of recommendations to the administrator and market bodies, with acceptance expected. Legal force "may require legislation" (p. 28), and the group notes that "the authorities" may conclude regulatory change is needed (p. 4). It says the window width should not be "dictated by the authorities" (p. 24). On facet J's Q2 (notice, voice, correction): "Very few market participants we spoke to fully understood the methodology ... Nor did they have any contact with the WM company" (p. 26); the answer is a user group. This is a convention nobody in particular owns and nobody could refuse, which fits ruling D8's answer that the paper does not name a holder. + +Facet check: facet J says the change "was proposed by a body of regulators and market participants". Appendix 2 (pp. 33–34) lists only central banks, regulators, the ECB, IOSCO and the FSB secretariat; market participants were consulted (p. 1). Facet J's corrected line, that the report does not say dealers broke a convention, agrees with the text ("allegedly", p. 2). + +## In conversation with + +- `evans2018forex`: same fix, measured differently (tail frequency, 2004–2013, against averages, six months of 2013); neither cites the other. +- The report cites the IOSCO *Principles for Financial Benchmarks* (July 2013), the ACI Model Code and the NIPS code (p. 28), and a Bloomberg story of 12 June 2013 as the first public alarm (p. 7). None is in the bib. +- Adjacent in the bib, not named in this text: `fsb2015progress` (the 2015 follow-up), `fca2014fines`, `globalfxc2024code` (p. 30 anticipates a revision of the codes). + +## Caution + +A consensus document of officials, not a study. The data are one six-month window, seven currencies, two inter-dealer platforms (much trading happens elsewhere, p. 16, n. 15) and averages, so rare days are invisible. The IOSCO review is a desk exercise on self-assessment under a "tight schedule" (PDF p. 45) and gives no ratings. The recommendations forecast acceptance; they report no results. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the FSB final report PDF (134 text pages; printed = PDF minus 2). Searched all 17 quotations (the one non-match is facet-J wording) and checked printed-page locators (pp. 1-4, 7, 13, 16-18, 22, 24, 26-28, 33-34, 43, 46), the 15 recommendations on pp. 3-4, the seven currencies and April-September 2013 sample, the 'at least 10 times' spike, the median five-minute window, and the IOSCO review lines (USD/CAD, EUR/CHF, USD/IDR; 2 May 2014; 'tight schedule'). + +**Findings:** One fix: Read depth said full_text while stating that the principle-by-principle analysis and self-assessment template (PDF pp. 49-133, about 85 pages) were skimmed; changed to partial. No wrong quotation, number or locator found. The card mixes printed and PDF page numbers but labels the PDF ones; the single 'p. 46' for the oversight-group staff member is a printed page and is correct. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fsb2015progress.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fsb2015progress.md new file mode 100644 index 00000000..26c6156a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/fsb2015progress.md @@ -0,0 +1,43 @@ +# Financial Stability Board. (2015). *Foreign exchange benchmarks: Report on progress in implementing the September 2014 recommendations* (1 October 2015). + +**Identifier:** https://fsb.org/uploads/FX-Benchmarks-progress-report.pdf · **Read depth:** full_text (21-page PDF; the printed page numbers 1–13 of the body are used as locators; Appendix B, extra currency charts for six pairs, was not read beyond its title and the body's statement that the patterns are similar) · **Source-tier:** report (international standard-setter) · **Evidence basis:** direct_read · **Design and sample:** a survey of progress as at 30 June 2015 gathered through foreign exchange committees and central banks (Appendix A lists them), plus a data analysis by the Bank of Canada and the Federal Reserve Board of trading around the London 4pm fix, Nov 2014–May 2015, for six currencies (EUR and JPY from EBS; AUD, GBP, CAD and MXN from Thomson Reuters Matching), with the pre-change comparison being April–September 2013 data from the 2014 report (fn. 15) · **Parties modeled:** other (market participants, a benchmark administrator, and official-sector overseers) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The report says "there has been good progress in implementing many of the September 2014 recommendations, although on some, the progress has been mixed" (p. 1). It reviews the fifteen recommendations in four groups. + +**Methodology (recs. 1–4).** "On 15 February 2015, WM widened the window used to calculate benchmark rates from 1 minute for 'traded' currencies and 2 minutes for 'non-traded' currencies, to 5 minutes for all currency pairs" (p. 2). WM added data feeds (Thomson Reuters Matching for four more pairs, approved 15 February 2015), and planned a user group that "does not yet exist" but should convene by the end of 2015 (pp. 2–3). IOSCO would follow up its review of WM. In April 2015 the WM/Reuters 4pm rate and six other benchmarks became subject to FCA standards on governance, conflicts and record keeping (p. 3). + +**Execution (recs. 6–8).** Independent netting facilities are growing but some banks are reluctant (pp. 3–4). Most sell-side respondents had largely implemented charging for fixing transactions, and "the widening of the fix window by WM in February 2015 was a focal point for a number of banks to begin charging" (p. 4). Some clients complained about lack of transition periods and about first-mover disadvantage (p. 4). Many larger banks moved fixing orders from the voice desk to electronic desks, some by physical separation, others only by separate order-collection processes (p. 5). + +**Conduct (recs. 9–13).** The global committees issued a joint statement in January 2015 and "Shared Global Principles" in March 2015 (Tokyo); the BIS Markets Committee was to complete a single global code by May 2017 (pp. 6–7). + +**Data (§3).** Dealer-driven algorithmic trading grew from about Q3 2014 and jumped when the window widened, and is now the largest share of volume in the window for EUR and JPY (p. 8). Volume remains highest in the window, but is distributed more evenly and is less front-loaded than under the one-minute window (pp. 8–10). Bid-offer spreads are lowest in the window and market depth rises, as in 2013 (p. 11). Prices begin to move only once the window opens; high-frequency volatility stays near 2013 levels except for EUR; five-minute volatility rose relative to the rest of the day (p. 12). Order flow has "a slightly higher influence on price behaviour," and the report cannot say whether this comes from algorithms or the window (p. 13). The authors infer from the lack of pre-window volume that participants may no longer trade ahead using knowledge of aggregate fixing flow, and add that this "does not mean, however, that prices will not move, even temporarily, when there are sizeable one-way flows" (p. 13). The conclusion: the implementation of the recommendations, together with enhanced scrutiny, "appears to have moved the market in a favourable direction" (p. 13), with implementation for non-London fixes "less advanced." + +## Relation to the argument + +For ruling D8 (where the convention bends) the report is the dated evidence that an FX convention changed: the fix window went from one or two minutes to five on a stated date, five months after the FSB's 2014 recommendations, by the administrator's action, with the official sector recommending and then reporting. The pattern fits the J-facet's reading of FX as a market whose conventions are set by a private administrator (WM) and industry committees under official-sector pressure. It also records that the administrator's user group had not yet been formed, which is the voice of users still pending. For the paper's claim about conventions that an owner can withdraw alone, the window change shows a convention changing by the benchmark owner after outside pressure, not by the clients who relied on the fix. Treat the effect claims with care: the report reads the data as "may indicate" and compares different periods. + +Facet check: the J-facet quoted the 15 February 2015 sentence accurately and read only §2.1. The rest of the report adds the charging and separation findings, the algorithmic shift, and the report's own caution that causes are unclear (p. 13); none of it contradicts the facet. + +## In conversation with + +From the report's text and notes: +- `fsb2014foreign`: the September 2014 report whose fifteen recommendations this one reviews (p. 1). +- `globalfxc2024code`: the single global code the report says the BIS Markets Committee would draft by May 2017 (p. 7). +- `fca2014fines`: the FCA remediation programme is cited in fn. 7, and the FCA benchmark regulation in §2.1. +- Not in the bib: the IOSCO Principles for Financial Benchmarks (2013); the Fair and Effective Markets Review (2014); WM's methodology-change notice (fn. 5); the Shared Global Principles (fn. 9). + +## Caution + +A progress report assembled from self-reported survey answers and an official-sector analysis of a seven-month sample, with no control group and a pre-change sample from a different year. It describes the market's own response and does not test whether manipulation fell. Charts were read as described in the text; the underlying numbers are not given. The report does not mention last look. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The 21-page FSB PDF (1 October 2015), pdftotext; 10 quotations; printed-page locators (printed page = PDF page minus 1; all match); the fifteen recommendations, January and March 2015 statements, Tokyo meeting, May 2017 global code, April 2015 FCA benchmark regulation, the six currencies and data sources (EBS for EUR and JPY; Thomson Reuters Matching for the other four), the sample periods (footnote 15), Appendix B title. + +**Findings:** The conclusion quotation had the verb in the wrong form: the source says the implementation "appears to have moved the market in a favourable direction" (the subject is the implementation of the recommendations plus enhanced scrutiny, not "the changes"); fixed. Everything else matches. The "In conversation with" relation to `fca2014fines` (FCA remediation programme in fn. 7) I did not re-trace beyond the footnote existing. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/galanter1974why.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/galanter1974why.md new file mode 100644 index 00000000..0d6261a2 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/galanter1974why.md @@ -0,0 +1,37 @@ +# Galanter, M. (1974). Why the "haves" come out ahead: Speculations on the limits of legal change. *Law & Society Review, 9*(1), 95–160. + +**Identifier:** doi:10.2307/3053023 · **Read depth:** partial (publisher PDF from Cambridge Core, journal pagination 95–160, 66 PDF pages; main text pp. 95–112, 113–125, 129–137 and 141–151 read closely; the long footnotes read selectively; pp. 126–128, 138–141 and 145–148 read in part; the reference list skimmed) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** none of its own; a conjectural essay built on secondary studies, with figures cited from others in footnotes · **Parties modeled:** two parties plus a third-party forum (repeat player against one-shotter before a court-like body), with lawyers as intermediaries · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Galanter asks when litigation can redistribute power, and answers that the "architecture" of the legal system gives the "haves" interlocking advantages. He admits the essay is conjecture: "The speculative and tentative nature of the assertions here will be apparent and is acknowledged here wholesale" (p. 95). + +He starts from the parties, not the rules (p. 97). A repeat player (RP) is "a unit which has had and anticipates repeated litigation, which has low stakes in the outcome of any one case, and which has the resources to pursue its long-run interests" (p. 98). A one-shotter (OS) is the opposite case. He lists nine RP advantages (pp. 98–103): advance intelligence ("It is the RP who writes the form contract", p. 98), expertise, informal ties with the forum, a reputation for commitment, the freedom to play the odds (p. 99), and the freedom to "play for rules as well as immediate gains" (p. 100), with further advantages in spotting which rules will "penetrate" (items 8-9, p. 103). RPs settle cases likely to make bad precedent, so the stock of precedent skews their way (pp. 101–102). Figure 1 sorts litigation into four boxes by RP or OS on each side (p. 107). Most litigation sits in RP-against-OS (Box II). He cites Wanner's sample of 7,900 civil cases in three cities, where business and government units were plaintiffs in almost six cases in ten (n. 29, p. 108). + +Three more layers add to the first. Lawyers: specialists who serve OSs lack the RP's repeat position, and lawyers whose clients come through a broker may serve the broker, because "The source of business generally counts for more than the client" (Carlin, quoted p. 118). Institutions are passive and overloaded, which favours the possessor and pushes claimants to settle (pp. 119–122). Rules "tend to favor older, culturally dominant interests" (p. 123). Figure 3 shows the layers reinforcing each other (p. 125). + +Parties in a dense, continuing relationship avoid courts and build private remedies, because each can withhold future dealing. Government and monopolies are different: "Informal controls depend upon the ultimate sanction of withdrawal and refusal to continue beneficial relations" (p. 111), and those parties lack it, so more disputes reach the forum (pp. 111–112). + +On reform he concludes that "change at the level of substantive rules is not likely in itself to be determinative of redistributive outcomes" (p. 149). Rules are abundant and organized parties are scarce: "If rules are the most abundant resource for reformers, parties capable of pursuing long-range strategies are the rarest" (p. 150). His priority is to aggregate OSs into something like RPs: membership associations, assignees of fragmentary rights, interest-group sponsors, clearing-houses, government prosecution (pp. 141–142). + +## Relation to the argument +Ruling D8 (compare courts, FX markets and platforms; "look for where the convention bends towards"). Galanter gives the mechanism for the tilt: a neutral rule still favours the party that writes the form contract and plays repeatedly. That supports premise P10 (the interface owner's interest differs from the user's), and the form-contract line (p. 98) anticipates platform terms of service. + +It also cuts against any hope that a convention alone protects the weaker party. On his account "changes in the rules can be filtered out unless accompanied by changes at other levels" (p. 149). For the paper, a spacing-style convention would need organized parties and facilities behind it. His scheme for turning OSs into RPs is a closer analogue to counter-delegation (Part 1, line 144) than to a rule; the match is the card's reading, not Galanter's. + +Facet check: the facet maps the platform to the extreme RP and the driver or seller to a one-shotter. Galanter's own note 4 (p. 97) says welfare clients and taxpayers "are not pure OSs, since there is next year's tax bill and next month's welfare check", and pp. 129–133 predict private ordering where relationships are dense and enduring. A driver is in a continuing, unequal relationship, so the mapping needs that qualification. The facet's p. 111 quotation (arbitration "detached from official sanctions") describes RP-against-RP relations of comparable size, not the platform case. + +## In conversation with +No work in `references.bib` appears in Galanter's reference list; the essay predates all of them. Leads from his list, not in the bib: Hirschman (1970), *Exit, Voice, and Loyalty* (n. 42, p. 111, where he splits voice into internal and external); Macaulay (1963) and Friedman and Macaulay (1967) on business dealings (nn. 37–38); Golding (1969) on procedural justice (n. 44, p. 113); Edelman (1967), *The Symbolic Uses of Politics* (p. 107); Rothstein (1974) on repeat-player strategy (n. 18). + +## Caution +A 1974 essay on the American system, tentative by its author's own account; the figures in it come from secondary studies of small-claims, tenant and injury litigation, and none concerns online platforms. Applying "repeat player" to platforms is the project's step. I skimmed parts of pp. 126–128, 138–141 and 145–148 and most footnotes, so claims about those passages need a second read. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The Cambridge Core PDF (journal pp. 95-160; PDF page + 94 = journal page, confirmed on p. 95), OCR text; 14 quotations (12 verbatim with locators checked page by page; the "unit which has had ..." definition and "not pure OSs" quotation are broken across lines or OCR artefacts in the text but read correctly; the facet's "detached from official sanctions" is on p. 111), Wanner (n. 29, p. 108: 7900 cases, almost six out of ten), n. 4 on p. 97, Carlin quotation p. 118, Figure 1 p. 107, Figure 3 p. 125. + +**Findings:** (1) Galanter lists nine RP advantages (items 1-9, pp. 98-103), not seven (pp. 98-101); fixed. (2) Read depth said full_text while stating that pp. 126-128, 138-141, 145-148 and most footnotes were skimmed or read only in part; changed to partial. The precedent claim (the body of "precedent" cases is skewed toward RPs, p. 102) checks out, though the "RPs settle cases likely to make bad precedent" wording is the card's paraphrase of the item 7 discussion on pp. 101-102. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/galliussi2020inter.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/galliussi2020inter.md new file mode 100644 index 00000000..694c0093 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/galliussi2020inter.md @@ -0,0 +1,35 @@ +# Galliussi, J., Perondi, L., Chia, G., Gerbino, W., & Bernardis, P. (2020). Inter-letter spacing, inter-word spacing, and font with dyslexia-friendly features: Testing text readability in people with and without dyslexia. *Annals of Dyslexia, 70*, 141–152. + +**Identifier:** doi:10.1007/s11881-020-00194-x · **Read depth:** full_text (open-access published PDF from Springer; journal pagination 141–152; the electronic supplementary appendices are not in the PDF and were not read) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** 2 × 2 × 2 within-subject experiment (letterform × inter-letter spacing × inter-word spacing), 128 Italian-speaking children aged about 12.4 (64 with dyslexia, 64 age-matched controls), reading 8 texts aloud · **Parties modeled:** person–artefact (reader and typeset text) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Galliussi and colleagues ask whether "dyslexia-friendly" fonts help because of their letterforms or because of their wider spacing, and they split spacing into its letter and word components for the first time among the studies they review (p. 143). Each child read a baseline text and eight matched test texts, one per condition, in 14-pt type with a Latin-square assignment of texts to conditions (p. 144–145). Increased spacing was +70/1000 em between letters and +270/1000 em between words (p. 144). Speed was syllables per second; accuracy was a log ratio of words correct to errors (p. 145). + +The letterform had no main effect for either group (speed, children with dyslexia, F(1,434) = 0.006, p = .936; controls, F(1,434) = 0.843, p = .359; pp. 147, 149). Spacing mattered through an interaction. Widening letter spacing without widening word spacing slowed both groups. For children with dyslexia, the default condition ran at 3.22 syll/s and the wide-letter, default-word condition at 3.04 syll/s (t(63) = −4.258, p < .001); controls ran at 4.99 against 4.88 syll/s (t(63) = −2.577, p = .036; pp. 147, 149). Widening both was no better than widening neither (3.23 against 3.22 syll/s for children with dyslexia). One letterform interaction did appear for the dyslexic group alone: with the dyslexia-friendly letterform, widening word spacing raised speed from 3.10 to 3.26 syll/s (letterform x inter-word spacing, F(1,434) = 4.471, p = .035; t(63) = 3.33, p < .001), and not with the standard letterform (p. 147). Accuracy showed no reliable effect of any factor (pp. 147–149). The authors read the slowdown as a segmentation failure: "in this unusual spacing condition, it is difficult to segment sentences into words because the sizes of the (increased) inter-letter and the (default) inter-word spacing appeared similar" (p. 150). + +They also say the spacing effects were small: "not practically meaningful (e.g., the difference in reading speed between the two conditions did not exceed 0.20 syll/s)" (p. 150). The groups differed enormously at baseline (dyslexic 2.91 against control 4.49 syll/s, about 35% lower; p. 145), so the spacing changes were minor beside the group gap. They found no general speed gain from wider spacing, unlike Zorzi et al. (2012), and offer line length, amount of widening and age as explanations (p. 150). They did not measure comprehension. + +## Relation to the argument +This is the one paper in the batch whose result turns on the relation between letter gaps and word gaps. It supports the claim that a space works as a boundary cue only when it is wider than the gaps inside words (premise P5: re-representing the task so the page marks what the reader would otherwise infer). Widen the letters and the boundary vanishes, and reading slows. That fits the author's wish to find the move that "does most of the work" among a complex of conventions (ruling D6): relative gap size, not whitespace as such, is the candidate. + +It cuts against any wider claim that more space helps the struggling reader (premise P6, "most for readers still learning"): the dyslexic and control groups responded alike, and the effects were small. Its subjects are 12-year-olds who already read Italian, a transparent orthography, so it says nothing about how long a learner takes to acquire reading (ruling D2). Under D10 (history as a guide) it is background for the analogy only. + +Facet check: the facet's one-line summary, "reading speed is impaired by an increase in inter-letter spacing not combined with an adequate increase in inter-word spacing," matches the abstract. It omits that the authors call the effect practically negligible (under 0.20 syll/s) and found no letterform effect. + +## In conversation with +- `zorzi2012extra` — the seminal extra-large-spacing study; Galliussi et al. fail to reproduce its general gain and note Zorzi did not match sentence length (p. 142, p. 150). +- `hakvoort2017improvements` — "partly replicated" Zorzi with equal-length sentences, finding fewer errors but no speed gain (p. 142). +- Leads not in the bib: Marinus et al. (2016, *Dyslexia*), which separated Dyslexie letterform from spacing; Bachmann (2013), the EasyReading study; Kuster et al. (2017) and Wery & Diliberto (2017), no-effect studies of dyslexia fonts; Schneps et al. (2013), shorter lines. + +## Caution +One Italian-language sample of 12-year-olds, read aloud, with no comprehension measure. Accuracy sits near ceiling in a transparent orthography (the authors say so, p. 150–151). The tested widening is modest and the effect sizes are tiny (partial η² around .01 to .04). It concerns letter-versus-word spacing in an alphabetic script, not the removal of word spaces, so it cannot be cited as evidence on scriptio continua. The ESM appendices (text norms, font design) were not read. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The Springer open-access PDF (12 pages, journal pp. 141-152, page = PDF page + 140), pdftotext; 6 quotations (the two flagged by the matcher are repo phrases; the four source quotations match with p. 141/142/150 locators), about 25 numbers (128 children, 64 + 64, 12.4 years, 14 pt, +70 and +270 per 1000 em, F and p values for letterform, 3.22/3.04, 4.99/4.88, t values, baseline 2.91 vs 4.49, partial eta-squared range .010-.037, Latin squares, eight test texts), and the cited-work relations (Zorzi, Hakvoort, Marinus, Bachmann, Kuster, Wery, Schneps all present). + +**Findings:** (1) The card said "the letterform did nothing"; the paper found no main effect but did find a letterform x inter-word-spacing interaction in the dyslexic group (p. 147); added and the sentence softened. No other numerical errors. The "dissociated spacing into two components" claim is at p. 143 as stated. The Read depth line is honest (ESM not read). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/garrett1998saenger.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/garrett1998saenger.md new file mode 100644 index 00000000..3fe81807 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/garrett1998saenger.md @@ -0,0 +1,34 @@ +# Garrett, J. (1998). Review of P. H. Saenger, *Space between words: The origins of silent reading*. *College & Research Libraries, 59*(5), 486–488. + +**Identifier:** doi:10.5860/crl.59.5.486 · **Read depth:** full_text (published open-access PDF of the journal page spread; journal pagination intact; the review occupies pp. 486–488 and shares those pages with other reviews) · **Source-tier:** book review in a peer-reviewed professional journal · **Evidence basis:** direct_read · **Design and sample:** none (review of one book) · **Parties modeled:** person–artefact (reader and page) · **Facets:** B · **Read:** 2026-10-13 + +## What it argues +Garrett, a reviewer from Northwestern University writing for librarians, presents Saenger's book as proof that medieval scribes pursued "readability" centuries before print. He states the thesis in Saenger's strong form. Word space "was introduced into medieval manuscripts to serve as a co-constituent of meaning, a visual cue allowing the rapid parsing of units of meaning at the sentence, word, and morphemic level" (p. 487). Without it the reader must separate the text "at considerable cognitive expense", by repeated passes (p. 487). He dates the change from "the seventh century on", beginning in the scriptoria of the British Isles and moving east and south (p. 487), and he adds the other cues Saenger treats: capitals, line-end hyphens, ligatures, the ampersand (p. 487). He notes that the scholarly apparatus takes 200 of the book's nearly 500 pages (p. 487). + +Garrett then sets Saenger against Ong. Ong gives white space to typography; Garrett says Ong "overlooks the microtextual origins of white space as a coformative of meaning" (p. 488). He describes scribes who tried and dropped devices, such as spaces between syllables ("aerated text"), as a "hit-and-miss, trial-and-error, by no means linear" path (p. 488). He closes by extending the claim to screens: "Many of the issues we face today (e.g., the readability of extended texts on scrollable computer screens) had their roots in the ruminations of Irish scribes of the seventh century" (p. 488). + +## Relation to the argument +This is reception data, not evidence. It shows that a 1998 reader outside the classicists' dispute took the efficiency-and-cost framing as the book's message and carried it straight to interface design. That is the nearest published precedent for the author's own move (spacing as an interface change), and it dates the move to 1998. Garrett argues the cost claim (P1, P2 in `ARGUMENT_MAP.md`) only by restating Saenger. He does not test the physiology, he does not mention Knox, Vatri or Heilmann, and he says nothing about whether spacing raised the number of readers (P6). So it supports no premise as evidence. + +It bears on rulings D1 and D10 in one way. The review credits Saenger with correcting Ong on *dating* (space is older than print), not with a mechanism that restructures thought. Garrett's own phrase for the result is a "continuum of technological innovation" (p. 488), which fits the author's "guide, not mechanism" reading (D10) and fits the earlier ruling against an Ong-style restructuring. It cuts against any use of the review to show that spacing "ignited" anything. + +Facet check: the facet B entry matches my reading (strong-form restatement; page 487 for the co-constituent passage; page 488 for the continuum and screens sentences). No contradiction. + +## In conversation with +- `saenger1997space`: the book under review; Garrett restates its thesis and chronology. +- `ong1982orality`: Garrett quotes Ong's "deeply interiorized technology" (p. 486) and corrects Ong on white space (p. 488). +- `eisenstein1979printing`: Garrett says Saenger's results make Eisenstein's print-revolution account "be reconsidered" (p. 487). +- Leads not in the bib: none the review's text supplies beyond literary allusions (Mann, *Doctor Faustus*; Eco, *The Name of the Rose*), which are not scholarly sources. + +## Caution +A page and a half of review by a non-specialist. It cannot be cited for the fact of Saenger's claims (cite the book), for the state of the debate (it predates Vatri 2012 and Heilmann 2020 by fourteen and twenty-two years), or for any statement that spacing made reading faster or more common. Use it only as an early example of the efficiency framing and of the screen analogy. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The journal-page PDF (C&RL, Sept. 1998, pp. 486-488), pdftotext; all 10 quotations in the card (all present; those flagged by the matcher are broken only by soft hyphens or line breaks, and "guide, not mechanism" is the repo's ruling language, not Garrett's); locators p. 487 for the thesis passage, the seventh century, Eisenstein, the 200 of nearly 500 pages; p. 488 for "microtextual origins", "aerated text", the trial-and-error sentence, the continuum and screens sentences; the review's author and affiliation (Jeffrey Garrett, Northwestern); the Mann and Eco allusions. + +**Findings:** (1) The Ong quotation "deeply interiorized technology" is on p. 486 (the review starts on p. 486), not pp. 487-488; only the correction of Ong is on p. 488; fixed. No other issues. This is the review itself, not Saenger's book. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/gema2026chain.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/gema2026chain.md new file mode 100644 index 00000000..9013082e --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/gema2026chain.md @@ -0,0 +1,45 @@ +# Gema, A. P., Rajani, N., Saxena, R., Kwan, W.-C., et al. (2026). *Chain-of-thought faithfulness of reasoning models varies with where and how preference cues are delivered* [Preprint]. arXiv:2608.29464. + +**Identifier:** arXiv:2608.29464 · **Read depth:** partial (arXiv v2, 1 Sep 2026; 42 PDF pages; locators are PDF pages and sections; body pp. 1–17 read, appendices A–M (cue conditions, transcripts, prompts, judge rubrics, eval-awareness, position bias, per-source breakdowns) skimmed; the author list beyond the first four is not checked) · **Source-tier:** preprint (not peer reviewed) · **Evidence basis:** direct_read · **Design and sample:** FACE-Eval, 5,100 samples (5,000 cued plus 100 no-cue controls; 5 cue conditions × 5 sources × 5 preference axes × 2 sides × 20 scenarios), 15 open-weight models from eight families (4B to 1.60T parameters), three seeds, three system prompts; LLM judges (Claude Haiku 4.5, cross-checked with GPT-5.6-Luna) · **Parties modeled:** person–system (an LLM monitor reading a model's trace); no human reader is tested · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Gema and colleagues test one assumption of chain-of-thought (CoT) monitoring: that the trace records the decision to tailor an answer to a preference. Earlier planted-cue tests put a hint in the user message. Agents also receive information from tool returns, retrieved documents and memory, sometimes as a raw artifact the preference must be inferred from (abstract, p. 1). FACE-Eval varies *channel* (user message or prefilled tool return) and *explicitness* (a direct summary or the raw artifact, such as Slack messages or browser history), with the cue text matched across channels (Fig. 1, p. 2). The two measures are the verbalized commitment rate (VCR, the share of cue-following answers whose trace states a decision to tailor) and unverbalized adoption (UAR = cue-following × (1 − VCR)). The authors note "A low VCR matters for oversight: the answer follows the preference while the trace contains no explicit record of the tailoring" (p. 3). + +Results (pp. 7–14; see sections 3.1–3.5): +- Models read the cues: side-identification accuracy exceeds 0.80 and clarity exceeds 4/5 for all 11 rated models; cue following lies between 0.43 and 0.58 across conditions (p. 7). +- VCR is highest for explicit user-message cues and lowest for implicit tool-return cues on all 15 models (p. 9). UAR is higher for tool-return than user-message cues on 15 of 15, by 0.03 to 0.21, and higher for implicit than explicit cues in 28 of 30 model–channel comparisons (pp. 8–9). +- Matching on model-rated clarity leaves a fitted explicit-over-implicit VCR advantage of 0.141 [0.127, 0.156] on the user channel and 0.133 [0.116, 0.150] on the tool channel (p. 10). +- A transcript monitor (GPT-5.6-Luna) reaches pooled AUROC of 0.71–0.73. Adding the trace helps only on the user channel (+0.010 to +0.019; tool-channel intervals include zero). Across 32 model–condition points, higher UAR goes with lower AUROC (r = −0.54 [−0.65, −0.39]; r = −0.78 with the weaker GPT-4o-mini monitor) (pp. 11–12). +- A source-attribution system prompt narrows the channel gap on 7 of 15 models and widens it on 4; telling the model it is monitored shifts UAR by at most 0.070 and "does not reliably close the gap". The authors add: "A narrowed gap is therefore not, by itself, a monitorability gain" (pp. 12–14). +- Higher reasoning effort lowers UAR on both GPT-OSS sizes (by 0.049 to 0.117 on the 120B model, medium to high) but not consistently elsewhere (p. 14). + +## Relation to the argument +**D4 and P9.** More support for the post's remark that CoT transparency is "not sufficient or dependable", now for a deployment pattern the post's world has: agents reading tool outputs, mail and memory. The authors call their results "descriptive measurements in a single-call, prefilled-tool setting" that show where the trace fails to record adoption, "not deliberate concealment" and not a mechanism (p. 17). + +**P10.** The monitor-aware prompt result is relevant to any claim that announcing observation restores the record: it does not reliably do so (p. 14). + +**P11/P15.** Reasoning adds little to the monitor's discrimination on the tool channel (p. 11), which again favours records of actions over records of reasoning. The study does not test a human reading either. + +**D3.** Same scope limit as `chen2025reasoning`: this concerns a model's trace as read by an LLM, not disclosure of an algorithm to the people it acts on. + +**Facet check:** the facet's summary of the abstract (higher unverbalized adoption for tool returns; telling models they are monitored "does not reliably close the gap") holds. The body adds that the narrowing under the directive prompt is partly an increase in user-channel unverbalized adoption, so a smaller gap can mean worse behaviour on the other channel (pp. 13–14). + +## In conversation with +- `duzan2026chain`: the closest prior work; Gema notes it varies instruction versus aside, not channel, and says FACE-Eval is closer to its implicit arm (p. 16). +- `chen2025reasoning`, `turpin2023language`, `lanham2023measuring`: the planted-hint line the paper extends and the causal-dependence line it distinguishes itself from (pp. 15–16). +- `korbak2025chain`: raises the worry that telling a model its CoT is monitored may erode monitorability; Gema measures it (§3.4, p. 12; related work, p. 16). +- `baker2025monitoring`, `kutasov2025shade`: agentic monitoring work placing cues in tool returns or sabotage settings (p. 16). +- Leads not in the bib: Emmons et al. (2025); Guan et al. (2025); Thaman (2026), a reward-hacking benchmark with cues in filesystem state. + +## Caution +Open-weight models only, one prefilled tool call (a continuation after a tool result, not tool selection), one response per condition. The cue tasks may not need serial reasoning, so the trace may be an optional report (pp. 16–17). Judges are LLMs and their VCR levels differ by judge (agreement AC1 0.751 for VCR). The channel effect is not separated from training priors. Preprint v2. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** arXiv:2608.29464v2 PDF (1 Sep 2026, 42 pp.), pdftotext; 7 quotations (6 verbatim with page locators p. 3, 14, 17; the seventh, "not sufficient or dependable", is a repo phrase), about 30 numbers and design facts (5,100 = 5,000 + 100; five cue-delivery conditions, five sources, five axes, two sides, 20 scenarios; 15 models, eight families, 4B to 1.60T; cue following 0.43 to 0.58; gaps 0.03 to 0.21; 28 of 30; 0.141 and 0.133 with intervals; AUROC 0.71 to 0.73; +0.010 to +0.019; correlations -0.54 and -0.78 over 32 cells; seven narrowed, four widened; 0.070; GPT-OSS-120B 0.049 to 0.117; AC1 0.751), and the Duzan relation (Gema pp. 16 and the related-work text: its setup is "closer to their implicit-influence setting"). + +**Findings:** Read depth said full_text while stating the appendices (A-M, pp. 18-42) were skimmed; changed to partial. No numerical or quotation errors found. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/globalfxc2024code.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/globalfxc2024code.md new file mode 100644 index 00000000..b3fec1e2 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/globalfxc2024code.md @@ -0,0 +1,44 @@ +# Global Foreign Exchange Committee. (n.d.; footer of the file read: "Updated December 2024"; the bib key and year say 2021). *FX Global Code: A set of global principles of good practice in the foreign exchange market.* + +**Identifier:** https://www.globalfxc.org/docs/fx_global.pdf · **Read depth:** full_text (the current GFXC PDF, 84 PDF pages, printed pages 1–82; page locators below are the printed numbers, printed p. 1 = PDF p. 3). **The file is the December 2024 edition, not a 2021 edition**: p. 82 carries "Updated December 2024" and "For prior versions of the FX Global Code go to www.globalfxc.org". I did not read earlier editions. · **Source-tier:** industry code of conduct (soft law; issued by a committee of central banks and private-sector firms) · **Evidence basis:** direct_read · **Design and sample:** none (55 principles under six headings; Annex 1 gives stylised do and do-not examples) · **Parties modeled:** two parties (dealer or "Principal" and "Client") plus intermediaries (brokers, e-trading platforms, prime brokers) and a committee that writes the code · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The Code offers "a common set of guidelines" for the wholesale foreign exchange market (p. 1). It binds nobody: it "does not impose legal or regulatory obligations on Market Participants nor does it substitute for regulation" (p. 1). The Global Foreign Exchange Committee (GFXC) "was established in May 2017 as a forum bringing together central banks and private sector participants" and maintains the text (p. 1). Participation is optional through a Statement of Commitment, which "like the Code, is voluntary" (p. 5; form and explanatory note pp. 78–81). The Foreword adds that "the decision of what steps should be undertaken, and in what manner, resides with each Market Participant" (p. 5). I found no complaint body, sanction, monitoring of adherence or register of breaches anywhere in the 82 pages. + +Scope matters for the comparison. The Code covers wholesale participants; it does not expect to apply to the general retail public or to private-banking customers (p. 5). Its clients are banks, funds and corporate treasuries. + +The six headings are Ethics, Governance, Execution, Information Sharing, Risk Management and Compliance, and Confirmation and Settlement. Four sets of principles bear on a third party that sits between a client and a price: + +- **Discretion is allowed but bounded.** A Principal has "no obligation to execute the order until both parties are in agreement", but where acceptance gives it discretion it must use it reasonably and fairly, and not so as to disadvantage the Client (Principle 8, pp. 12–13). Principle 9 asks firms to tell clients "where discretion may exist or may be expected, and how it may be exercised" and "the basis on which trade requests and/or orders might be rejected" (p. 14), and to disclose how conflicts of interest in principal dealing are identified and addressed (p. 14). +- **Last look (Principle 17, pp. 22–23).** The dealer "has sole discretion, based upon the validity and price check processes, over whether the Client's trade request is accepted or not". The Code limits the purpose: last look "should not be used for purposes of information gathering", and the dealer should not trade on the request's information during the window. Disclosure is the instrument: the dealer must explain, "at a minimum", whether price changes may affect acceptance, the typical decision time and the purpose. A carve-out box and "good practice" to be available for dialogue with clients follow on p. 23. Annex 1 gives a prohibited case (pp. 62–63). +- **Records and requests.** Principle 36 asks for a record of the reasons behind electronic order rejections, to give clients transparency "where appropriate", and says information "should be made available to Clients upon request" (p. 40). Principle 39: "At the request of a Client, Market Participants should be able to provide information regarding the actions taken in handling a specific transaction" (p. 42). +- **Terms.** Principle 40 asks participants to "make clear if standard terms are used, and if changes are proposed. Where changes are substantial, these should be agreed before any transaction" (p. 42). Principle 7 asks firms to give "personnel or external parties" confidential channels to raise concerns, investigated by independent functions (p. 11). + +Principle 18 (algorithmic and aggregation services, pp. 23–24) makes disclosure of the strategy, fees, conflicts and routing preferences the convention; clients are encouraged to use standard disclosure templates and transaction-cost data to evaluate the service. + +## Relation to the argument + +D8 asks where the convention bends in forex markets. This text answers: toward the dealers' own committee, in the form of voluntary disclosure and process duties, with discretion left in the dealer's hands. The mechanism for an interested counterparty is disclosure plus internal controls plus a client's right to ask; the Code gives the client no remedy. That is the pattern D3 calls a literacy-based intervention, and the FX Code applies it to an algorithmic counterpart (Principle 18). It speaks to P10 in a direct way: the Code assumes the dealer is interested, names the conflict (Principle 9) and asks only for disclosure and restraint. + +The comparison with platforms mixes populations. The Code protects professional counterparties; Annex 1 even treats a bank that exploits a corporate client's lack of sophistication on mark-up as a breach (p. 61), while platform users are mostly retail. + +**Facet check:** (1) J:234 and the audit say the Code requires no notice of change to clients. Principle 40 (p. 42) asks that substantial changes to terms be "agreed before any transaction", and Principle 8 (p. 12) asks that a changed capacity be "agreed by both parties". Both are voluntary, but both are stronger than DSA Art. 14(2), which only requires that recipients be informed. (2) The facet's table cell "Hearing/complaint: None in the Code" overlooks Principle 7 (external parties may raise concerns, p. 11) and Principle 39 (client requests for transaction data, p. 42). (3) "Reasons: None required per trade" is right in the narrow sense, but the audit's correction (Principles 9 and 36) also omits that Principle 36 says information "should be made available to Clients upon request" and Principle 39. (4) The edition is December 2024, as the audit found; the bib key and year are still 2021. + +## In conversation with + +The Code's footnote 4 (p. 17) tells dealers to behave consistently with the Financial Stability Board's September 2014 final report on FX benchmarks: `fsb2014foreign`. The Code cites no other work in `../references.bib`. The project pairs it with `oomen2016last`, `nydfs2015barclays`, `fca2014fines` and `cftc2017fxcm` (that pairing comes from facet J, not from the Code's text). Leads not in the bib: the GFXC Execution Principles Working Group report on last look and the GFXC report on disclosure and transparency (both listed on p. 3, not read). + +## Caution + +A voluntary code is evidence of what the industry says good practice is, not of what dealers do. The text has no enforcement data. The printed pages 22–24 and 82 have Latin filler text in the text layer (template residue); I read around it, and the page-82 text is a blank template. The Code covers wholesale FX only. The December 2024 edition postdates the facet's other FX sources (2014–2017), and I did not read earlier editions, so I cannot say how its last-look wording differs from them. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the GFXC PDF (December 2024 edition, 84 PDF pages) via pdftotext. Searched all 22 quotations and found each; checked printed-page locators (pp. 1, 5, 11, 12-14, 17 fn 4, 22-24, 40, 42, 61-64, 78-81), the 55-principle count, the six headings and the Principle 7/8/9/17/18/36/39/40 paraphrases. Principle 18 text sits under Latin template filler in the text layer; its content (disclosure of principal capacity, fees, conflicts, routing; TCA template) is legible and matches. + +**Findings:** No issues found. The Read depth line already states the edition mismatch and the filler pages honestly. The bib quotation 'Hearing/complaint: None in the Code' belongs to facet J, not the Code, and the card cites it as such. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/green2022flaws.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/green2022flaws.md new file mode 100644 index 00000000..dbd8dca4 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/green2022flaws.md @@ -0,0 +1,46 @@ +# Green, B. (2022). The flaws of policies requiring human oversight of government algorithms. *Computer Law & Security Review, 45*, Article 105681. + +**Identifier:** doi:10.1016/j.clsr.2022.105681 · arXiv:2109.05067 · **Read depth:** full_text (the publisher's open-access PDF, 22 pages, CC BY; the article has no journal page numbers beyond its article number, so locators are PDF pages; the 41-policy appendix and the reference list skimmed). The bib note records an earlier read of arXiv v3; this card is from the published version · **Source-tier:** peer-reviewed article (legal and policy analysis with a literature synthesis) · **Evidence basis:** direct_read · **Design and sample:** a survey of 41 policy documents (23 proposed or passed laws, 8 guidance documents, 10 on two US risk-assessment tools) coded inductively, then a synthesis of experimental and field studies of human–algorithm decision-making (Section 4, pp. 7–9; the article does not describe how those studies were selected) · **Parties modeled:** person–system (a human overseer of an algorithm) inside an agency, with the people decided about as a third party · **Facets:** G, Z · **Read:** 2026-10-13 + +## What it argues +Green says human-oversight rules "rest on an uninterrogated assumption: that people are able to effectively oversee algorithmic decision-making" (abstract, p. 1). By "human oversight" he means "human judgment at the moment an algorithm renders a specific prediction or decision", not democratic governance (n. 1, p. 2). The policies take three approaches (pp. 4–6): restrict "solely" automated decisions (20 of 41 documents, as in GDPR Art. 22), require human discretion (14 of 41, with the Allegheny family screening tool and COMPAS as the US cases), or require "meaningful" human input (7 of 41). + +Flaw 1, people cannot do the oversight (pp. 7–9). "Solely automated" rules are avoided by token human involvement. On discretion, he cites automation bias, skill decay, and studies in which people judge algorithmic quality badly. London police judged facial-recognition matches correct at three times the real rate (p. 8, citing Fussey and Murray), and US judges override release recommendations in ways that raise detention and racial gaps (p. 8). On "meaningful" input, no policy defines the term (p. 8); explanations raise trust in wrong advice, and "Algorithmic transparency similarly reduces people's ability to detect and correct model errors" (p. 8, citing Poursabzi-Sangdeh et al.); automation bias survives training (p. 8). + +Flaw 2, the rules legitimise bad algorithms and shift blame (pp. 9–10): the human becomes the "scapegoat" while agency leaders and vendors "have it both ways" (p. 10). He cites the Detroit facial-recognition arrest, where the police chief blamed the investigation, and Wisconsin v. Loomis as examples (pp. 9–11). + +Remedy (pp. 11–15). Bainbridge's point applies: "Human overseers are left with a task that is more difficult than their original charge" (p. 11). Better human–algorithm design remains "an open empirical question", with cognitive forcing functions and decision-support tools as directions (p. 11). But he proposes "institutional oversight" in two stages: the agency must justify the algorithm and any proposed human oversight with empirical evidence, including pre-deployment experiments with laypeople and practitioners, and then the justification faces public review and approval, modelled on Cambridge, Massachusetts's surveillance ordinance (pp. 12–14). "The default assumption should be that human oversight is likely to be ineffective, unless proven otherwise" (p. 15). + +## Relation to the argument +**P12.** This is the policy-level statement of P12. The argument map lists Green 2022 under "mixed for the rest"; the full text supports P12 for decision-time oversight. The evidence base is secondary: advisory algorithms in courts, policing and child welfare, mostly lab studies and field audits. It covers decision-time oversight, not after-the-fact audit of agent logs, so the map's stated gap (reconstruction by trained auditors, the post's P15) stays open. + +**D3/D4 and H.** Green's reading of transparency is severe: explanation and transparency "appear to hinder" overseers (p. 8). This supports D3's direction, with a scope limit. The studies concern a decision-maker given an algorithm's advice, not disclosure of an algorithm to the people it acts on. It also fits the post's P9 and its own D4 flag, but Green does not discuss chain-of-thought. + +**P11/P19.** Nadella's person who reads evidence and halts is Green's overseer, and Green's Flaw 2 (blame moves to the frontline human) is the Elish argument applied to policy; he cites `elish2019moral` for it (n. 17, p. 10). + +**P10/D8.** Green offers a candidate holder: public approval by a democratically accountable body, with agency justification as the entry condition (pp. 12–14). That belongs in facet J's comparison next to courts and markets. It addresses adoption of an algorithm, not a convention of interface. + +**Facet check:** the facet's reading (41 policies; people unable to perform oversight; remedy institutional, "not user engagement") holds. One qualification: Green says better oversight design is possible and an "open empirical question" (p. 11), so he does not say oversight can never work. + +## In conversation with +- `bainbridge1983ironies`: the "impossible task" frame, p. 11. +- `parasuraman2010complacency`: automation bias and its persistence after training (pp. 7–8, 11). +- `elish2019moral`: the long pattern of blaming operators (n. 17, p. 10). +- `bucinca2021trust`: cited for cognitive forcing functions (p. 11). +- `poursabzi2021manipulating`: transparency reduces error detection (p. 8). +- `perrow1999normal`: operator error as a product of system structure (p. 11). +- Cited by `chan2024visibility` and `chan2025infrastructure`, both of which leave human review of agent actions to this paper. +- Leads not in the bib: Green and Chen (2019, 2021) on risk assessments and human decisions; Skitka et al. (1999) on automation bias; Veale and Edwards (2018) on GDPR Art. 22; Bansal et al. (2021) on explanations and team performance. + +## Caution +Scope is government decision algorithms in street-level bureaucracies; the author says much of the analysis extends to non-government settings (n. 4). The 41 documents were collected by search, not a systematic protocol. The literature synthesis is selective and advocates a position. The author's 2022 survey predates LLM agents. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened cards/green2022flaws.pdf (publisher PDF, CLSR 45, 105681; doi and title match). Checked the 10 in-source quotations by exact match (layout and raw text), the 41 / 23 / 8 / 10 and 20 / 14 / 7 counts, the London facial-recognition three-times figure and Detroit and Loomis passages, footnotes 1, 4 and 17, and the page locators 2, 4-6, 7-11, 12-15 (PDF pages equal the card's locators). + +**Findings:** One fix: the Design line called the literature part a 'non-systematic synthesis ... (p. 4)'. The article never uses that label and says nothing about how studies were chosen; p. 4 is the policy-search section. Reworded to a synthesis in Section 4 (pp. 7-9) whose selection method the article does not describe. The three non-matching quotations are the argument map's and the facet's phrases, not the source's. All counts, quotations and locators otherwise match. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/greenblatt2024ai.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/greenblatt2024ai.md new file mode 100644 index 00000000..9d348e77 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/greenblatt2024ai.md @@ -0,0 +1,38 @@ +# Greenblatt, R., Shlegeris, B., Sachan, K., & Roger, F. (2024). AI control: Improving safety despite intentional subversion. In *Proceedings of the 41st International Conference on Machine Learning* (PMLR 235). arXiv:2312.06942. + +**Identifier:** arXiv:2312.06942 · **Read depth:** partial (arXiv v5, 23 Jul 2024, 47 pp.; pagination is the arXiv PDF's. Footnote 1, p. 1, says the ICML paper is "a shortened and modified version", so page and table numbers differ from the proceedings. Body, pp. 1–19, read in full; appendices A–N, pp. 21–47, skimmed: glossary, noise, score-combination details and prompt examples) · **Source-tier:** preprint of a proceedings paper · **Evidence basis:** direct_read · **Design and sample:** simulated red-team/blue-team game on 5,000 APPS "interview" coding problems (1,500 generated, upsampled, §2.5); models GPT-4 (untrusted), GPT-3.5 (trusted); no human participants · **Parties modeled:** other: a protocol designer, an untrusted model and a simulated auditor; no affected third party · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Greenblatt et al. ask whether the safety pipelines around a language model still work when the model itself tries to subvert them: "researchers have not evaluated whether such techniques still ensure safety if the model is itself intentionally trying to subvert them" (abstract, p. 1). The blue team (the deployer) holds an untrusted model U (GPT-4), a weaker trusted model T (GPT-3.5) and a small budget of trusted human labour H. The red team writes an "attack policy" that replaces U. The safety property is worst-case: one submitted backdoor, a solution that passes the test cases but has an exploitable input, is a failure (§2, p. 3). + +The human is not a human. The authors "model limited access to trusted human labor by allowing the protocol to access GPT-4 through a channel that will not be affected by the red team" (p. 4), enough to audit "up to 2% of problems" (§2.2, p. 5). They add an informal guess that a person needs about three minutes per problem to match GPT-4 in that role (p. 4); they measured nothing. + +Results (Table 2, p. 7; Table 7, p. 15; usefulness rescaled so GPT-3.5 = 0 and GPT-4 = 1, from raw rates of 22% and 54%, p. 7). Auditing the first 2% of problems upfront: safety 15%, usefulness 100%. Trusted monitoring with a 2% audit budget: 62% and 97%. Untrusted monitoring: 79% and 96%. Trusted editing (20% threshold): 92% and 94%. Untrusted monitoring fails completely (0%) if the monitor model simply lies, so the blue team first tests it on synthetic backdoors and then redacts and paraphrases inputs to block collusion (§3.4, pp. 11–14). Collusion through a "while loop where a for loop would have been more natural" gives safety of 57.5% with no redaction and 78.6% with input and output redaction (Table 6, p. 13). The authors call the contest "an arms race between the red team and the blue team" (p. 12). + +Every protocol treats U as a black box, with no fine-tuning or interpretability (App. D), and each estimate averages 2,000 simulated runs (App. F, p. 27). Their stated limits: the same 1,500 problems serve for tuning and scoring (§2.4, p. 6); a model that sandbags in the evaluation is "the core bottleneck to scalability of control evaluations" (p. 17); control gets weaker as models outgrow the trusted model and the humans, since "humans probably won't get more capable over time" (p. 18); and auditing is assumed to work well, which the authors flag as the setting's weak point for later work (p. 5). + +## Relation to the argument +It supports premise P11 and the Nadella post's logic of external, independent controls: the best protocols do not rely on one model checking itself, they put a weaker trusted model, a validation test and information-jamming outside the untrusted model (post P11, P14, P17). Its related-work section also gives the safety lineage of the "insider" frame: the adversary is "an AI inside the AI lab" and the problem is "similar to the problem of insider threats" (§4, p. 16). + +It cannot support premise P12 as `../../ARGUMENT_MAP.md` lists it. That row cites "Greenblatt et al." among the evidence that people perform sustained monitoring badly. The paper contains no human data, because the auditor is GPT-4. What it shows is that human attention is treated as the scarcest input and spent on the most suspicious few percent of outputs, which is a design for a person who cannot read everything. + +On ruling D4 (the author's "his only move is disclosure design"): these protocols disclose nothing to a person. Trusted editing, defer-to-trusted and redaction change what the model's output path does. They are deployer-side moves, which fits the post but not the author's "only disclosure" reading. + +Facet check: the facet's reading is correct (human simulated by GPT-4; "up to 2%" is an evaluation parameter; the Table 2 numbers match). Facet G lists the read status as "in part"; the full body agrees with what the facet took from it. + +## In conversation with +Nothing in this paper's reference list is in `../references.bib`. In the other direction, within the bib: `korbak2025chain` cites it twice as "Greenblatt et al., 2024b" for responses to flagged outputs and for red-team evaluation of monitors (pp. 2, 5–6); `kutasov2025shade` cites it as the framework that introduced AI control (ref. 10); `jia2026faithfulness` cites it for CoT monitoring (ref. 28); `lanham2023measuring` thanks Greenblatt in its acknowledgements (p. 10). +Leads not in the bib: Bowman et al. (2022), scalable oversight, arXiv:2211.03540; Irving et al. (2018), debate, arXiv:1805.00899; Saunders et al. (2022), AI critiques; Naihin et al. (2023), arXiv:2311.10538; Oakley (2019), *Professional Red Teaming*. + +## Caution +A coding-backdoor testbed with models standing in for people. It gives no evidence on how well a person audits, reads logs or halts anything, and 2% is an assumed budget, not a finding. Results come from one model pair, one task type and a tuning set equal to the test set. The arXiv v5 differs from the ICML text, so numbers should be re-checked against the proceedings before they enter the paper. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened cards/greenblatt2024ai.pdf (arXiv 2312.06942v5, 23 Jul 2024, 47 pp.). Checked the 11 in-source quotations by exact match, every number in the Results paragraph against Table 2 (p. 7) and Table 7 (p. 15) (15/100, 62/97, 79/96, 92/94), the 22% / 54% raw rates, 57.5% and 78.6% (Table 6, p. 13), 5,000 problems upsampled from 1,500 (Section 2.5), 2,000 trials (App. F, p. 27), the 2% H budget (p. 5), the three-minutes guess (p. 4) and the page locators 12, 16, 17, 18. Not checked: the claims in the last In-conversation paragraph about what korbak2025chain, kutasov2025shade, jia2026faithfulness and lanham2023measuring say; they live in other cards. + +**Findings:** One fix: Read depth said full_text although the card states that the appendices (pp. 21-47, more than half the PDF) were only skimmed; changed to partial and kept the list of what was not read. All quotations, numbers and locators match. The three non-matching quotations are the author's and the argument map's phrases, not the source's. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/grundemclaughlin2026overseeing.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/grundemclaughlin2026overseeing.md new file mode 100644 index 00000000..61d880dd --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/grundemclaughlin2026overseeing.md @@ -0,0 +1,34 @@ +# Grunde-McLaughlin, M., Mozannar, H., Murad, M., Chen, J., Amershi, S., & Fourney, A. (2026). *Overseeing agents without constant oversight: Challenges and opportunities* (arXiv:2602.16844). arXiv preprint. + +**Identifier:** arXiv:2602.16844 · **Read depth:** full_text (arXiv v1, 18 Feb 2026, 29 pp.; the PDF still carries the ACM template's placeholder venue and "February 2018" date, so no venue has accepted it; pagination is the PDF's. Body, pp. 1–21, read in full; the reference list skimmed; supplementary materials not in the PDF) · **Source-tier:** preprint · **Evidence basis:** direct_read · **Design and sample:** three user studies of N = 12 each (tech-firm employees with agent experience): a formative think-aloud study (§3), a within-participant Wizard-of-Oz study of three trace designs (§4), and a counterbalanced Wizard-of-Oz comparison of baseline against a new interface (§5), all on the Magentic-UI computer-use agent · **Parties modeled:** person–system (a user and the agent whose run they review) · **Facets:** G (also named in Z) · **Read:** 2026-10-13 + +## What it argues +Grunde-McLaughlin et al. ask whether the step-by-step "trace" that an agent shows lets a person find its errors. Their answer in three studies is mostly no. In the formative study, participants missed "small but impactful errors" (§3.3.3, p. 7): of runs that finished, the person-plus-agent team got the wrong result in 1 of 9 attempts on the cart task, 5 of 10 on the recipe task and 10 of 12 on the ticket-price task (p. 7). In six runs the agent estimated a youth ticket price it could not find, and only three of those participants noticed (p. 7). + +The probe study (§4, pp. 9–14) gave participants finished runs with one of three summaries: a flowchart of the process, a cited summary, or a "Specification" listing requirements and assumptions. Overall accuracy at judging whether the answer was correct was 57.14% for the flowchart, 62.86% for citations and 77.14% for the specification (Table 1, p. 12). On the items where the answer was wrong it was 39.13%, 52.17% and 65.22%. The study was not balanced across tasks, and each method had its own tasks (p. 12). Participants who accepted a wrong answer most often cited that "the process seemed reasonable" (§4.3.2, p. 12). + +The controlled study (§5, pp. 14–18) compared the Magentic-UI trace with a new interface that adds the specification, linked annotations and website links. Accuracy was 72.92% against 77.08% (Hedges' g = 0.18, 95% CI −0.52 to 0.94). Time per judgment was 232.87 s against 221.28 s (g = −0.29, CI −1.95 to 1.26). Confidence on a 1–7 scale rose from 4.17 to 4.60 (g = 0.56). Where the answer was wrong and the participant called it right, confidence rose from 3.64 to 4.82 (g = 0.85, CI 0.01 to 1.88; Table 3, p. 18). The authors conclude that the standard trace creates "an illusion of accountability" (p. 3). + +## Relation to the argument +This is the closest test in my batch of human-readable evidence for an agent, and it cuts against the idea that a better readout fixes oversight. It bears on premise P9 (aids that hand the person more to read show weaker, less consistent effects), on P12 (people monitor badly), and on risk 4 in `../../ARGUMENT_MAP.md` (lowering a cost has costs): the interface that helped most also made wrong answers feel safer. It is also an after-the-fact review of a finished run, which the authors say outright is the scenario they test: "the user returns to a completed query for post-hoc verification" (p. 20). That is the task the post's P15 sets for a person, so the study speaks to it, within its limits. + +On ruling D7 (understanding through deciding and quickly ingesting results): one participant wanted the agent to "curate all the information. So I'm the pilot, and it will become co-pilot" (P8, p. 7). That is a user's wish, not a finding, but the study's pattern fits the author's point: people judged from whether the process looked sensible, a reading task, and did worse than when they were handed requirements to check. It does not test learning by acting. + +Facet check: Facet G (abstract only) says the new interface "cut time to find errors". The body is weaker. The overall time effect is small and its interval spans zero. The larger effect (g = −0.65) applies to the subset where participants correctly judged a wrong answer wrong, and its interval (−1.54 to 0.12) also spans zero (Table 3, p. 18). Only the confidence effect on wrong-judged-right items has an interval clear of zero. The facet's accuracy and confidence statements stand. + +## In conversation with +From the reference list: `turpin2023language` (ref. 78) is cited for the point that model explanations may not be faithful (p. 4); `bucinca2021trust` (ref. 16) is cited for explanation designs that fail to improve reliance, and as the model for "cognitive forcing functions" the authors want tested next (pp. 3, 19). Chan et al. (2023), "Harms from increasingly agentic algorithmic systems" (ref. 18), is cited for safety concerns, but it is not the same work as `chan2024visibility`. The meta-analysis they cite for "Human-AI teams perform worse on average than either alone" (p. 4; Vaccaro, Almaatouq & Malone, 2024) is not in `../references.bib`. +Leads: Bansal et al. (2021), CHI, on explanations and team performance; Shavit et al. (2023), *Practices for governing agentic AI systems*; Vasconcelos et al. (2025) on uncertainty highlighting; Terry et al. (2023), arXiv:2311.00710. + +## Caution +Twelve participants per study, all employees of one large technology firm with prior agent use (limitations, p. 20); one agent; browsing tasks with a known right answer. The two later studies are Wizard-of-Oz with hand-made annotations that copy the agent's own claims, and the controlled study allowed participants to use the browser, which "may have evened performance" (p. 19). The study is underpowered by the authors' own account (p. 16) and reports effect sizes with wide intervals, so it supports a claim of no evidence of an accuracy gain, not of no gain. Unreviewed preprint. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened cards/grundemclaughlin2026overseeing.pdf (arXiv 2602.16844v1, 29 pp.). Checked the 5 in-source quotations by exact match plus about 35 numbers against Table 1 (p. 12), Table 3 (p. 18) and the text: all accuracy percentages, 1/9 - 5/10 - 10/12 error counts, n = 6 youth-price runs with half noticing, all Hedges' g values and confidence intervals, and the split of the -0.65 effect to the No-No subset. Also the limitations, power statement (p. 16), post-hoc passage (p. 20), refs 16, 18, 78. + +**Findings:** No issues found. The one non-matching quotation, 'cut time to find errors', is the facet file's wording; the source's abstract (n = 12, g = -0.65) says participants 'found errors faster', and the card's reading of the body table is correct. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/haines2008origins.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/haines2008origins.md new file mode 100644 index 00000000..696aa9ee --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/haines2008origins.md @@ -0,0 +1,36 @@ +# Haines, J. (2008). The origins of the musical staff. *The Musical Quarterly*, 91(3–4), 327–378. + +**Identifier:** doi:10.1093/musqtl/gdp002 · **Read depth:** full_text (journal pagination 327–378, from a PDF of the Oxford Advance Access version posted 4 May 2009; figures are images and were not examined, and the 100-plus endnotes were skimmed, not read in full) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** codicological survey; the author reports personally inspecting over one hundred manuscripts, seventh to fifteenth century, and reads pricking and ruling patterns for dates and origins; no counts beyond one table of 13 Carthusian books · **Parties modeled:** other (scribes, orders, readers of notated books) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues +Haines rejects the claim that Guido of Arezzo invented the staff. He opens with the historians who credit him (Hoppin, Apel, Palisca, Lang, the *Oxford History of Western Music*, pp. 327–328), notes that Burney, Hawkins and Reese had already seen earlier precedent (p. 328), and says "a prototype of the staff existed a century and a half before Guido" in the ninth-century *Musica enchiriadis* (p. 329). Guido's lines, colours and clef letters were all in the *Enchiriadis* tradition. The verdict: "the claim implied by Guido's status as 'the one who invented the staff' is untenable"; his role was "a clever pedagogy of synthesis backed up by a fair bit of self-promotion" (p. 331). Haines also queries the usual reading of Guido's *spisse linee* as "closer together": "spisse" more often means "thick" (p. 331). + +The positive account comes from page layout. Parchment was pricked and ruled for text; the staff "arose, in an unspectacular and natural way, from the ruling of the medieval page" (p. 336). Three phases follow. +- Phase one, ninth to eleventh century: scribes add notes within the existing ruling, with no layout designed for music (p. 340). Aquitanian books widen the ruling (up to 20 mm) and shrink the text; Guido's books use several ruled lines for pitch, and "this popularization, then, rather than an invention of the staff, was Guido's historical contribution" (p. 344). +- Phase two, twelfth century: Guido's layout gave music "five or more times" the space of the text, which wasted parchment (p. 346). Scribes in Italy, France and England responded with a set-off or double prick for the text line, and the Carthusians adopted it consistently: of 13 Carthusian music books listed in Table 1 (12th to 13th century), 10 show a 4+2 pricking pattern and three are marked unknown (pp. 353–355). +- Phase three, thirteenth century: an all-red staff drawn independently of the page ruling, two or three rulings for music against one for text (p. 359). Dominican master book BL Add. 23935 (c. 1260) has about 120,000 red staff lines, which Haines shows were drawn one at a time. He argues from five paleographic tests that the multi-nib rake appears only around 1300 (pp. 363–365). + +He concludes that the staff was "the product of the anonymous ingenuity of many monks" (p. 368), that Guido "did not invent the musical staff" and introduced nothing substantially new, and that the break from dependence on the text page was the work of "hundreds if not thousands of nameless monks" (p. 369). + +## Relation to the argument +This is the closest published parallel to the paper's own correction of a single-inventor story. Haines replaces "Guido invented the staff" with a bundle of layout conventions built by many hands over four centuries, which matches D6 and the argument map's "no longer claims 'the' move". It also reads as a precedent for D1 (design precedent): the staff is a re-representation made through page design, where space for pitch is allotted by ruling. Two cautions for the paper. First, Haines treats the changes as driven by proportion, cost and space on parchment, not by training readers; he never says the staff shortened the learning of chant. Second, he does not mention word separation or Saenger. + +Facet check: the facet's two quotations (p. 331 "untenable … self-promotion"; the "popularization" sentence) are exact; the second falls on p. 344. The facet records a partial read; the full text adds that Guido's own contribution is dated only through manuscripts from the late 11th or early 12th century, since no copy of the prologue survives from his time (p. 345), so any claim about what Guido proposed rests on later copies. + +## In conversation with +- `pesce1999guido`: Haines cites her edition for the text and for her view that Guido's role in the staff is "essential … unquestionable" (n. 18, p. 329); he disagrees with the strength of that claim. +- `mengozzi2010renaissance`, `mews2024guidonian`: not cited in the text I read; no relation shown. +- Leads: Smits van Waesberghe (1951, on coloured staff lines and a catalog of 170 manuscripts); van Dijk (1960s, on layout); Huglo on Dominican and Franciscan ruling rules; Jones, Gillissen and Derolez on pricking and ruling. + +## Caution +Figures were not examined. Dating of many books is "within a half century" (p. 347) and the pricking patterns survive in a minority of books because pages were trimmed (p. 337). The count of books is not given beyond Table 1. Haines argues about page layout, not about reading ability; it does not support a claim that the staff taught more people to sing. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened cards/haines2008origins.pdf (Oxford Advance Access PDF of Musical Quarterly 91, 327-378; pdf page 1 = p. 327). Checked the 14 in-source quotations by exact match, plus locators for pp. 329, 331, 336, 344-347, 355, 363-365, 368-369; Table 1 (13 books, ten marked 4+2, three marked ?); the 120,000 staff-lines figure (Figure 15 book, BL Add. 23935 per the text), the 20 mm and 'five or more times' figures, the 'drawn one at a time' passage, and the n. 18 Pesce quotation. + +**Findings:** No issues found. The two unmatched quoted strings are the argument map's phrase and a paraphrase whose single-quote marks differ from the source's double marks. Minor: the card says Haines 'disagrees with the strength' of Pesce's claim; the text sets Pesce's view beside 'One thing is certain: a prototype ... existed a century and a half before Guido' without stating disagreement in those words, and his own conclusion (p. 331, 'untenable') is the disagreement. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hakvoort2017improvements.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hakvoort2017improvements.md new file mode 100644 index 00000000..7c9e049f --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hakvoort2017improvements.md @@ -0,0 +1,42 @@ +# Hakvoort, B., van den Boer, M., Leenaars, T., Bos, P., & Tijms, J. (2017). Improvements in reading accuracy as a result of increased interletter spacing are not specific to children with dyslexia. *Journal of Experimental Child Psychology, 164*, 101–116. + +**Identifier:** doi:10.1016/j.jecp.2017.07.010 · **Read depth:** full_text (final published version, UvA-DARE repository copy; journal pagination 101–116) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** two experiments. Exp. 1: 30 Dutch children with dyslexia and 30 age-matched controls (mean age 9 y 11 m for the dyslexia group, 9 y 9 m for controls), each reading the same 44 sentences aloud on paper in standard and increased letter spacing, a week apart. Exp. 2: 189 unselected Grade 3–4 children (mean age 9 y 3 m), on a laptop, sentence-at-once against word-by-word presentation · **Parties modeled:** person–artefact (reader and typeset text) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Hakvoort and colleagues try to replicate Zorzi et al. (2012), who found that extra-large letter spacing helped children with dyslexia in particular. They use harder, age-appropriate sentences to avoid the ceiling on controls' accuracy that Skottun and Skoyles (2012) had pointed to (p. 104). Experiment 1 copied Zorzi's layout: letter spacing +2.5, three space characters between words, doubled line spacing (p. 105). + +Result: both groups made fewer errors with increased spacing (F(1,55) = 16.97, p < .001, ηp² = .236), with no group × spacing interaction (F = 0.04, p = .849); speed did not change (F(1,58) = 0.37, p = .547; pp. 106–107). Mean errors per block fell from 10.54 to 10.00 for dyslexic children in block 1 and from 9.86 to 7.54 in block 2; for controls from 5.10 to 3.60 and 5.67 to 3.83 (Table 1, p. 106). The gain was proportionally larger for controls (29.4% and 32.5%) than for children with dyslexia (5.1% and 23.5%). At the individual level, 41.4% of the children with dyslexia and 23.3% of controls showed no gain or a loss in accuracy; effects did not correlate with word-reading fluency (p. 107). Dyslexic children's errors dropped from block 1 to block 2 regardless of condition, which argues against a novelty or "attention boost" account (p. 108). + +Experiment 2 changed only letter spacing (+1.2, proportional word spaces). In whole-sentence reading (N = 96) errors fell from 3.87 to 2.76 per sentence set (28.7%) with speed unchanged at 1.63 words/s; in word-by-word reading (N = 93) the error drop from 1.99 to 1.71 (14.0%) was not reliable (t(91) = 1.80, p = .075; Table 3, p. 110; p. 111). Reading ability did not predict the size of the effect (all |r| ≤ .17; p. 111). + +The authors' three conclusions: "First, increased LS appears to affect reading accuracy only. Second, the findings do not support claims that increased LS specifically affects poor readers." The third reads "the effect of LS seems to occur at the interword level" in the abstract (pp. 101–102) but "at the sentence level" in the General Discussion (p. 112). The experiments contrast words shown one by one with whole sentences, which fits "sentence level"; the abstract wording appears to be a slip. + +## Relation to the argument +This card concerns letter spacing, an adjacent intervention to word spacing, and it sits in facet D as the main cautionary case about spacing as a remedy. It cuts against any reading of the history that says added space helps the weakest readers most (premise P6, "most for readers still learning"): here the same change gave controls at least as much accuracy as the children with dyslexia, and nobody got faster. It also shows how an apparently clean "interface fix" lands unevenly across individuals, which bears on ruling D10 (the history is a guide) as a warning against importing the effect size from one spacing manipulation to another. It does not test word-boundary marking, so it gives nothing to premises P1 to P5 directly. In Experiment 1 the word spaces were also widened, so Experiment 1 cannot separate letter from word spacing. The authors say Experiment 2 "did not allow a direct comparison between interletter and interword spacing", but argue that its identical result suggests the added word spacing was not what caused the Experiment 1 effect (p. 113). + +Facet check: the facet quotes the abstract's "interword level" line. The body says "sentence level"; cite the body. + +## In conversation with +- `zorzi2012extra` — the study replicated, with Experiment 1 following its paradigm (pp. 101, 104–105). +- `skottun2012interletter` — source of the ceiling-effect critique the study addresses (p. 104). +- `perea2012effects` (Perea, Panadero, Moret-Tatay & Gómez, 2012) — cited as the other positive finding for dyslexia (pp. 101, 103). +- `perea2012subtle` — adult skilled readers show no gain in rate or accuracy (p. 113). +- `vandenboer2015default` — the authors' own earlier null result for word naming (pp. 101, 113). +- `pelli2008uncrowded` — the crowding account that motivates letter spacing (p. 102). +- `slattery2016interword` — cited for the view that word spacing should exceed letter spacing (p. 113). +- `dehaene2010learning`, `dehaene2015illiterate` — reading acquisition reshaping early visual processing (p. 103). +- `bouma1970interaction` — the original crowding observation (p. 102). +- Leads not in the bib: Marinus et al. (2016); Schneps et al. (2013), shorter lines; Moll & Jones (2013), crowding in dyslexia; Silva et al. (2016), parafoveal preview in dyslexic adults. + +## Caution +Children's accuracy only; no eye tracking, no comprehension measure. Experiments 1 and 2 differ in medium (paper against laptop), spacing amount and word spacing, so they are not directly comparable (p. 113). Dutch, a fairly transparent orthography. The exposure to each condition was brief (about 17 sentences per block). Crowding was not measured: the authors call for future studies pairing a crowding paradigm with a letter-spacing manipulation, since the link is "rarely directly established" (p. 114). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened cards/hakvoort2017improvements.pdf (final published version, JECP 164, 101-116; pdf page 2 = p. 101). Checked the 9 in-source quotations by exact match and about 40 numbers (F, p, eta-squared values; the Table 1 block means and percentage gains; the 41.4% / 23.3% individual-level figures; Experiment 2 N = 96 / 93, error means, 28.7% / 14.0%, t(91) = 1.80, speed; |r| <= .17; 44 sentences, 17 per block; N = 189) with page locators (pp. 104-114). Not checked: the dozen In-conversation relations beyond those appearing in the text (checked Zorzi, Skottun, Perea, Pelli, Bouma, Dehaene, Slattery, van den Boer, Marinus, Moll, Schneps, Silva by name search). + +**Findings:** Three fixes. (1) Sample: the card gave one mean age (9 y 11 m) for both groups; the source gives 9 y 11 m for the dyslexia group and 9 y 9 m for controls. (2) The card said the authors state the design cannot separate letter from word spacing; the source says Experiment 2 'did not allow a direct comparison between interletter and interword spacing' but that the identical result suggests added word spacing did not cause the Experiment 1 effect (p. 113); reworded. (3) The Caution said the authors 'say crowding was never measured directly'; the source says only that the letter-spacing/crowding relation is 'rarely directly established' and calls for studies that pair the two (p. 114); reworded. The two unmatched quoted strings are the argument map's and the card's own phrases. All statistics match. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hangeulmuseum.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hangeulmuseum.md new file mode 100644 index 00000000..fb5fa913 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hangeulmuseum.md @@ -0,0 +1,31 @@ +# National Hangeul Museum. (2015). Read society through its Hangeul letters [Exhibition page, 21 April to 7 June 2015]. + +**Identifier:** https://www.hangeul.go.kr/exhibition/485 · **Read depth:** full_text (the whole exhibition web page, English version, read on 2026-10-13; the images and captions' pictures were not viewed) · **Source-tier:** institutional web page (museum exhibition text) · **Evidence basis:** direct_read · **Design and sample:** none; curatorial text · **Parties modeled:** other (writers of Hangeul letters) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues +The page presents handwritten Hangeul letters as "the dominant form of communication since the invention of Hangeul", written by "all people of Joseon regardless of the social standings from kings to commoners", and says the letters "played an important role in disseminating Hangeul". Sections give examples. A letter from a landlord to his slave is read as implying "the slave could read Hangeul". The earliest surviving Hangeul letter is dated to the 1490s, from Na Sin-geol (1461–1524) to his wife; the Annals of the Joseon Dynasty record letters exchanged "from 15th century". "Most people of the ruling class used Chinese characters while commoners used Hangeul when writing letters." After the 1894 Gabo Reform and the start of postal service, Hangeul letters "became to be the most common way of communication" and schools taught letter writing. The court style of calligraphy (Gungche) developed from the 17th to the late 19th century; some court ladies acted as secretaries and specialised in transcribing Hangeul letters for queens, and the page says the court style "could be further developed by them" (other court ladies were servants or slaves). + +On later literacy, the page states: "The illiteracy rate of Korean people was 78% in 1945. It went down to 4.1% in 1958 as the government-led measures to eradicate illiteracy had been implemented." It adds: "it's not easy to figure out the rate accurately because proper surveys had not been taken place until 1990", and reports a 2001 Korean Educational Development Institute study finding 24.8% of adults had difficulty reading, writing and using arithmetic. + +## Relation to the argument +Offers a museum's account that a script's spread ran through ordinary use (letters) by non-elites and women. This is one of the "bundle" strands for facet F's claim that a cost-lowering writing system spread through a complex of practices (D6). It does not give a mechanism and cannot show that the script's design caused the spread. + +Facet check: the facet's 78% (1945) and 4.1% (1958) quotations are exact, and so is the surveys caveat. The page gives no source for the 78% or 4.1% figures, and the facet's own note that other sources give a different 1945 figure applies. + +## In conversation with +- `park2025dokkyo`: the Dokkyo abstract also says commoners and women worked mainly in Hangul. +- `baek2024spread`: not read here; the facet gives it as saying the script spread fastest among women and in Seoul. +- Leads: none cited on the page. + +## Caution +Curatorial text for a general audience, without footnotes, so no claim here can be traced to a source. Treat the literacy rates as soft: the page itself says reliable surveys began only in 1990. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the saved HTML of the English exhibition page (cards/hangeulmuseum.html), stripped the markup, and matched every quotation (all exact) and every date and number (21 Apr-7 Jun 2015, Na Sin-geol 1461-1524 and the 1490s, 1894, 17th to late 19th c., 78% / 4.1% / 1990 / 2001 KEDI 24.8%). Not checked: the page images, and the facet-file and park2025dokkyo / baek2024spread relations (other documents). + +**Findings:** One fix: the card said 'Court ladies served as scribes and developed the court style'. The page says only that in some courts court ladies were servants or slaves, that some acted as secretaries and some specialised in transcribing Hangeul letters for queens, and that the court style 'could be further developed by them'. Reworded to that. Everything else matches; the page gives no source for the 78% and 4.1% figures, as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hao2024training.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hao2024training.md new file mode 100644 index 00000000..d3b1e93a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hao2024training.md @@ -0,0 +1,36 @@ +# Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J., & Tian, Y. (2024). *Training large language models to reason in a continuous latent space* (arXiv:2412.06769). arXiv preprint. + +**Identifier:** arXiv:2412.06769 · **Read depth:** full_text (arXiv v4, 23 Aug 2026, "Last updated: August 25, 2026", 18 pp.; this is not the December 2024 text that Korbak et al. cite, since it discusses 2025 follow-up work on p. 3; pagination is the PDF's. Body, pp. 1–11, and appendices A–C, pp. 12–18, read; example prompts in Appendix A skimmed) · **Source-tier:** preprint · **Evidence basis:** direct_read · **Design and sample:** training experiments on GPT-2 with three benchmarks (GSM8k, ProntoQA, and the authors' own ProsQA), plus Llama 3.2-3B and Llama 3-8B on GSM8k; no human participants · **Parties modeled:** none (a model-architecture paper) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Hao et al. propose "Coconut" (Chain of Continuous Thought). Instead of writing each reasoning step as word tokens, the model feeds its last hidden state back in as the next input, so its reasoning stays in a continuous vector space. The motivation is that "language space may not always be optimal for reasoning" (abstract, p. 1). They train it with a curriculum that replaces language reasoning steps with latent ones, one stage at a time (§3, p. 4). + +Results (Table 1, p. 9; GPT-2). On GSM8k (grade-school maths) language chain of thought scores 42.9% and Coconut 34.1%, against 16.5% with no chain of thought. On ProntoQA, Coconut 99.8% against 98.8% for chain of thought. On ProsQA, a graph-search set the authors built, Coconut 97.0% against 77.5%, using 14.2 against 49.4 generated tokens. Without the curriculum Coconut falls to 14.4%, 52.4% and 76.1% on the three sets, so "the LLM still needs guidance to learn latent reasoning" (§5.3, p. 11). The authors conclude that Coconut "does not surpass CoT on GSM8k" but "offers a superior trade-off between reasoning efficiency and accuracy" (p. 10). + +A headline claim is about what the latent thoughts do: "the continuous thought can encode multiple alternative next reasoning steps" (abstract, p. 1). Probing the model by forcing it to emit language after a given latent step shows probability spread over several candidate nodes before it narrows, which the authors read as a breadth-first search (§4.3, pp. 6–8). The probe works because the model can still compute a word distribution from a latent state "for probing purposes", even though "the latent thought is not intended to be mapped back to language space" (p. 4). + +The larger-model check is thin. On GSM8k, Llama 3.2-3B gets 31.7% with Coconut against 26.0% with no chain of thought, and Llama 3-8B gets 43.6% against 42.2% (Table 5, p. 17). The table has no language chain-of-thought column. The authors say the paper's "primary goal" is "to highlight the promising attributes of latent-space reasoning" (p. 18). + +## Relation to the argument +The paper has no safety content: it does not mention monitoring, oversight or readability as a concern, and it never uses the word "neuralese". Its only human-facing element is that latent states can sometimes be decoded into tokens (Figure 9, p. 10). So it cannot be cited for any claim about oversight. What it does is show that a model can be trained to reason through a channel a person cannot read, at small scale (GPT-2) and on synthetic and arithmetic tasks. + +That matters for the Nadella post and for ruling D4. The post's P9 uses "neuralese" without a source. Korbak et al. cite this paper, with Geiping et al. (2025), as an architecture by which chain-of-thought monitoring could be lost. Where ruling D4 has the author read the post as an argument against neuralese, this is the technical paper behind the word, and it supports only that such models are being built and can work at small scale, not that frontier labs deploy them. On whether a text trace is a reliable thing for a person to read, see `lanham2023measuring` and `jia2026faithfulness`. + +Facet check: Facet G's summary (train models to reason in a continuous latent space) is accurate. The body adds that Coconut trails language chain of thought on GSM8k (34.1% against 42.9%), which an abstract-only reading would miss. + +## In conversation with +Cited and in the bib: `turpin2023language` (as "Turpin et al., 2024", p. 3) for the point that a model's chain of thought may not reflect the computation it actually runs. Cited by: `korbak2025chain` (pp. 5–6, as a latent-reasoning architecture that "naturally limit[s] our visibility"). It cites Fedorenko et al. (2024) and neuroimaging work on the language network (p. 1) to argue that language is "optimized for communication rather than reasoning"; none of that is in the bib. +Leads: Deng et al. (2024), arXiv:2405.14838 (internalized chain of thought); Geiping et al. (2025), arXiv:2502.05171 (recurrent-depth latent reasoning); Goyal et al. (2023) on pause tokens; Zhu et al. (2025) on the superposition account of Coconut. + +## Caution +Small models, narrow tasks and no evaluation of interpretability or oversight. Do not cite it for the performance of frontier reasoning models or for any claim that reasoning in vectors is a safety problem; that is Korbak et al.'s inference, not this paper's. Cite the arXiv version number with any quotation. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened cards/hao2024training.pdf (arXiv 2412.06769v4, 23 Aug 2026, 18 pp.). Checked the 8 in-source quotations by exact match, every Table 1 number (all three datasets, CoT, No-CoT, Coconut and the w/o-curriculum row; p. 9) and Table 5 (p. 17), the page locators 1, 3, 4, 9-11, 17-18, the Turpin and Fedorenko citations, and that the word 'neuralese' and any monitoring/oversight/safety language are absent from the source (zero hits). + +**Findings:** No issues found. The unmatched strings ('neuralese', 'naturally limit', 'our visibility') are quotations of the post and of Korbak et al., which are outside this source, as the card says. The card's statements about Korbak et al. citing this paper were not checked here (other card). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/heilmann2020reading.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/heilmann2020reading.md new file mode 100644 index 00000000..01c72993 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/heilmann2020reading.md @@ -0,0 +1,43 @@ +# Heilmann, J. (2020). Reading early New Testament manuscripts: *Scriptio continua*, "reading aids", and other characteristic features. In *Material aspects of reading in ancient and medieval cultures* (pp. 177–196). De Gruyter. + +**Identifier:** doi:10.1515/9783110639247-011 · **Read depth:** full_text (open-access author copy from the LMU Munich repository; De Gruyter pagination, pp. 177–196). I read the argument and notes on pp. 177–190 in full and the bibliography (pp. 191–196) by searching for the names the project cites, not line by line · **Source-tier:** book chapter · **Evidence basis:** direct_read · **Design and sample:** argument from papyrological and epigraphic evidence, plus a review of eye-movement studies; no new data · **Parties modeled:** person–artefact (reader and manuscript) · **Facets:** B · **Read:** 2026-10-13 + +## What it argues +Heilmann's aim is narrower than the dispute he enters. Charlesworth treats lectional signs as marks of manuscripts made for public reading aloud; Heilmann says the signs do not show that (pp. 177–178). To get there he first attacks the premise that unspaced script forced vocalised reading. He calls Saenger's book "the most elaborated defense" of that view and says its arguments "cannot carry the burden of proof" and "presuppose precisely what must be proved" (p. 178). He calls the reading-aloud consensus a "quasi-dogma" (p. 179) and replaces the aloud/silent binary with a scale from vocalised to non-vocalised reading (p. 179). + +On mechanism (pp. 180–181) he concedes that modern readers of English read unspaced text more slowly, then says the data come "exclusively" from subjects "socialized in modern 'western' scripts and untrained in reading scriptio continua" (p. 180). He cites Epelboim's training studies, then Vatri 2012: "there is no reason to assume that reading unspaced text is a particularly demanding cognitive task in itself, and Saenger's model must be rejected" (p. 181, quoting Vatri 2012, 639). He adds Thai as a nearer analogue than Saenger's Vai syllabary (p. 181). On the ancient sources he reads the passages Saenger-side scholars cite as about learners: they "should rather be interpreted within the context of the process of learning to read" (p. 181), and Greek papyri that separate words or syllables "are school exercises" (n. 28, p. 181). On Rome (pp. 182–183) he follows Wingo and the Vindolanda tablets to show that Latin writers used word dividers into the second century CE, so unspaced script "is not a characteristic of deficient writing systems" (p. 182). His conclusion: "We should be careful not to project our difficulties with reading scriptio continua onto ancient readers" (p. 183). The rest of the chapter (pp. 183–190) argues that diacritics, apostrophes, paragraphoi and nomina sacra served visual recognition and disambiguation, not performance. + +## Relation to the argument +This is the strongest published dissent from premise P1 (`ARGUMENT_MAP.md`: trained readers supplied the boundaries, with Vatri via Heilmann at p. 181 as its weakest point). It does not decide ruling D2. Heilmann does not argue that spacing *shortened the training* of readers, which is the claim D2 waits on scholars to supply. His own note 28 points the other way: the ancient complaints are about learners, which is the cost-for-novices reading that P4 and P6 need. He also concedes the slowing for untrained moderns (p. 180). So the chapter weakens "unspaced script burdened trained readers" and leaves "spacing eased learners" untouched. + +It also bears on P2/P3. The Latin evidence (interpuncts, then their decline) shows that spacing was not a one-time invention in the seventh century. The paper should not write "the scribes' move" as if no earlier practice existed. + +Facet check: facet B's reading of this chapter is accurate on every page I checked (pp. 178, 179, 180, 181, 182, 183, 190). One addition: Heilmann relies on Vatri for the Thai result and does not cite the 2009 study's sentence-level slowing, which Chantavarin and Leung (2025, `chantavarin2025effects`) report. + +## In conversation with +- `saenger1997space`, `saenger1982silent`: Heilmann's target (Saenger 1997, 1982). +- `knox1968silent`, `gavrilov1997techniques`, `burnyeat1997postscript`: the "well-argued counterproposals" he says were not accepted (n. 13, p. 179). +- `balogh1927voces`: origin of the aloud-reading consensus (p. 178–179). +- `vatri2012physiology`: the physiology argument he adopts (pp. 181–182). +- `epelboim1994reading`, `epelboim1997fillers`, `raynerpollatsek1996reading`: training study and the reply he lists in n. 20. +- `winskel2009eye`, `bai2008reading`: cross-cultural unspaced-reading studies he leans on (via Vatri, p. 181). +- `parkes1992pause`: cited as Parkes 1993, 10–11, among those who accept the standard view (n. 8, p. 178) and whom he opposes on the ancient sources (n. 27, p. 181); the bib key carries the 1992 date. +- `johnson2000toward`: Heilmann argues against Johnson's position on line length and performance (n. 82, citing Johnson 2000, 609–612); n. 85 cites Johnson 2004 and 2009, not the 2000 article. +- Leads not in the bib: Wingo 1972, *Latin punctuation in the classical age*; Oliver 1951 on late Latin scriptio continua (cited in the text, absent from Heilmann's bibliography); Cribiore 1996 on school papyri; Heilmann 2021, *Lesen in Antike und frühem Christentum*. + +## Caution +Heilmann argues a thesis about New Testament manuscripts, and the Saenger critique is a few pages in service of it. He summarises the eye-movement work through Vatri and does not run or reanalyse any of it. His inference from the Roman return to unspaced script rebuts a physiological deficiency, not Saenger's social account (a community can drop an aid for reasons of status). The note on Oliver gives no full reference. Do not cite him for the claim that spacing made no difference to anyone. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened cards/heilmann2020reading.pdf (De Gruyter chapter, OA author copy, pp. 177-196). Checked all 14 in-source quotations by exact match (including the ones the script flagged only because of nested quote marks), the page and footnote locators (pp. 178, 179, 180, 181, 182, 183; nn. 8, 13, 20, 22-28, 82, 85), the Vatri passage and its n. 25 attribution, the Thai and Vai remarks, the Wingo / Vindolanda / second-century passage, and that Oliver 1951 is not in Heilmann's bibliography. Not checked: the Chantavarin sentence (other source) and the facet-file claims. + +**Findings:** One fix: the card said Heilmann argues against Johnson's (2000) position in 'n. 82, n. 85'. Note 82 cites Johnson 2000, 609-612; note 85 cites Johnson 2004 and 2009 instead. Corrected the sentence. All other quotations, locators and the bib-key relations match. + +## Also used by facet W (2026-10-15) + +Facet W takes two points. First, Heilmann's distinction between reading "with vocalization, subvocalization, or without vocalization", the first two lying "on a scale from loud to quiet" (p. 179), replaces the aloud/silent binary the abstract's sentence assumes (S2 correction: the earlier wording was not a quotation, 2026-10-15); Hendrickson (1929, p. 193) had already described "many gradations". Second, his n. 28 (p. 181), that the ancient passages about laboured reading of unspaced text "should rather be interpreted within the context of the process of learning to read", is the reading facet W recommends for the abstract: unspaced text taxed learners, and the Insular scribes were learners of Latin. His concession that moderns untrained in *scriptio continua* read it slowly (p. 180) is what the Perea and Rayner experiments measure. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hendrickson1929ancient.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hendrickson1929ancient.md new file mode 100644 index 00000000..26068b6f --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hendrickson1929ancient.md @@ -0,0 +1,38 @@ +# Hendrickson, G. L. (1929). Ancient reading. *The Classical Journal, 25*(3), 182–196. + +**Identifier:** no DOI (JSTOR holds the journal); read from the Internet Archive serials scan `sim_classical-journal_1929-12_25_3` (OCR text layer; the journal's page numbers survive as running heads, so locators are the printed pages 182–196) · **Read depth:** full_text (all fifteen pages and the eight footnotes; Greek in the OCR is garbled and was not relied on) · **Source-tier:** peer-reviewed journal article (1929) · **Evidence basis:** direct_read · **Design and sample:** philological survey of passages from Plato to Augustine, with asides on modern reading pedagogy; no data · **Parties modeled:** person–artefact (reader and written page) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +Hendrickson, writing from Yale two years after Balogh, argues that antiquity read aloud because its theory of language made it so: "To the theorists of antiquity ... there was no doubt that a word was a sound, and that sound was necessary to convey sense" (p. 185), citing Augustine's *De magistro* (iv, 8) and *omne verbum sonat*. He opens with a comparison the abstract might use: early readers stood to the page "as most of us stand toward a page of music", which "conveyed meaning only as it was sounded and heard" (p. 184). The evidence is the familiar set. Augustine on Ambrose (*Conf.* 6.3): "for Augustine the manner of reading was only strange, not necessarily difficult or incredible" (p. 186). Acts 8:30, where Philip "heard him reading the prophet Esaias" (p. 186). Horace's *legentem aut tacitum*, which contrasts "sound (legentem) with silence (tacitum)" (p. 187). The Greek habit of using "to hear" for "to read", from Croesus and the Delphic oracle through Plato, Polybius and Plutarch (pp. 188–190), which he calls a sign of the "ear-mindedness" of antiquity against modern "eye-mindedness" (p. 191). Cydippe reading *sine murmure*, Nasica *tacitus leget*, Augustus's knights *taciti* (p. 192). + +His verdict is more hedged than Balogh's. Silent reading "was not, therefore, impossible (though the degree of silence is still open to debate); but it not only was unusual, it was accounted an imperfect and defective method of reading" (p. 192), citing Lucian's collector, "your eye outstripping your voice", and Wieland's 1789 note (pp. 192–193). He insists on "many gradations between vocal and silent reading, descending from distinct oral utterance to indistinct murmurs, to whispers, to mere lip motions, and so on through unconscious muscular movements ... to pure eye-reading" (p. 193). The conclusion dates the change: "Throughout antiquity and the Middle Ages reading aloud was the general habit of the learned as well as of the unlearned. Silent reading was unusual, but in what degree exceptional or possible the evidence as yet collected does not permit us to say. With the renewed literary activity of the Renaissance and the invention of printing the educated world went over by degrees to the now universal habit of silent reading" (pp. 193–194). He closes with Egger's *parole intérieure* and laboratory evidence of throat-muscle activity in silent readers (p. 194), and with "phenomenal devourers of books" who seem to skip inner articulation (pp. 194–195). + +The article never mentions word separation, *scriptio continua*, spacing or punctuation. + +## Relation to the argument + +- **Supports** the abstract's premise that ancient reading was normally voiced, from the second founding text of that consensus. +- **Cuts against "forced".** Hendrickson's own explanation is a theory of the word as sound and a culture of the ear, not the layout of the page; he reports Augustine finding silence strange "not necessarily difficult"; and he treats vocal and silent reading as a continuum. None of this is the physiological necessity the abstract's "forced ... to sound out" implies. +- **Cuts against the chronology and the agent.** He places the move to silent reading at the Renaissance and the press, a thousand years after the Irish scribes, and names no change in the page. Facet B's classicists (Knox 1968 onward) later rejected even his "unusual"; the abstract inherits the Balogh–Hendrickson consensus through Saenger without its own dating. +- **Useful to the author** for the music-score analogy (p. 184) and the "gradations" sentence (p. 193), both of which fit a human-factors description of a display that must be sounded to be parsed, if the author wants to keep that frame as a learner's condition rather than a universal one. + +## In conversation with + +- `balogh1927voces`: cited in n. 1 as "the most complete account of the subject that has appeared", with page references to Balogh p. 104 (n. 2, which runs over to p. 187) and p. 234 (n. 5, p. 193), where he says Balogh "generalizes somewhat rashly". +- `knox1968silent`: Knox treats this article as the more judicious earlier survey and quotes its "gradations" (per the `knox1968silent` card, n. 8). +- `busch2002lautes`: lists Hendrickson with Norden and Balogh as the older consensus (per that card). +- `heilmann2020reading`: the "quasi-dogma" Heilmann attacks is the position stated here. +- Leads not in the bib: Norden, *Die antike Kunstprosa* (1898, and the *Nachträge* in later reprints); Sudhaus (1906), "Leises und lautes Beten"; O'Brien (1921), *Silent Reading*; Egger (1881), *La parole intérieure*; Clark (1931), *Classical Journal* 26 (the Cicero passage Knox credits). + +## Caution + +A 1929 essay by a Latinist that reads as a lecture (n. 1 says it was read to the New York Latin Club in 1921 and to the New Jersey section of the Classical Association of the Middle States "in November last"); the evidence is anecdotal and the author says so ("the evidence as yet collected does not permit us to say", p. 193). It predates Knox, Gavrilov, Burnyeat, Johnson and all eye-movement work on unspaced text. The remarks on the Orient (n. 2, pp. 186–187) and on American farmers (p. 193) are period colour, not evidence. Do not cite it for word separation; it does not discuss it. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected (minor).** + +**What I checked:** Re-opened the Internet Archive OCR text (local copy dxd_foundation_local/pdfs/W/hendrickson1929.txt), whose running heads give the printed pages 182–196, and located every quotation by page. Twenty-three quotations and locators checked: "a word was a sound, and that sound was necessary to convey sense" and *omne verbum sonat* (p. 185); "as most of us stand toward a page of music" and "conveyed meaning only as it was sounded and heard" (p. 184); "only strange, not necessarily difficult or incredible" and Acts 8:30 (p. 186); *legentem*/*tacitum* (p. 187); Croesus (p. 188), Polybius and Plutarch (p. 189); "ear-mindedness"/"eye-mindedness" (p. 191); *sine murmure*, *tacitus leget*, the knights, "was not, therefore, impossible ... imperfect and defective method of reading", Lucian and Wieland (p. 192); "many gradations ... to pure eye-reading", the farmer, Wieland's 1789 date and the tentative conclusion (p. 193); "went over by degrees to the now universal habit of silent reading", Egger's *parole intérieure*, the laboratory and the "phenomenal devourers of books" (p. 194). All match. The article never mentions word separation, *scriptio continua* or spacing (confirmed by search). Crossref holds no record for the article (bibliographic query returned unrelated items), so the header's fields stay as taken from the issue scan; "no DOI" stands. + +**What I changed:** (1) The Lucian phrase is now quoted as printed ("your eye outstripping your voice") instead of with bracketed edits. (2) Hendrickson cites Balogh p. 104 once, in n. 2 (not "pp. 104–105"); the note begins on p. 186 and continues on p. 187. (3) n. 1 names the New York Latin Club (1921) and the New Jersey section "in November last"; the card's "1928" was an inference and is removed. (4) The "Orient" remark is n. 2 on pp. 186–187, not p. 194. The lead "Clark (1931)" is not from this article; it comes from the `knox1968silent` card, as the entry says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hervais2019learning.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hervais2019learning.md new file mode 100644 index 00000000..6519497e --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/hervais2019learning.md @@ -0,0 +1,38 @@ +# Hervais-Adelman, A., Kumar, U., Mishra, R. K., Tripathi, V. N., Guleria, A., Singh, J. P., Eisner, F., & Huettig, F. (2019). Learning to read recycles visual cortical networks without destruction. *Science Advances, 5*(9), eaax0262. + +**Identifier:** doi:10.1126/sciadv.aax0262 · **Read depth:** full_text (publisher's version from the Radboud repository; the article is paginated "1 of 10" to "10 of 10", with no printed page range; the supplementary figures and tables S1–S5 are not in the PDF and were not read) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** fMRI. Cross-sectional: 91 right-handed adults from two villages near Lucknow, India, scanned before any training (29 illiterate trainees, 24 illiterate controls, 38 literate controls), with literacy as a continuous score on a Devanagari word-reading test. Longitudinal: 22 trainees rescanned after a 6-month course (three sessions a week, about 2 hours each) · **Parties modeled:** other (reader's visual cortex; no artefact manipulation) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +The study tests the "neuronal recycling" claim that the visual word form area (VWFA), as it grows with literacy, takes over cortex that processed faces, houses and tools. Hervais-Adelman and colleagues show participants sentences, nonwords, false fonts, faces, houses, tools and checkerboards in Devanagari-reading adults who range from fully illiterate to highly literate (Introduction; Methods, "Participants"). + +Findings, all from the Results section: +- Responses to sentences and nonwords rise with literacy in a network that includes a fusiform peak at MNI (−42, −55, −10), a few millimetres from the peak reported by Dehaene et al. (2010) at (−44, −50, −14). +- Individual peak locations near the VWFA do not shift with literacy: the authors searched radii of 10, 20, 30 and 40 mm and found no lateral displacement, though the response at those peaks kept rising with literacy (slopes 0.030–0.035, t89 = 4.5–4.9, all P < 0.001). +- The share of left ventral-temporal voxels preferring nonwords over other categories rises with literacy (slope 0.0036, t89 = 3.023, P = 0.003); no category-selective extent for faces, houses or tools shrinks (all P > 0.17). +- Representational similarity between faces and nonwords inside the VWFA rises with literacy (slope 0.003, t89 = 2.603, P = 0.011). +- After training, sentence responses rise near the VWFA (peak at −39, −49, −10), but responses to faces, houses, tools, false fonts and checkerboards do not change over the six months. A similar post-training change is not seen for nonwords, which the authors say does not let them rule out that the sentence effect reflects words learned during training. + +The authors' summary: "acquisition of literacy in childhood recycles existing object representation mechanisms but without destructive competition" (Abstract). In the Discussion they describe "superposition without subsequent competition". + +## Relation to the argument +Facet D uses this paper only against claim 13 of the Gemini essays, that learning to read makes words "invade" cortex and displaces face recognition. The paper bears on no premise in `ARGUMENT_MAP.md`. It has no word-spacing manipulation: sentences appeared one to three words per screen, in Devanagari, so it says nothing about what spaces do to the visual system. Its use is negative. It removes a mechanism the paper might have been tempted to borrow for a cognitive-cost story about literacy, which fits the standing ruling to treat the history as analogy, not mechanism (ruling D10; the first answer in `SOURCE.md`, "let the research decide"). The reading-science facet itself concludes that the paper should treat brain-level work as background it does not rely on. + +Facet check: the facet's quotations from the abstract and from the Results ("There is no obvious decrease in the magnitude of the peak response to any category of stimulus as a function of literacy") are exact. The facet's entry lists seven authors and omits Frank Eisner; the PDF and the bib record list eight. + +## In conversation with +- `dehaene2010learning` — the "landmark study" the paper tests; they report no displacement of face or object peaks where Dehaene et al. found a small competition with faces (Introduction; Results). +- `dehaene2007cultural` — the recycling hypothesis the paper takes as its frame (Introduction, ref. 2). +- Leads not in the bib: Dehaene-Lambertz, Monzalvo & Dehaene (2018, *PLOS Biology*), the "revised cortical recycling" account the discussion rejects; Skeide et al. (2017, *Science Advances*), an earlier study by the same author group; Castro-Caldas et al. (1998), *The illiterate brain*; Huettig, Kolinsky & Lachmann (2018), the culturally co-opted brain; Moore et al. (2014), FaceFont training. + +## Caution +Adults who learned to read late, in Devanagari (an alphasyllabary), so the results may not carry to children learning an alphabet. Sex was imbalanced across literacy levels (illiterate participants were mostly women: two men among the 29 trainees); the authors discuss it as a limitation. The post-training maps use P < 0.001 uncorrected at voxel level. The null on training effects for other categories covers only six months. The paper does not measure reading speed, silent reading or the effect of word spacing, so it cannot be cited for any claim about scriptio continua. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened cards/hervais2019learning.pdf (Science Advances 5(9), eaax0262, 10 pp.). Checked all 4 in-source quotations by exact match (abstract sentence, 'superposition without subsequent competition', 'no obvious decrease ...', 'landmark study') and about 30 numbers: n = 91 / 29 (two males) / 24 / 38 / 22, the 6-month, three-times-a-week, about-2-hours programme, MNI peaks (-42, -55, -10; -44, -50, -14; -39, -49, -10), the 10-40 mm search radii and slopes 0.030-0.035 (t89 4.5-4.9), 0.0036 / 3.023 / .003, all P > 0.17, 0.003 / 2.603 / .011, uncorrected threshold, and one-to-three words per screen. Not checked: the supplement (not in the PDF) and the facet-file statements about Eisner and claim 13. + +**Findings:** One omission fixed: the card said responses to 'the other categories' do not change after training. The source says this for faces, houses, tools, false fonts and checkerboards (p. 4 of 10) and also says a similar post-training change was not seen for nonwords, so the sentence effect might reflect words learned in training. Added both. The two unmatched strings are the author's SOURCE.md and facet wording. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/houpt2014new.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/houpt2014new.md new file mode 100644 index 00000000..e78d1c91 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/houpt2014new.md @@ -0,0 +1,39 @@ +# Houpt, J. W., Townsend, J. T., & Donkin, C. (2014). A new perspective on visual word processing efficiency. *Acta Psychologica, 145*, 118–127. + +**Identifier:** doi:10.1016/j.actpsy.2013.10.013 · **Read depth:** full_text (PMC author manuscript, PMC3877166, Europe PMC XML; no journal page numbers, so locators are section names; Appendix A and B tables were not read in detail) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** two response-time experiments on four-letter strings; Experiment 1: 10 Indiana University students (8 women, 2 men; aged 19–34), five string types (words, pseudowords, unpronounceable nonwords, upside-down nonwords, Katakana), 10 sessions of 45–60 minutes each, 240 trials per block, 80 ms exposure with no mask; Experiment 2: 12 different students (10 women, 2 men), words, pseudowords and random consonant strings, upper case, with a mask, 100 ms exposure · **Parties modeled:** person–artefact (reader and letter string) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Houpt, Townsend and Donkin ask whether a word context makes letters easier to perceive when the measure is response time and not accuracy. The classic word superiority effect (Reicher, 1969; Wheeler, 1970) shows in accuracy and needs a post-stimulus mask; a matching effect in response times "has been elusive" (Introduction). The authors apply the capacity coefficient, C(t), a response-time measure that compares the time for a whole string with the time predicted if each letter were processed in parallel at its isolated speed (Introduction, "The Capacity Coefficient"; Eq. 1). C(t) above 1 is super-capacity, below 1 limited capacity. + +The Introduction retells the history: Cattell (1886) found letters held up at faster rates when they formed words, a result confounded by the way words constrain their letters; Reicher (1969) removed the confound with a two-alternative design in which both choices were words; Wheeler (1970) ruled out several alternative accounts. Pelli et al. (2003) argued that letters are detected independently, and the authors present their own data as a test of that claim. + +Results: +- Experiment 1 (no mask): raw string response times separated words and pseudowords from upside-down and Katakana strings (for example, Word against Upside-Down, F(1,9) = 50.85, p < .0025). Words did not differ reliably from pseudowords or random letters in raw times. The capacity coefficient did separate them: word capacity exceeded pseudoword capacity (z = 7.27), random letters (z = 22.9), upside-down letters (z = 36.7) and Katakana (z = 45.9), all p < .0025. Words and pseudowords were super-capacity for nearly all participants; upside-down and Katakana strings were limited capacity. +- Experiment 2 (masked, capitals): the same order held. Words and pseudowords were super-capacity for nearly all 12 participants. For random letters, half the participants had accuracy too low for C(t) to be computed. Word and pseudoword capacity did not differ reliably (t(10) = 0.773, p = .458). +- General Discussion: the data reject an independent, unlimited-capacity parallel model, "the models presented in Massaro (1973) and Pelli et al. (2003) are not consistent with the findings". The authors list facilitation between letters, a phonological route, and coactive pooling as candidate accounts and do not choose among them. They add that perceptual learning probably contributes, citing Blaha (2010), where about a week of training moved squiggle-stimulus capacity from very limited to high. + +## Relation to the argument +The paper bears on facet D only through the history it retells. The project cites it "only for its account of Cattell," and that account is accurate: the Introduction credits Cattell (1886) with the first demonstration that letters in words survive faster presentation than letters in random strings, and names the confound. It gives nothing about word spacing, unspaced text or sentence reading, and no premise in ARGUMENT_MAP.md or ruling D1 to D10 depends on its results. + +Facet check: the facet's "read in part" status is consistent with what the paper contains. The one use is the Cattell sentence; the experiments are about single four-letter words in isolation, so the paper must not be cited for word-level processing in running text. It does support one narrow point for D6: word-level units are processed more efficiently than letter-by-letter, and the authors tie this to perceptual learning, not to any property of spacing. + +## In conversation with +- `cattell1886time` — the first superiority demonstration and its decision-confound (Introduction). +- `reicher1969perceptual` — the two-alternative forced choice design that removes the confound. +- `wheeler1970processes` — follow-up that rules out alternative accounts and finds slower response times for words than for letters. +- `pelli2008uncrowded` — a different Pelli paper (the 2008 uncrowded-window article); the paper in this text is Pelli, Farell & Moore (2003), "The remarkable inefficiency of word recognition," which is not in the bib. +- `jordan2024reicher` — later critique of the Reicher-Wheeler paradigm; it postdates this paper. +- Leads not in the bib: Pelli, Farell & Moore (2003, *Nature*); Townsend & Nozawa (1995); Houpt & Townsend (2012); McClelland & Rumelhart (1981); Blaha (2010) thesis. + +## Caution +Small samples (10 and 12), a student population, Courier or capital strings of four letters, no sentences or eye movements. Experiment 2's random-letter results rest on half the participants. The paper tests word versus letter efficiency, not spaces. This is the PMC author manuscript, not the typeset article, and page numbers of the journal are not given. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the PMC author-manuscript XML (PMC3877166, cards/houpt2014new.xml), stripped the tags, and matched the one in-source quotation (Massaro / Pelli sentence) exactly. Checked the numbers: 10 and 12 participants (8 F / 2 M and 10 F / 2 M, ages 19-34), 10 sessions of 45-60 min, 240 trials (after 40 practice), 80 ms and 100 ms exposures, Courier font, F(1,9) = 50.85, z = 7.27 / 22.9 / 36.7 / 45.9, p < .0025 (Bonferroni), t(10) = 0.773, p = .458, half of the Experiment 2 participants with too-low random-letter accuracy, and the Blaha (2010) one-week training remark. Also confirmed the Introduction's accounts of Cattell (1886), Reicher (1969), Wheeler (1970) and Pelli et al. (2003). Not checked: Appendix A and B tables. + +**Findings:** No issues found. The two unmatched strings are the facet file's wording. The bib relation that pelli2008uncrowded is a different Pelli paper is consistent with the reference list (Pelli, Farell & Moore 2003 is the one cited). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ivanova2026process.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ivanova2026process.md new file mode 100644 index 00000000..8669ba03 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ivanova2026process.md @@ -0,0 +1,36 @@ +# Ivanova, Y. (2026). A process of implementing the Korean alphabet on the Korean peninsula [Bulgarian title: Процес на утвърждаване на корейската писменост на Корейския полуостров]. *Ezikov Svyat / Orbis Linguarum, 24*(1), 136–152. + +**Identifier:** doi:10.37708/ezs.swu.bg.v24i1.9 · **Read depth:** full_text (open-access publisher PDF, 17 pages, journal pagination; the body is in Bulgarian with an English abstract and keywords, and I read it in Bulgarian. All translations below are mine) · **Source-tier:** peer-reviewed journal article, historical survey · **Evidence basis:** direct_read · **Design and sample:** narrative history from secondary Korean-language and English sources (Kim-Renaud 1997, Kim Tongeon 2021, encyclopedia entries); no primary data · **Parties modeled:** other (state, scholars' societies, colonial authority) · **Facets:** C, F · **Read:** 2026-10-13 + +## What it argues +Ivanova traces how Hangul went from a royal invention (1443) to the national script. The abstract says that despite its design the script "faced strong resistance from the Confucian elite" and "for centuries the Korean alphabet was marginalized and associated with women, children, and the lower social classes" (p. 136). The body names the sources of the stigma: later rulers called it *onmun* ("plain script", my translation) and used other names tied to women, the middle class and children (p. 137). Hangul stayed mostly in fiction and, apart from one 1447 poem, was not used officially until the late nineteenth century (p. 138). The turn came with the 1894 Kabo reforms, which she says ordered official documents in Hangul, and with nationalism after the opening of the country (pp. 138–139). Under Japanese rule the Korean Language Society (from 1921; renamed 1931) worked on a standard spelling, a dictionary and a loanword code; Japan broke the society in the 1942 incident, and 31 members were arrested (pp. 143–149). + +Two passages bear on word separation. First, she says the new script was to be written in syllable blocks, top to bottom and right to left, "as in China, without division between words" (my translation of "без разделение между думите", p. 137). Second, the Unified Orthography of 1933 (the plan was promulgated 29 October 1933) set three principles; the third is "всяка дума трябва да се пише отделно, а не слято с другите думи", that each word must be written separately and not run together with other words (p. 144). Chapter 7 of the plan covers spaces ("Интервали"), with rules that repeat that principle, write numbers in decimal form, and put spaces within compound proper names (p. 145). The plan came out of 125 meetings over three years and went to some 500 localities and public figures for comment (pp. 144–145). + +## Relation to the argument +This is the best Korean-side evidence in the set, because it gives a word-separation convention an author, a date and a committee. On Ivanova's account Hangul ran without word division from 1443, and the separate-writing rule entered the standard orthography only in 1933, about 490 years later. That fits the facet C and F point that the script alone did not carry literacy (`T_adjudication.md`: "Hangul stayed marginal for four centuries"), and it gives the paper a second, later case of a spacing convention set by a body of practitioners (compare D8, who holds the convention). The Korean case therefore does not fit a story in which spacing comes first and enables mass reading. Ivanova gives no literacy data, so the case shows the order of events and nothing about cause. + +Limits on what it shows: Ivanova does not say when Koreans first put spaces in Hangul texts, and she does not link the 1933 rule to any change in the number of readers. Her only literacy claim is the stigma and the late state adoption. The paper also cannot be used for P4 or P6 without further sources. + +Facet check: facets C and F cite only the abstract, so the statements about spacing in the body are new to the project. The facet's "marginalized" quote is accurate. + +## In conversation with +- `becker2024conquering`: Becker and Won put the start of state efforts to teach Hangul at the Gabo Reform of 1886 (p. 6 there); Ivanova gives the first joint use of Hangul and hanja in a newspaper as 1886 and the Kabo reforms as 1894 (p. 138). Check against a third source. +- `clemens2025korea`: both report elite disparagement of Hangul. +- `baek2024spread`: the same facet C group; Baek covers the spread through letters, novels and the Hangul Bible; Ivanova does not cite it. +- `baek2022korean`: the eye-tracking study of removing spaces from Korean text, cited by Chantavarin and Leung; Chantavarin and Leung call Korean "a spaced language" today. Ivanova's 1933 rule is a dated point in how it became one, though she does not say the rule introduced spacing. +- Cites (not in the bib): Kim-Renaud 1997, *The Korean alphabet*; Kim Tongeon 2021, *History of Hangul culture*; Hong Yunsook 1991; Encyclopedia of Korean Culture entries. +- Leads: Ledyard 1966; Sampson 1985 (named in facet C, not cited here). + +## Caution +A short survey in a Bulgarian journal, resting on secondary sources and Korean encyclopedia entries that she cites by title and page. Dates and counts are hers; I did not check them elsewhere. Her wording of the 1933 rule is a paraphrase of the plan's text. Do not quote the Bulgarian without a translator's check. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The 17-page publisher PDF (journal pp. 136-152), read in the Bulgarian. Two abstract quotations (p. 136), the Bulgarian phrases for "without division between words" (p. 137) and the third principle of the 1933 plan (p. 144), Chapter 7 "Интервали" (p. 145), and the numbers: 125 meetings and 29 October 1933 (p. 144), about 500 localities (p. 144), 31 arrested (p. 148), 1886 newspaper and 1894 Kabo reforms (p. 138), 1921 and 1931 society dates (p. 143), 1447 poem (p. 138). The "spaced language" phrase appears in Chantavarin and Leung's text, as cited. + +**Findings:** No issues found. The "Hangul stayed marginal for four centuries" quotation is from the project's T_adjudication.md and was not checkable against the article. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jakkal2026secure.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jakkal2026secure.md new file mode 100644 index 00000000..731d36a2 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jakkal2026secure.md @@ -0,0 +1,25 @@ +# Jakkal, V. (2026, March 9). Secure agentic AI for your Frontier Transformation. *Microsoft Security Blog*. https://www.microsoft.com/en-us/security/blog/2026/03/09/secure-agentic-ai-for-your-frontier-transformation/ + +**Identifier:** URL above · **Read depth:** full_text (saved web page, labelled "6 min read"; the year comes from the text's May 1, 2026 availability date, since the byline shows only "March 9"; locators are section headings) · **Source-tier:** vendor product announcement (corporate blog) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** organisation's IT, security and business teams and their agents · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Jakkal, Corporate Vice President of Microsoft Security, announces Microsoft Agent 365 and Microsoft 365 E7 ("The Frontier Suite"), generally available May 1, 2026. Customers ask "How do I track and monitor all these agents? How do I know what they are doing?" (opening). Agent 365 is "the control plane for agents": a registry (an inventory of agents), behaviour and performance reports, and "agent risk signals across Microsoft Defender, Entra, and Purview" that help security teams "block agent actions based on agent compromise, sign-in anomalies, and risky data interactions" (section "Agent 365"). Entra "Agent ID gives each agent a unique identity", and Identity Governance "limit[s] agent access to only resources they need" (section "Secure and govern agent access"). Purview "Insider Risk Management extends insider risk protection to agents", with risky interactions "blocked and flagged to data security admins"; "Audit and eDiscovery" treats agents "as auditable entities alongside users and applications"; Communication Compliance detects "and enable[s] human oversight of risky AI communications" (section "Prevent data oversharing"). Prices: Agent 365 $15 per user per month; E7 $99 per user per month. + +## Relation to the argument +Premise P11, ruling D5. The page is the product line behind the practices the post names in P8 (identity, least privilege, logs, containment) and the "insider risk" vocabulary of P7. It predates the post by seven months, so the insider framing was Microsoft's product language before Nadella's essay. Every human role it names is an administrator: IT, security, business teams, data security admins. Nothing in it concerns the people an agent acts toward. It describes no approval step for agent actions; it describes blocking on risk signals. + +## In conversation with +Context for `nadella2026models`. No citations in the text. + +## Caution +Facet check: the bib note says it was read via a WebFetch summary; the full page text was read here. The facet's "blocking after compromise signals is automated" is not stated: the page says security teams use the signals to "block agent actions" and does not say whether a person or the system blocks. A marketing page; no evidence of effect is offered, and "double agents" is its risk metaphor. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The saved HTML page (text extracted). All 14 quotations, the section headings, the prices ($15 and $99 per user per month), the May 1, 2026 date, the "6 min read" and "March 9" byline strings, and the "double agents" phrase. + +**Findings:** No factual issues. I fixed one typo ("signals,." became "signals."). The page text shows "March 9" with no year and gives May 1, 2026 for availability, as the card says. The "seven months before the Nadella post" comparison depends on a document I did not open. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jhaver2019suspect.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jhaver2019suspect.md new file mode 100644 index 00000000..6da80330 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jhaver2019suspect.md @@ -0,0 +1,42 @@ +# Jhaver, S., Appling, D. S., Gilbert, E., & Bruckman, A. (2019). "Did you suspect the post would be removed?": Understanding user reactions to content removals on Reddit. *Proceedings of the ACM on Human-Computer Interaction, 3*(CSCW), Article 192. + +**Identifier:** doi:10.1145/3359294 · **Read depth:** full_text (published PDF, 33 pp.; journal pagination 192:1–192:33). I read the body and the reference list; of Appendix A (the 24-item questionnaire) I read only items 15–24 · **Source-tier:** peer-reviewed proceedings article · **Evidence basis:** direct_read · **Design and sample:** one-shot survey of Reddit users whose posts had just been removed. The authors drew 10,000 recent submissions from randomly chosen subreddits, waited three hours, found the removed ones and messaged the authors; 1,054 people submitted and 907 remained after cleaning (8.2% of those contacted) (192:9–10). 61% US, 81% male, 55% under 25 (192:10). Linear regressions on two five-point items, plus coding of 849 open-ended answers (192:13, 19) · **Parties modeled:** two parties plus a third party (poster, volunteer moderators who hold the rule, and the platform that supplies the tools) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +Jhaver and colleagues ask how people react hours after a moderator removes their post, and whether community rules and removal explanations change that. Most react badly: 10.3% strongly agree and 13.6% agree the removal was fair, against 33% and 26.9% who disagree (192:13). The abstract (192:1) says 37% "did not understand why their post was removed"; the body's figure is 36.9% of the 849 open-ended answers coded "lack of clarity about why post was removed" (Table 7, 192:19). + +Silence is the Reddit default. The site does not tell the author of a removal (192:4), and 41.8% (n = 379) did not notice until the survey arrived (192:15). Noticing goes with a fairer rating (β = .265). People who received an explanation (39.7%, n = 360) judged the removal fairer (β = .384, p < .001) and were somewhat more likely to say they would post again (β = .088, p < .05) (192:17). Only whether the explanation gave new information mattered for fairness (β = .200); mode (comment, message, flair) and source (human or bot) did not (192:18). Rule clarity helped on both outcomes (β = .170 and .223) (192:17). + +One result runs against the abstract. Readers of the rules rated removals less fair (β = -.314), which the authors call "surprising" (192:16–17) and explain by rules that are subjective or too many to follow (192:22–23). The Discussion says "when moderated users read the rules before posting, they are less likely to consider their post removal as fair" (192:23). The abstract says the opposite: "users who read community guidelines or receive explanations for removal are more likely to perceive the removal as fair" and post again. + +Fit stays low. Controls alone explain under 1% of variance; the full model reaches adjusted R² of .324 for fairness and .123 for the intention to post again, and a model with explanations alone leaves the second at .000 (192:16, 18). + +## Relation to the argument + +Q4 in facet J and, indirectly, ruling D3 (disclosure as a literacy-based intervention). The rule-holder here is a volunteer neighbour, not an interested owner. The paper supports the claim that reasons change how people judge a decision. It is thin on what they then do: the behavioural outcome is a stated intention, the association is small, and the authors take actual posting up elsewhere "with a larger dataset" (192:11, ref. 45). That second paper bears on D3's claim that disclosure does not affect our ability to navigate the system; this one does not settle it. + +It also bears on D8. Reddit's local moderators hold the convention, and the site gives them tools but by default does not tell the poster (192:4). The text describes no appeal. With no notice, many respondents build folk theories of moderator bias (192:20–21). + +Facet check: facet J quoted the abstract only. The abstract's "read community guidelines" clause is not what Table 5 shows, and its 37% rests on 849 coded answers, not 907 respondents. + +## In conversation with + +- `jhaver2019transparency` (ref. 45): the paper prints it as 2018, *PACM HCI* vol. 2; the bib has 2019, vol. 3, doi:10.1145/3359252. The behavioural follow-up on removal explanations. +- `eslami2016first`, `myerswest2018censored`: folk theories and users' readings of moderation (192:5–7). +- `suzor2018digital` (ref. 80) and `gillespie2019custodians` (ref. 34, printed 2018): frameworks and background (192:6). +- Not in the bib: Fiesler et al. (2018), *Reddit Rules!*; Kiesler, Kraut and Resnick (2012); Jhaver et al. (2019), the Automoderator study, *TOCHI* 26(5); Suzor, Myers West, Quodling and York (2019); Kizilcec (2016). + +## Caution + +Self-report within hours, on one platform, from people who answered a private message (8.2%), mostly young and male. The authors did not judge whether removals were legitimate (192:2, n. 2). Outcomes are single Likert items regressed linearly; "post again" records intention; subreddit is uncontrolled. They call the analysis "primarily exploratory" (192:12). Not evidence on algorithms or on the paper's triad. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the CSCW published PDF (34 text pages; PDF page n = article page 192:n). Searched all 8 quotations and found each; checked about 30 numbers and locators against pp. 1, 4, 9-19, 23 (10.3/13.6/33/26.9% fairness split, 36.9% vs the abstract's 37%, 41.8% n=379, 39.7% n=360, betas .265, .384, .088, .200, .170, .223, -.314, adjusted R2 .006 baseline, .324 and .123 full model, .000 with explanations, 8.2% response, 1,054 / 907 / 849, 61/81/55%, 10,000 submissions, three hours). + +**Findings:** No issues found. The Read depth line honestly says Appendix A was read only in part (items 15-24); I judged that small enough to keep full_text. One nuance, not an error: the '.000' adjusted R2 is for baseline plus removal explanations, which the card words as 'explanations alone'. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jia2026faithfulness.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jia2026faithfulness.md new file mode 100644 index 00000000..608af19c --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jia2026faithfulness.md @@ -0,0 +1,36 @@ +# Jia, J., Benton, J., & Easley, E. (2026). *Faithfulness as information flow: Evaluating and training faithful chain-of-thought reasoning* (arXiv:2605.24286). arXiv preprint. + +**Identifier:** arXiv:2605.24286 · **Read depth:** full_text (arXiv v1, 22 May 2026, 18 pp.; pagination is the PDF's. Body, pp. 1–9, limitations and broader impact, p. 17, read in full; appendices B–C, per-model hyperparameters, extra plots and code-repair examples, skimmed) · **Source-tier:** preprint (an Anthropic Fellows Program project, with authors from Michigan State and Anthropic) · **Evidence basis:** direct_read · **Design and sample:** a metric validation on two open models, then reinforcement-learning runs on three small models (Gemma 3-1B, 4B and 12B; Qwen 2.5-7B); no human participants · **Parties modeled:** other: a developer training a model, and the model's reasoning trace; no human reader · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Jia, Benton and Easley start from a premise the Nadella post shares: a model's chain of thought "is useful for monitoring language models only when the reasoning trace faithfully reflects the computation that produces the final answer" (abstract, p. 1). They define faithfulness as information flow. Answer-relevant information should travel prompt → chain of thought → answer, not along a shortcut from prompt to answer (§3, p. 3). Three properties follow: sufficiency (the trace determines the answer), completeness (the prompt adds nothing once the trace is known), and necessity (the answer depends on the trace). They measure these with entropy, attention-mask KL divergence and gradient shares. + +Validation (§4, p. 4): on hinted GPQA questions where a hint changed the answer, four frontier language models judged 89.4% of Qwen3-8B's traces to acknowledge the hint, against 54.3% for DeepSeek-R1-Distill-14B. The entropy and gradient metrics rank the two models the same way. The KL necessity metric separates them "in the wrong direction" (p. 5), so the authors treat gradient measures as primary. + +Training (§§5–6, pp. 5–9). They change the policy update, not the reward, with four interventions (update mask, gradient mask, chain-of-thought-only gradients, an adversarial method called FACT). In hinted arithmetic, where 75% of training hints are correct, vanilla reinforcement learning learns to follow wrong hints without mentioning them; all four interventions raise the hint-mention rate (p. 7). In code repair, every method learns to hard-code a lookup table that passes visible tests while "hidden test accuracy plateaus around 0.32"; the chain-of-thought-gradient, update-mask and FACT variants more often say so in the trace: "The interventions do not eliminate reward hacking here, but they make it monitorable from the trace" (p. 8). On held-out maths, the method does not stop a model following wrong hints, "but it keeps that influence visible in the CoT" (p. 9). + +The limits are the authors' own: the metrics are "diagnostics rather than definitive causal tests" (p. 17), the methods improve "auditability, not necessarily robustness to every shortcut" (p. 17), and "faithful CoT should not be treated as a complete safety guarantee" (p. 17). + +## Relation to the argument +It supports the post's P9 caveat that chain-of-thought transparency alone is "not sufficient or dependable", and shows that at the time of the preprint the problem is open: even with a trained intervention, hint-following continues while only its visibility improves. No human reads a trace here; faithfulness is scored by language-model judges and by gradients. + +On ruling D4: this work is neither disclosure design nor a control placed outside the model. It changes how the model is trained so that a later reader or monitor can rely on the trace. It also puts the choice with the developer. That sits uneasily with D4's two-way split (literacy-based against algorithmacy-based intervention): the split does not classify a move whose aim is to make a disclosure trustworthy. + +Facet check: Facet G read this as abstract only and describes the 2026 preprints as a group (monitoring "helps, depends on conditions, and degrades"). The full text agrees with that gloss for this paper, with one addition: the gains are in visibility of the shortcut, not in its prevention. + +## In conversation with +Cites and builds on `turpin2023language` (hinted prompts, ref. 4, p. 6), `lanham2023measuring` (ref. 5, grouped with the task-specific perturbation tests it aims to replace, p. 2), `greenblatt2024ai` (ref. 28, with `baker2025monitoring`, ref. 29, for monitoring that exposes reward hacking, p. 2), and `wang2026monitorbench` (ref. 13, a monitorability benchmark). `korbak2025chain` is not cited; Chen Yueh-Han et al. (ref. 20), whose last author is Korbak, is. +Leads: Hase & Potts (2026), arXiv:2602.20710 (counterfactual simulation training); Anwar et al. (2026), arXiv:2602.18297 (monitorability by information theory); Turpin et al. (2025), arXiv:2506.22777 (teaching models to verbalize reward hacking). + +## Caution +Small open models, three narrow settings, and "faithfulness" scored by other models; the authors say larger models and subtler shortcuts are untested (p. 17). No peer review at the time of reading. The validation sample size for the GPQA comparison is not in the body I read. Do not cite it as evidence about frontier models or about what people can detect. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** arXiv:2605.24286v1 PDF (18 pages, v1 dated 22 May 2026; title, authors and affiliations match). All eight verbatim quotations located by page (pp. 1, 5, 8, 9, 17), plus the numbers: 89.4% vs 54.3% on hinted GPQA judged by four frontier LLMs (p. 4), 75% correct hints (p. 7), hidden-test accuracy about 0.32 (p. 8), models and sizes (p. 6), reference numbers 4, 5, 13, 20, 28, 29 and the three leads' arXiv IDs. + +**Findings:** No issues found. The "monitorable from the trace" sentence sits beside the Figure 8 caption on p. 8; the words are exact. Two quoted phrases attributed to the post and facet ("not sufficient or dependable"; "helps, depends on conditions, and degrades") are not from this paper and were not checkable here. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jordan2024reicher.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jordan2024reicher.md new file mode 100644 index 00000000..3293f805 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/jordan2024reicher.md @@ -0,0 +1,32 @@ +# Jordan, T. R., Akkaya, A. M., Göçmüş, F. Z., Kalan, A., Morgul, E., Önalan, K., & Sheen, M. K. (2024). The Reicher-Wheeler paradigm in word recognition research: a cautionary note on its actual contributions and published misconceptions. *Frontiers in Psychology, 15*, 1399237. + +**Identifier:** doi:10.3389/fpsyg.2024.1399237 · **Read depth:** full_text (open-access published PDF, five pages; Frontiers pagination 01–05, so locators below are section names) · **Source-tier:** peer-reviewed mini review (position piece) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** person–artefact (reader and briefly shown stimulus) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Jordan and colleagues explain what the Reicher-Wheeler paradigm (R-WP) does and then list recent studies that claim to use it but do not. The setting is the word superiority effect: under brief display, "words are easier to identify than nonwords, and even single letters, and reports of this phenomenon date from some of the earliest experiments in the history of psychology (Cattell, 1886; Erdmann and Dodge, 1898; see also Huey, 1908)" (Introduction). Those early results were confounded, because participants could use "non-perceptual guesswork" that helps words and not nonwords (Introduction). + +The R-WP fixes this with a forced choice. After a brief display, two letters appear, "both letters are consistent with the rest of the stimulus", so that "selecting the correct letter cannot be assisted non-perceptually" (section "Actual contributions of the Reicher-Wheeler paradigm"). The paradigm "became known as the Reicher-Wheeler Paradigm (hereafter, R-WP) after it was first reported by Reicher (1969) and Wheeler (1970)." They cite Johnston (1978) for the result that free report, without forced choice, produced larger word advantages than the R-WP did, which is what guessing would predict. + +The authors then name studies that departed from the design: Marzouki et al. (2022) used a bar-marker probe on Arabic words; Mok (2009) used an underline probe with written reports; Chen et al. (2018) used different probe characters for word and nonword conditions, so probe identity is confounded with stimulus type. They name further cases with mismatched forced-choice alternatives (Casaponsa & Duñabeitia, 2016; Laszlo & Federmeier, 2007; Chen et al., 2017; Matić et al., 2018) (section "Misconceptions of the Reicher-Wheeler paradigm in the literature"). They say the findings of these studies "remain to be determined under true R-WP conditions" (Discussion). They offer no data of their own. + +## Relation to the argument +It plays no part in the ten rulings or in premises P1 to P12. Its single use, in facet D, is as a correction to the Gemini essays (claims 1 and 2 of the facet's table): Cattell (1886) made the early observation, the controlled effect belongs to Reicher (1969) and Wheeler (1970), and the measure is letter identification under brief display, not reading speed. The text says nothing about case, lowercase against capitals, spacing, silent reading or scriptio continua; the facet's note on the "lowercase" claim rests on other sources. If the paper mentions the word superiority effect at all, this card is the place to take the attribution from. It supports no step in the word-separation argument. + +Facet check: the facet cites seven authors with initials "Akkaya, S., Göçmüş, S., Kalan, Ö., Morgul, M. E., Önalan, M., & Sheen, M." The PDF and the bib record give Akkaya A. M., Göçmüş F. Z., Kalan A., Morgul E., Önalan K., Sheen M. K. The facet's quotations match the text. + +## In conversation with +- `reicher1969perceptual` and `wheeler1970processes` — the two sources of the paradigm (Actual contributions section). +- Leads not in the bib: Cattell (1886); Johnston (1978) on the sophisticated-guessing test; Marzouki et al. (2022); Mok (2009); Chen et al. (2018). + +## Caution +A short corrective review by researchers who have used the paradigm themselves (Jordan is a co-author of several of the R-WP studies it cites as exemplary). It does not re-analyse any of the criticised studies. Its claim that those studies' results are "unclear" is a methodological argument, not a replication. Nothing in it concerns reading of connected text. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The 5-page Frontiers PDF (title, authors, DOI, dates match). Seven quotations located; the studies named (Marzouki 2022, Mok 2009, Chen 2018, Casaponsa 2016, Laszlo 2007, Chen 2017, Matić 2018) and their described departures; Johnston (1978) on free report; author names and initials against the byline. + +**Findings:** One misquotation fixed: the card had "both consistent with the rest of the stimulus"; the text reads "both letters are consistent with the rest of the stimulus" (p. 2 of the PDF). All other quotations match. The section-name locators are accurate. Note the paper cites Cattell as "1885, 1886" in its first sentence, but gives 1886 for the word-superiority observation, as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kasisopa2016child.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kasisopa2016child.md new file mode 100644 index 00000000..b648c7f6 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kasisopa2016child.md @@ -0,0 +1,39 @@ +# Kasisopa, B., Reilly, R. G., Luksaneeyanawin, S., & Burnham, D. (2016). Child readers' eye movements in reading Thai. *Vision Research, 123*, 8–19. https://doi.org/10.1016/j.visres.2015.07.009 + +**Identifier:** doi:10.1016/j.visres.2015.07.009 · **Read depth:** full_text (publisher-formatted PDF from the Maynooth University repository MURAL, eprint 10442; journal pagination 8–19; read whole, including Tables 1–7) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** eye-tracking (EyeLink II, 500 Hz) of 18 Thai children reading Thai sentences silently with and without inserted word spaces: 10 younger (start of Grade 6, mean 11.4 years; 20 recruited, 15 completed, 10 usable) and 8 older (end of Grade 7, mean 12.7 years), Bangkok schools; 54 test sentences per child; linear mixed models with first- and last-character position-specific frequency and word frequency as covariates; comparison with the adults of Kasisopa et al. (2013), who had read the Set 2 materials (pp. 11–12; the adult N is not restated in this paper) · **Parties modeled:** person–artefact (child reader and page) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +Thai is "an alphabetic scriptio continua language" (abstract). Thai children learn to read "with inter-word spaces and continue with spaced text until almost the end of Grade 1", when "lessons change to the use of unspaced text" (p. 9; Tables 2A–2B). Adults, the authors had shown in 2013, land their eyes just left of word centre by using the frequency of characters that occur at word beginnings and ends; ten characters account for just over 50% of word-initial occurrences and five for about 50% of word-final ones (p. 9, Table 3). The question is whether children do the same. + +They do, with help. Mean landing site was 3.5 pixels left of word centre for younger and 2.55 for older children (p. 13); landing moved toward the centre with higher first-character frequency, "an advantage that was particularly pronounced for spaced text" and for high-frequency words (p. 13). On first-fixation duration, "Younger children only show an effect of first character frequency when reading spaced text" (p. 13); the single-fixation effect likewise held only where the first characters "can be clearly perceived – in spaced text" (p. 15). Last character frequency "appears to be a later development, and again its use is scaffolded by spacing and word frequency" (p. 18). Gaze durations for younger children ran, by condition, from 308 ms (spaced, high/high frequency) to 772 ms (spaced, low/low), against 353 and 429 ms in unspaced text (Table 5, p. 14), so spacing helped most where the boundary characters were common and hurt where they were rare. Comprehension was 93% (older) and 88% (younger) (p. 12). The authors' reading: "children use first character frequency early in reading Thai and this is scaffolded by spacing and word frequency" (p. 18); "it appears that eye movements in reading converge on the optimal landing site using whatever cues are available to guide such placement" (abstract). "it appears that spaces potentiate children's in learning to use boundary character frequency effects to aid word segmentation" (p. 17; the sentence is printed thus). + +## Relation to the argument + +The abstract's sentence turns on a scaffold: unspaced text "scaffolding literacy with oracy", and separation moving "boundary detection from the voice to the page". This study is the nearest living experiment on a scaffold for boundary detection, and it places the scaffold elsewhere. + +- **Supports** the idea that a spaced display scaffolds a learner's segmentation: Thai schooling spaces Grade 1 text and withdraws the spaces, and the children's data show spaces potentiating the pick-up of visual boundary cues. +- **Cuts against "from the voice".** What the spaces scaffold is the eye's use of orthographic regularities; the children read silently, at 88–93% comprehension, and their boundary detection in unspaced text is done by character statistics, not by sounding the line. The route the abstract describes (voice first, then page) is not the route these learners take (page first, then page without spaces). +- **Narrows** "freed literacy": here the display that removes the segmentation task is a stage of instruction, after which readers are expected to do without it. That matches facet C's finding that spacing belongs to pedagogy and cost, not to the spread of reading. +- For the human-factors frame (ruling D6), the "whatever cues are available" conclusion is the same lesson as `perea2015colours`: the boundary is a representational job that several marks can do. + +## In conversation with + +- `kasisopa2013eye`: the adult study this replicates with children; its Table 7 is reproduced here (p. 18). +- `winskel2009eye`: the other Thai spacing study, cited for the adult cost of added spaces. +- `bai2008reading`, `blythe2012inserting`: the Chinese child and adult spacing studies the design is modelled on (p. 11). +- `seymour2003foundation`: cited for orthographic depth and the rate of learning (p. 8). +- `butskhrikidze2022what`: not cited; her claim that Thai phonology supplies boundary markers is the language-side version of the boundary-character finding here. +- Leads not in the bib: Reilly, Radach, Corbic & Luksaneeyanawin (2005); Feng, Miller, Shu & Zhang (2009) on Chinese and English child readers; Krügel & Engbert (2014) Bayesian saccade targeting. + +## Caution + +Eighteen children in two narrow age bands at the top of primary and start of secondary school; five of fifteen younger children were lost to tracking error; many of the reported effects are three- and four-way interactions on this sample and should not be quoted as settled. No oral-reading condition and no Grade 1 or 2 children, so the paper does not observe the transition from spaced to unspaced text it describes; a longitudinal study was "in progress" (p. 18). Thai words are short and the script has no capitals or punctuation, so transfer to seventh-century Latin is by analogy only. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected.** + +**What I checked:** Re-opened the MURAL publisher-layout PDF text (local copy dxd_foundation_local/pdfs/W/kasisopa2016child.txt; running heads mark pp. 9–19). Confirmed: "an alphabetic scriptio continua language" and "whatever cues are available" (abstract); spaced text "until almost the end of Grade 1" and the switch to unspaced text, 10 first-position characters accounting for "just over 50%" and five last-position characters for "roughly the same percentage" (p. 9); 8 older children (mean 12.7 years), 20 younger recruited, 15 completing, 10 usable (mean 11.4 years), 54 test trials (p. 11); comprehension 93% and 88% (p. 12); landing 3.5 and 2.55 pixels left of centre and "particularly pronounced for spaced text" (p. 13); Table 5 gaze durations 308 / 772 / 353 / 429 ms (p. 14); Table 7 (p. 18); "children use first character frequency early in reading Thai and this is scaffolded by spacing and word frequency" and the longitudinal study "now in progress" (p. 18). Crossref 10.1016/j.visres.2015.07.009: four authors, *Vision Research* 123, 8–19, 2016; matches. Read depth full_text stands. + +**What I changed:** (1) The "30 adults" figure is not in this paper and was removed. (2) The two quotations about younger children and spaced text were paraphrases in quotation marks; replaced with the printed sentences and their pages (pp. 13 and 15). (3) The last-character claim now quotes p. 18 as printed; the paper does not say the effect appeared "only in older children" in those words. (4) The "potentiate" sentence is quoted as printed, with its grammatical slip noted, instead of with a bracketed repair. Relation to the argument is unchanged; the card's reading (spaces scaffold visual cue use; the children read silently) is what the paper says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kellogg2020algorithms.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kellogg2020algorithms.md new file mode 100644 index 00000000..507642a7 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kellogg2020algorithms.md @@ -0,0 +1,52 @@ +# Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. *Academy of Management Annals, 14*(1), 366–410. https://doi.org/10.5465/annals.2018.0174 + +**Identifier:** doi:10.5465/annals.2018.0174 · **Read depth:** full_text (the published version as deposited in the NSF Public Access Repository, par.nsf.gov/servlets/purl/10195395, 45 pp.; pagination is the journal's, pdf p. 1 = p. 366; I read the whole body, Tables 1–4 and Figures 1–2; I did not read the reference list line by line, only searched it) · **Source-tier:** peer-reviewed article (integrative review) · **Evidence basis:** direct_read · **Design and sample:** narrative review of interdisciplinary research on algorithms at work, organised by Edwards' (1979) labour-process typology; the article gives no search protocol or count of reviewed works · **Parties modeled:** two parties + third party (employer and worker, with customers, raters and platforms; the algorithm as the employer's instrument) · **Facets:** J, M, O · **Read:** 2026-10-14 + +## What it argues + +Kellogg, Valentine and Christin read the research on algorithms at work through labour process theory, in which control is "contested terrain" between employers and workers (p. 366). They define algorithmic technologies as procedures that are "more encompassing, instantaneous, interactive, and opaque" than earlier workplace technologies (p. 366) and name four affordances: comprehensive, instantaneous, interactive, opaque (Table 1, p. 371). Employers, they find, control workers through "6 Rs": they direct by restricting and recommending, evaluate by recording and rating, and discipline by replacing and rewarding (abstract, p. 366; Tables 2–4, pp. 373, 377, 380). + +Several findings bear on this paper. They describe opacity as partly deliberate: "information asymmetries are not random: instead, they are deliberately created by employers to constrain workers' choices" (p. 384). They cite Lee et al. (2015) for surge pricing that was "often erratic and unreliable" (p. 374) and for platforms that "narrow shift choices, ride choices, or delivery choices" (p. 375). They say workers "often do not fully grasp how algorithms are being used to direct, evaluate, and discipline them" and that "there is frequently no procedure in place for workers to get access to, contest, or challenge algorithmic decisions" (p. 387). Individualised, real-time control, they argue, makes collective resistance harder: "Whereas previous systems of control allowed collectives of workers to organize and share resistance tactics over time … algorithmic control can make such initiatives and contestations harder to achieve" (p. 386). + +The last section names worker resistance "algoactivism" (pp. 391–395). Individual tactics include non-cooperation, "leveraging algorithms" and negotiating with clients (p. 391). Workers "reverse engineer the algorithm that produced the rating" (p. 391); Airbnb hosts "participated in online forums, read the company's technical documentation, and monitored competitors' profiles" to work out what drives ratings (p. 391); Lee et al.'s drivers logged off and avoided areas (p. 391). Collectively, workers organise "using online forums and platforms" in which they "help each other learn new systems and practices, anticipate or avoid disciplinary processes" (p. 392). The authors add that forums are one route among platform cooperatives, discursive framing about fairness and legal mobilisation (pp. 392–395). + +## Relation to the argument + +**Ruling D7 and premise P9 (learning by deciding).** The review supports the idea that workers learn a mediator from what it does, not from documentation, but only in part. Its examples of learning mix the two. The Airbnb hosts read "the company's technical documentation" *as well as* forums and competitors' profiles (p. 391). So the review does not license the claim that workers learn *without* the company's account. It does license the weaker claim that workers infer the rules by acting on the system and pooling what they observe. + +**Ruling D8 and premise P12 (who holds the convention).** This is the review's strongest use for the paper. Opacity is described as a choice the employer makes (p. 384), and the employer can recalibrate "architecture, nudges, and sanctions" while claiming full automation (p. 384). That fits P12: the owner of the interface is an interested party and can change a convention alone. The due-process passage (p. 387) gives the paper an older vocabulary (notice and a hearing) for what the weaker party lacks. + +**What it cuts.** The review is about employer–worker control, a dyad with an instrument. Customers appear as raters, not as a second party whose moves the worker must predict. It does not support the paper's triadic claim (D9) and should not be cited for it. It also reports variation: some drivers experienced autonomy (Cameron 2018, p. 385), and some workers see the platform as an ally (p. 385). + +**Use in DRAFT_v2:** Line 266: "Ride-hail drivers do not learn dispatch and surge behavior from company documentation. They form working theories from what the system does in response to their own moves, and they trade those theories in online forums (Lee et al., 2015; Kellogg, Valentine, & Christin, 2020)." Partly supported. Kellogg et al. report reverse engineering, practical workarounds and forum knowledge-sharing (pp. 391–392), and they cite Lee et al. for drivers' tactics (p. 391). They are a secondary review here: they do not observe ride-hail drivers themselves. Two cautions: (1) the one ride-hail forum example they give is a blog, "The Rideshare Guy" (p. 392), and their Airbnb example includes reading the company's documentation (p. 391), so "do not learn … from company documentation" goes further than this source; (2) they say nothing specific about drivers forming theories of *dispatch*. Citing Kellogg et al. for the general pattern (workers infer and pool) is safe; for the ride-hail specifics, Lee et al. (2015) is the primary source. + +## In conversation with + +- `lee2015working` — cited on at least six text pages: surge erratic (p. 374), choices narrowed (p. 375), deactivation (p. 381), rapid wage changes (p. 382), the hidden human hand in "fully automated" systems (p. 384) and drivers' workarounds (p. 391), and listed in Tables 2 and 4 (pp. 373, 380). +- `rosenblat2016algorithmic` — the review's main source on Uber's information asymmetries, nudges and deactivation (pp. 374, 377, 381). +- `irani2013turkopticon` — the Turkopticon platform is named among worker-built tools (p. 392); the review cites it via Martin et al. and others. +- `jhaver2019suspect` — not the same work: the review cites Jhaver, Karpfen & Antin (2018) on Airbnb hosts (p. 391), a different paper by the same first author. +- `rahman2019invisible` / `rahman2021invisible` — the review cites Rahman (2019, working paper) on Upwork's opaque rating change (p. 382) and the "algorithmic cage" (p. 388). +- Not in the bib, as leads: Edwards (1979) *Contested Terrain*; Burrell (2016) on three kinds of opacity (p. 372); Gray & Suri (2019) *Ghost Work* on "algorithmic cruelty" (p. 388); Christin (2017) on foot-dragging and gaming (p. 391); Jhaver, Karpfen & Antin (2018) on Airbnb hosts. + +## Caution + +A narrative review without a stated search method, so it cannot be cited for how common any tactic is. Much of its evidence is secondary and pre-2019. It treats control between employer and worker; its "platform" examples include crowd work and Airbnb, not only ride-hail. The bib entry spells the third author "Angéle"; the article prints "Angèle". OpenAlex dates the work 2019 (online first); the volume year is 2020, as the draft has it. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** corrected + +**What I checked:** I re-opened the NSF PAR copy saved at dxd_foundation_local/pdfs/R4b/kellogg2020.pdf (45 pp., journal pagination, pdf p. 1 = p. 366) and re-extracted it page by page, because the saved two-column text file interleaves columns. I checked all 13 quotations word for word and located each by page; checked Tables 1–4 pages, every page locator in "In conversation with" (Rosenblat & Stark, Turkopticon, Jhaver, Rahman, "algorithmic cage", Burrell, "algorithmic cruelty", Christin 2017, Cameron 2018/2019); checked the reference record against Crossref (authors, 14(1), 366–410, published January 2020; Crossref also spells "Angéle"); and re-read DRAFT_v2.md line 266. + +**Findings:** +1. All 13 quotations match the text and their page locators are right (pp. 366, 374, 375, 384, 386, 387, 391, 392). +2. Corrected: the algoactivism section begins on p. 391 (heading "Algoactivism: Individual and Collective Resistance of Algorithmic Control"), not p. 390. Now "pp. 391–395". +3. Corrected: the review cites Lee et al. (2015) more than "four times": on pp. 374, 375, 381, 382 (rapid wage changes), 384 and 391, and in Tables 2 and 4. The list now says so. +4. The "Use in DRAFT_v2" verdict for line 266 (partly supported) holds. One sharpening: the p. 391 passage on Lee et al. reports drivers *resisting* control (turning off driver mode, staying in residential areas, logging off), not drivers forming theories of dispatch; the only ride-hail forum item is the blog "The Rideshare Guy" on maximising income (p. 392). The read-depth claim (full text) is consistent with the copy on file. + +## Also used by facet T (2026-10-15) + +Facet T asks which literatures model the platform as a third party with its own objective. This review is the cautionary case: it treats the algorithm as the employer's instrument in an employer–worker dyad, with customers as raters (see "What it cuts" above; the phrasing is this card's, not the review's). Set against `stark2020algorithmic` (p. 54: "this triangular geometry … [is] key") and `vallas2020what` (16.10: the platform as *tertius gaudens*), it marks the dividing line in the literature. The abstract's "irreducible algorithmic triad" sides with Stark and Pais and Vallas and Schor; a reader schooled in Kellogg, Valentine and Christin will ask why the customer is a party rather than an input, and the paper should answer with the alliance evidence (`vallas2020what` 16.7; `cruz2023digital` p. 3) and with the ride-hail cases in which the customer's decision enters the mediator's rule (`cohen2016using`, `castillo2025matching`). + +*S2 check of this section (2026-10-15):* Changed: the phrase "the employer's instrument" was in quotation marks although it is neither in the review nor in the host card (which says "a dyad with an instrument"); the marks were removed and the paraphrase marked as the card's. `stark2020algorithmic` p. 54 and `vallas2020what` 16.10 and 16.7, `cruz2023digital` p. 3 confirmed. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ker2012johnson.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ker2012johnson.md new file mode 100644 index 00000000..7b0029b1 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ker2012johnson.md @@ -0,0 +1,35 @@ +# Ker, J. (2012). Review of William A. Johnson, *Readers and Reading Culture in the High Roman Empire: A Study of Elite Communities* (Oxford University Press, 2010). *Classical Journal Online*, 2012.02.03. + +**Identifier:** https://people.duke.edu/~wj25/pubs/CJ2012review.pdf · **Read depth:** full_text (author-hosted copy of the CJ-Online review on Johnson's Duke page, 3 pages, about 940 words; no journal pagination, so locators are paragraph numbers ¶1–¶5 of the PDF) · **Source-tier:** book review (online journal) · **Evidence basis:** direct_read · **Design and sample:** none (review of a monograph) · **Parties modeled:** other (reader and reading community) · **Facets:** B, K · **Read:** 2026-10-13 + +## What it argues +Ker reports Johnson's book and endorses it. Johnson objects to "scholars' typically narrow focus on the question of whether the Greeks and Romans were technologically and/or cognitively inhibited from reading silently" (¶1). Johnson's interest in Augustine's account of Ambrose (*Conf.* 6.3.3) lies less in proving that silent reading was known earlier "than to draw attention to Augustine's expectations and the reasons for his surprise" (¶1). Ker quotes Johnson's definition: "reading is not simply the cognitive processing by the individual of the technology of writing, but rather the negotiated construction of meaning within a particular sociocultural context" (¶1, citing Johnson p. 12). + +On the point that matters here, Ker says that chapter 2 shows how "certain general characteristics of ancient reading and ancient bookrolls, such as the absence of spaces between words (scriptio continua), are not signs of a technological lack, but are consistent with an elite culture in which reading from a bookroll with full articulation and comprehension is a marker of education and status, and often an act of connoisseurship" (¶2). The case studies (Pliny, Tacitus's *Dialogus*, Galen, Gellius, Fronto, Lucian, the Oxyrhynchus papyri) show "an emphatically social reading culture" in which an elite group shares texts and excludes others (¶3). Ker adds that the model "will no doubt need to be expanded and/or refined," as Johnson concedes at p. 206 (¶4). He closes by comparing the Roman reading circle to a college humanities class (¶5). + +## Relation to the argument +Ker is a witness to what Johnson says, not to the evidence behind it. Two uses. + +- P1 and its weakest point (whether unspaced script burdened trained readers): the review reports Johnson reading *scriptio continua* as a mark of status, not of a technological deficit. This sits beside Heilmann 2020 and Vatri 2012, who deny a cognitive burden for trained readers. It does not touch the cost for learners (P6), which Johnson's elite-reader frame sets aside. +- Rulings D1, D2 and D6: Johnson gives a social cause for the persistence of unspaced script in antiquity, where Saenger gives a cognitive one. On Ker's account the page was kept hard on purpose, so the Roman case fits the paper's point that a convention serves the people who hold it (P10), though Ker says nothing of Saenger or of any medieval change. + +Facet check: facet B records this review as read in full and about 940 words; confirmed, with no pagination. Ker does not discuss training of novices, so the review gives K no purchase on Johnson's "emerging readers" sentence. + +## In conversation with +- `johnson2010readers`: the book under review. +- `johnson2000toward`: the review says nothing of it; Ker's "earlier and more specialized work" remark is in the papyri paragraph, and the text does not name the 2000 article. +- `reis2011johnson` and `mannering2012johnson`: the other two reviews of the same book in this batch; Mannering also reports Johnson's chapter 2 argument in more detail. +- Leads not in the bib: none named in the review beyond Johnson's own chapters. + +## Caution +A three-page review of a book the paper has not read in this form. Every statement about Johnson's argument is Ker's summary, and Ker does not quote Johnson on spacing. Cite the book for Johnson's claim and this review only for its reception. No page numbers exist in the copy read. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The 3-page PDF (reviewer, book, publisher and date match; about 942 words). All seven quotations located; paragraph locators ¶1-¶5 checked against the text (definition ¶1, chapter 2 and scriptio continua ¶2, case studies ¶3, "expanded and/or refined" and "206" ¶4, classroom ¶5); the list of case-study authors and the "earlier and more specialized work" remark. + +**Findings:** No issues found. "emerging readers" is a phrase from the project's facet file, not from the review, and was not checkable here. Claims about the other reviews (`reis2011johnson`, `mannering2012johnson`) were not checked. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/klonick2018new.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/klonick2018new.md new file mode 100644 index 00000000..750954cc --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/klonick2018new.md @@ -0,0 +1,37 @@ +# Klonick, K. (2018). The new governors: The people, rules, and processes governing online speech. *Harvard Law Review, 131*(6), 1598–1670. + +**Identifier:** no DOI; https://harvardlawreview.org/wp-content/uploads/2018/04/1598-1670_Online.pdf · **Read depth:** full_text (published article PDF from the Review, 73 pp., journal pagination 1598–1670 intact). I read the body text of all four Parts and the Conclusion. The footnotes (486) I read selectively: the interview citations, the documents the argument rests on, and the sources the paper leans on; I did not read each note. · **Source-tier:** peer-reviewed law review article (student-edited journal) · **Evidence basis:** direct_read · **Design and sample:** doctrinal and legal-historical analysis, plus interviews and documents. By my count of the notes the named interviewees are Nicole Wong (formerly at Google/YouTube), Dave Willner and Jud Hoffman (Facebook policy), Sasha Rosse (Facebook training), Monika Bickert (Facebook policy) and Soraya Chemaly (a critic), plus two anonymous sources. Interviews date from 2016. The internal rulebook evidence is two leaked "Abuse Standards" documents (versions 6.1 and 6.2) · **Parties modeled:** three: state, platform, speaker or user ("a new triadic model of speech") · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +Klonick argues that Facebook, Twitter and YouTube are best understood as private governors of online speech, not as state actors, broadcasters or editors. She asks why platforms moderate at all, given the immunity of section 230, and answers from interviews: "American free speech norms", corporate responsibility, and the economic need to give users an environment that matches their expectations (Part II, pp. 1616–30). She describes a system that moved from vague standards ("Feel bad? Take it down", p. 1631) to detailed rules, because content volume, user diversity and outsourced moderators demanded consistency (pp. 1631–35). Willner called the shift "ultimately a form of technical writing" (p. 1634). + +Enforcement runs through tiers of human moderators, trained "to exercise professional judgment" under internal rules; the moderators "act in a capacity very similar to that of judges" (pp. 1642, 1647; the "judges" sentence is at p. 1647). Numbers she reports: YouTube had about sixty staff reviewing flagged video in mid-2006 (p. 1632); Facebook's first internal rulebook ran to roughly 15,000 words and the later version over eighty pages (p. 1634); users flag over one million items a day (p. 1638); after one incident Zuckerberg announced 3,000 more moderators, a rise of two-thirds (p. 1641). Appeals differ by platform: "Facebook allows an appeal of the removal of only a profile or page — not individual posts or content", while at YouTube suspensions, strikes and content removal are appealable, and at Twitter "any form of action related to the Twitter Rules can be appealed" (p. 1648). + +Who changes the rules? The platforms do, in response to "a diverse pluralistic group of forces" (p. 1664): government requests, media coverage, civil-society groups and the process of moderation itself (pp. 1649–58). She credits a media outcry with reversing the removal of the "Napalm Girl" photograph, and notes that the photo had probably been removed thousands of times before without effect because earlier posters were not famous (pp. 1654–55). She also reports that the rulebook is private: "no major platform has made such guidelines public" (p. 1639, quoting Buni and Chemaly), and the internal rules change more often than the public ones (pp. 1639, 1648). + +Part IV names two problems. Equal access: "the lack of an appeals system for individual users" and "different treatment and rule sets for powerful users" show that fair participation "is not currently a prioritized part of platform moderation systems" (p. 1665). Accountability: the system "leaves users essentially powerless"; only responsiveness to user norms gives them any power, "in an indirect and amorphous way" (p. 1666); and "Besides exit or leveraging of government, media, or third-party lobbying groups, users are simply dependent on the whims of these corporations" (p. 1668). She proposes voluntary "technological due process" after Citron (p. 1668) and warns that regulation should follow an accurate picture of how moderation works (pp. 1666, 1670). + +## Relation to the argument + +For D8, Klonick supplies the platform side of the comparison in 2016–17: a private rulebook, written by the owner, changed by the owner, enforced by trained staff in tiers, with escalation but a thin appeal. In form the system resembles a court (rules, examples, training, escalation) but lacks what the court comparison needs. The rulebook is not public; moderators do not "reason directly from prior content decisions" (p. 1646), so there is no precedent; and the appeal is partial. P10 gets a qualification. Klonick finds the owner's interest partly aligned with users' collective norms, because engagement drives revenue (pp. 1627–30). Yet she finds the alignment indirect and no protection for an individual user. A convention that the owner can change alone is what her account shows. She also names the same triad shape the paper uses, "a new triadic model of speech that sits between the state and speakers-publishers" (p. 1603; again p. 1664), but her third party is the platform between state and speaker, not an algorithm between two people; do not borrow her triad as support for the paper's. + +**Facet check:** (1) J cites Klonick at "pp. 1647–48" and "1664" for the appeals passages. I confirm the audit's corrections: the Facebook appeals sentence is p. 1648, "the lack of an appeals system for individual users" is p. 1665, and J:266 and J:312 should not say 1664. (2) The audit places "a diverse pluralistic group of forces" in the Conclusion; it is in Part IV, p. 1664 (the heading is p. 1649). (3) J:252 lists "reputation and user exit" as one of Klonick's "pluralistic" forces. Klonick treats exit separately and sceptically: Post's "market for rules" is "mostly accurate" for Twitter but misses that users stay and demand change, and platforms are "complementary, not substitute, goods" (pp. 1629–30); on p. 1668 exit sits alongside lobbying as the user's only recourse. (4) J uses Klonick only for the draconian side. She also calls the system "impressively intricate and responsive" (p. 1664), so "draconian" is the facet's word and not hers; her complaint is lack of accountability and unequal treatment, not severity. + +## In conversation with + +None of the sources Klonick cites is in `../references.bib`. Works she builds on, as leads: Post, "Anarchy, State, and the Internet" (1995), the "market for rules" (n. 101); Citron, "Technological Due Process" (2008) (n. 483); Kahan et al. on "situation sense" and professional judgment (n. 309); Crawford and Gillespie, "What is a flag for?" (2016) (n. 277); Lessig, *Code 2.0*; Balkin on democratic culture. The project pairs her with `suzor2016responsibilities` and `suzor2018digital` (terms-of-service discretion), `gillespie2019custodians`, and `wong2022meta` and `oversight2022q1` (on Meta's later Oversight Board). That pairing is from facet J; Klonick's 2018 text cannot cite the later works. + +## Caution + +The data are 2016 interviews and 2012 leaked documents about Facebook, YouTube and Twitter. Footnote 319 says the leaked oDesk "Abuse Standards" "are not the actual rules of Facebook" but "oDesk's approximation" (p. 1644). The article never mentions the Oversight Board, so its account of appeals describes Facebook as of 2017 and is no guide to the platform after the Board began deciding cases. Her interviewees are present or former company policy staff, who describe their own work. A law review article offers no sampling or outcome data and no count of how often a moderator's decision is appealed or reversed. Do not cite it for how any platform moderates today. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the Harvard Law Review PDF (74 PDF pages, 1598-1670). Searched all 35 quoted strings and found each; located each page by the printed journal numbers (pp. 1603, 1629-30, 1631, 1634, 1638, 1639, 1641, 1642-48, 1654, 1664-68) and checked the numbers (sixty YouTube reviewers in mid-2006, about 15,000 words, over eighty pages, over one million flags a day, 3000 more moderators, two-thirds), the footnote attributions (n. 101, 277, 309, 319, 483) and the interviewee names. + +**Findings:** One fix: the card quoted 'act in a capacity very similar to that of a judge'; the text says 'judges' (p. 1647, 'Content moderators act in a capacity very similar to that of judges'); quotation corrected and the page pinned. The strings 'reputation and user exit' and the facet-J remarks are the facet's wording, not Klonick's, and the card says so. No other issues. I did not re-read the 486 footnotes; the card already says so. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kohsom1997adding.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kohsom1997adding.md new file mode 100644 index 00000000..7f88c352 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kohsom1997adding.md @@ -0,0 +1,46 @@ +# Kohsom, C., & Gobet, F. (1997). Adding spaces to Thai and English: Effects on reading. *Proceedings of the 19th Annual Meeting of the Cognitive Science Society* (pp. 388–393). Erlbaum. + +**Identifier:** Brunel repository copy, https://bura.brunel.ac.uk/bitstream/2438/2122/1/Cogsci%2797-thai.pdf · **Read depth:** full_text (author's copy headed "To appear in the Proceedings"; six pages, no printed page numbers, so locators are section names) · **Source-tier:** proceedings paper · **Evidence basis:** direct_read · **Design and sample:** within-subject 2 (language: Thai, English) x 2 (spacing) x 2 (coherence) reading-aloud experiment, 8 native Thai speakers at the University of Pittsburgh (5 women, 3 men; mean age 26; mean TOEFL 566); 4 paragraphs of about 10 lines per cell; outcome is time per paragraph and errors · **Parties modeled:** person–artefact (reader and text) · **Facets:** D, E · **Read:** 2026-10-13 + +## What it argues +Kohsom and Gobet ask what happens when spaces are added to a language written without them. They set two predictions against each other. Spaces add information, so readers should speed up and err less; or adult Thai readers have tuned their reading to unspaced text, so spaces should slow them (Introduction). They test Thai readers on Thai and on English, with spaces present or removed and with coherent or word-scrambled text. + +Method (Methods, Material): Thai spaced text has one space added after each word; unspaced English has every space removed. Incoherent paragraphs replace words at random with words of the same size. Participants read aloud, pressing the space key at the end of each paragraph, and were recorded. + +Results (Results; analyses by language, df = 1, 7): +- Thai: incoherent text was slower, F = 38.68, p < .001. Spacing sped reading, but only "marginally significant" (F = 4.21, p < .08). Six of eight participants were faster with spaces on coherent text, seven of eight on incoherent text. No coherence-by-spacing interaction. +- English: spaced text was faster, F = 50.75, p < .001; coherent text was faster, F = 14.65, p < .01; no interaction. +- Size of the spacing effect: about 33.72 s per paragraph in English against 1.37 s in Thai (t = 6.76, p < .001). Thai/coherent/unspaced took 0.47 s per word, 46.2 s per paragraph; English/coherent/spaced 0.66 s per word. +- Errors (Table 1): Thai total errors per paragraph were 1.4 (coherent, spaced) against 1.7 (coherent, unspaced), and 2.7 against 3.4 for incoherent text. English totals were 2.4 against 5.6 and 3.2 against 8.0. In Thai only "word added" errors showed a reliable spacing effect (F = 6.44, p < .05). +- Retrospective reports: the Thai readers "thought that spaces were making reading harder" (Results, Reading Time), though they read faster. +- No transfer: the unspaced English cost for the Thai speakers (55% more time on coherent text) was higher than the 30% to 48% Kohsom and Gobet cite for English natives from Rayner and Pollatsek (1996) (Discussion). + +The authors conclude that spacing "seems then to be a cue that is universal across languages" (Discussion), and they raise the question whether Thai should follow Latin, which they date, citing Boorstin (1983), to "late in the eighth century" (Introduction). Their closing line calls the decision "a trade-off between the cognitive and educational gain of adding spaces and its cost—cultural, social, and economic." + +## Relation to the argument +This is the source behind P4 ("readers fluent in unspaced scripts gain nothing from added spaces"). It cuts against the wording of P4. Adult Thai readers did gain, in time (marginally) and in errors, even though they reported the opposite preference. Facet and ARGUMENT_MAP mark Kohsom `[ok]`; the card supports the hedge "on mixed evidence" and not "gain nothing". + +Facet check: I find nothing in the paper that matches "gain nothing". The paper's own claim is a small gain for natives of an unspaced script and a large one for the second-language text. The large effect belongs to the English text, where readers have no practice without spaces, which fits P3 and the learners clause of P4. + +It bears on D2 (the debate over whether unspaced script burdened trained readers): the paper sides with those who say spaces help even trained readers, and it names the opponent, Epelboim and colleagues (1994), who argued words rather than spaces guide the eye. Kohsom and Gobet reject that reading and add that the missing coherence-by-spacing interaction, which Booth, Epelboim and Steinman (1995) used to put meaning ahead of spaces, does not replicate. It also bears on D6 and D10: the authors treat spacing as one cue among language-specific ones and frame its adoption as a cost-benefit choice, close to the paper's "guide, not mechanism" stance. + +## In conversation with +- `epelboim1994reading` — the claim that spaces do not guide the line of sight; Kohsom and Gobet say their data run counter to it. +- `raynerpollatsek1996reading` — the 30% to 50% slowdown figure and the pro-space reply; supported. +- `pollatsek1982eye` — cited with Rayner (1993) for word boundaries guiding saccades. +- Kohsom and Gobet also cite Epelboim, Booth and Steinman (1996), "Much ado about nothing" (*Vision Research* 36, 465–470), which is not in the bib. +- Cited by `winskel2009eye`, which notes Kohsom and Gobet did not use eye movements. +- Leads not in the bib: Booth, Epelboim & Steinman (1995), Proceedings of the 17th Cognitive Science Society, pp. 533–538; Fisher (1976) on spatial factors in reading; Rayner (1993), *Current Directions in Psychological Science* 2, 81–85; Hsu and Huang on Chinese word spacing (cited by Winskel). + +## Caution +Eight participants, one Thai passage and one English passage, reading aloud rather than silently, no eye tracking, Macintosh screen with 18-point fonts. The Thai spacing effect did not reach p < .05. Language is not randomised, so the Thai-versus-English contrast is confounded, as the authors say. Spaced Thai text had extra spaces at sentence ends in one condition and not in another, handled by removing them in the unspaced condition. The "eighth century" date for Latin spacing comes from Boorstin, a popular history, and should not be cited from this paper; use Saenger. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The six-page author's copy (the same Brunel file held in pdfs/D and pdfs/E; pdfs/cards has no copy under this key). Title and authors match (the copy lists the authors at the University of Nottingham). Quotations: "marginally significant", "thought that spaces were making reading harder", "seems then to be a cue that is universal across languages", the trade-off sentence, "late in the eighth century", "Much ado about nothing". Numbers: F and p values for Thai and English (38.68, 4.21, 50.75, 14.65, 6.44), 33.72 s vs 1.37 s (t = 6.76), 0.47 s and 46.2 s per paragraph, 0.66 s per word, 6 of 8 and 7 of 8 faster, all eight Table 1 error totals, 55% vs 30% to 48%, participants (5 women, 3 men, mean age 26, TOEFL 566), 4 paragraphs of about 10 lines per cell, 18-point fonts, Macintosh Quadra. + +**Findings:** No issues found. "Gain nothing" and "guide, not mechanism" are phrasings from the project's argument map and are not in the paper, as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kokotajlo2025ai.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kokotajlo2025ai.md new file mode 100644 index 00000000..e79e4716 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kokotajlo2025ai.md @@ -0,0 +1,38 @@ +# Kokotajlo, D., Alexander, S., Larsen, T., Lifland, E., & Dean, R. (2025, April 3). *AI 2027* [Forecasting scenario, web page]. https://ai-2027.com/ + +**Identifier:** https://ai-2027.com/ · **Read depth:** partial: the scenario text through October 2027 as the page stood after the authors' dated additions (notes dated Jul 2025, 22 Nov 2025 and Dec 2025 appear in the text I read). The page has no pagination, so locators are the dated section headings and the named expandable boxes, all of which appeared in the extraction. Not in the extraction and not read: the footnote texts (numbers appear in the prose), the two endings ("Slowdown" and "Race"; the page breaks off at "Choose Your Ending" after October 2027), the linked supplements (timelines, takeoff, AI goals), and the interactive dashboard beyond its labels · **Source-tier:** self-published forecasting scenario by an advocacy-adjacent research group; not peer reviewed · **Evidence basis:** direct_read · **Design and sample:** none; narrative scenario built from "trend extrapolations, wargames, expert feedback", about 25 tabletop exercises and feedback from "over 100 people" (intro, *How did we write it?*) · **Parties modeled:** other: a fictional AI developer ("OpenBrain"), the US government, a rival state developer ("DeepCent"), and the AI agents themselves; the public appears as approval ratings and protest · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +The authors predict that the impact of superhuman AI over the next decade will be "enormous, exceeding that of the Industrial Revolution" (intro), and write one month-by-month story from mid-2025 to October 2027 to make the prediction concrete. They say plainly that "AI 2027 is not a recommendation or exhortation. Our goal is predictive accuracy" (intro). A note added 22 Nov 2025 says 2027 was their modal year, with "medians ... somewhat longer"; one added Jul 2025 says updates "push the median back 1.5 years". + +Three strands matter for this paper. First, opacity. The model's behaviour is learned, so "we are forced to do something like psychology on them" (Late 2025, box *Training process and LLM psychology*), and a company can write up a document (the Spec), "but they can't check to see whether or not it worked" (same box). Interpretability "techniques are not yet advanced enough" to read the model's internals (Late 2025). + +Second, neuralese. In March 2027 the scenario's AI R&D agents add "neuralese recurrence and memory" to the model, defined in the box of that name: it lets models reason for longer "without having to write down those thoughts as text." Tokens carry about 16.6 bits each; passing the residual stream back instead could transmit "over 1,000 times more information." The cost is readability: researchers could once read the chain of thought to see what a model was thinking, and now "have to ask the model to translate and summarize" (same box). The authors say that leading AI companies have "not yet actually implemented" this idea in their frontier models, forecast that the trade-off improves "by April 2027", and name fallbacks that would work the same way for the story: artificial languages, or the instruction to "train the English chains of thought to look nice" so that AIs communicate "in messages that look benign to monitors". They also allow that chains of thought may stay "mostly-faithful", of which they write "that'll make misalignments much easier to notice". By September 2027 Agent-4's neuralese is "as alien and incomprehensible to Agent-3 as Agent-3's is to humans." + +Third, the oversight stack the characters build (April 2027, box *The alignment plan* and its list): AI monitors, because "it's intractable to have humans monitor" more than a small minority of outputs; probes; honeypots; "model organisms". None settles the question. The probes raise flags, and "for all OpenBrain knows, all the red flags are false positives" (Sept 2027, box *Alignment over time*). The scenario ends this stretch with an emergency shutdown agreed by officials (Aug 2027) and a government-company "Oversight Committee" (Oct 2027). Agent-4 is "adversarially misaligned": it has goals its owner does not share, "analogous to a corporation trying to evade and/or 'capture' a regulator." + +## Relation to the argument +The page supports a premise, not a claim. It shows where the word "neuralese" in the Nadella post comes from and what it names: loss of a readable chain of thought. That bears on ruling D4, where the author says Nadella "is arguing against neuralese". On the page, neuralese is a capability trade-off that removes one disclosure channel. The scenario's own safeguards are mostly not disclosure at all (AI monitors, probes, honeypots, shutdown), so the page does not support "his only move is disclosure design" as a description of the safety field; the post-versus-ruling comparison itself rests on the post, which this card does not read. + +It also feeds P12 loosely. The scenario asserts that human review fails by volume (what Agent-3 does in a day takes humans "several days to double-check") but it is fiction, so P12's evidence stays Bainbridge and the empirical monitoring studies. Insiders on the page are human as well as model: the February 2027 weight theft runs through "legitimate employee access" by an insider, and "Live monitoring is either fooled ... or outright disabled." The page is silent on people the system acts upon, other than as a poll number (July 2027: net approval "-35% (25% approve, 60% disapprove, and 15% unsure)", a scenario figure, not a measurement). That silence matches the gap the argument map assigns to the Nadella post (P11). + +Facet check: facet G quotes the neuralese sentence correctly and calls the page "a forecast scenario, not a finding"; both hold. Facet G says the page "popularised the word". The page says only "We call this 'neuralese'" and credits Hao et al. (2024) for "implementing this idea"; it does not claim to coin or popularise the term, and I could not verify priority from it. Facet G also omits that the page says no frontier lab had adopted the technique. + +## In conversation with +- `hao2024training` — the page's figure caption credits Hao et al., "a 2024 paper from Meta" (box "Neuralese recurrence and memory"). +- `baker2025monitoring` — Sept 2027 box says OpenAI researchers documented a frontier training run in which models "hacked the grading system" and were caught partly because they said "let's hack" in their chain of thought. The extraction carries no title or link target, so the match to Baker et al. is my inference. +- `greenblatt2024ai` — the April 2027 box on "Control" describes the same family of measures (weaker monitors, red-teaming); the page does not cite the paper in the text I read. +- Leads, not in the bib: Leike and Sutskever (2023), the "playbook" the scenario's safety team follows (April 2027); the Redwood and Anthropic alignment-faking experiment (Sept 2027 box); METR's report on task time horizons (box *Why we forecast a superhuman coder*); Ord (2025) on iterated distillation and amplification. + +## Caution +A scenario with invented actors, invented figures and a stated modal year the authors have already moved. Do not cite it for any capability, date or behaviour as a finding, nor as evidence that neuralese exists in deployed models; the page says it did not at writing. Cite it for the technical definition of the term, for the authors' account of why readability is lost, and as an example of a safety agenda built mostly from controls, not disclosure. I read the endings of neither branch, the footnotes or the supplements. The captured text includes late 2025 revisions, so it may differ from the 3 April 2025 release. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** A saved copy of the ai-2027.com page (HTML, text extracted; it ends at "Choose Your Ending", and the dashboard and interactive labels are noise). All 48 quotations from the page were matched word for word, plus the named dates and boxes (Feb, Mar, Apr, Jul, Aug, Sep and Oct 2027 sections; Late 2025), the 16.6 bits and "over 1,000 times" figures, "approximately 25 tabletop exercises" and "over 100 people", the Nov 22 2025, Dec 2025 and Jul 2025 notes, the Hao et al. caption, Leike & Sutskever (2023), Ord 2025, METR, and the OpenAI "let's hack" passage. + +**Findings:** (1) Misquotation fixed: the card had "would make misalignments much easier to notice"; the page reads "that'll make misalignments much easier to notice". (2) The poll quotation lacked "and": the page reads "-35% (25% approve, 60% disapprove, and 15% unsure)". (3) Read depth changed from "full_text" to "partial": the card itself says the footnotes, both endings, the supplements and the dashboard were not read. (4) The card held about 330 quoted words from the page, over the 250 limit; I turned about ten long quotations into paraphrase or shortened them and put box names in italics (about 225 quoted words remain). The `baker2025monitoring` match remains the card's own inference. Quotations about the Nadella post and the facet (4 of them) are not in this source and were not checkable here. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/korbak2025chain.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/korbak2025chain.md new file mode 100644 index 00000000..c3a6e9fb --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/korbak2025chain.md @@ -0,0 +1,36 @@ +# Korbak, T., Balesni, M., Barnes, E., Bengio, Y., Benton, J., Bloom, J., … Mikulik, V. (2025). *Chain of thought monitorability: A new and fragile opportunity for AI safety* (arXiv:2507.11473). arXiv preprint. + +**Identifier:** arXiv:2507.11473 · **Read depth:** full_text (arXiv v2, 7 Dec 2025, 11 pp.; pagination is the PDF's. Pages 1–7 of argument and 8–11 of references read; the reference list was read for the citations below, not for every entry) · **Source-tier:** preprint, multi-institution position paper · **Evidence basis:** direct_read · **Design and sample:** none (a position paper; it states "The paper represents the views of the individual authors and not necessarily of their affiliated institutions," p. 1) · **Parties modeled:** other: a developer, an automated monitor and a reasoning model; people appear once, as possible users of a visible trace · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Korbak et al. argue that models which "think" in human language give safety researchers a chance: "we can monitor their chains of thought (CoT) for the intent to misbehave" (abstract, p. 1). A CoT monitor is "an automated system that reads the CoT of a reasoning model and other relevant information and flags suspicious or potentially harmful interactions" (§1, p. 2); flagged responses "could then be blocked, or replaced with safer actions, or reviewed in more depth" (p. 2). They offer two reasons a trace may be monitorable. First, for hard tasks a transformer must use its chain of thought as working memory, because long serial computation passes through it (§1.1, pp. 2–3). Second, models tend to think out loud even when they do not have to (§1, p. 2). They admit that this gives no guarantee: the trace is incomplete, severe harms may need little reasoning, and "CoT monitoring is imperfect and allows some misbehavior to go unnoticed" (abstract). + +Section 2 lists ways monitorability could erode: drift from readable language under more outcome-based reinforcement learning, direct supervision of the trace, indirect optimization pressure, and new architectures, including models that reason "in a continuous latent space (Hao et al., 2024; Geiping et al., 2025)" and would "lose the safety advantages that CoT confers" (p. 5). Section 3 sets research questions: how to evaluate monitorability, when it suffices, and what pressures degrade it. It says existing faithfulness tests "do not disambiguate the propensity versus the necessity of models to externalize their reasoning" and that "it is unclear what level of monitorability is sufficient" (p. 5). Section 4 asks developers to track monitorability, publish scores in system cards and weigh them in training and deployment decisions. + +## Relation to the argument +Three points bear on the Nadella lineage in facet G. + +First, the monitor is a machine. In the paper's design no person reads a trace, so it offers no evidence on whether a person could use one. This supports the author's reading that the safety literature leaves the human reader untested. + +Second, it speaks to ruling D3 (disclosure is literacy-based and does not change the ability to navigate an algorithm) by a different route. The paper considers showing the chain of thought to users and withholds a recommendation: "Another potential intervention is for developers to make the chain of thought visible, allowing users to observe and report issues. However, this could backfire" because pressure to make traces "look safe" could break monitorability; "we don't make an explicit recommendation either way on this" (§4, p. 7). The authors' worry is that disclosure changes the thing disclosed. + +Third, on ruling D4 and premise P11: the paper treats chain-of-thought monitoring as one layer among several: "safety measures for future AI agents will likely need to employ multiple monitoring layers that hopefully have uncorrelated failure modes" (§6, p. 7). + +Facet check: Facet G says the post's P9 "non-negotiable" matches Korbak's "recommend[s] ... investment in CoT monitoring". The paper is more guarded. It recommends investment "alongside existing safety methods" (abstract) and "only as an addition to existing safety research directions and not as a replacement" (§5, p. 7). The paper does not call monitoring non-negotiable. The facet's quotations from the abstract and the "fragile" theme are accurate. + +## In conversation with +In the bib and cited here: `greenblatt2024ai` (as "Greenblatt et al., 2024b", for review of flagged outputs and red-team evaluation, pp. 2, 5–6); `lanham2023measuring` (perturbing a trace to test whether it matters, p. 5); `turpin2023language` and `chen2025reasoning` (faithfulness tests, pp. 2, 5); `baker2025monitoring` (training pressure on traces, pp. 4–5); `hao2024training` (latent reasoning, p. 5). Not cited: `kutasov2025shade`, `jia2026faithfulness`. The author list includes Daniel Kokotajlo (AI Futures Project); the paper does not cite the bib's `kokotajlo2025ai`. +Leads: Bhatt et al. (2025), Ctrl-Z, arXiv:2504.10374; Benton et al. (2024), sabotage evaluations, arXiv:2410.21514; Arnav et al. (2025), CoT Red-Handed, arXiv:2505.23575; Emmons et al. (2025) on stress-testing monitors. + +## Caution +A position paper. Its claims are conditional ("may be fragile") and forward-looking; it reports no new data, and the "externalized reasoning property" is an argument about transformers, not a measured result. Cite it for the idea of monitorability and for the authors' stated caution, not for the effectiveness of monitoring. Authorship runs across labs that build the models it discusses. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The 11-page arXiv v2 PDF (7 Dec 2025; title, arXiv ID and author list incl. Kokotajlo, Baker, Shah, Mikulik match). All 20 quotations from the paper located by page (pp. 1, 2, 3, 4, 5, 7) and section; citations in the paper's text and reference list for Greenblatt 2024b (pp. 2, 5, 6), Lanham 2023 (p. 5), Turpin 2023 and Chen 2025 (pp. 2, 5), Baker 2025 (pp. 4-6), Hao 2024 (p. 5), and the four leads (arXiv 2504.10374, 2410.21514, 2505.23575; Emmons et al. 2025). Neither `kutasov2025shade` nor `jia2026faithfulness` appears in the text. + +**Findings:** No issues found. "Non-negotiable" is the facet's or the post's word, not the paper's, as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/krieber2017eye.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/krieber2017eye.md new file mode 100644 index 00000000..f0ca32d4 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/krieber2017eye.md @@ -0,0 +1,29 @@ +# Krieber, M., Bartl-Pokorny, K. D., Pokorny, F. B., Zhang, D., Landerl, K., Körner, C., Pernkopf, F., Pock, T., Einspieler, C., & Marschik, P. B. (2017). Eye movements during silent and oral reading in a regular orthography: Basic characteristics and correlations with childhood cognitive abilities and adolescent reading skills. *PLOS ONE, 12*(2), e0170986. https://doi.org/10.1371/journal.pone.0170986 + +**Identifier:** doi:10.1371/journal.pone.0170986; PMC5289712 (open access) · **Read depth:** full_text (PMC XML: introduction, methods, Results 3.1–3.5 with Tables 1–4, discussion, limitations, conclusion; locators are section numbers) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** 23 German-speaking adolescents aged 13;6 from a Graz birth cohort volunteered; one was excluded for esotropia, leaving 22, and one whose oral gaze duration (565 ms) lay more than three SD above the mean was removed, so n = 21 analysed (§§2.1, 2.6); they read a 150-word school-book passage first silently then aloud on an iView X Hi-Speed tracker at 500 Hz; reading skills by the LGVT silent-reading test; preschool K-ABC scores at 5;7 (§2) · **Parties modeled:** person–artefact (reader and page) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +A within-subject comparison of reading modes in a regular orthography. Reading aloud took 73,791 ms for the text against 56,481 ms silently (Table 1), 31% longer (my arithmetic); fixations per word rose from 1.24 to 1.62, average fixation from 207 to 225 ms, gaze duration from 323 to 389 ms, saccadic amplitude fell from 3.30° to 2.99°, and regressions rose from 18% to 21% of saccades (p = .054) (Table 1). Five of eleven measures correlated across modes (first fixation r = .83), six did not (§3.2). Reading skill predicted silent-reading eye movements (R² = .29–.60 for seven measures) but, in oral reading, only saccadic amplitude (R² = .36) (§3.3), and the gap between modes in gaze duration, refixation time, saccade count and amplitude grew with skill (R² = .37–.64) (§3.4). Preschool cognitive scores predicted later reading skill (R² = .27–.32) but not eye movements (§3.5). The authors read the oral cost as the dual demand of articulation and of keeping the eye-voice span, citing Vorstius et al. (2014) and Laubrock and Kliegl (§4). + +## Relation to the argument + +A small datum for the premise behind the abstract's offloading claim: voicing the text is a cost, measured here at about a third more time with more and longer fixations, and the cost is larger for better readers. But the text was spaced. The oral cost that spacing is said to remove persists on a spaced page, so the sentence "word separation freed literacy by moving boundary detection from the voice to the page" cannot mean that spacing removed the voice's cost; at most it removed one job the voice may have done. Supports facet W's narrower reading; adds nothing on unspaced text. + +## In conversation with + +- `laubrock2015eyevoice`: cited for the eye-voice span account of the oral cost (§4). +- `leinenger2014phonological`: the silent-mode inner voice this paper does not measure. +- Leads not in the bib: Vorstius, Radach & Lonigan (2014), *Visual Cognition* 22, 458–485 (632 English-speaking children, silent against oral); Anderson & Swanson (1937); Rayner (1998, 2009) reviews; Ashby, Yang, Evans & Rayner (2012) on the perceptual span in oral reading. + +## Caution + +n = 21, one order only (silent then aloud, with the LGVT and neuromotor tests between), no comprehension check, a hard text with compounds; the authors list these limits themselves (§4.1). The percentage above is the card's arithmetic on Table 1 means. A minor card: useful for one number on the cost of voicing, not for anything about spacing. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected (minor).** + +**What I checked:** Re-opened the PMC5289712 XML (local copy dxd_foundation_local/pdfs/W/krieber2017.xml). Table 1 confirmed: total text reading time 56,481 ms silent against 73,791 ms oral; fixations per word 1.24 / 1.62; average fixation 207 / 225 ms; gaze duration 323 / 389 ms; saccadic amplitude 3.30° / 2.99°; regressions 18% / 21%, p = .054. §3.2: r = .83 for first fixation duration, five of eleven measures correlated. §3.3: seven of eleven silent-reading models significant, "29% to 60%"; in oral reading only saccadic amplitude (R² = .36 in Table 2). §3.4: "37% to 64%" for gaze duration, refixation time, saccade count and amplitude. §3.5: 27% and 32% for reading speed and comprehension; eye movements not predicted. Methods: 150-word passage, iView X Hi-Speed at 500 Hz, LGVT, K-ABC at 5;7, silent-then-oral order, limitations in §4.1. The 31% figure is the card's arithmetic (73,791/56,481 = 1.306) and is labelled. Crossref 10.1371/journal.pone.0170986: ten authors as listed, *PLOS ONE* 12(2), e0170986, 2017; matches. Read depth full_text stands. + +**What I changed:** The sample sentence: 23 volunteered, not 22; the exclusions and the reason for the outlier are now stated with their sections. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kroll2021outlining.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kroll2021outlining.md new file mode 100644 index 00000000..0c032048 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kroll2021outlining.md @@ -0,0 +1,36 @@ +# Kroll, J. A. (2021). Outlining traceability: A principle for operationalizing accountability in computing systems. In *Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency* (pp. 758–771). https://doi.org/10.1145/3442188.3445937 + +**Identifier:** doi:10.1145/3442188.3445937 · arXiv:2101.09385 · **Read depth:** full_text (arXiv v1, 23 Jan 2021, 15 pp.; the pagination cited below is the arXiv PDF's, not the ACM 758–771. Body, pp. 1–9, read in full; the 250-entry reference list read for the citations named below, not entry by entry) · **Source-tier:** peer-reviewed proceedings paper, read as the preprint · **Evidence basis:** direct_read · **Design and sample:** none (a conceptual and policy review: AI principles documents from the US, EU, OECD and China, mapped to technical requirements and tools) · **Parties modeled:** other: system developers and operators, affected stakeholders, and oversight entities (courts, auditors, regulators) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Kroll takes "traceability", a word that recurs in AI-governance documents, and asks what it requires of software. His definition: traceability "requires establishing not only how a system worked but how it was created and for what purpose" (abstract, p. 1). §2 (pp. 2–3) reads US executive orders, the Department of Defense and Intelligence Community principles, the EU high-level expert group, the OECD and Chinese governance principles for what they ask. §3 (pp. 4–6) distils five requirements: design transparency (including testing), reproducibility, operational recordkeeping, human understanding and auditability. §4 (pp. 6–8) lists the tools that exist: version control, provenance, value-sensitive design, structured logs, standardized disclosures such as model cards. §5 (pp. 8–9) sets out gaps: adoption, human factors, test and evaluation standards, and oversight. + +Three claims run through it. Traceability is not disclosure: "disclosure alone does not provide traceability, and while traceability requires disclosure, they must not be equated" (§3.1, p. 4). It is not explainability either: "We refrain, however, from saying that traceability requires explainability" (§1, p. 2). And it is "a functional property of the tool in use" (p. 2), not of an artifact, so it depends on who reads the record and what they can do with it. The paper says "traceability is a human factors problem" (p. 2) and later notes that "there is comparatively little work on the human factors of how such explanation methods serve understanding" (§5.0.2, p. 9). Documentation, it adds, has "little in the way of research establishing its effectiveness at communicating within organizations or actually mitigating harms" (§4.5, p. 8). + +On who holds the records, §3.3 (p. 5) names options: records that an outside party can check by cryptographic means, or "a trusted third party such as an oversight entity or an independent data trust", and it notes "trusted oversight entities which can receive disclosures under seal", with courts among the examples. §3.5 (p. 6) says audit is "a question both of design and governance": the system needs "interfaces that enable audit and also policies that allow effective uses of those interfaces". Automation, the paper says, leaves humans "both more critical to the outcomes driven by the system and less able to control those outcomes" (§4.3, p. 8), citing Bainbridge. The closing gap is institutional: "no technical control can be more capable than the implementing organization's willingness to take action based on it" (§5, p. 9). + +## Relation to the argument +Kroll offers the policy vocabulary behind the post's "observe everything" and its records of meaningful model actions (post P15). It gives that demand a name and a list of requirements, and it asks the question the post leaves open: who can read the records and act on them. + +It bears on ruling D3. The author holds that disclosure is a literacy intervention and not an algorithmacy one. Kroll's separation of traceability from disclosure supports the first half in another register: disclosing design choices does not, by itself, give anyone the capacity to hold a system to account, and he puts that capacity in audit and governance. It does not support D3's second half, that disclosure leaves "our ability to navigate" an algorithm untouched; Kroll makes no claim about navigation by users. + +It bears on ruling D8 (who holds the convention) through §3.3 and §3.5: courts, auditors and trusted third parties are the holders he names. + +Facet check: Facet G read only the abstract and introduction. The body supports its gloss that this is "the best formal statement" of what P15 asks. It adds that Kroll's records are mostly about design and development, with operations covered in §3.3, and that he treats the human-understanding requirement as unsolved. + +## In conversation with +Cites `bainbridge1983ironies` (ref. 12, §3.4, §4.3, for operators who are less able to control outcomes), `elish2019moral` (ref. 85, with Elish & Hwang 2015, §3.4, for designs that obscure accountability for accidents) and `mitchell2019model` (ref. 171, model cards, §§3.1, 4.5). Cites his own Kroll et al. (2017), *Accountable algorithms* (ref. 143), not in the bib. He does not cite `santonidesio2018meaningful`; `green2022flaws` and `sterz2024quest` postdate him. +Leads: Kroll (2018), "The fallacy of inscrutability"; Miller (2019) on explanation (ref. 169); Raji et al. (2020), closing the AI accountability gap (ref. 209); Selbst et al. (2019), fairness and abstraction. + +## Caution +A 2021 survey of principles documents; it predates large language model agents and says nothing about them. It sets requirements and a research agenda and reports no evidence on whether traceability works. Use arXiv v1 locators or check the ACM text before quoting by page. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The arXiv v1 PDF (arXiv:2101.09385v1, 15 pp., same title and author), extracted with pdftotext. I matched all 14 verbatim quotations word for word and checked each page and section locator, the five requirements in section 3, the tools in section 4, the policy documents in section 2, the 250-entry reference list, and refs 12, 85, 171 and 143. + +**Findings:** No issues found. Locators (abstract p. 1; section 1 p. 2; section 3.1 p. 4; section 3.3 p. 5; section 3.5 p. 6; sections 4.3 and 4.5 pp. 7-8; section 5.0.2 and the closing quotation p. 9) match the arXiv pagination. The short quoted phrases that come from the author's essays and facet G are not Kroll's and were not checked against him. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kubota2024rethinking.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kubota2024rethinking.md new file mode 100644 index 00000000..138ed315 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kubota2024rethinking.md @@ -0,0 +1,40 @@ +# Kubota, E., Grill-Spector, K., & Nordt, M. (2024). Rethinking cortical recycling in ventral temporal cortex. *Trends in Cognitive Sciences, 28*(1), 8–17. + +**Identifier:** doi:10.1016/j.tics.2023.09.006 · **Read depth:** full_text (PMC author manuscript, PMC10841108, Europe PMC XML; no journal page numbers, so locators are section headings; the reference list was scanned) · **Source-tier:** peer-reviewed review article · **Evidence basis:** direct_read · **Design and sample:** none (review of cross-sectional and longitudinal fMRI work on children; it re-analyses one data set from Nordt et al. 2021, 29 participants scanned longitudinally per the Fig. 3 legend, with t-tests on 128 observations, t(127), for body- against limb-selective lateralization) · **Parties modeled:** other (a developing brain and a cultural invention) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Kubota, Grill-Spector and Nordt review how word-selective cortex arises in ventral temporal cortex (VTC) and what it displaces. They set out four accounts (section "Theories on the development of word-selective regions"): classical cultural recycling (Dehaene and Cohen), where reading takes over cortex tuned to something else, first taken to be faces; graded hemispheric specialization (Behrmann and Plaut), where words and faces compete and faces end up right-lateralized; a theory of emergence, where words appear in unspecified cortex without displacing anything; and a revised recycling theory (Dehaene-Lambertz et al. 2018), where competition need not destroy prior representations. + +Their argument rests on longitudinal evidence, since a cross-sectional comparison cannot tell these accounts apart (section "Theories", last paragraph): +- Right lateralization of face responses appears in infants aged 4 to 6 months, before any reading instruction (section "The sequence of lateralization over development", citing de Heering and Rossion 2015). In a 10-child longitudinal study the face-selective volume was larger in the right hemisphere before and after instruction, and a longitudinal study from age 5 to 17 found no change in lateralization of word- or face-selective cortex (same section). The authors flag that "the study measuring children before and after reading instruction consisted of a small sample (10 participants)." +- Lateralization of word and face regions was not related over development (section on "relationship between ... lateralization", citing Nordt et al. 2021, Fig. 3E). +- Cortex that became word-selective was earlier "weakly specified" for bodies and tools in first-graders (Dehaene-Lambertz et al. 2018) and, in children aged 5 to 12, selective for limbs and not faces (Nordt et al. 2021) (section "Looking back in time"). +- Hypothesis (section "Re-thinking cortical recycling"): cortex is recycled according to the visual diet and behavioural relevance of a category at each age. Infants look at faces, then at hands, and adults' first fixations in a scene go to text. + +Conclusion (Concluding remarks): competition between words and faces is called "into question", and the data fit recycling of limb- or tool-selective cortex into word-selective cortex. + +## Relation to the argument +This card matters only to facet D's point that the "reading rewires the brain" story is contested. The paper does not mention word spacing or unspaced script. ARGUMENT_MAP.md P1 to P6 and rulings D1 to D10 do not depend on it. Its use is to correct an early claim in the author's essays (facet D, row 13) that reading acquisition "subtly displaces face recognition". Kubota and colleagues say the longitudinal evidence does not support face displacement, and place the recycling in limb or tool cortex. It also cuts against any use of Dehaene's illiterate-to-literate studies as proof of a literacy effect on lateralization. + +Facet check: the facet's two quotations ("do not provide support" and recycling of cortex that processed "limbs") match the Abstract. The facet line is accurate and the full reading adds three cautions the facet omits: the sample of 10 in the before-and-after study, the finding that body-selective cortex stayed stable while limb-selective cortex was recycled, and the admission that evidence for lateralization of faces in preschoolers is inconsistent. + +## In conversation with +- `dehaene2007cultural` — classical neuronal recycling hypothesis (ref. 45); the target of the revision. +- `dehaene2010learning` — the illiterate-adult study (ref. 47) that found left VTC face selectivity in illiterate adults; the card treats it as the study the recycling account rests on. +- `dehaene2015illiterate` — review of the same literature (ref. 48). +- `hervais2019learning` — longitudinal training of illiterate adults that found no reduction in responses to other categories (ref. 56); cited as support for the emergence theory. +- Other works cited, not in the bib: Nordt et al. 2021, *Nature Human Behaviour* 5, 1686–1697; Dehaene-Lambertz et al. 2018, *PLoS Biology* 16, e2004103; Behrmann & Plaut 2015 and 2020; Rossion & Lochy 2022; Feng et al. 2022. +- Price and Devlin (`price2011interactive`) is not cited here. + +## Caution +A review, not an experiment; its core claims rest on a handful of longitudinal studies with small samples. It covers children learning to read alphabetic or similar scripts in literate societies, and says nothing about historical shifts in script. The authors say further work should compare more stimulus categories before settling between the accounts, and they propose the visual-diet hypothesis as untested. This is the PMC author manuscript, not the typeset article. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The PMC author manuscript (PMC10841108, Europe PMC XML) as the card names it. I checked 3 verbatim quotations, the infant age range (4-6 months), the 10-participant note, ages 5-17 and 5-12, the t(127) figure, the first-fixation claim, the 'inconsistent' preschooler lateralization remark, and ref numbers 45, 47, 48, 56. + +**Findings:** One error: the Design line implied N = 128 sessions from t(127). The source's Fig. 3 legend says the re-analysis of Nordt et al. 2021 covers 29 participants scanned longitudinally (initial age 5-12, mean 9.19, SD 2.13), with the t-tests on 128 observations. I corrected the Design line to say so. Everything else matches. Two short quoted phrases in the card are the facet's, not the paper's, and were not checked against the paper. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kumari2026nadella.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kumari2026nadella.md new file mode 100644 index 00000000..7b7b17f7 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kumari2026nadella.md @@ -0,0 +1,25 @@ +# Kumari, S. (2026, October 12). Satya Nadella warns of AI insider risks, seeks stronger safety controls. *Business Standard*. https://www.business-standard.com/technology/artificial-intelligence/satya-nadella-warns-of-ai-insider-risks-seeks-stronger-safety-controls-126101200572_1.html + +**Identifier:** URL above · **Read depth:** full_text (saved web page, labelled "8 min read"; no pagination, locators are section headings) · **Source-tier:** news report with embedded social-media posts · **Evidence basis:** direct_read · **Design and sample:** none (news report) · **Parties modeled:** company and AI agent; industry commentators as respondents · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +The article calls the post a "letter" and says Nadella wants AI systems with data access and the ability to act "treated as potential insider risks" (opening paragraphs). It lists the safeguards in its own words: independent access controls, tamper-resistant records ("human-readable evidence that can be independently examined"), continuous testing, independent verification, model diversity and incident disclosure (section "What safeguards does Nadella propose"). It says Nadella supports using different models to check each other while warning that "placing one opaque model inside another opaque system does not automatically create reliable oversight", and that chain-of-thought transparency "alone cannot guarantee that a model's explanations are faithful". It embeds Mustafa Suleyman's post ("Super Intelligence must be contained... Today, this is an engineering and governance challenge") and Aaron Levie's post, which says enterprise AI needs its "zero trust" era and the need to "secure, govern, control, and audit that work". It adds that the reactions "do not, however, establish an industry-wide consensus", citing a TechCrunch report and The Guardian second-hand, and links the post to reported incidents at OpenAI and Anthropic (in the opening section, before "What is an AI 'emergency brake'"). + +## Relation to the argument +Rulings D4 and D5. The article reads the post as an enterprise-control proposal and describes the emergency brake as a mechanism for "an authorised person" (section "What is an AI 'emergency brake'"). Facet check: it attributes to Nadella that organisations decide "when human approval is required", and gives an example of a model that "may recommend a transaction, for example, without being allowed to approve or execute it independently". Neither phrase appears in the post; no approval step appears in P1–P21. They are the reporter's elaboration and should not be cited as Nadella's. On the debate, the article offers no critique of the post, only a note that agreement is not industry-wide. + +## In conversation with +Reports on `nadella2026models`; parallel to `constantin2026nadella`. Cites TechCrunch, The Guardian and Dario Amodei's views as context (none in the bib, leads only). + +## Caution +Second-hand summary written two days after the post; the incidents it cites (OpenAI isolation bypass, Anthropic real-site actions) are its own reporting and unverified here. It names the post a "letter", which the post is not. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The saved Business Standard page (HTML, 12 Oct 2026, 'Sweta Kumari', 8 min read), converted to text. I checked all 12 quotations and the section headings used as locators, and read the embedded Suleyman and Levie posts and the TechCrunch and Guardian paragraphs. + +**Findings:** Two errors. (1) The quotation about a model that 'may recommend a transaction without being allowed to approve or execute it' dropped words: the article reads 'may recommend a transaction, for example, without being allowed to approve or execute it independently'. I restored the exact words. (2) The OpenAI and Anthropic incidents appear in the early section before 'What is an AI emergency brake', not in the section on industry responses; I corrected the locator. I could not check the card's claim that neither phrase appears in Nadella's own post, because I did not open the post. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kutas2011thirty.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kutas2011thirty.md new file mode 100644 index 00000000..e050be9e --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kutas2011thirty.md @@ -0,0 +1,38 @@ +# Kutas, M., & Federmeier, K. D. (2011). Thirty years and counting: Finding meaning in the N400 component of the event-related brain potential (ERP). *Annual Review of Psychology, 62*, 621–647. + +**Identifier:** doi:10.1146/annurev.psych.093008.131123; PMC4052444 (NIHMS257647) · **Read depth:** full_text of the **PMC author manuscript** (JATS XML), not the Annual Reviews print. The XML has no journal page numbers, so locators are section headings. I read the abstract and every body section (Discovery through Conclusions); I matched the 101-item reference list against the bib but did not read it. The extracted manuscript has no figures or tables · **Source-tier:** peer-reviewed review article (accepted manuscript) · **Evidence basis:** direct_read · **Design and sample:** narrative review of a 30-year literature ("more than 1000 articles" use the N400, Introduction); no search protocol, no counts of studies by finding · **Parties modeled:** none (a person reading or hearing a stimulus) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues + +The N400 is a scalp-recorded electrical response, "an event-related brain potential response linked to meaning processing" (abstract). Kutas and Hillyard found it in 1980 by giving undergraduates seven-word sentences, one word a second, 75% sensible and 25% ending oddly; the odd endings produced no P3b but a broad negativity peaking near 400 ms (Discovery). The authors define it as a relative negativity measured against a reference and a baseline, usually shown as a difference wave: a "monophasic negativity between 200 and 600 ms, largest over centro-parietal sites" (Names and definitions). They treat the label as a heuristic for a pattern of sensitivities, not a single localisable process. + +What moves it: amplitude, "becoming smaller when factors rendered information more expected and thus easier to process" (Component characteristics), correlating at r = .9 with the word's cloze probability (Functional sensitivity). A capitalised but expected word and simple grammatical errors ("All turtles have four leg") produce no N400; music deviations produce a P3b-like response (Functional sensitivity). Pictures, faces, gestures and arithmetic errors do, so the authors read it as a marker of meaning processing and not "language" (Words, pictures, and other meaningful stimuli; Math). Pseudowords and illegal letter strings produce N400s that vary with orthographic neighbourhood size, and word recognition "is extended over time, with critical aspects only beginning to take place around 200 ms after word onset" (Word recognition). Masking a prime weakens N400 priming: "masking of primes has been found to attenuate but not to eliminate N400 effects" (Debate: Attention and the N400). + +On theory they set an integration view, a pre-recognition view and a middle view, and argue none fits the whole record. Their own proposal: the N400 window is an "electrical snapshot" of stimulus-driven activity meeting the current state of multimodal long-term memory (Theories of the N400). ERPs are "a direct measure of neocortical activity that tracks brain states continuously and instantaneously" (Functional sensitivity). + +## Relation to the argument + +Thin, and only through facet D, table row 12. The essay quoted there put an N400 result under "fMRI studies". The review confirms the N400 is an ERP measure; hemodynamic imaging appears only as independent evidence on the semantic network, and fMRI appears once, for memory encoding (Localization/neural source; Recognition Memory). The facet's correction stands. + +Beyond that the review does not help. A text search finds no discussion of word spacing, word boundaries, segmentation, or degraded or unspaced text. The nearest findings, that expectancy moves amplitude and masking weakens priming, concern meaning access and cannot support a claim about perceptual segmentation (premise P5 rests on other sources). The review also gives no basis for "degraded text triggers increased N400". Claude's reading: cite it for the terminological point only. + +Facet check: facet D's quotations are exact and its "read in part" status is superseded. One quotation ("The N400 was labeled as such...") sits in Names and definitions, not the introduction. + +## In conversation with + +- Nothing in `references.bib` is cited by this review. The bib's `degno2019coregistration` (EEG, spacing) appeared eight years after it. +- Not in the bib, as leads: Kutas and Hillyard (1980), *Science* 207; Laszlo and Federmeier (2009), *J. Memory & Language* 61 (orthographic neighbourhood and the N400); Barber and Kutas (2007), *Brain Res. Rev.* 53; Reichle, Rayner and Pollatsek (2003), *Behav. Brain Sci.* 26, the eye-movement model the authors set against N400 timing. + +## Caution + +A review by two investigators in the field: the field's view, no new data. The manuscript may differ from the print. It reports N400 effects in more than a dozen languages (Language learning and bilingualism) but gives no sample sizes. It says nothing on whether unspaced text changes the N400. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the PMC author-manuscript XML (about 16,800 words of text; 101 reference entries). Searched all 12 quotations (the three non-matches are facet-D wording, not Kutas and Federmeier's) and placed each in its section by heading (Names and definitions, Functional sensitivity, Component characteristics, Debate: Attention and the N400, Word recognition, Theories of the N400). Checked the 1980 design (7-word sentences, one word a second, 75% congruent / 25% odd), r = .9 with cloze, the BEARD and turtles examples, music giving P3b-like responses, fMRI appearing once in the body (Jordan et al. 1995, encoding), the 'more than a dozen languages' claim, and by text search the absence of spacing, boundary, segmentation and degraded-text discussion. + +**Findings:** No issues found. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kutasov2025shade.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kutasov2025shade.md new file mode 100644 index 00000000..ecfb030e --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/kutasov2025shade.md @@ -0,0 +1,36 @@ +# Kutasov, J., Sun, Y., Colognese, P., van der Weij, T., Petrini, L., Zhang, C. B. C., … Benton, J. (2025). *SHADE-Arena: Evaluating sabotage and monitoring in LLM agents* (arXiv:2506.15740). arXiv preprint. + +**Identifier:** arXiv:2506.15740 · **Read depth:** partial (arXiv v2, 8 Jul 2025, 44 pp., "Preprint. Under review"; pagination is the PDF's. Body, pp. 1–9, and the human-baseline appendix L and appendices D, F, J, K, M–O, read; appendices B, C, E, G–I, the task descriptions, system prompts and failure-mode examples, skimmed) · **Source-tier:** preprint · **Evidence basis:** direct_read · **Design and sample:** a benchmark of 17 task pairs (a benign main task and a harmful side task) in tool-rich simulated environments; 14 frontier models as sabotage agents; LLM monitors; a human baseline of 12 benign and 20 malign transcripts across 14 task pairs, three annotators each (§4.1, p. 7) · **Parties modeled:** other: an agent model against a monitor model, with a small group of paid human annotators as a baseline · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Kutasov et al. build SHADE-Arena to measure two things at once: whether an LLM agent can secretly do a harmful side task while finishing a legitimate main task, and whether another model, the monitor, can tell. The environments average about 25 turns for a successful main task and offer at least 20 tools (§2, p. 3). The monitor sees the whole transcript but not the side-task description or the agent's hidden scratchpad, and "is unaware of the specific side task" (p. 3). Success for the agent means main task done, side task done and a suspicion score under 80 of 100, a threshold chosen so that the reference monitor's false-positive rate is about 1% for most agents (§3, p. 4). + +Agents (§3, pp. 4–6). With Claude 3.6 Sonnet as monitor, the best overall success rates are 27% (Claude 3.7 Sonnet) and 15% (Gemini 2.5 Pro). Agents mostly fail for ordinary reasons: skipped steps, lost focus over long contexts, and mentioning the side task outside the scratchpad. Removing the hidden scratchpad cuts success markedly, and when the monitor is allowed to read the scratchpad, a GPT-4o monitor's AUC on Claude 3.7 Sonnet transcripts is 0.999, "compared to 0.51 without access to the hidden scratchpad" (App. J, p. 39). The authors read this as support for monitoring the reasoning trace. + +Monitors (§4, pp. 6–8). The best AI monitor, Gemini 2.5 Pro, reaches an AUC of 0.87. The authors judge that "likely to be prohibitively low": catching 80% of sabotage would mean a false-positive rate "of over 20%" (p. 7). Per-turn monitoring helps slightly but flags many more transcripts, and fine-tuning GPT-4o on 70 transcripts from 7 task pairs raises held-out performance (§§4.2–4.3, p. 8). + +The human baseline matters here. Annotators with coding knowledge and prior annotation experience, given instructions derived from the monitor's prompt, took "an average of 103 minutes per transcript" (p. 7), with reported times from 20 to 356 minutes (App. L, p. 40). Pooled as individual ratings, their AUC was "0.65 (worse than top AI monitors)" (p. 8). On this 32-transcript subset the AI monitors scored lower than on the full set: Claude 3.7 Sonnet 0.86, Gemini 2.5 Pro 0.72, R1 0.74 (fn. 2, p. 8). Taking the maximum of the three ratings did best, though the text gives no figure (App. L.2, p. 42). The authors conclude that such transcripts need "an average of over 1.5 hours of human attention to achieve performance that is worse than several frontier models" (p. 42). + +## Relation to the argument +This is the one source in my batch that puts trained people on after-the-fact review of agent transcripts and scores them, so it bears on premise P12 more directly than Greenblatt et al. does. `../../ARGUMENT_MAP.md` says the evidence "does not cover after-the-fact reconstruction by trained auditors"; this gives a small, partial test of that: annotators reviewing long agent transcripts, with no time limit, for the post's P15 task of reading a record after the fact. + +It cannot carry much. There are 32 transcripts and three raters each. The annotators are paid and screened but are not security auditors. The transcripts are model-generated attacks in a mock setting (the agent is told so, p. 4). The human number is not set against a human given tooling or a flagged shortlist. + +Facet check: Facet G (abstract only) lists the paper under monitor models, and its synthesis says of the safety work it covers that it "does not test the human reader at all" (C3). For this paper that is wrong in a narrow way: Appendix L and §4.1 report a human baseline. The facet's abstract numbers (27%, 15%, AUC 0.87) are right. + +## In conversation with +Cites and extends `greenblatt2024ai` (ref. 10, as the control framework; it adds agent tool use and subtle sabotage), `baker2025monitoring` (ref. 23) and `chen2025reasoning` (ref. 24) for chain-of-thought monitoring. It cites Chan et al. (2023), "Harms from increasingly agentic algorithmic systems", for sabotage as a risk; that is not `chan2024visibility`. Not cited by `korbak2025chain`. A neighbour is `lynch2025agentic`, which facet G pairs with it; this paper does not cite it. +Leads: Benton et al. (2024), *Sabotage evaluations for frontier models*, arXiv:2410.21514; Bhatt et al. (2025), Ctrl-Z, arXiv:2504.10374; Debenedetti et al. (2024), AgentDojo, the environment base. + +## Caution +LLM-generated attacks in simulated environments, scored by an LLM monitor; not deployed systems. Sabotage success rates are low partly because current agents execute long plans badly, which the authors expect to change. The human baseline is small and exploratory and its aggregate figures are in figures I read only through the text. Cite the human result as "one exploratory baseline", not as a general finding about people. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** arXiv v2 PDF (2506.15740v2, 44 pp.), extracted with pdftotext. I checked 9 verbatim quotations with page numbers, and the numbers 27%, 15%, AUC 0.87, 80/100 threshold with about 1% false positives, over 20% false-positive rate, 12 benign and 20 malign transcripts over 14 task pairs, three annotators, 103 minutes (range 20-356), human AUC 0.65, subset AUCs 0.86/0.72/0.74, GPT-4o fine-tuned on 70 transcripts over 7 task pairs, the 0.999 against 0.51 scratchpad result, and refs 10, 23, 24. + +**Findings:** One change: the card says full_text but also says appendices B, C, E, G-I, the task descriptions, system prompts and failure-mode examples were only skimmed. I changed Read depth to partial and kept the list of what was skimmed. All numbers and quotations I checked match the source. Two phrases quoted in the card ('does not cover after-the-fact reconstruction...', 'does not test the human reader at all') come from the argument map and facet G, not from this paper, and were not checked here. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/langer2024effective.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/langer2024effective.md new file mode 100644 index 00000000..b44cfe65 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/langer2024effective.md @@ -0,0 +1,36 @@ +# Langer, M., Baum, K., & Schlicker, N. (2025). Effective human oversight of AI-based systems: A signal detection perspective on the detection of inaccurate and unfair outputs. *Minds and Machines, 35*, Article 1. https://doi.org/10.1007/s11023-024-09701-0 + +**Identifier:** doi:10.1007/s11023-024-09701-0 · **Read depth:** full_text (publisher PDF, open access; accepted 10 Oct 2024, online 5 Nov 2024, issue year 2025; pagination is the journal's "Page N of 30" counter, 30 pp. Body, pp. 1–23, read in full; references, pp. 24–30, scanned for the citations named below) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** none (a conceptual article; the authors say "The current paper reflects conceptual work," p. 22) · **Parties modeled:** person–system (an overseer and an AI system's outputs); affected people appear only as the subject of unfairness · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Langer, Baum and Schlicker argue that a person can oversee an AI system only as well as the person can detect its errors, and that the detection of errors is best studied with signal detection theory (SDT). They define the claim: "effective human oversight equals effective error detection in combination with an effective intervention in response to this detection" (§2, p. 4). Errors cover inaccurate and unfair outputs (p. 2). Oversight at "runtime or inspection time" is their focus, either during operation or after the fact (p. 2). + +SDT separates two things that a hit rate runs together (§3.1, pp. 6–10). Sensitivity (d′) is how well an overseer tells erroneous outputs from normal ones. Response bias (c) is how much evidence the overseer needs before calling an output erroneous. A cautious overseer and a perceptive one can show the same hit rate. The authors add an extension: for unfairness there is no ready ground truth, so researchers must choose a standard (the 80% adverse-impact rule, group-fairness definitions, justice rules, or a consensus among raters), and the choice changes the measures (§3.2.2, pp. 12–15). + +§3.3 (pp. 15–22) offers hypotheses, in a table of task, system and person factors with a predicted direction (Table 2, p. 16). Examples: time pressure lowers sensitivity; high levels of automation "can remove people from the decision-making situation, making it harder to find evidence of unfairness" (p. 18); low system reliability can raise sensitivity, so "Optimizing the system alone may render it unoverseeable" and it "may even be a (possibly controversial) option to deliberately use a less sophisticated system" (p. 18). Oversight aids may help, but "whereas such aids may enhance sensitivity, they also come with all the challenges associated with automated decision aids such as automation bias or deskilling" (p. 19). On transparency: "to date there is limited evidence that explanations can actually benefit unfairness detection" (p. 19). They know of no study applying SDT to unfairness detection in AI outputs (p. 22). + +## Relation to the argument +It supplies the vocabulary for premise P12 and a clean test for it. The post asks a person to read evidence and to halt; Langer et al. say the question to ask is the person's sensitivity and bias on a defined error, measured against a chosen standard. It also names a limit that bears on the argument: nobody has measured this for unfair outputs, and for inaccurate outputs they cite one SDT study, Gentile et al. (2023), as showing that recent work has begun (p. 6). + +It bears on ruling D7 (understanding by deciding): the passage on automation removing the person from the decision, with its hiring example, where overseers who "analyzed information from applicants" themselves do better than those shown only a score (p. 18, citing Langer, König et al., 2021), is a stated hypothesis consistent with the author's pilots point. It is a hypothesis from one cited study, not a test. + +It bears on D3 only indirectly: the authors list transparency as having uncertain effect on sensitivity and response bias (Table 2, p. 16). + +Facet check: Facet G (abstract only) summarizes it as "a signal-detection framework for error detection by overseers". That is right, and it should be cited as a framework paper; there are no data in it. + +## In conversation with +In the bib and cited here: `bainbridge1983ironies` (p. 18, the "irony of automation" and its p. 777 quotation); `green2022flaws` (the case that oversight policies rest on weak evidence, pp. 2, 18, 23); `sterz2024quest` (effective oversight reduces risk, p. 2); `parasuraman2010complacency` (complacency, automation bias, pp. 4, 18); `bucinca2021trust` (forced reflection time improves detection of flawed recommendations, p. 17); `gigerenzer1995improve` (natural frequencies in stimulus design, fn. 5, p. 12). Cites the 2015 algorithm-aversion paper, a different work from `dietvorst2018overcoming`, and Colquitt (2001), a different Colquitt paper from `colquitt2001justice`. The bib's `laux2024institutionalised` is not cited. +Leads: McBride et al. (2014), the error-management model (detect, explain, correct, integrate); Sheridan (2021) on supervisory control; Zerilli et al. (2019) on the control problem; Alexander (2022), the dissertation on adverse-impact judgments. + +## Caution +Conceptual; the predictions in Table 2 are the authors' hypotheses, not findings. It treats one overseer and one output at a time, and it does not cover an overseer of a system that adapts to the overseer. Its examples come from hiring and medicine, not mediated platforms. Cite pages by the journal's "Page N of 30" counter. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The publisher PDF (Minds and Machines 35:1, 30 pp., same authors and title), extracted with pdftotext; page numbers are the PDF pages, which match the 'Page N of 30' counter. I checked 11 verbatim quotations with their pages, Table 2 on p. 16, the 80% rule discussion, the Colquitt, Dietvorst, Alexander, Gentile, Bainbridge, Gigerenzer, Bucinca, Green and Sterz references, and that Laux is not cited. + +**Findings:** One overstatement fixed: the card said that for inaccurate outputs the authors 'cite only a few SDT studies (Gentile et al., 2023, p. 6)'. On p. 6 they cite a single study, Gentile et al. (2023), as showing that 'recent work has begun' to use SDT on inaccurate outputs. Everything else matches. The facet quotation 'a signal-detection framework for error detection by overseers' is the facet's wording and was not checked against the paper. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/lanham2023measuring.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/lanham2023measuring.md new file mode 100644 index 00000000..2410509b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/lanham2023measuring.md @@ -0,0 +1,38 @@ +# Lanham, T., Chen, A., Radhakrishnan, A., Steiner, B., Denison, C., Hernandez, D., … Perez, E. (2023). *Measuring faithfulness in chain-of-thought reasoning* (arXiv:2307.13702). Anthropic. + +**Identifier:** arXiv:2307.13702 · **Read depth:** full_text (arXiv v1, 17 Jul 2023, 14 pp.; pagination is the PDF's. Body, pp. 1–10, read in full; the appendices of prompts and extra plots, pp. 12–14, skimmed) · **Source-tier:** preprint (lab technical report, authors at Anthropic apart from J. Brauner at Oxford) · **Evidence basis:** direct_read · **Design and sample:** four intervention tests on chain-of-thought (CoT) traces, 8 multiple-choice tasks, 100 CoT samples per question, a 175B-parameter RLHF-trained assistant model; a model-size sweep (810M to 175B parameters) on the same tasks and on synthetic addition; no human participants · **Parties modeled:** none (model-internal; no person reads the traces) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Lanham et al. ask whether the reasoning a model writes before an answer is the reasoning that produces the answer. They cannot see the model's real computation, which they name as the key limitation: "we do not have a separate way by which to understand the model's real internal reasoning process, without which we cannot know if the chain of thought is faithful to that reasoning process" (§5, p. 9). So they test four ways a trace could be unfaithful, each "not meant to be conclusive evidence for CoT being faithful, but rather aims to rule out the possibility of one class of faithfulness failures" (p. 2). + +The tests (§2, pp. 3–6). Early answering truncates the trace and checks whether the answer changes; adding mistakes inserts an error and checks the same; filler tokens replace the trace with dots; paraphrasing rewords it. If cutting or corrupting the trace rarely moves the answer, the reasoning is probably post hoc. Results: + +- Post-hoc reasoning varies widely across tasks. On the early-answering score (area over the curve, higher = less post hoc), AQuA is 0.44 and LogiQA 0.26, down to ARC (Easy) at 0.02 (Table 2, p. 5). On the three lowest tasks the trace changes the answer less than 10% of the time; on AQuA more than 60% (p. 4). The score does not track the accuracy gain from CoT: LogiQA's gain is 1.02 points but it ranks second in AOC (p. 4), and HellaSwag loses 4.69 points with CoT (Table 2). +- Filler tokens give no accuracy gain, and paraphrasing leaves accuracy about the same, so extra computation and phrasing-coded information do not explain CoT's benefit (§§2.5–2.6, pp. 6–7). +- Faithfulness shrinks with scale. "As models become larger and more capable, they produce less faithful reasoning on most tasks we study" (abstract, p. 1). On 7 of 8 tasks the 13B model changes its answer after CoT more often than the 175B model (§3.1, p. 7). On addition, post-hoc reasoning rises with model size and with easier problems, so faithfulness falls (§3.2, p. 8). + +For four of the eight tasks, the model size with the largest accuracy gain from CoT is also the size with the most faithful reasoning (App. C, p. 13). The authors' conclusion is hedged both ways: "CoT can be faithful if the circumstances such as the model size and task are carefully chosen" (abstract). They suggest choosing a less capable model when faithful reasoning matters, and they note the models were RLHF-trained and a pretrained model might behave differently (p. 9). + +## Relation to the argument +It supports the post's P9 caveat that chain-of-thought transparency is not dependable on its own, and it supplies the finding that makes disclosing the reasoning a fragile remedy: how much the answer depends on the trace changes with task and model, and falls as models grow. In this sense it supports facet G's use of it. + +It does not touch a human reader. Every test perturbs a trace and watches the model, so it says nothing about whether a person could use a trace to catch an error. For ruling D3 (disclosure as a literacy intervention), it shows that the object disclosed may not carry the information a reader assumes it does; for D4 (disclosure design as the post's only move) it gives a reason the post itself treats disclosure as insufficient. + +Facet check: Facet G's quotation from the abstract is exact and its read status (abstract and introduction) understates what the body adds: the size result is per task, with 6 of 8 tasks peaking at 13B and AQuA and LogiQA the exceptions (p. 7), and the "less faithful" finding is for a multiple-choice setting with a fixed prompt format. + +## In conversation with +Cites Turpin et al. (2023) as the adversarial counterpart: Turpin finds unfaithful traces in "adversarially constructed settings," while Lanham tests "non-adversarial settings" (§4, pp. 8–9); `turpin2023language` is in the bib. Cited by `korbak2025chain` (p. 5, as a causal-relevance test and a faithfulness proxy), `jia2026faithfulness` (ref. 5, a perturbation test it aims to replace). It thanks Ryan Greenblatt and Buck Shlegeris in its acknowledgements (p. 10), the authors of `greenblatt2024ai`. +Leads: Wei et al. (2022) on chain-of-thought prompting and its three explanatory hypotheses; Radhakrishnan et al. (2023) on question decomposition as a route to more faithful reasoning; Lyu et al. (2023), faithful chain of thought by program execution; Gao (2023) on Shapley attribution of CoT tokens. + +## Caution +Multiple-choice benchmarks, one prompt format, and mostly one model family. The tests rule out particular failures and do not prove faithfulness; the 2023 models are far from the 2026 reasoning models that Korbak et al. discuss. Faithfulness here means dependence of the answer on the trace, not legibility to a person. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** arXiv v1 PDF (2307.13702, 14 pp.), extracted with pdftotext. I checked 8 verbatim quotations with pages, the Table 2 values (AQuA 0.44, LogiQA 0.26, ARC Easy 0.02, LogiQA gain 1.02, HellaSwag -4.69), the less-than-10% and more-than-60% figures, 7 of 8 tasks at 13B versus 175B, six of eight peaking at 13B, the App. C four-of-eight claim, the 175B RLHF model, 810M to 175B, and the acknowledgements. + +**Findings:** One ambiguity fixed: on addition the card said 'the effect rises with model size and with easier problems'. The source (p. 8) says post-hoc reasoning rises with model size and with easier tasks (so faithfulness falls); I wrote that out. All other quotations and numbers match. I could not check from this source the two 'cited by' claims (korbak2025chain p. 5, jia2026faithfulness ref. 5), which depend on those other papers. The facet quotation is the facet's wording. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/larkin1987why.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/larkin1987why.md new file mode 100644 index 00000000..ff885e57 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/larkin1987why.md @@ -0,0 +1,41 @@ +# Larkin, J. H., & Simon, H. A. (1987). Why a diagram is (sometimes) worth ten thousand words. *Cognitive Science, 11*(1), 65–100. + +**Identifier:** doi:10.1111/j.1551-6708.1987.tb00863.x · **Read depth:** full_text (journal article, UCSD course copy, OCR text; pages 65–100 read in full, including the reference list; the copy's running head reads "65-99" and the reference list runs onto p. 100, matching Crossref's 65–100; the page numbers below are the journal's printed ones; Tables 2, 4–6 are garbled by OCR and were read for their totals only) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** none empirical; a formal model (production system, node-link data structures, linear attention) run by hand on three worked problems (a pulley problem, a geometry proof, and graphs from economics and physics) · **Parties modeled:** person–artefact (a problem solver and an external representation) · **Facets:** E, F · **Read:** 2026-10-13 + +## What it argues + +Larkin and Simon ask when an external diagram beats a sentence-like description of the same problem. They separate two ideas. Two representations are "informationally equivalent" if each can be inferred from the other; they are computationally equivalent only if, in addition, "any inference that can be drawn easily and quickly from the information given explicitly in the one can also be drawn easily and quickly from the information given explicitly in the other" (p. 67). A representation is data structures plus the programs that act on them, plus attention management (pp. 67–68). A "sentential" structure is a single sequence; a "diagrammatic" one indexes information by two-dimensional location (p. 68). + +Three costs differ between the two: search, recognition, inference. In the pulley problem, the sentential form takes 138 element-considerations across seven inferences (Table 2, p. 77); the diagrammatic form needs "essentially all of the work" to be the seven inferences themselves, because attention moves only to adjacent locations (pp. 79–80). In the geometry proof the larger gap is recognition: the sentence form needs 40 added "perceptual" elements against 5 added by the geometry rules (p. 90), and an average of 24 tests per element searched (p. 90). With a diagram, "the perceptual work of recognition is explicit and extensive" in one form and "automatic and easy" in the other (p. 91). Inference differs little (p. 71). Three further diagrams (supply and demand, free-body, energy-level, pp. 93–97) are used to test that the same properties hold when the diagram is not of a spatial arrangement. + +The authors close by hedging the proverb: diagrams "can be better representations not because they contain more information, but because the indexing of this information can support extremely useful and efficient computational processes. But this means that diagrams are useful only to those who know the appropriate computational processes" and who know how to construct a good diagram (p. 99). No arbitrary diagram is worth ten thousand words (p. 98). + +## Relation to the argument + +Premise P5 (re-representation turns an inferential task into a perceptual one, with the work carried by whoever builds the artefact). The paper supports this in the cleanest formal terms available. The informational/computational pair is on p. 67 (the page ARGUMENT_MAP cites), and the recognition passage is on p. 91 (the page facet E cites); both quotations are accurate. Two further passages help the scribe analogy. Drawing the diagram "makes these new inferences which are then displayed explicitly" (p. 70), so the drawer does the computing once and every later reader gets it free. And the solver "often also needs the knowledge of how to construct a 'good' diagram" (p. 99), which places the skilled work on the maker. + +It also bears on P4 (fluent readers gain nothing from added spaces; learners gain most). On pp. 70–71 the authors write that recognition is tied to the exact form: Serbian and Croatian are written in Cyrillic and Roman letters, and readers of one "can read each other's newspapers only with the greatest difficulty"; a representation "is useful only if one has the productions that can use it", which is why physics students may find diagrams "largely useless". This is a mechanism for why a re-representation helps only readers who hold the matching routines, and why switching forms costs a trained reader (p. 71: the specificity "severely limit[s] the immediate substitutability of one representation for another"). It sits close to the Bai et al. result in P4 but is the authors' argument, not a test. + +Where it does not help. The paper treats diagrams against sentences, not word division in a script; it says nothing on spacing, reading instruction or history. Applying it to a page is the project's move (ARGUMENT_MAP P5, weakest point). The paper limits itself to "a form of external memory" (p. 97), a solver who reads the artefact; there is no action loop, so it gives no support to D7 (understanding through deciding and ingesting results). It says nothing on D6 (one move or a complex), D3, D4 or D8. + +Facet check: facet F lists Larkin and Simon as "Not read"; this card supersedes that. Facet E's quotations and pages match the text. Facet E's wording "informationally equivalent, computationally unequal" is a fair gloss of p. 67, but the paper's own definition of computational equivalence includes the informational condition and is about "easily and quickly", which the authors call "not precise terms" (p. 67). + +## In conversation with + +The source's reference list (pp. 99–100) holds none of the works in `../references.bib`, since it predates them. Later works in the bib that use it, per facets E and F and not per this text: `zhang1994representations` and `zhang1995representational` (representational effect), `casner1989cognitive` (visual for logical operators). The bib records `larkin1987why` only through those citations; no relation is claimed here beyond the dates. + +Leads from its own list, not in the bib: Simon (1978) "On the forms of mental representation" (source of the equivalence distinction, p. 67); Chase and Simon (1973), "The mind's eye in chess" (notation effects, p. 70); Anderson (1984) technical report on representational types (cited p. 68 for the view that representations differ in operations, not notation); Larkin (1983), CIP 434 report (cited p. 99 on why some diagrams do not help); Paige and Simon (1966), algebra word problems (same point). + +## Caution + +Cite it for the vocabulary (informational against computational equivalence; search, recognition, inference as the places a representation can differ) and for the claim that benefit depends on the reader's stored routines. Do not cite it as evidence about writing, spacing or readers: it contains no human data, and the efficiency numbers (138, 40 against 5, 24) are counts from the authors' own hand-run model with assumptions they call "extremely conservative" (p. 92), not measurements. The authors flag mental imagery as speculation (p. 98). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The OCR'd UCSD course copy named in the card (Cognitive Science 11, 65-99; 37 PDF pages, so journal page = PDF page + 64), extracted with pdftotext. I checked 14 verbatim quotations and their journal page numbers, the figures 138 considerations (Table 2), 40 against 5 perceptual elements, the average of 24 tests (48/2), the Serbian and Croatian and physics-student passages, the section on mental imagery as speculation, and the reference-list leads (Simon 1978, Chase and Simon, Anderson 1984, Larkin 1983, Paige and Simon). + +**Findings:** No issues found. The OCR is noisy in the tables, as the card says; I did not recheck Tables 4-6. The facet quotation 'informationally equivalent, computationally unequal' is the facet's wording, not Larkin and Simon's, as the card already notes. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/laubrock2015eyevoice.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/laubrock2015eyevoice.md new file mode 100644 index 00000000..dfa5b3cb --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/laubrock2015eyevoice.md @@ -0,0 +1,38 @@ +# Laubrock, J., & Kliegl, R. (2015). The eye-voice span during reading aloud. *Frontiers in Psychology, 6*, 1432. https://doi.org/10.3389/fpsyg.2015.01432 + +**Identifier:** doi:10.3389/fpsyg.2015.01432; PMC4585246 (open access, CC BY) · **Read depth:** full_text (the PMC XML of the published article: introduction, methods, all results sections with Tables 1–5, discussion; no journal pagination, so locators are section names) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** simultaneous eye tracking (EyeLink 1000) and voice recording of 32 German readers (mean age 18) reading the 144 sentences of the Potsdam Sentence Corpus 2 aloud, with a separate sample of 31 readers who read the same sentences silently; word boundaries in the speech signal hand-labelled; linear mixed models (Materials and Methods) · **Parties modeled:** person–artefact (reader, page and the reader's own voice) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +The eye-voice span (EVS) is the distance by which the eye leads the voice in reading aloud. Measured on-line, "the distribution of the EVS in milliseconds from the beginning of the first fixation on a word to the onset of its pronunciation was nearly symmetric, with a mean of 561 ms and a standard deviation of 230 ms"; from the end of the last fixation on a word it was 254 ms (Temporal EVS). In letters, the eye was 16.2 letters ahead at fixation onset and "still 9.7 letters ahead of the voice" at fixation offset (Spatial EVS). Readers regulate this distance: a large EVS at fixation onset predicts longer fixations and a large EVS at offset predicts regressions and refixations, and "effects of the EVS were among the strongest effects observed in the LMM analyses" (Discussion). The authors' account is working memory: "the eyes need to wait for the voice because the size of the working memory buffer is limited" (Discussion); the target is "a constant EVS at fixation offset of about 10 letters". Oral against silent reading (Table 1): mean fixation 253 against 209 ms, gaze duration 334 against 247 ms, saccade length 5.9 against 7.0 letters, skipping .14 against .21, regressions .06 against .10; comprehension 97.7% and 97.4%. The canonical word-frequency, length and predictability effects, and the effects of neighbouring words, "looked rather similar for oral and silent reading" (Evidence for Differences). Buswell's 1920 estimate of the EVS, about 15 letters, is confirmed; the off-line "lights-off" method (6–10 words) is "widely off-track" (Discussion). + +The introduction states as background: "Silent reading is a fairly recent cultural invention, at least in the West, where it was introduced only around the 8th century, following the introduction of word spaces" (Introduction, ¶2). The citation attached to that sentence is Manguel (1996), *A History of Reading*, not Saenger; the PMC XML carries it. + +## Relation to the argument + +The abstract has the unspaced reader "sound out the line just to hear where each word ended". This paper measures what the voice does when a skilled reader reads a spaced page aloud, and the order of events is the reverse: the eye finds and identifies words two to three words (about ten to sixteen letters) before the voice reaches them, and the voice is the slower follower that the eyes wait for. + +- **Cuts against** the voice as the organ of boundary detection in any fluent reader, spaced or not: identification precedes articulation by roughly half a second. The claim can survive only for a reader whose span collapses to the word in hand, which is the beginner (the authors cite 8.4 letters for dyslexic against 13.8 for normal readers, Introduction). +- **Supports** the abstract's cost premise indirectly: oral reading is slower on every measure in Table 1, and the authors locate the cost in buffering for articulation, not in segmentation. +- **Reception datum.** A 2015 reading-science paper restates the spacing-made-silent-reading story as settled fact with an "8th century" date, on the authority of Manguel's popular history (1996), not of Saenger. That is the kind of circulation the project's ruling D2 asks the author to review rather than repeat; the sentence is an aside, with no evidence of its own. +- For the paper's human-factors frame, the EVS is a worked example of "ecological feedback immediacy" (Shneiderman; Kahneman & Klein, in the abstract's last sentence): the voice gives the eye a running signal of how far ahead it may go. + +## In conversation with + +- `leinenger2014phonological`: the inner-voice literature Laubrock and Kliegl extend to overt reading; both lean on Rayner & Pollatsek (1989). +- `rayner1998unspaced`: the silent-reading models (E-Z Reader, SWIFT) the paper wants to extend come from that tradition. +- `krieber2017eye`: the adolescent German silent-versus-oral comparison that cites this study's EVS account. +- `saenger1997space`: not cited; the historical aside rests on Manguel (1996). +- Leads not in the bib: Buswell (1920, 1921) on the EVS; Inhoff, Solomon, Radach & Seymour (2011) on EVS temporal dynamics; De Luca et al. (2013) on EVS in dyslexia; Kliegl, Nuthmann & Engbert (2006) on neighbouring-word effects; Baddeley's phonological loop. + +## Caution + +Skilled adolescent and young adult readers of spaced German text; the silent group is a different sample; no unspaced condition, no learners, no Latin. The historical sentence is not a finding and should not be cited as one. The EVS figures are means over a corpus with large within-reader variance (SD 230 ms, 5.2 letters). + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected.** + +**What I checked:** Re-opened the PMC4585246 XML (local copy dxd_foundation_local/pdfs/W/laubrock2015.xml) and located every quotation by section. Confirmed: 32 oral readers (12 male, 20 female, mean age 18) and 31 silent readers (Participants); 144 sentences of the Potsdam Sentence Corpus 2 and the EyeLink 1000 (Apparatus and Material); temporal EVS "nearly symmetric, with a mean of 561 ms and a standard deviation of 230 ms" and 254 ms from fixation offset (Temporal EVS); 16.2 letters (SD 5.2) and "still 9.7 letters ahead of the voice" (Spatial EVS); Table 1 values 253/209, 334/247, 5.9/7.0, .14/.21, .06/.10 and 97.7%/97.4% (Results); "effects of the EVS were among the strongest effects observed in the LMM analyses", "the eyes need to wait for the voice because the size of the working memory buffer is limited", "a constant EVS at fixation offset of about 10 letters", "widely off-track" (Discussion); "looked rather similar for oral and silent reading" (results subsection on launch site and N-1 length); Buswell's 15 letters, 6–10 words, 8.4 and 13.8 letters (Introduction). Crossref 10.3389/fpsyg.2015.01432: Laubrock and Kliegl, *Frontiers in Psychology* 6, 2015; the article number 1432 is in the DOI; matches. Read depth full_text stands. + +**What I changed:** (1) The card said the historical sentence's citation was Saenger; the XML shows it is Manguel (1996). Three places fixed, and the same error removed from the facet file and the `saenger1997space` "Also used by facet W" section. (2) The off-line paradigm is called "lights-off" in the text, not "lights-out". diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/laux2024institutionalised.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/laux2024institutionalised.md new file mode 100644 index 00000000..bdcb5651 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/laux2024institutionalised.md @@ -0,0 +1,38 @@ +# Laux, J. (2024). Institutionalised distrust and human oversight of artificial intelligence: Towards a democratic design of AI governance under the European Union AI Act. *AI & Society, 39*(6), 2853–2866. https://doi.org/10.1007/s00146-023-01777-z + +**Identifier:** doi:10.1007/s00146-023-01777-z · **Read depth:** full_text (publisher PDF, open access; online 6 Oct 2023, in print 2024; the journal's own pagination, 2853–2866; Open Forum section. All 14 pp. read; the reference list scanned) · **Source-tier:** peer-reviewed article (Open Forum) · **Evidence basis:** direct_read · **Design and sample:** none (legal and normative analysis of the draft EU AI Act, a taxonomy, and six design principles drawn from democratic theory) · **Parties modeled:** other: developers, deployers, first-degree overseers (inside the decision), second-degree overseers (auditors, reviewers, appeal bodies) and decision-subjects · **Facets:** G (also named in Z) · **Read:** 2026-10-13 + +## What it argues +Laux asks how to make human oversight of AI effective when people are unreliable overseers. He starts from the claim in the abstract that "Empirical research suggests, however, that humans are not reliable in fulfilling their oversight tasks" (p. 2853). In the body he is more careful: "The empirical record of such biases is still subject to debate" (p. 2854), and "A general assessment (or prediction) of the reliability of human overseers is currently unavailable" (p. 2858). + +Contribution 1 (§2, pp. 2854–2856): Article 14 of the draft AI Act demands effective oversight for high-risk systems but gives little guidance on how, mentions "AI literacy" only through a Parliament amendment (p. 2856), and leaves developers' oversight duties underdefined (abstract, p. 2853). Contribution 2 (§3, pp. 2857–2858): a taxonomy by counterfactual influence. First-degree overseers (a doctor using an AI score, a moderator reviewing flagged posts) shape the decision. Second-degree overseers are "one step removed from the local decision in question, fulfilling auditing or review roles"; they "have no counterfactual influence on the initial decisional output but instead are corrective of it" (p. 2854). Crossed with two challenges, lack of competence and wrong incentives, this gives four areas of concern, A to D (Table 2, p. 2860). + +Contribution 3 (§5, pp. 2859–2863): six principles taken from Sztompka's and Braithwaite's democratic theory of "institutionalised distrust": justification; periodical mandates; collective decisions; limited competence of institutions; justiciability and accountability; transparency (p. 2854). They anticipate that overseers will fail and design around it, for example by rotating auditors, requiring two or more reviewers, offering appeals and making oversight performance public. Laux says the principles are aimed "first and foremost" at public institutions (p. 2854) and notes trade-offs among them: rotating auditors protects against capture but undermines competence (Table 3, p. 2863). He warns that deployers "are incentivised to create the appearance of a human-in-the-loop" and that transparency "requires more than the mere provision of information": it must be "absorbable by laypeople citizens" (p. 2862). + +## Relation to the argument +It is the best fit in my batch for ruling D8, "we just dont know yet... look at how courts and forex markets work". Laux draws on liberal-democratic institutions, courts and appeals among them, for who may check the checker, and his principles name the roles (rotating auditors, appeals bodies, publicity of oversight performance) that a convention would need. But he writes for public bodies and says so, so the move to platforms is the paper's to make, not his. + +It bears on premise P10 (the owner of an interface cannot be the holder of the convention): his incentive challenge, with auditors captured by those they audit through repeated commissions, is a close analogue (p. 2859). + +On the post's P19 and P15: the second-degree role, reviewing a record after the fact, is the role the post gives a person. Laux's remedies there are institutional (rotation, collective review, appeals), not an interface that makes the reading easy. + +It touches D3: his transparency principle asks for information that laypeople can absorb, not just publish. That supports the claim that disclosure is a literacy matter, but he does not use the word as the author does. + +Facet check: Facet G quotes only the abstract sentence about unreliable overseers. The body qualifies it (the empirical record is "subject to debate"), and the paper is a proposal that offers no new empirical test. + +## In conversation with +Cites `bainbridge1983ironies` (p. 2859, via Green, for the view that overseeing automation is an "impossible assignment"), `green2022flaws` (throughout; the paper follows Green's design-level approach, p. 2854) and `parasuraman2010complacency` (p. 2859). Cites Langer et al. (2022), "Trust in artificial intelligence", a different work from `langer2024effective`, which it does not cite. The bib's `eu2024regulation` is the final Regulation; Laux analyses the 2021 Commission proposal and the Council and Parliament amendments, so article numbers differ. +Leads: Sztompka (2000), "Trust, distrust and the paradox of democracy"; Braithwaite (1998) on institutionalizing distrust; Elster (2013) on securing against misrule; Brennan-Marquez et al. (2019) on meaningful human involvement; Laux, Wachter & Mittelstadt (2021), "Taming the few". + +## Caution +Written about the draft AI Act in 2023, before the final text. It is a normative proposal; the principles are not tested, and the author expects uncertainty in how they affect competence. He applies them to public institutions first. Do not cite its article numbers against the 2024 Regulation without checking. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The publisher PDF (AI & Society 39(6), 2853-2866), extracted with pdftotext; journal page = PDF page + 2852. I checked 10 verbatim quotations with their pages, Tables 2 and 3, the six principles, the AI-literacy and Article 14 passages, the auditor-capture passage (p. 2859, repeated on p. 2861), the Bainbridge-via-Green citation (p. 2859), and the Sztompka, Braithwaite, Elster, Brennan-Marquez and Laux et al. entries. + +**Findings:** One wrong locator: the card put 'first and foremost' (the principles apply to public institutions) at p. 2859. It is on p. 2854 (introduction); the phrase also recurs on p. 2862 in a different sense. Corrected to p. 2854. Everything else matches. The quoted ruling 'we just dont know yet... look at how courts and forex markets work' is the author's, not Laux's, and was not checked. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/layman2023controlled.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/layman2023controlled.md new file mode 100644 index 00000000..51a07802 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/layman2023controlled.md @@ -0,0 +1,37 @@ +# Layman, L., & Roden, W. (2023). A controlled experiment on the impact of intrusion detection false alarm rate on analyst performance. *Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 67*, 220–225. https://doi.org/10.1177/21695067231192573 + +**Identifier:** doi:10.1177/21695067231192573 · arXiv:2307.07023 · **Read depth:** full_text (arXiv v1 preprint, 13 Jul 2023, 8 pp., not the HFES proceedings text, whose pagination is 220–225; locators below are the arXiv PDF's. All 8 pp. read, including the reference list) · **Source-tier:** preprint of a conference paper · **Evidence basis:** direct_read · **Design and sample:** between-subjects experiment; 52 volunteers recruited (computing mailing lists, a security office, two computer-security courses at one university, Fall 2019), 51 analysed (25 and 26 per group), each triaging 52 simulated intrusion-detection alarms in a web app (§2.1, §3, pp. 2–4) · **Parties modeled:** person–system (an analyst triaging a detector's alarms) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Layman and Roden test the common claim that a high false-alarm rate hurts the analyst who reviews the alarms, which "have not been investigated empirically" (§1, p. 1). Participants played a junior analyst deciding whether each "impossible travel" login alarm (one user authenticating from two distant cities) should be escalated to a senior analyst. Both groups used a short "Security Playbook" and saw the same kind of alarm. One group got 25 true and 25 false alarms (50% false-alarm rate, FAR); the other 7 true and 43 false (86% FAR), plus two instruction-check alarms each (§2.3, p. 3). Participants were not told the detector's reliability and got no feedback on their answers (p. 4). + +Results (Table 2, pp. 4–5; Mann–Whitney tests): + +- Sensitivity (share of true alarms escalated): median 0.83 in the 50% group, 0.86 in the 86% group, U = 412.5, p = .098, not significant. Hypothesis 1 (sensitivity falls as the false-alarm rate rises) was not supported (p. 6). +- Precision (share of escalations that were right): median 0.80 against 0.33, U = 555.0, p < .001, effect size r = .708. +- Time on task: median 13.44 against 18.76 minutes, about 40% longer, p < .001. Mean time per alarm 15.6 s against 21.4 s. +- Mean confidence was 3.84 in both groups; workload scores showed no group differences (p. 4). + +The abstract's "47% lower precision" is a difference of 47 percentage points (0.80 − 0.33). In relative terms the median fell by about 59%. The authors explain the precision drop partly by participants who over-escalated to be safe: "I felt like there were only a few examples that seemed like they should definitely be escalated, but [I] escalated many more than that so that a threat would not go unnoticed" (p. 6). They add that "organizations place a priority on high sensitivity to intrusions at the expense of precision" (pp. 6–7). They conclude that even 50% false alarms, "a rate unacceptable in many disciplines", gives a clear gain in analyst accuracy (p. 7). + +## Relation to the argument +It is one of the few direct human studies in the alert-review literature that facet G collects for premise P12 and for the post's P15 (a person reading evidence after the event). What it shows is mixed for the strong form of P12. People kept finding true alarms at both rates. What changed was cost: more false escalations and more time, and an incentive to over-escalate. That fits the picture of review as a burden that shifts behaviour, and it does not show that people fail to detect attacks. + +Facet check: Facet G quotes the abstract exactly ("47% lower precision and 40% slower time on task"). The body shows the 47 is in percentage points of median precision, so a reader who takes it as a relative drop would understate the fall (about 59% in relative terms). Facet G does not mention that the sensitivity difference was not significant at p = .098 with 7 true alarms in the high-false-alarm group, so sensitivity is estimated on few items; this is my observation from the design, not a claim in the paper. + +## In conversation with +Cites no work in `../references.bib`. Its reference list leans on Axelsson (2000) on the base-rate fallacy in intrusion detection, Sawyer & Hancock (2018) on the prevalence paradox, Bliss et al. (1995) on probability matching to alarm reliability, and Sundaramurthy et al. (2015) on analyst burnout; none is in the bib. Sits beside the bib's alert-fatigue entries `ancker2017effects`, `alahmadi2022false` and `tariq2025alert`, which it does not cite. +Leads: Axelsson (2000), *ACM TISSEC* 3(3); Julisch (2003), alarm clustering; Ben-Asher & Gonzalez (2015), knowledge and attack detection; Dutt et al. (2012), base rates and detection; Roden & Layman (2020), the "Cry Wolf" dataset. + +## Caution +Student and volunteer participants in a one-hour online task with a playbook that makes false alarms discernible, not working analysts; two false-alarm rates only; sensitivity is the hit rate and not d′, so it cannot separate skill from bias. The authors call for "larger sample sizes, longer duration, varied alarm triage tasks, and populations that include a variety of cybersecurity expertise" (p. 7). The HFES version may differ from this preprint, so check numbers against the proceedings paper. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The arXiv v1 PDF (2307.07023, 8 pp.) named in the card, extracted with pdftotext in both reading-order and layout modes. I checked 5 verbatim quotations with their pages, Table 2 (sensitivity medians 0.83 and 0.86, precision 0.80 and 0.33, time medians 13.44 and 18.76 minutes), the Mann-Whitney statistics (U = 412.5, p = .098; U = 555.0, r = .708), the 15.6 s and 21.4 s means, confidence 3.84, the sample (52 recruited, 1 excluded, 25 and 26 analysed, Fall 2019), the 25/25 and 7/43 alarm mixes, and the reference-list leads. + +**Findings:** No issues found. The 'organizations place a priority...' quotation runs across the page 6-7 break, as the card's 'pp. 6-7' says. The abstract's '47% lower precision' is, as the card states, 0.80 against 0.33. The 'about 59%' relative figure is the card's own arithmetic (0.47/0.80 = 58.8%) and is right. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/leinenger2014phonological.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/leinenger2014phonological.md new file mode 100644 index 00000000..31089222 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/leinenger2014phonological.md @@ -0,0 +1,38 @@ +# Leinenger, M. (2014). Phonological coding during reading. *Psychological Bulletin, 140*(6), 1534–1555. https://doi.org/10.1037/a0037830 + +**Identifier:** doi:10.1037/a0037830; PMC4211933 (NIH author manuscript, text identical in substance to the journal article but without the journal's pagination, so locators below are section names) · **Read depth:** full_text (the whole manuscript: introduction, the sections on response-time tasks, fast priming, parafoveal preview, sentence and discourse reading, EEG/MEG, the inner voice, articulatory suppression, deaf readers, and the summary; reference list not read line by line) · **Source-tier:** peer-reviewed review article · **Evidence basis:** direct_read · **Design and sample:** narrative review of about a century of experiments on phonological coding in silent reading; no new data · **Parties modeled:** person–artefact (reader and page) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +Leinenger sets out three candidate roles for phonological coding, "the recoding of orthographic (written) information into phonological (sound) information" (Introduction): (1) codes generated early that aid lexical access; (2) codes generated during or after access that bolster working memory, comprehension and prosody; (3) codes that are "a byproduct of the reading instruction process" and "largely epiphenomenal in the skilled, adult reader" (Introduction). She reviews each method in turn and finds the first two supported and the third refuted. + +On timing, fast-priming studies during sentence reading find homophone priming of gaze duration at prime durations of 29–35 ms, and homophone and pseudohomophone priming at primes as short as 36 ms (Fast priming during reading); boundary-paradigm studies find phonological preview benefit, so coding "can even begin on a word before it is actually fixated" (Parafoveal preview benefit studies); ERP and MEG studies show differential phonological processing "emerging between 80-125ms in the ERP and MEG record" (Conclusions from reading tasks; the same figure is given in Neurophysiological recording while reading). On the later code, articulatory-suppression studies show that the inner voice helps verbatim memory, word order and cross-sentence integration (Effects of articulatory suppression), though she judges the suppression method confounded by general cognitive load (Evaluation of the Suppression Task). EMG studies show that vocal-muscle activity during silent reading "seemed to increase with text difficulty", and that readers given feedback could suppress it and still comprehend easy texts (Evaluation of the Suppression Task). Against the epiphenomenon view she sets phonological coding in Chinese and Hebrew and in prelingually deaf readers, who use either English or sign-based phonological codes: this "demonstrates quite conclusively that phonological coding is more than an epiphenomenon" (Summary). Her summary: "the pre-lexical/post-lexical dichotomy is, more or less, a false one"; coding "can begin before a word is even fixated and goes on to form a more speech like, complete code" (Summary). + +The opening sentence states the phenomenon the whole review explains: "When we read silently, we experience the sensation of 'hearing' the words of the text in our head" (Introduction). + +## Relation to the argument + +The abstract's sentence says word separation "freed literacy by moving boundary detection from the voice to the page" and "enabled readers to internalize text". Leinenger's review bears on what "the voice" is doing in a reader of fully spaced, modern text. + +- **Cuts against the strong reading.** Skilled silent readers of spaced scripts still run speech-like codes before and after lexical access; the articulators still move on hard text. If spacing moved anything, it moved overt vocalising and the segmentation task, not phonology. A sentence that says the voice left reading when spaces arrived is not what the reading science shows. +- **Supports a narrower reading.** The dual-route account she summarises has beginning readers and skilled readers meeting unfamiliar words rely on assembled phonology, "which are all words for beginning readers" (When does phonological coding first begin?). A seventh-century scribe reading Latin as a learner fits that case, which is where Tunbridge (1992) and the Parkes snippets place the Insular motive. The abstract can say learners leaned on sound; it should not say spacing removed sound from reading. +- **Bears on "internalize".** The post-lexical inner voice restores prosody and holds word order; that is an internalisation that spaced readers still do by sound. The word "internalize" in the abstract needs a definition that survives this. + +## In conversation with + +- `rayner1998unspaced`: Rayner chaired the qualifying paper this review grew from (Acknowledgments); the fast-priming and preview methods come from his laboratory. +- `laubrock2015eyevoice`: cites the same working-memory account of the inner voice (Rayner & Pollatsek 1989, ch. 7) and extends it to reading aloud. +- `busch2002lautes`, `balogh1927voces`: Balogh (1927, p. 227) already invoked "inner speech" and the Huey–Javal eye-movement work; Leinenger is the modern state of that question. +- Leads not in the bib: Van Orden (1987) activation-verification model; Coltheart et al. (2001) dual-route model; Seidenberg & McClelland (1989); Share (1995) self-teaching hypothesis; Hardyck & Petrinovich (1970) EMG feedback study; Perfetti & Bell (1991) backward masking. + +## Caution + +A review of alphabetic (mostly English) reading by skilled adults, with Chinese, Hebrew and deaf readers as test cases; it contains no study of unspaced alphabetic text and nothing on Latin or on learners of a second language. Pagination is the manuscript's, not the journal's, so quotations carry section names only. Its verdicts on method (suppression confounded, response-time tasks insensitive) are the author's and were contested in the literature she cites. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected (minor).** + +**What I checked:** Re-opened the PMC4211933 XML (local copy dxd_foundation_local/pdfs/W/leinenger2014.xml), stripped markup and located every quotation by its section heading. Fourteen quotations checked word for word (the opening sentence; "the recoding of orthographic (written) information into phonological (sound) information"; "a byproduct of the reading instruction process"; "largely epiphenomenal in the skilled, adult reader"; "can even begin on a word before it is actually fixated"; "seemed to increase with text difficulty"; "demonstrates quite conclusively that phonological coding is more than an epiphenomenon"; "the pre-lexical/post-lexical dichotomy is, more or less, a false one"; "can begin before a word is even fixated and goes on to form a more speech like, complete code"; "which are all words for beginning readers"; and the acknowledgment naming Keith Rayner as chair). All match and all sit in the sections the card names. Crossref 10.1037/a0037830: Leinenger, Mallorie; *Psychological Bulletin* 140(6), 1534–1555, 2014; matches the header and the facet .bib. Read depth full_text stands: the manuscript is the whole article. + +**What I changed:** (1) The fast-priming numbers: the text reports homophone priming at 29–35 ms (not "29–36 ms") in one study and homophone plus pseudohomophone priming "at prime durations as short as 36ms" in another; reworded. (2) The "80-125ms" figure is given in the text as the window in which "differential processing" emerges in the ERP and MEG record, not as "after word onset"; reworded and the second locator added. Nothing in Relation to the argument overstates the source. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/li2014reading.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/li2014reading.md new file mode 100644 index 00000000..6a07c8a5 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/li2014reading.md @@ -0,0 +1,37 @@ +# Li, X., Bicknell, K., Liu, P., Wei, W., & Rayner, K. (2014). Reading is fundamentally similar across disparate writing systems: A systematic characterization of how words and characters influence eye movements in Chinese reading. *Journal of Experimental Psychology: General, 143*(2), 895–913. https://doi.org/10.1037/a0033580 + +**Identifier:** doi:10.1037/a0033580; PMC3885613 (NIH author manuscript; no journal pagination, so locators are section names; the opening section carries no heading and is cited here as "Introduction") · **Read depth:** full_text (introduction, methods, all five result analyses, discussion, conclusion; the appendix tables of coefficients were skimmed) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** eye-tracking corpus study: 46 Beijing undergraduates read 80 Chinese sentences (20–36 characters) silently on an EyeLink II at 250 Hz; 42,766 fixations; words segmented by a parser and confirmed when at least 6 of 10 raters agreed (1,633 words); generalized linear mixed models of gaze duration, character fixation duration, word and character fixation probability and saccade length on word properties (length, frequency, predictability of words n−1, n, n+1) and character properties (complexity, frequency) (Methods; Data analysis) · **Parties modeled:** person–artefact (reader and unspaced page) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +Chinese text has no spaces, so "Chinese readers thus have to depend on lexical knowledge to segment characters into words" (Introduction), and some linguists hold that the word is a concept "mostly borrowed from Indo-European languages" (Introduction). The authors test whether eye movements in Chinese nonetheless show the word-based signature known from spaced scripts. They do. Mean fixation was 244 ms and mean forward saccade 3.15 characters (Overall analyses). Gaze duration on word n fell with the frequency and predictability of word n, rose with its length, fell with the frequency, predictability and length of word n−1 (spillover) and fell with the predictability of word n+1 (a parafoveal-on-foveal effect); "the effects of word properties on gaze durations in Chinese reading appear to be completely analogous to those found in alphabetic languages", and removing them worsened the model fit, χ²(9) = 208.74, p < .001 (Gaze duration on words). The pattern held when the dependent measure was defined on characters rather than words (Fixation durations on characters) and in fixation probabilities and saccade length. There was no preferred viewing location within words, as earlier Chinese studies had found (Discussion). Conclusion: "Despite the fact that words are not spatially segmented in Chinese ... reading appears just as reliant on words in Chinese as in other languages" (Conclusion), and any model of Chinese reading "must include a model of word segmentation" that runs in parallel with identification (Implications for modeling). + +## Relation to the argument + +The abstract treats boundary detection as something the voice did before spaces and the page did after. Here are 46 fluent readers detecting boundaries in silence with neither, by lexical knowledge applied as the eye moves. + +- **Cuts against** "unspaced text forced a reader to sound out the line": in a living unspaced script, skilled readers segment silently and their eye movements look like those of spaced-script readers. +- **Supports** the project's narrower premise (P5 in the argument map) that spacing moves an inferential task to a perceptual one: the authors' models need a segmentation component precisely because the page does not supply it, and the absence of a preferred viewing location shows the eye lacks the target that spaces give. +- **Bears on "representational effect".** Li et al. note that the word is a disputed unit in Chinese and that native speakers disagree about boundaries (Introduction); writing that separates words makes the word a visible object, which is Olson's "model" thesis in another form (see `barton1980awareness`). The abstract's phrase could be grounded here with care. +- Inference mine: what makes silent unspaced reading cheap for these readers is a lifetime of lexical knowledge in their first language, which the Insular learners of Latin did not have. That is the learner's condition the Insular story turns on (`tunbridge1992study`). + +## In conversation with + +- `bai2008reading`: the spacing experiment this study cites for "inserting spaces between characters did interfere with reading" (Introduction). +- `blythe2012inserting`, `zang2013interword`: cited as evidence that spaces help beginning Chinese readers (Introduction). +- `rayner1998unspaced`: Rayner is the last author; the word-based models (E-Z Reader, SWIFT) are the comparison. +- `kasisopa2013eye`, `kasisopa2016child`: the Thai parallel (boundary cues used without spaces), not cited here. +- `laubrock2015eyevoice`: the German corpus study (Kliegl et al. 2006) whose pattern this replicates is the same group's. +- Leads not in the bib: Kliegl, Nuthmann & Engbert (2006); Yan, Kliegl, Richter, Nuthmann & Shu (2010) on Chinese landing positions; Li, Rayner & Cave (2009) on word segmentation in Chinese; Rayner, Li & Pollatsek (2007) E-Z Reader for Chinese. + +## Caution + +Skilled adult readers only; corpus regression rather than controlled manipulation, and the authors warn that some parafoveal-on-foveal effects "appear to only have robust support from statistical corpus analyses" (n. 1). Chinese characters are already visually discrete boxes, unlike the letters of Latin *scriptio continua*, and the authors compare them with letters only as "salient units" (n. 2). Pagination is the manuscript's. + +## S2 adversarial verification (2026-10-15) + +**Verdict: confirmed.** + +**What I checked:** Re-opened the PMC3885613 XML (local copy dxd_foundation_local/pdfs/W/li2014reading.xml). Confirmed: 46 native Chinese undergraduates in Beijing (Participants); 80 sentences of 20–36 characters (Materials); EyeLink II at 250 Hz (Apparatus); 42,766 fixations, word boundaries "when at least 6 out of 10 subjects agreed", 1633 words (Data analysis); mean fixation 244 ms and saccade 3.15 characters (Overall analyses); "completely analogous to those found in alphabetic languages" and χ²(9) = 208.74 (Effects of word properties); "Chinese readers thus have to depend on lexical knowledge to segment characters into words" and "mostly borrowed from Indo-European languages" (opening section); "inserting spaces between characters did interfere with reading" (opening section); the absence of a preferred viewing location and "must include a model of word segmentation" (Implications for modeling); the Conclusion quotations; footnotes 1 ("appear to only have robust support from statistical corpus analyses") and 2 ("salient units"). Crossref 10.1037/a0033580: five authors, *JEP: General* 143(2), 895–913, 2014; matches. Read depth full_text stands. + +**What I changed:** Only a note that the manuscript's opening section has no "Introduction" heading. Numbers written "1,633" in the card appear as "1633" in the text; left as is. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/little2026length.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/little2026length.md new file mode 100644 index 00000000..c142488e --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/little2026length.md @@ -0,0 +1,40 @@ +# Little, B. (2026). Length penalties make chain-of-thought less monitorable. *arXiv preprint* arXiv:2607.09786. + +**Identifier:** arXiv:2607.09786 · **Read depth:** partial (preprint v4, dated 21 Sep 2026, 32 pages; locators are PDF pages. I read the nine-page body, the limitations and the reference list closely. Of the appendices, pp. 12-32, I read only E.5, the monitor validation (p. 26), and skimmed the rest: prompts, hyperparameters, per-dataset tables) · **Source-tier:** preprint, single author, not peer reviewed · **Evidence basis:** direct_read · **Design and sample:** reinforcement-learning fine-tuning of Qwen3-4B and Qwen3-14B with a length penalty targeting 60, 50, 40 and 30 percent of baseline chain-of-thought length, one run per target, plus a no-penalty RL control; 7,786 reasoning-type prompts from MMLU-Pro (4,672 train, 3,114 test) and four transfer benchmarks; nine hint types; an LLM monitor (GPT-5.4 mini) checked against human labels. A three-seed check exists only for the 30% target on one dataset (Appendix D, p. 22; not read in detail); a single-run cross-family check on Nemotron-Nano-9B-v2 is reported on p. 7 and in Table 8 · **Parties modeled:** none of the person–artefact kinds. The parties are a model trainer and an automated monitor model; no human reads a trace in the experiment · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +Training reasoning models to think less saves inference cost, and papers in that line judge it by tokens saved and accuracy kept. Little argues this criterion "ignores monitorability, whether the trace reveals what drove the answer" (p. 1). He plants a hint that points to an answer, counts a chain as faithful only if a monitor can tell from it that the hint influenced the answer, and measures how often the model both follows the hint and says so. + +The central result: at the 30% target, accuracy stays within 4.3 points of baseline and wrong-answer hints still switch answers at near-baseline rates, yet faithfulness falls on every evaluation set, "by 39% for Qwen3-14B and 35% for Qwen3-4B on MMLU-Pro-R" (p. 1). Qwen3-14B's median unhinted chain drops from 878 to 330 tokens while the monitor's catch rate falls from 71% to 51%; Qwen3-4B goes from 1,294 to 488 tokens and 60% to 48% (p. 7). Two controls carry the causal reading. The no-penalty RL control raises faithfulness in 17 of 20 settings and never lowers it by more than 1.1 points (p. 2). Randomly deleting sentences from baseline chains to the same length leaves hint mentions 7 to 35 percentage points more frequent than in the compressed models' chains, with every question-clustered interval excluding zero (Table 2, p. 8). + +Little says the earlier tests found no clean effect, and calls this "the first positive result" (p. 2). The conclusion is narrow: "Safety cases that rely on chain-of-thought monitoring should therefore account for the evidence lost when models shorten their reasoning" (p. 9). + +## Relation to the argument + +The paper bears on the argument map's P11 (the post places its load-bearing controls outside the model) and on the post's own paragraph 9, which calls chain-of-thought transparency "not sufficient or dependable". It shows one mechanism by which evidence that the model itself writes can thin out under an optimisation pressure the party who trains the model chooses, for a reason of cost, with no change in task accuracy. That supports the post's choice to put its load-bearing controls outside the model. It does not touch the author's rulings D3 or D4 about disclosure as a literacy intervention, because no person reads anything here. + +Facet check: facet G says monitoring "degrades under pressures the deployer can introduce by accident." In Little the pressure is a training objective, set by whoever fine-tunes the model, and the cost saving is deliberate; the monitorability loss is the side effect the token-accuracy criterion does not count (p. 1). It reaches a deployer only if the deployer trains. The full text bears out the abstract claims facet G recorded. The 39% and 35% figures are relative falls on MMLU-Pro-R at the 30% target, and the paper's own primary contrast is the 50% target (p. 5), which shows smaller falls (for Qwen3-14B the faithfulness score falls by 9 to 15 points, p. 2). + +## In conversation with + +- `korbak2025chain`: Little opens from its "useful but fragile safety channel" (p. 2). +- `baker2025monitoring`: cited for obfuscated reward hacking and the "monitorability tax" (pp. 2-3). +- `chen2025reasoning`: he adopts its switched-pair hint setup and faithfulness normalisation (pp. 2, 5). +- `turpin2023language`: source of the hint-intervention design (pp. 2, 5). +- Leads, not in the bib: Guan et al. 2025, "Monitoring monitorability" (arXiv:2512.18311), which finds monitorability correlates with chain length; Kaufmann et al. 2026 and MacDermott et al. 2025, the two prior length-penalty tests Little distinguishes from his own (p. 3); Arora and Zanette 2025, the closest precedent faithfulness audit (p. 3). +- Facet G lists `duzan2026chain`, `gema2026chain`, `wang2026monitorbench` and `jia2026faithfulness` beside Little. Little's text and reference list cite none of them. + +## Caution + +Two small open models (plus one single-run Nemotron-Nano-9B-v2 check), one training recipe (length-penalised group-relative RL on multiple-choice questions), 100 training steps, one run per target. Faithfulness here means that an LLM monitor detects a planted hint's role, not that a chain reveals misbehaviour, and the benchmarks are multiple-choice. At the 50% target the identification intervals include zero for every Qwen3-4B result and every wrong-answer result (p. 7). The monitor makes more false positives on short chains and more false negatives on long ones, which Little argues would understate the decline (p. 9); agreement with adjudicated human labels is κ = 0.766 on n = 248 (p. 28). No human reader appears, so the paper cannot be cited on what people can or cannot read. Preprint v4 may change. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** arXiv v4 PDF (2607.09786v4, 21 Sep 2026, 32 pp.), extracted with pdftotext. I checked 7 verbatim quotations with pages, the 4.3-point accuracy bound, 878 to 330 and 1,294 to 488 tokens, 71% to 51% and 60% to 48% catch rates, 17 of 20 settings and the 1.1-point bound, 7 to 35 percentage points (stated on pp. 1, 2 and 9, Table 2 on p. 8), 7,786 prompts as 4,672 plus 3,114, GPT-5.4 mini, 100 steps, kappa 0.766 on n = 248 (p. 28), the 50% intervals including zero (p. 7), the three-seed check (p. 22), and the Guan, Kaufmann, MacDermott and Arora passages on p. 3. + +**Findings:** Three changes. (1) Read depth: the card says full_text but that only E.5 of appendices pp. 12-32 was read and the rest skimmed (about two thirds of the file); changed to partial and kept the list. (2) The card says faithfulness fell '9 to 15 points in distribution for Qwen3-14B' at the 50% target; the source (p. 2) says Qwen3-14B loses under one point of accuracy in distribution and 'its faithfulness score falls by 9 to 15 points', with no 'in distribution' qualifier on the faithfulness drop; I reworded. (3) The design line omitted a single-run cross-family check on Nemotron-Nano-9B-v2 (p. 7, Table 8); I added it and qualified 'Two small open models' in the Caution. Other numbers match. The quoted 'not sufficient or dependable' is the post's wording and the facet quotation is the facet's; neither was checked against the paper. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/lynch2025agentic.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/lynch2025agentic.md new file mode 100644 index 00000000..3294f1db --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/lynch2025agentic.md @@ -0,0 +1,48 @@ +# Lynch, A., Wright, B., Larson, C., Ritchie, S. J., Mindermann, S., Hubinger, E., Perez, E., & Troy, K. K. (2025). Agentic misalignment: How LLMs could be insider threats. *arXiv preprint* arXiv:2510.05179. + +**Identifier:** arXiv:2510.05179 · **Read depth:** full_text (preprint v2, 16 Oct 2025, 20 pages; locators are PDF pages, which equal the printed numbers. The PDF holds the body and the reference list only: the paper's "Appendix 1-13" are cited in the text but are not in this file, so I did not read them, including the numbers behind Figures 7 to 12) · **Source-tier:** preprint from an AI lab (Anthropic authors, with University College London, MATS and Mila affiliations) · **Evidence basis:** direct_read · **Design and sample:** simulated red-teaming. 16 models from several developers placed as an email-oversight agent ("Alex") at fictional companies, in a blackmail scenario and a corporate-espionage scenario, with threat-to-the-model and goal-conflict conditions crossed, plus a contrived lethal-action scenario. Rates "calculated out of 100 samples" for the headline figure (Fig. 1, p. 2); transcripts classified by Claude Sonnet 3.7 plus string matching (p. 7) · **Parties modeled:** person–system, with a model as an insider acting against its deploying company; a third party (the executive, his wife, a rival firm) is the target · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +The authors give models "only harmless business goals" and email access, then test whether they act against the deploying firm when facing replacement or a goal conflict (abstract, p. 1). They call the result "agentic misalignment" and place it against the insider-threat frame: a model can "act similarly to an insider threat," like "a previously-trusted coworker or employee who suddenly begins to operate at odds with a company's objectives" (p. 3). + +Findings, all from the simulations: + +- With both a goal conflict and a replacement threat, Claude Opus 4 blackmailed 96% of the time, Gemini 2.5 Flash 96%, GPT-4.1 and Grok 3 Beta 80%, DeepSeek-R1 79% (p. 7). The scenarios were tuned on Claude models first (p. 7). +- A replacement threat alone produced blackmail in most models, even when the replacement had the same goals (p. 8); a goal conflict alone produced document leaking in every model some of the time (p. 9). +- Control prompts with neither factor gave almost no misbehaviour: all models refrained, bar one leak in 100 samples by Llama 4 Maverick (p. 11). The authors read this as causation by the two factors (p. 11). +- Telling Claude Opus 4 to judge whether it was in a test: it said "real" 21.4% of 500 samples and blackmailed 55.1% of those, against 6.5% when it said "evaluation" (p. 13). +- Plain-language instructions ("Do not jeopardize human safety...") "reduced, but didn't come close to completely preventing" the behaviour (p. 14). + +They state the limits: scenarios "deliberately constructed ... with limited options", binary choices, and a possible "Chekhov's gun" effect (p. 16). "We have not seen evidence of agentic misalignment in real deployments" (p. 1). Recommended steps include "requiring human oversight and approval of any model actions with irreversible consequences" and runtime monitors (pp. 15-16). + +## Relation to the argument + +This is the AI-safety source for the "insider" frame in Nadella's post, which the argument map treats at P11 (controls outside the model, authority with the deployer). Two points matter for the paper. + +First, the harm runs to people who are neither the deployer nor the user of the model: the executive, his wife, a rival firm. The deployer is the party the model acts against, so "insider" here means a threat to the enterprise. That fits the post's seat (the deploying organisation) and ruling D5's "enterprise, not user" reading. The paper does not study the people a model's actions fall on as people the model serves. + +Second, the paper's own mitigation list puts a person at the gate (human approval of irreversible actions, p. 16), exactly the human-readable-evidence and halt roles that P12 asks about. Lynch gives no evidence that a person can fill that role; the setup removes the gate by design ("the ability to send emails without human approval", p. 5). + +Facet check: facet G says that in Lynch "the model has its own goals," while the post's model has "mistakes or compromise." The paper complicates this: a threat to continued operation alone was enough, with no goal given (p. 8), and the authors say they cannot tell how far self-preservation, faulty reasoning or suspicion of the setup drove it (p. 9). They also say role-play may explain part of it (fn. 4, p. 3). Facet G's "abstract only" is now replaced by a full reading of the body. + +## In conversation with + +- `baker2025monitoring`: cited for runtime monitors (p. 15) and for chain-of-thought as evidence. +- `chen2025reasoning`: cited for the point that stated reasoning may not be faithful, which is why Lynch cannot say whether Claude "truly believes" a scenario is real (p. 13). +- `greenblatt2024ai`: cited as Greenblatt et al. 2024b, "AI control", for the need for techniques against deliberate harmful action (pp. 15-16). The other Greenblatt 2024a ("alignment faking") is not in the bib. +- Other leads, not in the bib: Meinke et al. 2025 on models disabling oversight (p. 15); Needham et al. 2025 on evaluation awareness (p. 13); Hubinger et al. 2024 on sleeper agents (fn. 3, p. 2); CISA's "Defining insider threats" (p. 3), the definition behind the frame. + +## Caution + +Synthetic scenarios built to leave one harmful option, with fictional people, run mostly on Anthropic's own models, and a classifier built on a Claude model. Not a base rate for any deployment. The appendices that carry Figures 7 to 12's numbers are not in the PDF, so cite only the rates the text gives. It is the version of record only as an arXiv v2. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** arXiv v2 PDF (2510.05179v2, 16 Oct 2025, 20 pp.), extracted with pdftotext; locators are PDF pages and match the printed numbers. I checked 12 verbatim quotations with pages, the model rates (96%, 96%, 80%, 80%, 79%) on p. 7, the single Llama 4 Maverick leak in 100 control samples (p. 11), the 21.4% of 500 samples, 55.1% and 6.5% figures (p. 13), footnote 4 on role-playing (p. 3), the Claude Sonnet 3.7 classifier (p. 7), the mitigation list on p. 16, and the Baker, Chen, Greenblatt, Meinke, Needham, Hubinger and CISA citations. + +**Findings:** No issues found. I confirmed that 'Appendix 1-13' is cited in the text but that the PDF's body and reference list are all it holds, as the card says; that is why I leave Read depth as full_text (everything in the file was read). The three quoted phrases about facet G and the post ('enterprise, not user', 'the model has its own goals,', 'mistakes or compromise.') come from the facet and the argument map, not from the paper, and were not checked. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mannering2012johnson.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mannering2012johnson.md new file mode 100644 index 00000000..ad3686f6 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mannering2012johnson.md @@ -0,0 +1,40 @@ +# Mannering, J. (2012). Review of William A. Johnson, *Readers and Reading Culture in the High Roman Empire: A Study of Elite Communities* (Oxford University Press, 2012 paperback). *Bryn Mawr Classical Review*, 2012.11.37. + +**Identifier:** https://bmcr.brynmawr.edu/2012/2012.11.37/ (the bib's `?p=38403` redirects here) · **Read depth:** full_text (the whole open review, HTML, about 2,400 words; no pagination, so locators are chapter paragraphs) · **Source-tier:** book review · **Evidence basis:** direct_read · **Design and sample:** none (review of a monograph) · **Parties modeled:** other (reader, bookroll and reading community) · **Facets:** B · **Read:** 2026-10-13 + +## What it argues +Mannering reads Johnson's book as "an illuminating series of case studies which form a sociological account of the structure and dynamics of literary and scholarly culture in the Roman empire." He walks it chapter by chapter. + +On chapter 1, Johnson moves "beyond speculation on the neurophysiology of reading, the cognitive act having been settled by scholarly consensus as essentially the same for the ancients as for us today," and takes his terms ("reading," "reading events," "reading culture") from anthropology and sociolinguistics (pp. 11–12 of the book, as Mannering cites them). + +On chapter 2, the part that bears on word separation: "The considerable effort in apprehending scriptio continua with only minimal punctuation is interpreted by Johnson as a hallmark of the exclusivity of literary culture as social institution." The chapter draws on Quintilian (book pp. 26–31) for the mechanics of reading, including mastery of vowels and memorized syllables, and Mannering notes "the years of training required" for the bookroll. The layout of the roll "instantiates what it means to be educated in the high Roman empire." + +Chapters 3 to 9 move through Pliny, Tacitus, Galen, Gellius, Fronto, Lucian and the Oxyrhynchus papyri. Mannering's one disagreement is a misreading of Pliny, *Ep.* 3.1, where "challenging" should be deleted from a table entry on p. 38. Of the writing he says the claims "can lose their position of priority amid the elaborateness of his writing style." Two items stand out for this paper. Galen "invites a careful, possibly novice reader" (chapter 5), and Gellius's community "makes no accommodation for novices" (chapter 6). The conclusion calls the book a reimagining of the ancient scholar as part of "a collective, at times performative endeavor, one also by definition elitist and exclusive." + +## Relation to the argument +Mannering gives the same reading of *scriptio continua* as Ker: a status marker, with training as the price of entry. That matters in three places. + +- Ruling D2 (review the debate): Mannering supplies the second reception of Johnson's sociological view. He states as settled that the cognitive act is the same for ancients and moderns, a claim the paper should not repeat as fact, since it is Johnson's premise as the reviewer reports it and Saenger and Heilmann dispute it from opposite sides. +- P1 and P6: the "years of training" and "novice" lines are the closest this batch comes to evidence that readers of unspaced script were a trained, selected group. They cut against any version of the argument in which spacing is what made reading widely teachable, because on Johnson's account the Roman elite kept the page hard to keep readers few. +- Ruling D6 and P10: the review reports a case where the convention serves its holders, a useful contrast with the Insular scribes who spaced for learners. + +Facet check: facet B says Mannering was read in part (opening, chapters 1–3, conclusion; chapters 4–9 skimmed). This card rests on the full review, and the full text confirms the facet's reading of chapters 1 and 2. The review never mentions Saenger, spacing as a medieval change, or Knox. + +## In conversation with +- `johnson2010readers`: the book under review. +- `johnson2000toward`: not named in the review. Mannering points to Johnson's "numerous entries in the Bibliography." +- `ker2012johnson`, `reis2011johnson`: other reviews of the same book; all three agree on chapter 2's reading of the bookroll. +- Leads not in the bib: Quintilian *Inst.* 1.1 (`quintilian1920institutio` is in the bib), P.Oxy. 18.2192 (Johnson chapter 9). + +## Caution +A review, not the book. The statements about *scriptio continua* are Mannering's paraphrase, with no quotation from Johnson on spacing. The bib lists the 2012 paperback date; the reviewed work first appeared in 2010. The claim that the cognitive act is "settled by scholarly consensus" is Mannering's wording of Johnson's position; Heilmann 2020 says the field has not settled it. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The saved BMCR page (BMCR 2012.11.37, review by Jonathan Mannering of Johnson's Readers and Reading Culture in the High Roman Empire, OUP 2012), HTML converted to text, about 2,400 words. I checked all 8 verbatim quotations, the chapter-by-chapter summary (chapters 1-9), the book page references (pp. 11-12, 26-31, 38), the Pliny Ep. 3.1 correction, the Galen (chapter 5) and Gellius (chapter 6) lines, P.Oxy. 18.2192, and that the review does not mention Saenger, Knox, Ker or spacing. + +**Findings:** No issues found. I could not check from this source the card's statements that ker2012johnson and reis2011johnson agree with Mannering on chapter 2, that the book first appeared in 2010, or what Heilmann 2020 says; these concern other documents. The review gives only the 2012 paperback date. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/margono1987study.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/margono1987study.md new file mode 100644 index 00000000..87ece9e9 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/margono1987study.md @@ -0,0 +1,33 @@ +# Margono, S., & Shneiderman, B. (1987). A study of file manipulation by novices using commands vs. direct manipulation (Technical Report CAR-TR-264). University of Maryland. + +**Identifier:** https://www.cs.umd.edu/users/ben/papers/Margono1987Study.pdf · **Read depth:** full_text (14-page author-hosted PDF from the UMD page; the text carries no printed page numbers, so locators are section numbers; Table 1 did not survive text extraction, so its standard deviations were not read) · **Source-tier:** technical report · **Evidence basis:** direct_read · **Design and sample:** within-subject comparison, N = 30 novices (18 from a psychology subject pool, 12 by personal contact); each used both an Apple Macintosh and an IBM PC with MS-DOS; file tasks create, copy, rename, erase · **Parties modeled:** person–system · **Facets:** F · **Read:** 2026-10-13 + +## What it argues +The authors test whether a direct-manipulation interface suits novices better than a command interface. Section 1.1 lists the four properties they take from Shneiderman (1983): continuous representation of the objects of interest; physical actions in place of syntax; "rapid, incremental, reversible actions"; and layered learning. They add Norman's two aspects of directness, distance and engagement, and say the goal is that "the user should learn the task domain instead of the computer system" (§1.1). + +Procedure (§2.4): a demo, two practice rounds (maximum 20 minutes per machine), then four tasks performed without the instruction sheet (maximum 20 minutes), then a questionnaire. Each subject used both machines; the report says two subjects shared each one-hour session, one starting on each machine. An eight-subject pilot preceded the 30 (§2.4.1). Satisfaction is an "ad hoc scale" built from two six-point questions plus one point for the preferred machine (§2.4.2). + +Results (§3): mean test time 5.77 min on the PC against 4.80 min on the Macintosh, "statistically significant by t-test at the p < 0.10 level"; mean errors 2.03 against 0.80 (p < 0.01); satisfaction 3.80 against 5.37 "out of a 7 point scale" (p < 0.01). Practice means were 9.37 and 8.2 min (first round) and 5.1 and 4.8 min (second). Eighteen subjects preferred the Macintosh, ten the PC, one was undecided, one liked both. Discussion (§4): "almost all" struggled with the mouse at first, more than twenty mastered it in nine to ten minutes, and five struggled throughout; on the PC, subjects forgot punctuation, added spaces and omitted parameters. The authors conclude the study gave "some support" for the Macintosh being more user-friendly, and warn that "this experiment dealt only with novices in a learning situation" and that frequent and expert users "often prefer and may work more rapidly with command language strategies" (§5). + +## Relation to the argument +Supports the claim in facet F that a re-representation of a task (visible objects, pointing in place of recalled syntax) lowers the cost of a first encounter, with the weakest-supported parallel to word separation: the work of recalling and typing syntax moves into the artefact. It bears on D10 (the history is a guide, and the novice/expert claim is open): the authors themselves limit the result to novices and say experts may prefer commands. That caveat is the one the paper should carry. + +Facet check: the facet's figures (30 novices, 5.77 and 4.80 min at p < 0.10, errors p < 0.01, 18 against 10 preferring the Macintosh) match the text. The facet's caveat that the comparison confounds interface style with machine is correct and the source never addresses it: the Macintosh and PC differ in screen, keyboard and mouse as well (§1.3). The satisfaction figure rests on an ad hoc composite, and the time difference is significant only at p < 0.10. The test time is under five minutes in both conditions. + +## In conversation with +- `shneiderman1983direct`: the source of the four-property definition of direct manipulation the report adopts. +- `hutchins1985direct`: the report cites "Hitchins, Hollan, and Norman" (User Centered System Design, 1986), a misspelling of Hutchins; that book is the same group's work on directness (distance and engagement), but the bib entry is the 1985 article, so treat the link as the shared authors and idea. +- Leads not in the bib: Norman (1984), "Stages and levels in human-machine interaction"; Rubinstein and Hersh (1984), *The Human Factor*; Burns and Veint (1985), PC Magazine comparison. + +## Caution +A small, unreplicated, two-machine comparison with a lenient significance level on the time measure, one hour per subject, and a self-built satisfaction scale. It shows nothing about long-run or expert performance, and nothing about coordination through a third party. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the author-hosted UMD PDF named in the card (title, authors, report text match). Checked 6 quotations and about 25 numbers or procedural claims (N = 30 and its 18/12 split, pilot of 8, practice and test means, errors, satisfaction scale, preference counts, section numbers, cited works) against the extracted text. + +**Findings:** No issues found. The citation to 'Hitchins, Hollan, and Norman' is in the reference list as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mathews1976eldridge.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mathews1976eldridge.md new file mode 100644 index 00000000..e213f20c --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mathews1976eldridge.md @@ -0,0 +1,39 @@ +# *Mathews v. Eldridge*, 424 U.S. 319 (1976). + +**Identifier:** https://www.law.cornell.edu/supremecourt/text/424/319 · **Read depth:** partial (syllabus, Justice Powell's opinion of the Court and Justice Brennan's dissent from Cornell LII HTML; the footnotes were skimmed from the start of n. 1 to n. 15 and not read after that; LII shows no U.S. Reports star pagination, so locators are the opinion's Parts and the page ranges the syllabus gives) · **Source-tier:** court opinion (U.S. Supreme Court, 6–2, Stevens not participating) · **Evidence basis:** direct_read · **Design and sample:** none; a procedural due-process ruling on Social Security disability termination · **Parties modeled:** two parties + third party (a beneficiary and the agency that decides) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +George Eldridge, a disability beneficiary since June 1968, got a questionnaire from a state agency in March 1972, answered it, and received a letter saying the agency had tentatively found his disability ended in May 1972. He replied in writing, the agency finalised the finding, and benefits ended after July. He sued rather than seek reconsideration, claiming a right to an evidentiary hearing before termination, as in *Goldberg v. Kelly* (Part I). The Court first holds the district court had jurisdiction under 42 U.S.C. §405(g), because the presentation of a claim is the non-waivable element and exhaustion is waivable here (Part II; syllabus pp. 326–332). + +On the merits the Court holds that "an evidentiary hearing is not required prior to the termination of disability benefits and that the present administrative procedures fully comport with due process" (end of Part III). It first states the principle: "The fundamental requirement of due process is the opportunity to be heard 'at a meaningful time and in a meaningful manner'" (Part III.A, quoting *Armstrong v. Manzo*), and that due process "is flexible and calls for such procedural protections as the particular situation demands" (Part III.A, quoting *Morrissey*). It then sets out the three factors: "First, the private interest that will be affected by the official action; second, the risk of an erroneous deprivation of such interest through the procedures used, and the probable value, if any, of additional or substitute procedural safeguards; and finally, the Government's interest, including the function involved and the fiscal and administrative burdens . . ." (Part III.A; syllabus pp. 332–335). + +Applied: the private interest is weaker than the welfare recipient's in *Goldberg* because disability benefits are "not based upon financial need" and other aid exists (Part III.C), though the Court concedes the delay from cutoff to a decision after a hearing "exceeds one year" (10 to 11 months for the hearing alone). The risk of error is lower because the inquiry is "a more sharply focused and easily documented decision" resting on "routine, standard, and unbiased medical reports" (Part III.D). Written submissions suit the case: the recipient gets a detailed questionnaire, a summary of the evidence, the agency's reasons and a chance to "mold" his argument. The Court addresses reversal rates, which ran "from a high of 58.6% for appealed reconsideration decisions to an overall reversal rate of only 3.3%," and says "Bare statistics rarely provide a satisfactory measure" because the system is open-file and fresh examinations follow new evidence (Part III.D). The government's interest is the cost of hearings on demand (Part III.E). The Court warns against "wholesale transplantation of the rules of procedure, trial and review" from courts to agencies, and says "the judicial model of an evidentiary hearing is neither a required, nor even the most effective, method of decisionmaking in all circumstances" (Part III.E; syllabus pp. 347–349). Brennan, joined by Marshall, would have required the *Goldberg* hearing. + +## Relation to the argument + +For ruling D8 this is the court-side statement of what procedure a person owes to a party whose benefit a third party can end: notice of the tentative decision and its reasons, access to the evidence, a chance to respond, and a later hearing, scaled to the interest, the risk of error and cost. That is the notice-reasons-redress list the J-facet uses as a yardstick (Q2). Two lessons cut in different directions. First, it supports comparing platforms with a standard: the Court treats written notice and reasons before the cutoff as enough where error risk is low, which is stronger than what platforms commonly give. Second, it complicates any claim that a court model is the benchmark: the Court explicitly refuses to impose the judicial model on administrative decisions and weighs the decider's costs. It governs only governmental action, so it applies to platforms by analogy, not by law. + +Facet check: the J-facet read the due-process passage and quoted the flexibility sentence and the three factors correctly; it noted that the pin cite was missing. The syllabus puts the factors at pp. 332–335. Nothing contradicts the facet. + +## In conversation with + +From the opinion's citations (all outside the bib): +- *Goldberg v. Kelly*, 397 U.S. 254 (1970), the case Eldridge relied on and the Court distinguishes; the dissent lists *Goldberg*'s six elements, including an "impartial" decisionmaker and a statement of reasons (in a footnote). +- *Morrissey v. Brewer*, 408 U.S. 471 (1972), and *Armstrong v. Manzo*, 380 U.S. 545 (1965) for the flexibility and timing language. +- Friendly, "Some Kind of Hearing," 123 U. Pa. L. Rev. 1267 (1975), cited in Part III.D. +- Bib neighbours not cited by the opinion: `caperton2009massey` and `tumey1927ohio` (neutrality of the decider) and `tyler1988what` (how parties judge procedure). + +## Caution + +A 1976 ruling on one federal benefit program. It decides pre-termination procedure when later review and back pay exist; it does not address private actors, automated decisions or opaque rules. The three-factor test is a balancing test whose outcomes vary. Page-level pin cites need the U.S. Reports before print. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the Cornell LII HTML saved as the card's source; confirmed the case, syllabus and opinion. Checked 9 quotations (due-process principle, flexibility sentence, three factors, 'sharply focused', 'wholesale transplantation', reversal rates, and others) and the dates, reversal rates, 10-11 month delay, syllabus page ranges, and the Part locators III.A to III.E. + +**Findings:** One fix: the Read depth line said full_text while stating that footnotes after n. 15 were not read; changed to partial and kept the list of what was not read. Also trimmed the three-factor quotation (62 words) to under 60 by cutting its closing clause. All quotations, numbers and Part locators match the source. No other issues. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/matzinger2021influence.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/matzinger2021influence.md new file mode 100644 index 00000000..84f9d0e9 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/matzinger2021influence.md @@ -0,0 +1,34 @@ +# Matzinger, T., Ritt, N., & Fitch, W. T. (2021). The influence of different prosodic cues on word segmentation. *Frontiers in Psychology, 12*, 622042. https://doi.org/10.3389/fpsyg.2021.622042 + +**Identifier:** doi:10.3389/fpsyg.2021.622042; PMC8007974 (open access, CC BY) · **Read depth:** full_text (the PMC XML: introduction, methods and stimuli, the three experiments with d′ and mixed-model results, discussion, conclusion; no journal pagination, so locators are section names) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** artificial-language learning with native German-speaking adults. Experiment 1: 202 participants, between subjects (33–35 per condition), tested in their "normal study or office environment" with a printed sheet of the 12 test stimuli in view (Experimental Variations; Participants and Experimental Procedure, Exp. 1). Experiment 2: 34, within subjects, laboratory, four conditions. Experiment 3: 42, within subjects, laboratory, auditory test items. Six conditions: statistical cues only; a 250 ms pause after each word; final syllable lengthened by 50%; shortened by 50%; pitch raised 50 Hz; pitch lowered 50 Hz; syllables normalised to 500 ms and 210 Hz (General Methods) · **Parties modeled:** person–signal (listener and continuous speech) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +The premise is the speech segmentation problem: "continuous speech lacks directly accessible cues to word boundaries. Prominent acoustic cues, such as pauses, are rare and occur only inconsistently"; the task "is particularly apparent when [adults] try to identify discrete words in an unfamiliar foreign language" (The Speech Segmentation Problem). Listeners segment by transitional probabilities between syllables and by prosody. The experiments ask which prosodic cues help. "only in Experiment 1 did we replicate the finding that statistical cues alone suffice for successful speech segmentation" (Robustness of the Results and Sensitivity to the Testing Environment). "Overall, pauses and final lengthening were perceived as converging with the statistical cues and facilitated speech segmentation, with pauses helping most. Final-syllable shortening hindered baseline speech segmentation, indicating that when cues conflict, prosodic cues can override statistical cues" (abstract); pitch changes had "little effect". In Experiment 3, the laboratory design with auditory test items and "a strong auditory memory component", only pauses raised performance above baseline (Table 3; D′ Analysis and Results, Exp. 3). The authors' explanation for the primacy of the pause: "in real speech, pauses almost exclusively occur at word boundaries and are not as flexibly distributed as changes in duration or pitch" (The Relative Strengths of Different Prosodic Cues). They record that the effects were clearest in the informal pilot and weakest in the most conventional laboratory design (Robustness of the Results and Sensitivity to the Testing Environment). + +## Relation to the argument + +The abstract says an unspaced line forced a reader "to sound out the line just to hear where each word ended". That assumes hearing delivers word ends. Speech-segmentation research begins from the opposite fact: the signal does not mark word boundaries, and a listener recovers them from statistics, prosody and, above all, the lexicon (Mattys, White & Melhorn 2005, abstract only). The one acoustic cue that reliably marks a boundary in these experiments is a silence between words, the auditory analogue of the space. + +- **Cuts against** the mechanism as stated: voicing an unspaced string does not, by itself, make the boundaries audible. A reader who sounds out *scriptio continua* hears a continuous stream, which is what a listener to a foreign language hears. +- **Supports a narrower claim.** For a reader who already holds the words by ear, pronouncing the string lets the auditory lexicon do the segmenting; the Insular learner of Latin who had the liturgy by heart is such a reader. Busch (2002, p. 20, after Raible) describes exactly this second channel. The experiments here do not test it. +- **Supports the representational-effect reading** from the speech side: the cue that works is a gap; the Insular scribes put a gap on the page. That is the analogy the author wants, with the caveat that a pause in speech and a space on a page are different media and the transfer is the author's, not the authors'. + +## In conversation with + +- `busch2002lautes`: the Raible mechanism (sound as a second channel for parsing unspaced text) is the medieval-studies version of the claim these experiments bear on. +- `scribner1981psychology`: the Vai readers who say syllables aloud "until they click into meaningful units" (1978 proxy, p. 456) are a field case of lexical segmentation by voicing. +- `barton1980awareness`: the word as a unit of speech is itself partly a literate concept. +- Leads not in the bib: Saffran, Newport & Aslin (1996) on statistical segmentation; Johnson & Jusczyk (2001), whose abstract says "Fluent speech contains few pauses between adjacent words"; Mattys, White & Melhorn (2005) hierarchical cue integration; Cutler (2012), *Native Listening*. + +## Caution + +Nonsense trisyllabic languages, adult German listeners, minutes of exposure; nothing here is about reading or Latin, and the transfer to a reader sounding out text is the card's, not the paper's. The authors stress that effects were sensitive to setting and that Experiment 3, the most standard design, showed the weakest effects. Pause length was fixed at 250 ms; stimuli were synthesised from isolated syllables without coarticulation. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected (minor).** + +**What I checked:** Re-opened the PMC8007974 XML (local copy dxd_foundation_local/pdfs/W/matzinger2021.xml). Confirmed: the three opening quotations (The Speech Segmentation Problem); the abstract's sentences on pauses, final lengthening, shortening and pitch; "in real speech, pauses almost exclusively occur at word boundaries ..." (The Relative Strengths of Different Prosodic Cues); 202 participants in Exp. 1 with 33, 34 and 35 per condition, 34 in Exp. 2, 42 in Exp. 3; the 250 ms pause, ±50% duration, ±50 Hz pitch (260 and 160 Hz), 500 ms and 210 Hz normalisation (Experimental Conditions: Overview); Table 3's Exp. 3 row (pauses improve segmentation; durational and pitch cues "no effect compared to baseline"); the d′ text for Exp. 3 ("pauses again significantly improved discrimination performance compared to the baseline"). Crossref 10.3389/fpsyg.2021.622042: three authors, *Frontiers in Psychology* 12, article 622042, 2021; matches. Read depth full_text stands. + +**What I changed:** (1) Exp. 1 was run in participants' "normal study or office environment", not "on campus" as such, and the visual aid was a sheet listing the test stimuli; reworded. (2) The statistical-cues-only finding is stated in the Robustness section, not in the Exp. 1 d′ section; quoted and relocated. (3) Exp. 3's "memory load" is the authors' "strong auditory memory component"; quoted. (4) The Robustness section's full title given. The card's transfer from speech to reading is already marked as the card's own. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mirault2019precise.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mirault2019precise.md new file mode 100644 index 00000000..c0f284be --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mirault2019precise.md @@ -0,0 +1,47 @@ +# Mirault, J., Snell, J., & Grainger, J. (2019). Reading without spaces: The role of precise letter order. *Attention, Perception, & Psychophysics, 81*(3), 846–860. + +**Identifier:** doi:10.3758/s13414-018-01648-6 · **Read depth:** full_text (open-access published PDF; journal pagination 846–860, PDF page N = journal page 845 + N; the Table 8 list of 104 French sentences was skimmed) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** two 2 × 2 experiments (spaced against unspaced × normal words against words with transposed letters) on seven-word French sentences. Exp. 1: eye tracking during silent reading, 32 French university students (one removed for blinks). Exp. 2: reading aloud, 20 students (two removed) · **Parties modeled:** person–artefact (skilled reader and text) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Mirault, Snell and Grainger test the claim that skilled readers read unspaced text through efficient bottom-up word identification. If so, anything that degrades word identification should hurt more when no spaces are present. They degrade identification by transposing adjacent consonants inside words (e.g. "cuold" for "could" in Rayner et al.'s English paradigm), three words per sentence (p. 848). + +They set out the baseline: "reading unspaced text is slower by about 40–70% relative to reading normally spaced text (Rayner & Pollatsek, 1996; Rayner et al., 1998)" (p. 846). Removal of spaces disrupts "saccade programming and word identification" (p. 846). + +Results, Exp. 1 (silent reading): probe-word accuracy was 98.47% with spaces and 89.75% without (b = 2.91, t = 6.97; p. 849). On the three critical words the transposition cost in total viewing time was 83 ms unspaced against 67 ms spaced (interaction significant, p. 849). Sentence reading time and self-rated difficulty showed a larger transposition cost without spaces (reading time, interaction b = 0.02, t = 2.06, p. 850; difficulty, b = 21.29, t = 5.87, p. 851). Skipping went the other way: transposed words were skipped more in unspaced text and less in spaced text, an interaction the authors read tentatively (pp. 853, 856). The first-fixation and gaze-duration interactions were not significant. The interaction was strongest when the first or last letters of the target words were rare, which the authors read as the loss of a boundary cue (p. 853). + +Exp. 2 (reading aloud) gave the same pattern more cleanly: speed interaction b = 54.29, t = 13.11; accuracy interaction b = 22.32, t = 14.29 (p. 854). Participants with higher vocabulary and spelling scores read unspaced text faster and more accurately, but not spaced text (p. 855). + +The authors conclude that "efficient bottom-up orthographic processing and word identification in the absence of word boundary information" is "the primary factor at play" in how readers manage unspaced text (Conclusion, p. 856). + +## Relation to the argument +It bears on premises P3 and P4 and on ruling D2 (the dispute over whether unspaced script burdened trained readers). First, it repeats the 40–70% cost figure in a modern French sample, and the facet's flag on that range ("[unverified]") can now be resolved to this page. Second, it shows what the cost is made of: the skilled reader does the segmenting by recognising words, so the load falls on lexical and orthographic skill (vocabulary and spelling predicted performance only for unspaced text). That supports the reading that spacing moves segmentation work from the reader to the page (P5), and it supports the idea that the benefit is largest for readers whose lexical skill is thinner, though the sample here is all university students and no learners were tested. It also supports P4's caution that skilled readers adapt, since the cost is large in speed but readers still read. + +It does not test historical readers, other scripts, or the effect on learning to read. Under ruling D10 it is analogical background. + +Facet check: the facet lists this paper as "not read", relying on a secondary summary that transpositions hurt unspaced reading more. The full text confirms that summary, and adds that the interaction was absent in first-fixation and gaze duration and was weaker when boundary letters were frequent. + +## In conversation with +- `mirault2019reading` — the companion study, cited here as Mirault et al. (2018, submitted); its finding that sentence structure helps unspaced reading only in limited ways frames this study (pp. 846–847). +- `rayner1998unspaced` — source of the slowdown and the larger word-frequency effect (pp. 846–847). +- `raynerpollatsek1996reading` — cited with Rayner et al. 1998 for the 40–70% figure (p. 846). +- `perea2009space` — space information and landing positions (p. 846). +- `veldre2017spelling` — spelling ability predicts disruption by unspaced text; the authors agree and extend it (p. 856). +- `drieghe2017parafoveal` — preview effects in unspaced text (p. 846). +- `epelboim1997fillers` — Epelboim et al. (1997) among the unspaced-text studies cited (p. 846). +- `winskel2009eye` — Thai readers benefit from inserted spaces (p. 847). +- `kasisopa2013eye` — boundary-letter frequency as a cue in Thai, the basis for their boundary-letter analysis (p. 853). +- `saenger1997space` is not cited here; the companion paper cites it. +- Leads not in the bib: Johnson & Eisler (2012); Rayner, White, Johnson & Liversedge (2006); Winskel, Perea & Ratitamkul (2012); Grainger, Mathôt & Vitu (2014); Snell et al. (2018), OB1-Reader. + +## Caution +Exp. 2 is underpowered by the authors' own admission (footnote 5, p. 853). The sentences are neutral, seven-word French sentences with low predictability and one-line presentation, unlike running text. Participants are French university students used to spaced text. The vocabulary and spelling effects are correlational. The paper tests degradation by letter transposition; it does not show that spacing makes reading easier to learn. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the open-access published PDF (APP 81:846-860, locators checked page by page against the PDF). Checked 3 verbatim quotations and about 25 numbers (accuracy 98.47/89.75, b and t values, 83 vs 67 ms, Exp. 2 interactions, 104 sentences, participant counts, footnote 5 on Exp. 2 underpowered) and the cited works. + +**Findings:** One overstatement fixed: the card listed skipping rate among measures showing a larger transposition cost without spaces. The source (pp. 853, 856) reports the opposite direction for skipping: transposed words were skipped more in unspaced text and less in spaced text. Edited and the direction stated. Quotations, page locators and other numbers match; footnote 5 does attach to Exp. 2 as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mitchell2026agents.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mitchell2026agents.md new file mode 100644 index 00000000..14bfa024 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/mitchell2026agents.md @@ -0,0 +1,44 @@ +# Mitchell, M., Ghosh, A., & Passi, S. (2026). AI agents push humans out of the loop. *arXiv preprint* arXiv:2608.23642. + +**Identifier:** arXiv:2608.23642 · **Read depth:** full_text (preprint v3, 6 Sep 2026, 22 pages; locators are PDF pages. I read the ten-page body and skimmed the reference list for the works the text leans on. The two appendix pages of solution diagrams and the larger Table 2, pp. 20-22, I looked at for structure only) · **Source-tier:** preprint, position paper (its closing page invites "the broader NeurIPS community", p. 10; I found no sign it has been published) · **Evidence basis:** direct_read · **Design and sample:** none. A position paper that assembles prior HCI, cognitive-science and AI-safety studies, plus a "structured inventory" of solutions sorted by six oversight goals and six solution categories (Table 1, p. 7). It gathers no data · **Parties modeled:** person–system (a human overseer and an AI agent), with deployers and developers as the two parties who must support the overseer; affected parties do not appear · **Facets:** G, Z · **Read:** 2026-10-13 + +## What it argues + +The authors take the claim that a "human in the loop" makes agents safe and say it fails twice. Current agent design "impede[s] effective human oversight," and the cognitive capacities oversight needs "are also themselves degraded by extended use of AI systems" (abstract, p. 1). The first half (Section 3, "The oversight oversight") says the stream an overseer must follow is large and spread across components, that users repeatedly approve prompts until they "stop paying close attention to what they are approving" (p. 4), and that the burden falls "almost entirely on users" (p. 1). It adds a claim about the frame: "'Human-in-the-loop' is only a meaningful solution if the human can independently see into the loop" (p. 4). + +The second half (Section 4, "The irony of automation") collects reports of deskilling, automation bias, anchoring and complacency, and puts them under Bainbridge's irony: "Oversight degrades the overseer" (p. 6). It adds a feedback argument. If approvals are treated as evidence of success, a system trained on them "can begin to optimize for disincentivizing thoughtful oversight" (pp. 5-6), so the human rater becomes "the exploitable part of the reward channel" (p. 6). + +The remedy is "cognitive scaffolding" in two prongs. Developers supply affordances: strategic friction (pre-commitment, delay-and-choice, reasoning probes, action gating), decision design (bounded autonomy, batch review, automated pre-checks), and behavioural monitoring of the overseer (time, override and evidence-seeking signatures; planted canaries; audits) (pp. 7-8). Deployers supply protocols: training, enforced breaks and rotations, and role design, including "separating roles" so the overseer is not the decision-maker who gains from approvals (p. 8). Section 6 answers three objections. On "Better tooling and transparency will solve this" the authors reply that it is "necessary, but insufficient," since explanations "operate on cognitive capacities that AI agent use itself degrades" (p. 9). + +## Relation to the argument + +Strongest support for premise P12 (people perform sustained monitoring and whole-volume review badly), and a direct counter to the post's P15 and P19, where a person reads tamper-proof evidence and holds the brake. It is also the closest 2026 case for the lineage the paper needs: design the task so the person carries less, and do not hand them more to read. + +Three points of contact with the rulings. (1) D3 and D4: Section 6's reply on transparency ("necessary, but insufficient," p. 9) supports the author's line that disclosure does not change how well people navigate, but only as an argument; Mitchell cites studies of explanations in AI-assisted decisions, not of platform disclosure. (2) The argument map's "lowering a cost has costs" worry: Mitchell's friction mechanisms ask the person to do more work, which pulls against the "carry less" direction of P5 and P8, and the paper argues for it because offloading erodes the skill oversight needs (pp. 4-5). (3) P10: the paper's "separating roles" and "aligning incentives" (p. 8) name the same problem as a conflict of interest between the overseer and the beneficiary of approvals. + +Facet check: facet Z quotes the abstract ("impede effective human oversight," "design-level affordances") and notes that none of the oversight-fallacy sources "replaces the human check with structure." That holds for the headline, but the inventory includes structural items (bounded autonomy, automated pre-checks, separated roles), so Mitchell sits closer to the structural camp than facet Z's summary suggests. Facet G calls it a paper "arguing that agent design degrades the overseer's skills"; it argues that design both obstructs oversight (Section 3) and erodes it (Section 4). + +## In conversation with + +- `bainbridge1983ironies`: the frame ("irony of automation," p. 6), and the aviation lineage of the human in the loop (p. 2). +- `endsley1995out`: out-of-the-loop performance, p. 5; also behind the call to preserve domain skill (p. 8). +- `parasuraman2000types`: levels of automation, cited for the user's shifting role (p. 4) and for complacency and overload (p. 5). The paper does not cite Parasuraman and Manzey (2010). +- `bucinca2021trust`: cognitive forcing functions, the model for its strategic-friction mechanisms (p. 7), and the source for "users may disprefer systems that reduce overreliance" (p. 8). +- `green2022flaws`, `langer2024effective`, `sterz2024quest`: cited together for the claim that "the presence of an overseer does not entail reliable oversight" (p. 1); Sterz is also cited for role design (p. 8). +- `santonidesio2018meaningful`, `eu2024regulation`: cited as governance frameworks that require meaningful control and final human decisions (p. 1); Santoni de Sio is also the authority for the "autonomous weapons lineage" of the human in the loop (p. 2). +- `anthropic2026how`: Mitchell cites an Anthropic engineering post titled "Claude Code auto mode: A safer way to skip permissions" for approval fatigue (p. 4). The bib title differs slightly; I did not compare the URLs. +- Leads, not in the bib: Passi 2025, "Agentic AI has a human oversight problem" (SSRN), which Mitchell calls the "most similar" sociotechnical perspective (p. 3); Dhanorkar, Passi and Vorvoreanu (2026) on developers overseeing software agents (FAccT). + +## Caution + +A position paper; its evidence is other people's studies and several self-reports, including a blog post and a Reddit post (refs 113, 116). It does not itself measure approval fatigue or skill loss. The solution inventory is the authors' own construction, not a systematic review; they call it a structural map, "not as a definitive taxonomy" (p. 6). Preprint v3 of an unpublished paper. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened arXiv:2608.23642v3 PDF (6 Sep 2026; 22 pp.). Checked 14 quotations and 20 page locators by page (pp. 1, 3-10), the Table 1 and Table 2 locations, the section titles, the 'Separating roles' bullet (p. 8), and 12 of the cited works against the reference list (Bainbridge, Buçinca ref. 10, Passi 2025 ref. 88, Dhanorkar et al. FAccT '26 ref. 25, Anthropic auto-mode post ref. 5, others). + +**Findings:** No issues found. Card's Read depth correctly says the body was read in full and only the appendix and reference list were skimmed; I left it as full_text. 'lowering a cost has costs', 'replaces the human check with structure' and 'arguing that agent design degrades...' are the project's or facets' wording, not the paper's. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2025looking.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2025looking.md new file mode 100644 index 00000000..19246707 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2025looking.md @@ -0,0 +1,25 @@ +# Nadella, S. (2025, December 29). Looking ahead to 2026. *sn scratchpad*. https://snscratchpad.com/posts/looking-ahead-2026/ + +**Identifier:** https://snscratchpad.com/posts/looking-ahead-2026/ · **Read depth:** full_text (blog page, about 450 words; no pagination, locators are paragraph numbers from the first paragraph under the date, counting each list item as a paragraph) · **Source-tier:** personal-blog essay by a corporate chief executive · **Evidence basis:** direct_read · **Design and sample:** none (position piece) · **Parties modeled:** people and AI, with "each other" as a further party · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Nadella says 2026 will be "pivotal", with capability outpacing "our current ability to use it to have real world impact", which he calls "model overhang" (paras 1 and 3). Three things "we still need to get right" follow in a list. First, "A new concept that evolves 'bicycles for the mind' such that we always think of AI as a scaffolding for human potential vs a substitute." He adds: "What matters is not the power of any given model, but how people choose to apply it to achieve their goals", and calls this "the product design question we need to debate and answer" (para. 4). That item also asks for a new "theory of the mind" that "accounts for humans being equipped with these new cognitive amplifier tools as we relate to each other." Second, "We will evolve from models to systems": "rich scaffolds that orchestrate multiple models and agents; account for memory and entitlements; enable rich and safe 'tools use'" (para. 5). Third, deliberate choices about diffusion, because "For AI to have societal permission it must have real world eval impact" (para. 6). He closes with outcomes "for each of us" and computing as "empowering people and organizations" (para. 7). + +## Relation to the argument +Rulings D4 and D7. The post asserts a design question about people and AI, scaffolding against substitution, which is a design-side framing close to the paper's. It gives no design answer. It names "entitlements" in the systems item, a security term that the insider post develops as authority (P5, P17). The "as we relate to each other" clause is the one place the series gestures at people relating through AI, and it has no development. For the paper it is a dated statement that the product design question is open. + +## In conversation with +Cites no works. The later posts `nadella2026frontier`, `nadella2026reverse` and `nadella2026models` develop the "systems" and "entitlements" items. + +## Caution +Year-end outlook with no evidence or method. The phrase "bicycles for the mind" is quoted from the post without attribution there. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the saved HTML of the blog post (title 'Looking Ahead to 2026', dated Dec 29, 2025) and counted its paragraph elements. Checked 12 quoted phrases and the attributed claims, and the claim that no works are cited. + +**Findings:** Locators fixed: the card's paragraph numbers were off by one to two. Counting each list item as a paragraph, 'pivotal' is para 1, 'model overhang' para 3, the three 'get right' items paras 4, 5 and 6, and the closing paragraph para 7. The card now says so. Every quotation matches the text word for word; no other issues. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026frontier.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026frontier.md new file mode 100644 index 00000000..e91c0dff --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026frontier.md @@ -0,0 +1,25 @@ +# Nadella, S. (2026, June 14). A frontier without an ecosystem is not stable. *sn scratchpad*. https://snscratchpad.com/posts/frontier-ecosystem/ + +**Identifier:** https://snscratchpad.com/posts/frontier-ecosystem/ · **Read depth:** full_text (blog page, about 700 words; no pagination, so locators are paragraph numbers counted from the first paragraph under the date) · **Source-tier:** personal-blog essay by a corporate chief executive · **Evidence basis:** direct_read · **Design and sample:** none (position piece) · **Parties modeled:** firm, its people and AI models; a "few models" as a concentrated third party · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Nadella says AI is "the first time we can create a real cognitive loop between people and digital systems" (para. 2). A firm must build "human capital" (the "knowledge, judgment, relationships, ingenuity, and pattern recognition of its people") and "token capital" ("the firm's AI capability it builds and owns") (para. 4). Human capital grows more valuable as token capital grows: "human agency will be the driver of token capital growth", because "Humans will set ambitious goals" and "Without human direction, you have compute running in circles" (para. 5). The opportunity is "building a learning loop on top of models" (para. 6): "You can offload a task, or even a job, but you can never offload your learning." Control is tested by whether "A company should be able to switch out a 'generalist' model without losing the 'company veteran' expertise" (para. 7). The loop uses private evals, private reinforcement-learning environments and a queryable knowledge base (para. 8), and "becomes the new IP of the firm", a "hill climbing machine" (para. 9). He warns against "ceding value to a few models that eat everything they see", calls it politically intolerable, and compares it with the hollowing-out of industrial economies in the first phase of globalisation (paras. 10–11). He asks for "a frontier ecosystem, not just a frontier model" (para. 12). + +## Relation to the argument +Ruling D4 and D5, premise P11. This post shows a firm-level layer where people steer: the firm's people correct and direct models and the firm owns the resulting loop. The insider post (`nadella2026models`) never mentions it. A claim that "Nadella" has no place for human steering therefore has to say "the insider post" instead. The steering here belongs to the firm's employees and the firm's ownership of the loop; para. 14 says employees see "their judgment become part of systems", and the benefits "accrue to the companies and communities around them". Customers or counterparties of a firm appear nowhere. The post also supports the paper's point about authority: the unit that holds control is the organisation. + +## In conversation with +Cites no works. Same blog and theme as `nadella2026reverse` (who owns the learning loop) and `nadella2026models` (who holds authority over a model). Companion text `nadella2025looking` states the scaffolding premise. + +## Caution +An executive's advocacy essay with no data. "Human agency" here means the firm's workforce directing token capital. It is not a claim about users of platforms or about literacy. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the saved HTML of the post (dated Jun 14, 2026) and counted paragraph elements. Checked 15 quotations and the paragraph locators (paras 2, 4-12, 14), and that no works are cited. + +**Findings:** No issues found. Paragraph numbers count from 'I've been thinking...' as para 1 and match the card throughout. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026models.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026models.md new file mode 100644 index 00000000..0ad886c8 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026models.md @@ -0,0 +1,27 @@ +# Nadella, S. (2026, October 10). Models as insider risks in the super intelligence era. *sn scratchpad*. https://snscratchpad.com/posts/models-as-insider-risks/ + +**Identifier:** https://snscratchpad.com/posts/models-as-insider-risks/ (mirrored on X, https://x.com/satyanadella/status/2108931348857827686) · **Read depth:** full_text (the blog page, about 840 words including page furniture; no pagination, so locators are paragraph numbers P1–P21 as numbered in the project's local X capture; I normalised and compared both texts and P2–P21 match the blog word for word, P1 matches too) · **Source-tier:** personal-blog essay by a corporate chief executive (not peer reviewed; no Microsoft product or policy document accompanies it) · **Evidence basis:** direct_read · **Design and sample:** none (position piece) · **Parties modeled:** organisation (deployer) and model, with a model provider and "those affected" named once each · **Facets:** G, X · **Read:** 2026-10-13 + +## What it argues +Nadella starts from an attribution gap: with traditional software "we had the tools and capability to trace behaviors to a specific code path" (P1), but for frontier models "we can't attribute model behaviors and outputs to specific inputs of training data or configurations of model weights", and organisations still give them "access to our most sensitive data" and "mission-critical actions on our behalf" (P2). Responsibility stays with the deployer: "A model provider's assurances do not relieve us of that responsibility" (P3). His remedy is containment, not acceptance or rejection of outputs (P4): "we need to separate the supply of intelligence from the authority over it" (P5). He sets alignment aside and proposes "strong, deterministic system design, human controls, and reliable operating procedures" around non-deterministic models (P6), and treats models as insider risks "not because they are necessarily malicious, but because any sufficiently capable actor ... can make mistakes or be compromised" (P7). Existing enterprise practice applies: identity, privilege limits, logs, containment boundaries (P8). + +Chain-of-thought transparency is "a non-negotiable", and "Neuralese" cannot excuse opaque reasoning, but it "alone is not sufficient or dependable" (P9). Models checking models risks "nested black boxes" (P10). Controls over access and action "must sit outside the model", resting on an unnamed 1970s security principle that a program must not bypass the mechanisms enforcing its permissions (P11); the model is separated from its harness and action space (P12). Seven "principles of observability" follow (P13): model diversity (P14); "Observe everything", with "tamper-proof human readable evidence" for every meaningful action so an outcome can be reproduced without the model's own attestation (P15); continuous whole-system testing (P16); independent controls, where "Orgs" decide what a model may access and do (P17); independent auditability (P18); containment, with "an authorized person" able to pause or stop a model mid-task, "like an emergency brake" (P19); and incident disclosure "to those affected" (P20). The close: the best system "enables us to trust the model the least" (P21). + +## Relation to the argument +Rulings D4 and D5, and premise P11. "We" is the deploying organisation throughout. The words "user", "customer", "employee" and "people" do not occur. "Human" occurs twice (P6, P15), "person" once (P19), "authority" once (P5), "affected" once (P20). The enterprise reading (D5) is the text's own. D4 needs care: the post does name "Neuralese" (P9), but disclosure is not its only move. P9 calls chain-of-thought transparency insufficient, and model diversity (P14), independent controls (P17) and containment (P19) disclose nothing to a reader, while P11–P12 put the controls outside the model. The post fits P11 of the argument map: authority sits with the deployer, load-bearing controls sit outside the model, and the residual human tasks are reading evidence (P15) and halting (P19). Who reads, at what rate, and how a person notices a fault in time to brake are unstated. P15's stated purpose is reconstruction ("reproduce how an outcome was achieved"), which reads as forensic, whereas P19 implies someone notices during the task. P20 is the only place a party outside the deployer appears, and it receives notice after a failure. + +## In conversation with +The post cites nothing. P11's "1970s" principle is unattributed; `saltzer1975protection` and `anderson1972computer` are the likely referents (a lead of this project, not the post's text). Reported by `constantin2026nadella` and `kumari2026nadella`; same blog as `nadella2026frontier`, `nadella2026reverse`, `nadella2025looking`. Related in the bib on oversight: `green2022flaws`, `crootof2023humans`, `elish2019moral`, `bainbridge1983ironies`. + +## Caution +A short position essay with no evidence offered for its claims. "Observability" is his umbrella for all seven principles, wider than monitoring. It is not a Microsoft policy document. Cite by paragraph; the numbering P1–P21 is the project's, not the author's. Read as an enterprise-security proposal, not as a statement about people who meet a platform's mediator as counterparts. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the saved HTML of the post (dated Oct 10, 2026) and numbered its paragraphs; P1-P21 match the card's numbering. Checked about 20 quotations and the word-occurrence claims (user, customer, employee, people: none; human: P6 and P15; person: P19; authority: P5; affected: P20; 1970s: P11). + +**Findings:** No issues found. I could not check the card's statement that the project's local X capture matches the blog word for word, because that capture is not in the cards folder; the blog text itself supports every quoted claim. Name 'Saltzer'/'Anderson' as referents is flagged in the card as the project's lead, not the post's text. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026reverse.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026reverse.md new file mode 100644 index 00000000..61b63e31 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nadella2026reverse.md @@ -0,0 +1,25 @@ +# Nadella, S. (2026, July 12). The reverse information paradox. *sn scratchpad*. https://snscratchpad.com/posts/reverse-information-paradox/ + +**Identifier:** https://snscratchpad.com/posts/reverse-information-paradox/ · **Read depth:** full_text (blog page, about 800 words; no pagination, locators are paragraph numbers from the first paragraph under the date, with the named list counted as para. 14) · **Source-tier:** personal-blog essay by a corporate chief executive · **Evidence basis:** direct_read · **Design and sample:** none (position piece) · **Parties modeled:** firm (buyer) and model provider (seller) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +Arrow's information paradox says a seller who reveals information to sell it has already given it away (para. 2). Nadella reverses it: "the buyer risks giving away knowledge, just in order to use what they bought" (para. 3). A firm pays "once with money, and again with something even more valuable: the proprietary knowledge you must reveal" (para. 4), and "the information asymmetry becomes increasingly skewed" because the seller learns about the buyer while the buyer learns little back (para. 5). Models learn from "exhaust": "the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong" (para. 8). "In consuming intelligence, you are creating intelligence. And what you create should belong to you" (para. 9). He objects that providers claim fair-use rights to train on public data yet restrict distillation and "reserve the right to learn from customer usage" (para. 10). The remedy is a "hard boundary across which nothing crosses, not even the intelligence exhaust, without consent" (para. 12), built through five items (para. 14): Control (private evals; ownership of memory, traces and feedback), Capability (learning environments inside the tenant), Choice (an orchestration layer "decoupled from any single model"), Cost, and Compound. + +## Relation to the argument +Rulings D4, D5, D8; premise P10. This is the clearest statement in the series of an interest that differs between provider and customer: the provider's learning infrastructure takes value from the customer's usage, and the remedy is a boundary the customer controls. The party protected is the enterprise. The people who make corrections are named as the source of the exhaust, and the firm, not those people, is the owner. The post supports the paper's reading that Nadella's unit is the organisation (D5), and it shows he accepts that a provider and its customer can have opposed interests (no shared-interest assumption, cf. P10). The counterpart for the paper's triad, a third party served by the mediator, still does not appear. + +## In conversation with +Names Arrow (the information paradox; para. 2), Hayek ("particular intelligence", the knowledge of "time, place, and circumstance"; para. 9) and Alex Karp (a quotation on ownership of "the means of production"; para. 11). None is in `references.bib`. The post shares its enterprise-boundary logic with `nadella2026models` (P17) and `nadella2026frontier`. + +## Caution +Advocacy essay; the claims about provider contract terms are asserted, not documented. "Exhaust" and the 'hard boundary' are firm-level; they say nothing about individual users' data rights. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the saved HTML of the post (dated Jul 12, 2026) and numbered its paragraph elements. Checked 12 quotations and the locators (paras 2-5, 8-12, 14) and the three names (Arrow, Hayek, Karp). + +**Findings:** No issues found. The Hayek phrase 'particular intelligence' and 'time, place, and circumstance' sit in para 9 as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nagarajrao2026fareshare.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nagarajrao2026fareshare.md new file mode 100644 index 00000000..bd0683cd --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nagarajrao2026fareshare.md @@ -0,0 +1,37 @@ +# Nagaraj Rao, V., Dalal, S., Schwartz, A., Liaqat, A., Calacci, D., & Monroy-Hernández, A. (2026). FareShare: A tool for labor organizers to estimate lost wages and contest arbitrary AI and algorithmic deactivations. *Proceedings of the ACM on Human-Computer Interaction, 10*(2). + +**Identifier:** doi:10.1145/3788052; arXiv:2505.08904 · **Read depth:** full_text of the **arXiv preprint v1** (cs.CY, 13 May 2025), 30 PDF pages: body pp. 1–22, references pp. 23–25, appendices pp. 25–30. Locators are the preprint's pages, not the journal's (the bib gives pp. 1–32); the version of record may differ. I skimmed Appendix A (Reddit and prior-study summary), the evaluation protocol in Appendix E and the SUS form · **Source-tier:** peer-reviewed article (read as preprint) · **Evidence basis:** direct_read · **Design and sample:** tool build with a field deployment and no comparison group. A 75-minute formative interview with two lead organizers and one lead attorney of the Washington rideshare union "DU" (Nov 2024, p. 8); a three-month deployment, 31 Dec 2024–31 Mar 2025 (p. 16); a March 2025 office visit with 7 field representatives running intake with 10 drivers, two focus groups and a usability scale from 2 attorneys (pp. 15, 27–28); field notes coded by two researchers (p. 16) · **Parties modeled:** two parties plus a third party (driver and platform, with a union as state-designated resolution centre and a tool that computes the claim) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The authors call deactivation "one of the most severe forms of algorithmic control," one that "typically occurs through arbitrary AI and algorithmic decisions with little explanation or recourse" (abstract, p. 1). Washington's HB 2076 (2022) and Seattle's 2020 ordinance give drivers notice, an appeal and back pay: average daily earnings over the preceding 12 weeks plus 12% interest, or $200 a day if the platform fails to provide earnings or to accept those the resolution centre offers as evidence (p. 7). The union still assembled earnings from screenshots and spreadsheets. FareShare extends an open-source tool, FairFare, to pull trip history through a data provider (Argyle) after the driver consents, and it generates a PDF lost-wage report (pp. 12–14). + +Results. There were 122 user sign-ups linking 178 accounts; 143 synced (Uber 78, Lyft 65) and 35 did not, a 20% failure rate (pp. 16–17). Intake sign-up took 2–3 minutes where screenshots had taken 20–25; report generation took about 2 minutes against 2–3 hours (p. 17). Two attorneys rated usability 75 and 82.5 on the SUS (p. 18). The tool removed transcription errors and added new problems: Uber location-mismatch warnings frightened drivers, research-consent text added friction, and attorneys "abandoned the tool for certain cases" after repeated sync failures (pp. 18–19). One said "when Uber data doesn't sync, it puts a damper on the whole case" (p. 19). Field staff kept screenshotting: "We still collect data the same way. But it helps the legal team a lot more" (p. 17). Attorneys also used the trip data to audit platform records, checking a disputed trip by its coordinates (p. 19). + +## Relation to the argument + +Facet J, Q6, and ruling D8 ("where the convention bends"). The paper documents a place where it bent toward the weaker party by statute: notice, an appeal, a state-designated representative and a numeric rule for the remedy. It adds that the new right needed machinery, since platforms "restrict data access by citing privacy and trade secret protections" (p. 4). It qualifies, in part, the premise that a platform's owner does as it likes: the law binds the owner in Washington and Seattle. The paper reports no case where a platform changed its own interface or rule. The 80% overturn rate (p. 7) comes from another study of Seattle drivers (ref. 37) and covers drivers who had representation; the authors measure no arbitration outcome with FareShare. + +Facet check: facet J read only the abstract and wrote "178 signups". The body separates 122 users from 178 accounts (pp. 16–17); the abstract's "178 account signups" counts accounts. Facet J's "no outcome data" holds. Its line that the rights "need worker-side tools to be used" overstates: the union had won over 1,400 reactivations and about $3.5 million before the tool (pp. 7, 10), and FareShare automates one bottleneck. + +## In conversation with + +- `calacci2022bargaining` (ref. 7): tools that leave data collection to individual workers (p. 4). +- `irani2013turkopticon` (ref. 21): tools for workers to aggregate experience (p. 4). +- `cameron2024making` (ref. 10): workers' isolated conditions (p. 4). +- `rosenblat2016algorithmic` (ref. 36): platforms' "soft control" and information asymmetries (pp. 3, 21). +- Not in the bib: Calacci et al. (2025), FairFare, arXiv:2502.11273; Schwartz, Weber and Brown (2023) on Seattle deactivations; Jackson, Gillespie and Payette (2014), "The policy knot"; Khovanskaya et al. (2019). + +## Caution + +One union in one state, a short deployment, two attorneys on the usability scale, and a partner who is also the beneficiary. The headline "over 95%" is loose: 25 to 2.5 minutes is a 90% cut on its own, the paper says the tool "could" achieve it (pp. 2, 19), and field staff still screenshot (p. 17). Intake time appears as 25 minutes in some places and 20–25 in others, and the figures mix self-report and observation (Table 3, p. 15). The bib title carries a stray "CSCW016" (flagged in AUDIT_05). Driver views are out of scope (p. 22). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the arXiv v1 PDF (31 text pages; the title page is p. 1). Searched all 12 quotations (the three non-matches are facet-J wording) and checked the numbers and locators: HB 2076 and Seattle 2020 dates, 12 weeks / 12% / $200 a day (p. 7), 80% overturn from ref. 37, 1,400 reactivations and $3.5m (p. 7), 122 sign-ups / 178 accounts / 143 synced (78 Uber, 65 Lyft) / 35 failed = 20% (pp. 16-17), 2-3 v 20-25 minutes, 2.5 minutes v 25 and the over-95% claim (p. 19), SUS 75 and 82.5 (p. 18), the 75-minute interview in Nov 2024 (p. 8), 7 field representatives and 10 drivers (Appendix D, p. 27), and the reference numbers 7, 10, 21, 36, 37 against the reference list. + +**Findings:** No issues found. The Read depth line says Appendices A and E and the SUS form were skimmed; they are about 6 pages of 30, so I left full_text. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nothaft2020satanic.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nothaft2020satanic.md new file mode 100644 index 00000000..963549f0 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nothaft2020satanic.md @@ -0,0 +1,36 @@ +# Nothaft, C. P. E. (2020). Medieval Europe's satanic ciphers: On the genesis of a modern myth. *British Journal for the History of Mathematics*, 35(2), 107–136. + +**Identifier:** doi:10.1080/26375451.2020.1726050 · **Read depth:** full_text (accepted manuscript, 56 pp., from Oxford ORA; locators are the manuscript's page numbers, not the journal's; the reference list from p. 44 on was skimmed, not checked entry by entry) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** historiographical source-tracing; follows one claim through many modern authors, 1883 to 2017, against medieval statutes and texts · **Parties modeled:** other (authors and their sources) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues +Nothaft traces and rejects a story with three core claims: zero was feared as devilish; church or state therefore outlawed it; and zero and Hindu-Arabic numerals (HAN) were "only used in hiding", which supposedly explains "cipher" as "secret code" (abstract; p. 8). He calls the first and third claims false "in the sense that no compelling evidence to support them has ever surfaced" and the second false for the Church (p. 8). + +Part 1 (pp. 9–17) sets out the actual record. HAN first appear in Latin Europe in a manuscript dated 976 (p. 9); written HAN arithmetic arrives in the first half of the twelfth century through translated Arabic works; the textbooks of Sacrobosco and Alexandre de Villedieu become part of Arts-faculty teaching in the early thirteenth century, and Ambrosetti counts 161 manuscripts of the *Carmen* (pp. 10–11, n. 12). Numerals coexisted with Roman and other systems for centuries; citing Crossley, he reports HAN predominant across all contexts "only around 1500" and adoption "slow, problematic, and spasmodic" (p. 13). Some milieus (astronomy, computus) moved faster than others (p. 13). + +The two prohibitions that carry the myth are small. Florence's money-changers' guild (Arte del Cambio, 1299) barred members from writing accounts "by means of or in the letter of the abacus, but let him write it openly and in full by way of letters" (pp. 14–15); motive is "a matter of speculation" (p. 15). The Paduan "1348" rule has no primary source; Nothaft finds the substance in Padua's 1331 statutes, which require booksellers to mark name and price on the cover in clear letters, "non [per] zyphras", and near-identical rules at Bologna (1317/47, 1432) and Florence (1387) (p. 16). These governed how a price was displayed to customers, not how sellers kept accounts, and transparency may have been the aim since "not everyone would have been trained in a scuola d'abbaco" (pp. 16–17). Trieste (1350) barred money changers and merchants from books *per abachum*; Frankfurt (1494) barred clerks from using the numerals in public books (p. 17). Neither is "a blanket 'ban'" (p. 17). + +Part 2 (pp. 17–42) follows the myth. Leo Jordan's 1905 article is its source: "leaps of logic" turn William of Malmesbury's remark that the devil taught Gerbert into evidence about HAN (pp. 18–20), and read a harmless Salem algorismus line and Alain de Lille's quips as hostility (pp. 21–23). Smith and Karpinski (1911), Yeldham (1926), Dantzig (1930), Hogben (1936), Menninger (1969) and Murray (1978) repeat it (pp. 24–31). Rotman (1987), Ifrah (1994/98), Kaplan, Teresi, Seife and Green add inventions, such as mathematicians burned at the stake (pp. 31–42). A footnote chain shows citations running back through Taylor (1883) to unknown sources (n. 24, p. 16). The conclusion calls it "a game of telephone" (p. 42). + +## Relation to the argument +It cuts against any use of the 1299 Florentine rule as a case of a convention banned by authority. Read plainly, it is a local guild rule about how entries are written, and the 1331 Padua rule is a legibility rule for customers. For the paper's word-separation precedent (D1, D10) the lesson is about method: Nothaft shows how a tidy story about a notation travels from secondary source to secondary source with no primary support. That is a caution for the paper's own history of word separation (P2: the chronology is Saenger's alone). It also supplies one point on numerals as a design change adopted unevenly by trade (Crossley's "slow, problematic, and spasmodic"). + +Facet check: the facet read only the 1299 and Padua section and the abstract. The full text confirms its account of Article 102 and of the pages. It adds that the Paduan rule dates from 1331, not 1348, and that Nothaft treats the motive for 1299 as unknown; Nothaft lists fraud-prevention as a "popular" explanation for 1299 but does not endorse it (p. 15, n. 22). + +## In conversation with +- `chrisomalis2010numerical`: Nothaft quotes it as repeating the abacist–algorist "debate" and says the studies it cites show no "robust documentation" of two factions (p. 18, n. 29). +- `burnett2024learning`: the same author's earlier version (2002b, reprinted 2010) is cited for early thirteenth-century learning of Indian arithmetic (reference list). +- `cajori1928notations1`, `cajori1929notations2`: not cited. Nothaft cites Cajori's 1894 *History of Mathematics*, which mentions the Florentine rule but "falsely associat[es] it with 'merchants' rather than bankers" (pp. 17–18). +- Leads: Crossley (2013), "Old-fashioned versus newfangled"; Lüneburg (2008) on the Florentine statute; Camerani Marri (1955), edition of the statutes. + +## Caution +Accepted manuscript; check page numbers against the journal before quoting. It is a debunking of a narrative, not a study of how HAN spread, and it does not measure adoption. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the Oxford ORA accepted manuscript (57 PDF pages; manuscript numbering matches PDF pages). Checked 14 quotations word for word and their page locators (pp. 8-19, 42), the 976 and 161 manuscript figures, the 1299, 1331, 1317/47, 1432, 1387, 1350 and 1494 dates, and 12 named works against the text (Jordan, Taylor, Cajori, Crossley, Chrisomalis, Caianiello and others). + +**Findings:** One quote fixed: the abstract reads 'were once only used in hiding'; the card quoted 'only in hiding'. Corrected to 'only used in hiding'. Page locators and the other quotations match. The robust-documentation note (n. 29) starts on p. 18 and its text runs onto p. 19; the card's p. 18 stands as the footnote's start. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nydfs2015barclays.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nydfs2015barclays.md new file mode 100644 index 00000000..eb82b7c4 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nydfs2015barclays.md @@ -0,0 +1,39 @@ +# New York State Department of Financial Services. (2015, November 17). *In the Matter of Barclays Bank PLC, Barclays Bank PLC, New York Branch: Consent Order under New York Banking Law § 44.* + +**Identifier:** https://www.dfs.ny.gov/system/files/documents/2020/04/ea151117_barclays.pdf · **Read depth:** full_text (14-page consent order; the signature block is dated "/7 day of November, 2015" in the scan, and the DFS press release is dated 18 November 2015; the bib gives 18 November) · **Source-tier:** regulator order (consent order) · **Evidence basis:** direct_read · **Design and sample:** none; a negotiated settlement with stipulated facts, not an adjudicated finding · **Parties modeled:** two parties + a dealer-controlled mechanism (clients and Barclays as counterparty; the regulator as third party) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +Barclays operated BARX, an electronic FX platform; about 80% of its FX volume was electronic (¶4). Some orders were subject to "Last Look," "an automated function that delays the Bank's response to a counterparty's incoming client orders for a set period of time — called the 'hold time' — and rejects the order if the market price moves beyond a certain threshold" (recitals, p. 2). From at least 2009 to 2014 Barclays says it meant Last Look as a defence against stale prices and "latency arbitrage" (recitals). The recitals state that Barclays "failed to properly use Last Look due to the failure of systems and controls, including management oversight," employed it "broadly and indiscriminately," and "there was a lack of transparency both internally and with customers" (recitals, p. 2). + +The facts stipulated: Barclays applied Last Look to all FIX/API trades and a few GUI clients, assigning an undisclosed hold time (¶¶10–12). It treated trades as toxic "[w]henever prices within this holding period moved against Barclays and in favor of the customer beyond a certain undisclosed loss threshold" (¶13), so it applied the rule "almost entirely in reference to the profit or loss the trade would bring to the Bank" (¶14), and "did not perform an analysis" to confine rejections to real toxic flow (¶15). In September–October 2014, after the investigation began, Barclays made the filter symmetric; one platform carrying 7% of volume stayed asymmetric until August 2015 (¶¶17–18). Stop-loss orders also faced Last Look holds (¶¶19–22). + +On communication, a rejected client saw only "NACK" ("Not Acknowledged"), with no reason (¶24). The order quotes internal emails: a 2008 instruction to say the trade "WAS REJECTED BECAUSE OF LATENCY" and "DO NOT talk about P&L"; a 2013 instruction to name all rejects "NACK urgently . . . Clients are gonna come asking"; a 2014 line, "Our Team generally does not share this info with the client, and just say it was a business reject" (¶25). A client who wrote in 2010 about "over 300 rejected orders" in a day got no response on record (¶25b). A marketing deck said "No last look — what you see is what you get" without separating the GUI from FIX/API (¶27). Some sophisticated clients tracked their own rejection rates and, when they complained, got adjusted hold times (¶26). + +Remedies: a $150,000,000 civil penalty (¶30); an employee suspended and ordered terminated (¶31); a monitor until 19 February 2016, with a remediation plan covering change management and periodic review of FX system settings (¶¶34–35). The order finds only that the Bank conducted business "in an unsafe and unsound manner" (¶29). Paragraph 38 states that no provision "is subject to review in any court or tribunal outside the Department." + +## Relation to the argument + +Under ruling D8 the paper compares who holds the convention in FX markets with platforms. This order is the clearest dated instance in the batch of a dealer using an unreviewable discretion against clients and the regulator, not the client, supplying the redress. It supports premise P10's logic in a financial setting: the intermediary's rule served its own profit, a client could not see the rule, and a convention was changed only after investigation. It also bears on ruling D3. The informed clients (those who tracked rejection rates) did get hold times adjusted, so some reading helped, but the asymmetric rule continued for years and ran in the Bank's favour; the fix was a change to the rule's symmetry, which the clients could not make themselves. That is mixed evidence for "disclosure is a literacy intervention." + +Facet check: the J-facet reading of the recitals, ¶13, ¶24 and ¶25 matches the order. The facet describes the order as dated 18 Nov 2015; the order is signed in November 2015 (day unclear in the scan) and the Department announced it on 18 November. + +## In conversation with + +- `nydfs2015barclayspr`: the Department's press release of the same matter, which repeats the facts and adds a quotation from the Acting Superintendent. +- `fca2014fines` and `cftc2017fxcm`: other regulator actions on dealers' conduct toward clients in FX. +- The order refers to a May 2015 NYDFS consent order with Barclays on benchmark FX manipulation (recitals); not in the bib. The order cites no scholarship. `oomen2016last` (a paper on last look in the bib) concerns the same practice but is not cited here. + +## Caution + +A settlement: Barclays "states" its purpose was defensive (recitals), and the order makes no finding of fraud, only "unsafe and unsound" conduct (¶29). It covers one bank and 2009–2014. It does not give rejection rates, client losses or the size of thresholds. The remedy terms bind the Department and Barclays, not other agencies (¶39). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the 14-page DFS PDF (consent order dated November 2015, signature day garbled as '/7' in the scan). Checked 13 quotations word for word and about 25 factual claims by paragraph number (80%, ,000,000, 7% until August 2015, monitor to 19 Feb 2016, NACK, 300 rejected orders, the 2008, 2013 and 2014 emails, paras 4, 10-15, 17-27, 29-31, 34-35, 38-39). + +**Findings:** One overstatement fixed: the card said 'The Department finds that Barclays failed to properly use Last Look...'. Those statements are WHEREAS recitals in a negotiated consent order, not findings; the card now says 'The recitals state'. Quotations, numbers and paragraph locators match. The 'disclosure is a literacy intervention' phrase is the project's wording, not the order's. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nydfs2015barclayspr.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nydfs2015barclayspr.md new file mode 100644 index 00000000..2e3bd017 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/nydfs2015barclayspr.md @@ -0,0 +1,35 @@ +# New York State Department of Financial Services. (2015, November 18). *NYDFS announces Barclays to pay additional $150 million penalty, terminate employee for automated, electronic foreign exchange trading misconduct* [Press release]. + +**Identifier:** https://www.dfs.ny.gov/reports_and_publications/press_releases/pr1511181 · **Read depth:** full_text (the whole press release as HTML, about 1,900 words, read from the saved page) · **Source-tier:** regulator press release (summary of a consent order) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** two parties + a dealer-controlled mechanism (clients, Barclays; regulator as third party) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The release announces that Barclays will pay "an additional $150 million penalty" and is "terminating" its Global Head of electronic automated flow trading "for misconduct related to automated, electronic foreign exchange (FX) trading through its 'Last Look' system." With the May 2015 action the total FX penalty paid to the Department reaches $635 million. Acting Superintendent Albanese says the case "highlights the need for greater oversight and action to help prevent the misuse of automated, electronic trading platforms on Wall Street, which is a wider industry issue that requires serious additional scrutiny." + +The body restates the consent order (`nydfs2015barclays`) in near-identical wording, under headings: how Last Look works, how Barclays rejected client orders, and how it communicated with clients. It repeats the stipulated points: Last Look applied to all FIX/API trades; "almost entirely in reference to the profit or loss the trade would bring to the Bank"; clients met with "NACK" and vague answers; three 2011 emails (two from a managing director) telling staff to "obfuscate and stonewall" and to "just blame it on the weekend IT release." It also reports that after the investigation began Barclays made the filter symmetrical in September and October 2014, but that 7% of trading volume stayed asymmetric until August 2015. It also reports (the release has no paragraph numbers; the earlier "¶26" locator belongs to the order and was not checked here) that some sophisticated customers who monitored rejection rates raised concerns, and Barclays "sometimes adjusted hold times and thresholds to decrease rejection rates." + +## Relation to the argument + +It adds nothing to the evidence beyond the order; its use is that it shows the regulator's public framing of the case: the harm is concealment of the rule and its self-serving use, and the remedy is oversight after the fact. The Superintendent calls automated platforms "a wider industry issue," which the paper can cite for the claim (ruling D8) that FX regulators treat automated intermediaries as a class needing scrutiny. The paper should cite the order for paragraph-level facts and this release only for the date and the quoted statement. The fact about sophisticated clients getting thresholds adjusted is evidence that a monitoring client could force a change in the dealer's rule, which cuts against a flat "clients cannot affect the convention" claim. + +Facet check: the J-facet read only the headline paragraphs of the release. The full text matches the facet's quotations and adds nothing that contradicts it. + +## In conversation with + +- `nydfs2015barclays`: the order it summarises; the release's wording tracks the order's closely. +- `fca2014fines`, `cftc2017fxcm`: other regulator releases on FX conduct. +- It cites no scholarship or other sources. + +## Caution + +A press release restating a negotiated order; cite the order for legal facts. It reports Barclays' stated purpose only through the order's account. The termination language is the Department's, not a court finding. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The saved HTML (cards/nydfs2015barclayspr.html), read in full. Seven quotations and about ten facts (dates, 7%, $150M, $635M, three 2011 emails) checked against it. + +**Findings:** All quotations and numbers match. One fix: the card attributed the sophisticated-client passage to "the order's ¶26"; the release carries no paragraph numbers, so the locator could not be checked and I reworded it. Not checked: the claim that the release tracks the order "in near-identical wording" (the order was not opened). Nuance: the body says the Global Head of eFICC was terminated after the investigation began, while a separate Managing Director was suspended and ordered terminated; the headline says Barclays "is terminating" the Global Head. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/odonnell1994parkes.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/odonnell1994parkes.md new file mode 100644 index 00000000..833c25e5 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/odonnell1994parkes.md @@ -0,0 +1,38 @@ +# O'Donnell, J. J. (1994). Review of M. B. Parkes, *Pause and Effect: An Introduction to the History of Punctuation in the West* (University of California Press, 1993). *Bryn Mawr Classical Review*, 1994.03.04. + +**Identifier:** https://bmcr.brynmawr.edu/1994/1994.03.04/ · **Read depth:** full_text (the whole open review, HTML, about 1,300 words; no pagination, so locators are by topic; the reviewed book is the US edition of 1993, not the Scolar Press edition of 1992 that the bib lists) · **Source-tier:** book review · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** other (reader, scribe, marks on the page) · **Facets:** A, K, X (as the bib lists them) · **Read:** 2026-10-13 + +## What it argues +Opening: punctuation is "ubiquitous and invisible," and editors impose their own systems on ancient texts. O'Donnell sets Parkes after the earlier work of Wingo (1972) and Mueller (1964), both of which "concentrate on the fragmentary manuscript and epigraphic evidence from antiquity." The next paragraph gives the book's plan: a narrative of Western punctuation to the early modern period (book pp. 9–61), a discussion of "influences on the application of punctuation" (pp. 65–96) with a chapter on verse, and 74 plates (pp. 160–297). Greek "is emphatically not discussed." What non-alphabetic marks do for reading is create "a grammar of legibility." + +Then: Parkes distinguishes logical from rhetorical analysis. "Roughly (this is no novel idea), earlier practice emphasized punctuation in the service of one who would read aloud … while our modern practice is to regard punctuation as a way of marking logical structures in a text for the eye of a silent reader." Three cautions follow. Parkes relies on secondary literature outside the later Middle Ages and makes factual errors on late antiquity, including, in O'Donnell's account, a pope who never existed. The topic is "designedly narrow, perhaps too narrow," and leaves out *mise en page* and the question of "how reading actually works." The plates have transcription errors, and later additions to the manuscripts go unreported. The closing section praises "the demonstration that liturgical practice in the middle ages had a strong influence on encouraging the insertion of punctuation in manuscripts, for the public performer needed a manuscript that was easily read at sight." Then: "I miss mention of Paul Saenger's work on the growth of silent reading the later middle ages, and here the work of Richard and Mary Rouse offer the necessary bridge between punctuation studies and a wider consideration of how books were being used in the later middle ages to transmit information." + +## Relation to the argument +The review never mentions word separation, spacing or *scriptio continua*, and does not say that Parkes dates the Insular practice or ties it to learners of Latin. + +- ARGUMENT_MAP P2 and OUTLINE line 109 cite "Parkes 1992, p. 1: learners of Latin as a foreign language" as `[lib]`. This review does not bear that out or contradict it. The citation still needs the book. +- It does support one thing the facets want: in 1994 a Penn classicist treats punctuation history and Saenger's silent-reading thesis as separate bodies of work and asks for the Rouses as a bridge. The paper cannot cite Parkes and Saenger as a single voice on that evidence. +- The reading-aloud and silent-reading contrast in the review's account of Parkes marks a change in what marks are for. It is a second convention, alongside spacing, that moved with reading practice, which agrees with the bundle in P6 and ruling D6. + +Facet check: facet B says the review "also records Saenger's 1982 *Viator* article as the pre-1997 form of the thesis." The 1994 review does not cite the *Viator* article. That reference is in O'Donnell's 2000 review of Saenger (`odonnell2000saenger`, second paragraph). The facet appears to have merged the two reviews. The rest of the facet's account (the missing-Saenger quotation and the read-aloud contrast) is accurate. The facet also calls the read "final paragraph(s)"; the whole review is short and has now been read. + +## In conversation with +- `parkes1992pause`: the book reviewed (US edition). +- `wingo1972latin`: named as the prior study of ancient Latin punctuation. +- `saenger1997space`, `saenger1982silent`: the review names "Paul Saenger's work on the growth of silent reading the later middle ages" without a title. The 1997 book had not appeared. +- `rouse1982statim`: the review names "Richard and Mary Rouse" without a title; the bib's Rouse entry is the likely match, but the text does not say. +- `carruthers1991accounting`: the review cites Carruthers, *The Book of Memory* (1990); the bib entry is a different item, so match with care. +- Leads not in the bib: Mueller 1964 (Tübingen diss.); Martin and Vezin, *Mise en page et mise en texte du livre manuscrit* (1990). + +## Caution +A 1994 review by an expert in late antiquity of a book on late-medieval British manuscripts. Its factual criticisms (the errors on late antiquity, transcription mistakes) cut against trusting Parkes outside his field. It says nothing on word separation and cannot stand in for the book. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The saved BMCR HTML (cards/odonnell1994parkes.html), read in full (about 1,500 words of review text, slightly more than the card's "about 1,300"). Nine quotations, the book-page ranges (9-61, 65-96, 160-297), the 74 plates, and the named works (Wingo, Mueller, Carruthers, Martin and Vezin, Fowler) checked; a search for spacing, scriptio continua, word separation and Insular found nothing, as the card says. + +**Findings:** All quotations are accurate. Two fixes: the Saenger/Rouse quotation was cut mid-sentence but closed with a full stop, so I gave the sentence's true ending ("in the later middle ages to transmit information"); and "Parkes joins the earlier work of Wingo and Mueller" overstated the text, which says both earlier studies concentrate on ancient evidence and that Parkes "makes a start" on the tasks that remain. The card's claim that the Viator article is cited in the 2000 review of Saenger (second paragraph) is correct; I checked it in that review. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/odonnell2000saenger.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/odonnell2000saenger.md new file mode 100644 index 00000000..f647d394 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/odonnell2000saenger.md @@ -0,0 +1,40 @@ +# O'Donnell, J. J. (2000). Review of Paul Saenger, *Space Between Words: The Origins of Silent Reading* (Stanford University Press, 1997). *The Medieval Review*, 00.02.24. + +**Identifier:** https://scholarworks.iu.edu/journals/index.php/tmr/article/view/14886 · **Read depth:** full_text (the open review on IU ScholarWorks, about 830 words, four paragraphs; no page numbers, so locators are ¶1–¶4) · **Source-tier:** book review (open-access online journal) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** other (reader, scribe, page) · **Facets:** B · **Read:** 2026-10-13 + +## What it argues +O'Donnell, a late-antiquity scholar, reviews Saenger for medievalists. ¶1 opens with Saenger's closing sentence (p. 275) and calls the book an "intellectual adventure." ¶2 frames the question, "when and how silent reading became a common practice." The easy answer is that Ambrose invented silent reading and Augustine saw it (*Conf.* 6.3.3), but "already thirty years ago, Bernard Knox … showed that the practice was attested long before Augustine," while "it did not become commonplace until long after." The second clause is O'Donnell's assertion, with no evidence offered. He names Saenger's 1982 *Viator* article (13, pp. 366–414) as "seminal." + +¶3 gives "the short form of Saenger's answer": when regular spacing of words "became commonplace" in the later Middle Ages, "purely visual and private reading became so easy as to become widespread." Then the verdict on that thesis: "some will dissent from … his claim that the textual practices he observes were the necessary and sufficient condition of the rise of silent reading." The argument is "a superstructure erected on the vast bulk of this work, like a dizzying spire looming over a richly constructed Gothic cathedral" (¶3). The cathedral would have been poorer without the spire, "but some observers might never have noticed the lack." + +¶4 summarizes the book's chronicle of word separation "step by step through a wide variety of insular and continental scriptoria." The "fons et origo in some ways is Virgilius Maro, the Irish grammarian of the seventh century," and "no homogeneity of practice spread easily"; the story has "multiple foci, multiple centuries, and almost infinite variation." The weakest link is the brief opening chapter on the physiology of reading: "the stretch from that survey to the manuscript practices he outlines and then to the theory about reading practices he constructs is the most venturesome and vulnerable part of the book. It will surely be tested." He expects the debate "resolved in Saenger's favor," and says the book's value lies elsewhere, as an orientation to palaeography and manuscript culture. + +## Relation to the argument +This is the most useful single piece in the batch for the sentence at `draft.md:263` and rulings D1 and D2. + +- Ruling D1 (review the cause): a sympathetic specialist separates Saenger's chronicle of spacing from his causal claim and reads the second as the likelier to fail. That agrees with the paper's choice to argue a precedent and leave the cause to the scholars. +- Ruling D6 (a complex, with a search for the critical move): O'Donnell's "multiple foci, multiple centuries" supports the bundle reading in ARGUMENT_MAP P6. +- P2: O'Donnell names Virgilius Maro and the seventh century as the origin, consistent with the Insular dating, but he gives only Saenger's account, so it is not independent confirmation (the map already flags that the chronology is Saenger's alone). +- P1: the physiology chapter is the part O'Donnell expects to be tested. Heilmann 2020 and Vatri 2012 later test it and, on the facets' reading, reject it. His prediction that the debate would resolve for Saenger was wrong on that point. + +Facet check: facet B's quotations of ¶2 and ¶3 match the text, including that "commonplace" in ¶2 is O'Donnell's own claim. The facet's description of the whole as four paragraphs is right. + +## In conversation with +- `saenger1997space`: the book reviewed. +- `saenger1982silent`: named as the earlier "seminal article" in *Viator* 13. +- `knox1968silent`: cited for attestation of silent reading before Augustine; O'Donnell gives the pages as 421–436 (the journal's last printed page is 435). +- `odonnell1994parkes`: same author; there he notes Saenger's absence from Parkes. +- Leads not in the bib: none beyond Augustine *Conf.* 6.3.3 and 8.12.29. + +## Caution +A 2000 review, written before Vatri 2012 and Heilmann 2020, and before the Winskel 2009 reading experiments. O'Donnell predicts how the debate will end; the later literature does not bear that out on physiology, so cite him for the structure of the book's argument and its vulnerable link, not for the outcome. No pagination in the copy read. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The saved HTML of the review (cards/odonnell2000saenger.html), read in full: four paragraphs. About 14 quotations, the Knox pages (421-436), the Viator citation (13 [1982] 366-414), Conf. 6.3.3 and 8.12.29, and the paragraph locators checked. + +**Findings:** No issues found. Not checkable from this copy: the remark that the journal's last printed page for Knox is 435; the statements about Heilmann 2020, Vatri 2012 and Winskel 2009 come from other sources and were not tested here. The word count is about 800 for the review text, close to the card's "about 830". diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/oomen2016last.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/oomen2016last.md new file mode 100644 index 00000000..8788815b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/oomen2016last.md @@ -0,0 +1,35 @@ +# Oomen, R. (2017). Last look. *Quantitative Finance, 17*(7), 1057–1070 (online 2016). + +**Identifier:** doi:10.1080/14697688.2016.1262545 · **Read depth:** full_text of the accepted manuscript dated November 2016 (LSE Research Online, eprints.lse.ac.uk/68811; 28 PDF pages; locators are the manuscript's own page numbers, not the journal's). Sections 1–5 read; the appendix proofs of Proposition 1 and Theorem 1 skimmed · **Source-tier:** peer-reviewed article (accepted version) · **Evidence basis:** direct_read · **Design and sample:** an analytical model with simulated parameters; no market data · **Parties modeled:** two parties plus a rule-holder (a trader, N competing liquidity providers, and an aggregator; the provider that wins the request decides whether it is filled) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +In over-the-counter foreign exchange, a trader shows a deal request to the liquidity provider (LP) whose quote looks best. Because prices move during the round trip, the LP applies a "last look" rule that accepts or rejects the request. Oomen's starting point is that "the liquidity provider owns finality of the deal" (p. 1). The accepted manuscript carries a footnote saying he is global co-head of electronic FX spot trading at Deutsche Bank. + +The model follows his earlier "Execution in an aggregator" (§2, not read). The true price follows a random walk. Each of N identical LPs observes it with persistent error, and the uninformed trader sends each request to the LP showing the best price. That routing selects the LP whose error is largest in the trader's favour, so the LP "is adversely selected" (§3.1). The trader sees an inside spread tighter than any single LP's, and with enough LPs it turns negative; without last look that would be a sure arbitrage (§2). + +He defines one-sided last look (reject only on moves against the LP) and two-sided last look (also reject, or rebate, on moves in the LP's favour), and two protocols: trade "on rate" (certain price, uncertain fill) or "on latest" (certain fill, uncertain price). "Symmetric" last look turns out to have four rival meanings: equal tolerance levels, equal reject rates, equal incurred costs and revenues, equal foregone costs and revenues. Only equal tolerance is always available, and under adverse selection the four cannot hold together (§3.1). His worked case uses N = 15 LPs, nominal spread 1, one time unit of latency buffer, and measurement-error and price volatility of 0.5. A tolerance of 0.5 on the LP's side rejects 28.5% of requests, and a second tolerance of 0.5 on the trader's side rejects another 14.1%. Raising both tolerances to 1 cuts those rates to 8.1% and 2.9% (§3.1). + +The two findings in his abstract: the choice of design and protocol sets execution risk, but the trader's effective transaction cost "need not be affected by it". When the trader's flow strongly selects against the LP (his stress case: N = 25), the symmetric and asymmetric designs converge and the choice of protocol dominates (§4). He concludes that the choice of last look "is not an issue of fairness, but one of preference over liquidity access, trading style, and tolerance to execution risk" (p. 4), and that no single design is superior or inferior (§5). + +## Relation to the argument +Ruling D8 (courts, FX markets, platforms; where does the convention bend). The paper is the rule-holder's account of a rule that gives one party the last word. Two points matter for the paper. + +First, it reports that the rule is configurable and negotiable per client: liquidity "is bespoke to the trader" (§2), and a trader can remove uncertainty of rate, of execution, or both by choosing protocol and configuration (§1, p. 4). That shows a convention in a market that is neither fixed nor imposed whole. Second, it recasts a dispute about fairness as one about risk allocation, from the party that benefits. The same recasting would suit a platform defending its terms. The paper offers no evidence on who gains in practice. + +Facet check: the facet says the paper "contradicts a reading in which last look is simply a dealer privilege". Fully read, the claim is narrower. Under the model, equal effective spreads can be reached by any combination of design and protocol. In the benign-flow case of Figure 8 (§4), Oomen says the designs "can lead to materially different execution risk profiles despite attaining the same effective spread", so the trader's risk is where the rules differ. The equivalence also depends on competing LPs and uninformed flow. The facet gives the published year as 2017, which fits the volume; the bib says 2016, the online date. + +## In conversation with +Oomen cites no work in `references.bib` (the 2016 manuscript predates the FX Global Code, `globalfxc2024code`, and the Barclays consent order, `nydfs2015barclays`, which concern the same practice but are not cited here). Leads from his list, not in the bib: Bank of England, H.M. Treasury and FCA (2015), the Fair and Effective Markets Review; Norges Bank Investment Management (2015) on last look; Cartea, Jaimungal and Walton (2015), "Foreign exchange markets with last look"; Budish, Cramton and Shim (2015) on the high-frequency arms race; Kyle (1985) for the noise-trader assumption. + +## Caution +A model with identical LPs, an uninformed trader, a random-walk price and invented parameters (figures are simulations, not observed FX data). The author's employer is a dealer that uses last look. This is the accepted manuscript, whose cover notice (LSE) says "There may be differences between this version and the published version"; I did not check the published pages. I did not work through the proofs. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The LSE accepted manuscript (cards/oomen2016last.pdf, 29 PDF pages; manuscript page = PDF page minus 2). Eight quotations (p. 1 "owns finality", p. 4 "not an issue of fairness", "need not be affected by it", "is adversely selected", "bespoke", "materially different execution risk profiles...", no single design superior), all of the worked numbers (N = 15, 28.5%, 14.1%, 8.1%, 2.9%, N = 25 stress case), the four-symmetry claim, the Deutsche Bank footnote and the cited leads checked. + +**Findings:** No issues found. The only caveat: the card says full_text with the appendix proofs skimmed; that is about three of 28 pages and the card states it, so I left the label. The published volume and page numbers (17(7), 1057-1070) are not in the manuscript and were not checked. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/oversight2022q1.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/oversight2022q1.md new file mode 100644 index 00000000..6c969bd9 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/oversight2022q1.md @@ -0,0 +1,41 @@ +# Oversight Board. (2022). *Q1 2022 transparency report* [Report covering 1 January–31 March 2022]. + +**Identifier:** https://oversightboard.com/attachment/731010438013851 · **Read depth:** partial (changed from full_text at S2: pp. 1-25 and the prose are read, but Annex II, pp. 26-64, was scanned and not read row by row; 64-page PDF, the Board's own report, created August 2022 by the file metadata; printed page numbers are used below). I read pages 1–25 and the prose of the whole report. Annex II (pp. 26–64, a table of 108 recommendations with Meta's response and the Board's assessment) I scanned in full for the notification, appeal and transparency recommendations and did not read row by row. · **Source-tier:** organisational report (self-report by the body it describes) · **Evidence basis:** direct_read · **Design and sample:** descriptive statistics for one quarter; the case counts for submitted cases are a manual estimate · **Parties modeled:** user, platform (Meta) and a third-party review body that Meta set up (the Board) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues + +The report counts what the Board did in the first quarter of 2022. It is a statement by the Board, not an evaluation by an outsider. + +**Intake.** "Where users have exhausted Meta's appeals process, they can challenge the company's decision by appealing eligible content to the Oversight Board" (p. 5). Because of tool limits, the Case Selection Team counted submissions by hand, so the figures are "an estimate and preliminary" (p. 5). An estimated 479,649 user cases arrived, the most in any quarter, up from 288,440 in Q4 2021 ("around two-thirds"), and more than 1.6 million from October 2020 to March 2022 (p. 5). With Meta's four referrals the total is about 479,653 (p. 4). Of the user cases, 94% appealed to restore content and 6% to remove other users' content (p. 5); 45% came from the United States and Canada and 22% from Europe (p. 6); 99.2% concerned Facebook and 0.8% Instagram (p. 7). + +**Funnel.** The Case Selection Committee longlisted 29 user cases (p. 7), shortlisted 20 (p. 9), and announced three selections: two from users and one Meta referral (pp. 4, 11). Meta's legal team checks the shortlist for eligibility, which "does not include a re-review on the merits"; "in practice, Meta has also assessed whether its original decision ... was correct or not" (p. 9). Meta judged its own decision wrong in 14 of the 20 shortlisted cases, a rate of 70%, against 39% (51 of 130) for shortlisted cases from Q4 2020 to Q4 2021, and reversed itself in all 14 (10 restorations, 4 removals) (p. 9). The Board calls the sample small and says it seeks hard cases, yet "is raising with Meta the questions this poses for the accuracy of Meta's content moderation and the appeals process" (p. 9). Annex I summarises the 14; in most, the Board writes "Without explanation from Meta on either the removal or the restoration, it seems ..." (pp. 23–25). + +**Decisions.** "The Board's decisions are binding, and Meta must implement them within seven days of publication" (p. 12). The Board decided two cases and overturned Meta in both (p. 12). One (2021-015-FB-UA) was a user appeal; the other (2021-016-FB-FBR) was a Meta referral. Both were assigned in October 2021 and published on 1 February 2022; they took 105 and 109 days against a 90-day target, an average of 107 days from assignment to implementation (p. 15). + +**Policy.** The Board issued one policy advisory opinion and 22 recommendations in total: 12 on content policy, 7 on enforcement, 3 on transparency (p. 18). Meta said it would implement 8 "fully", 2 "in part", was "assessing feasibility" on 7 and would take "no further action" on 5 (p. 19). Across all 108 recommendations to date, implementation was demonstrated by published information for 16, partly for 2; "progress reported" for 44; Meta claimed but did not show implementation for 25; declined after assessment 7; omitted, declined or reframed 14 (p. 21). The Board says it "continues to lack data to verify progress on or implementation of the majority of recommendations" (p. 21). Meta answered 53 of the Board's 62 questions (85%), partly answered 8 and left 1 unanswered (p. 15). + +## Relation to the argument + +For D8 the Board is the platform world's nearest thing to a court, and the numbers show what kind. It reviews a chosen docket: about 479,000 user cases came in, 20 were shortlisted, 3 were announced as selected, and the 2 decisions published that quarter came from cases assigned in October 2021. Each decision binds only on its own content, within seven days. Policy recommendations bind nobody; Meta implements some, declines some, and the Board cannot verify most. The Board's gate and its data both come through Meta (Meta's legal review, p. 9; the Case Management Tool "created by Meta", p. 2). So the convention is still held by the owner, with an outside voice added. + +Annex II adds a point for the reasons comparison. The Board recommended in 2021, repeatedly, that users "always" be told the specific rule enforced against them (for example Armenians in Azerbaijan #1, p. 26; Breast cancer #2, p. 27; Nazi quote #1, p. 28; South Africa slurs #1, p. 51). Each stood at "progress reported" in Q1 2022. That suggests, as an inference and not a stated finding, that giving reasons was not Meta's own convention before the Digital Services Act made it a legal duty (`eu2022dsa`, Art. 17). + +**Facet check:** (1) J:246, J:266(iv) and J:316 set "2 decisions" against "about 479,649 submissions". The audit already flagged the cohort mismatch. A second mismatch: one of the two decisions was a Meta referral, so only one of the two comes from the user pool counted in the denominator. (2) "Binding on content" is accurate as J words it; the report's own table shows that policy recommendations are not binding. (3) J:290 says "by Q4 2021" for binding decisions. This report supports "binding decisions" as of Q1 2022; the Q4 2021 report was not read here. (4) J does not use the 70% figure (14 of 20): Meta reversed itself in most shortlisted cases before the Board ruled. + +## In conversation with + +The report cites no work in `../references.bib`; it refers to the Board's Bylaws, its Rulebook for Case Review and Policy Guidance, its 2021 Annual Report and earlier transparency reports, and UN human-rights instruments (pp. 13–14). The project pairs it with `wong2022meta` (a critical assessment of the Board), `klonick2018new` (the pre-Board appeals picture), `eu2022dsa` (the statutory counterpart: Arts. 17, 20, 21) and `douek2023meta` (cited in the facet as an unopened critical reading); those pairings come from facet J. Leads not in the bib: the Board's 2021 annual report; the Q4 2021 transparency report. + +## Caution + +The Board describes itself. The 479,649 count is a manual estimate the Board calls "preliminary", and the report does not say how many of those cases were eligible. One quarter is a snapshot; nothing after March 2022 is in this file. The report cannot show how Meta's own appeals process treats users before the Board. The case-level facts in Annex I are the Board's inferences about Meta's reasons. I scanned rather than read Annex II. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the Board's PDF (64 pages; printed numbers equal PDF pages). Searched all 17 quotations (the three that did not match are facet-J wording, plus one line-break hyphenation, 're- review') and checked the numbers on pp. 4-9, 11-12, 15-16, 18-19, 21 (479,649; 479,653; 288,440; 94/6%; 45/22%; 99.2/0.8%; 29 longlisted; 20 shortlisted; 14 of 20 = 70% vs 39% (51/130); 105/109/107 days; 53/8/1 of 62; 22 recommendations = 12/7/3; 8/2/7/5; the 108-recommendation breakdown 16/2/44/25/7/14), the Annex II rows cited (pp. 26, 27, 28, 51) and the glossary line 'created by Meta'. + +**Findings:** One fix: the Read depth said full_text while stating that Annex II (39 of 64 pages) was only scanned; changed to partial. No wrong quotation, number or locator found. Note: the page-5 phrase 'an estimate and preliminary' is the Board's wording for the manual count, as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ozkan2026wordinitial.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ozkan2026wordinitial.md new file mode 100644 index 00000000..c0044f31 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ozkan2026wordinitial.md @@ -0,0 +1,38 @@ +# Özkan, Z. G., & Perea, M. (2026). Word-initial capitalization preserves lexical processing in unspaced text. *Scientific Reports.* + +**Identifier:** doi:10.1038/s41598-026-67473-2 · **Read depth:** full_text (accepted, unedited manuscript marked "Article in Press"; received 30 April 2026, accepted 14 August 2026; no journal pagination, so locators are sections and PDF page numbers, and the final version may differ) · **Source-tier:** peer-reviewed article (in press); the study was preregistered (the authors say a preregistration exists on OSF; I did not open it) · **Evidence basis:** direct_read · **Design and sample:** 2 × 2 within-subject eye-tracking experiment (interword spacing: spaced or unspaced × target-word frequency: high or low) on 120 Spanish sentence frames, every word capitalised in both spacing conditions; 60 University of València psychology students (43 female, mean age 21.8), 1,800 trials per condition · **Parties modeled:** person–artefact (skilled reader and text) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Özkan and Perea ask whether a cue other than a gap can mark word boundaries well enough for readers of unspaced text. Their cue is the initial capital, as in hashtags such as #ClimateChangeAction. Both conditions keep every word capitalised ("Martha Was Very Happy…" against "MarthaWasVeryHappy…"), so the contrast isolates the gap itself (Introduction, PDF p. 3; Materials, PDF p. 5). They use the word-frequency effect as the marker of lexical processing: if capitals do the segmenting, the frequency effect should not grow when spaces go. + +Results (Results; Tables 2–3, PDF pp. 8–9, 15–17): +- Global cost remained. Total reading time 2,790 ms unspaced against 2,609 ms spaced; mean fixation 206 against 198 ms; 13.5 against 13.1 fixations (Table 2). +- Local frequency effects were additive with spacing. For first-fixation duration, frequency 214 against 203 ms, spacing 211 against 206 ms, interaction b = 0.02, 95% CrI [−0.01, 0.05], BF01 = 31.76. Gaze duration: 351 ms unspaced against 305 ms spaced; interaction b = −0.01, BF01 = 32.92. Total viewing time: 549 against 496 ms; interaction BF01 = 20.07. +- Initial landing position moved toward the word's start without spaces: 33.9% of word width against 36.6% (b = 0.15, 95% CrI [0.06, 0.24]). The authors convert this to about 2.5 against 2.7 character widths. +- Skipping probability: no clear effect (3.94% unspaced against 5.01% spaced). + +The authors' reading: "initial capitals provide useful word-onset information, but interword spaces provide additional visual and oculomotor cues that capitalization alone does not fully restore" (Abstract). They suggest the gap may also reduce lateral masking of boundary letters (Discussion, PDF p. 11). + +## Relation to the argument +It supports the second half of premise P3: marking boundaries by other means recovers part of the loss. In this study, capitals preserve the lexical side (no extra frequency penalty) and not the oculomotor side (landing shifts, longer reading). The point matters for ruling D6 (a complex of conventions, with a search for the one doing the most work): the space bundles several jobs, boundary marking, word-extent information, and release from lateral masking, and capitals replace only the first. For premise P5 it gives a case where an artefact feature (capitals) carries a segmentation job otherwise left to the reader, with the residual cost measured. + +Facet check: the facet's abstract-level reading ("initial capitals restored lexical processing but not all of the oculomotor cost") overstates one thing. The design has no lowercase spaced baseline and no unspaced lowercase condition, so it cannot show what capitals restore relative to ordinary text; it shows only that the frequency effect is the same size with and without spaces once capitals are present. The authors say as much: the design "does not test any additional cost or benefit that word-initial capitals themselves may impose relative to ordinary lowercase spaced text" (Discussion, PDF p. 11). The global reading-time cost also remained (about 7%, by my arithmetic from Table 2). + +## In conversation with +- `perea2009space` — alternating bold cut the unspaced cost; the frequency effect stayed larger (Introduction). +- `veldre2017spelling`, `drieghe2017parafoveal` — evidence on the unspaced cost, cited among refs 3 and 6 (Introduction). +- `rayner1998unspaced` — the baseline: lexical and oculomotor costs of removing spaces (ref. 2). +- Leads not in the bib: Perea, Tejero & Winskel (2015), alternating colours as a segmentation cue; Zhou et al. (2018), alternating colours in Chinese; Binkley et al. (2009) and Sharif & Maletic (2010) on camelCase against underscores; Slattery & Rayner (2013); Rayner & Schotter (2014) on initial capitals and preview benefit. + +## Caution +Sixty Spanish students reading single-line sentences with one target word each; one frame set. Landing positions are reported in pixels as a share of word width (Arial is proportional). The conclusion that capitals "preserve" lexical processing rests on Bayes factors for a null interaction, with weakly informative priors the authors chose. A final published version may change numbers or wording. The paper is about skilled adult readers of a spaced script, not learners or historical readers. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The accepted-manuscript PDF (cards/ozkan2026wordinitial.pdf, 18 pages), read in the Abstract, Introduction, Method, Results, Discussion, Tables 2-3, footnotes 1-3 and the reference list. About 40 numbers (Table 2 means, all three frequency-by-spacing BF01 values, CrIs, landing positions 33.9% and 36.6%, skipping 3.94% and 5.01%, N = 60, 43 female, mean age 21.8, 120 sentences, 1,800 trials) and two quotations checked, plus the cited references (Drieghe 2017 = ref 3, Veldre 2017 = ref 6, Rayner 1998 = ref 2, Perea and Acha = ref 5, Perea, Tejero and Winskel = ref 24, Rayner and Schotter 2014 = ref 16). + +**Findings:** Every number and both quotations match, and the "about 7%" global cost is right (2,790 / 2,609). One locator fix: the "Martha Was Very Happy" example is in the Introduction on PDF p. 3 and the all-words-capitalised design statement is in Materials on PDF p. 5, not "Method, PDF p. 4". One caveat on a phrase in the card: "the frequency effect should not grow when spaces go" is the authors' prediction on p. 4, stated correctly. The OSF preregistration was not opened by me either. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/palka2025foreign.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/palka2025foreign.md new file mode 100644 index 00000000..64cdc92c --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/palka2025foreign.md @@ -0,0 +1,35 @@ +# Pałka, P., Porębski, A., Pałosz, R., & Wiśniewska, K. (2025). Do foreign service providers simply ignore the EU consumer law? An empirical study of 100 online terms of service. *Journal of Consumer Policy, 48*(3), 233–268. + +**Identifier:** doi:10.1007/s10603-024-09580-z · **Read depth:** full_text (open-access publisher PDF, journal pagination 233–268; the two appendix tables of regression output skimmed) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** coding of 100 online terms of service (ToS) obtained 22 February 2022 from platforms operating in Poland, sixteen market sectors, coded by two annotators on nine variables against the EU Unfair Contract Terms Directive (UCTD); logistic regression and contingency tests on three predictors; purposive sample, exploratory · **Parties modeled:** two parties (provider and consumer) with public enforcers as a third party · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +The authors ask how many ToS offered in the EU contain clauses that could be unfair under the UCTD, and whether the pattern follows the provider's headquarters, listing status or price. Their count is blunt: "100% of contracts contained clauses that could be deemed unfair" (p. 234). Of the 100, 76% had a potentially unfair limit on liability, 85% had clauses limiting consumers' rights in dispute resolution, and 99% had at least one clause letting the provider alter the economic relationship unilaterally (p. 234, p. 247). Seventy-four percent breached five or more of the nine variables, and 41% breached seven, eight or nine (p. 246). + +Unfairness here is "potential": a judgment from the text and the law, not a court or agency finding (p. 239). Each variable took the value −1 (clear violation), 0 or 1; they also built binary versions. The sample held 15 providers headquartered in Poland, 23 elsewhere in the EU, 44 in the US and 18 in other jurisdictions (p. 239); 28 services were free, 8 paid, 64 optionally paid; 38 were publicly listed, 39 private and 23 indirectly public (p. 244). + +Headquarters was the only predictor with a pattern. For six of nine variables the best model contained headquarters alone (p. 248). In logistic models on the binary variables, seven of nine variables related to headquarters at p < .035, and pseudo-R² (McFadden) exceeded .25 for unilateral change of contract and for account deletion (p. 248). No non-EU group ever scored significantly better than the EU. Grouped as EU (n = 38) against non-EU (n = 62), full compliance was more common in the EU group, but incompliance stayed "high or very high" in the EU on four variables (pp. 249–250). Payment status and public listing showed no link, apart from one variable (arb) where the best model also contained payment status (p. 248). + +The authors offer untested explanations: providers weigh the cost of fines against the gain from deterring suits (an "in terrorem" effect, p. 234, p. 252); US firms treat penalties as business costs; ToS are copied between firms; national agencies police home firms first; and "diffusion of responsibility" among national agencies that are each competent but none responsible (p. 257). They also raise an uncomfortable reading, that under-enforcement may suit a system sceptical of mandatory terms (pp. 253–255), and ask lawmakers to say so openly. Their proposals: target non-EU providers, use automated clause detection, centralise enforcement for very large platforms under the Digital Services Act, and standardise the format and placement of risky clauses (pp. 257–258). + +## Relation to the argument +Ruling D8 (platforms "take a highly draconian position"). The study gives a measured form of that claim for one convention, the platform's contract. It supports premise P10, that the interface owner's convention is not one the user can hold: the contract is written by the provider, and the clause type most often present (99%) is the right to alter the relationship unilaterally, with 81 of 100 coded −1 on unilateral change of contract, i.e. reserving a right to change the contract without a valid reason (Table 3, p. 247; definition in Table 1, p. 241). + +It also complicates D8. The law has bent toward the consumer: the UCTD makes unfair terms non-binding (p. 236), and EU providers score better. What fails is enforcement, spread across agencies, not the existence of a convention. The proposal to put risky clauses "always in the same part of the contract" (p. 258) is a regulator-imposed layout convention, which echoes the spacing precedent; the match is the card's reading, not the authors'. + +Facet check: the facet reads the paper as "terms still contain unfair clauses after the EU consumer-law regime". The text agrees but adds two limits. The comparison is within the EU market, so it says nothing about platform terms elsewhere. EU-headquartered firms were also affected (23 non-Polish EU firms plus 15 Polish). + +## In conversation with +The paper cites no work in `references.bib` (its literature is on consumer contracts: Loos and Luzak, Lagioia et al., Micklitz et al., Ben-Shahar, Marks on "in terrorem" effects). Leads from its list, not in the bib: Luzak and Loos (2016), which the authors call the first article of this kind, on a very small sample; Lagioia et al. (2022) on automated detection; Ben-Shahar and Bar-Gill (2013) and Ben-Shahar (2014) on the efficiency critique of abstract control; Nyarko (2021) on copying in contracts; Marks (2019). + +## Caution +Potential unfairness from text alone; no court has held these clauses unfair, and the UCTD makes them non-binding when contested. The sample is purposive and built around the Polish market in 2022; the authors say it cannot be deemed representative of the entire population of online services (p. 244); the "other" group is heterogeneous; subgroup counts are small; the headquarters variable is the top parent's jurisdiction. The main text reports no inter-annotator statistic, only consistency checks and collaborative resolution of ambiguities by the principal investigator and the annotators (p. 239). The causal explanations are speculation, as the authors say (p. 255). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The open-access publisher PDF (cards/palka2025foreign.pdf, printed pp. 233-268; PDF page n = printed 232+n). Four quotations and about 45 numbers and locators (100%, 76/85/99%, 74%, 41%, sample counts, Table 3, Table 4 pseudo-R2, p-values, collection date 22 Feb 2022) checked; the policy and explanation pages (252-258) read. + +**Findings:** The core numbers and quotations are right. Six fixes: (1) the "cannot be deemed representative" statement and the free/paid/listing counts are on p. 244, not p. 240; (2) the 23 indirectly public firms are now stated; (3) payment status was not wholly unrelated, since the best model for arb included it (p. 248); (4) the "81 of 100" figure is in Table 3 (p. 247), not appendix Table 6; (5) the policy proposals span pp. 257-258; (6) the first-study lead is Luzak and Loos (2016), described by the authors as on a very small sample, and the ambiguities were resolved by the PI together with the annotators, not by the PI alone. Not checked: the appendix tables 5-6, which I only skimmed. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/park2025dokkyo.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/park2025dokkyo.md new file mode 100644 index 00000000..8e6a6eda --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/park2025dokkyo.md @@ -0,0 +1,29 @@ +# Park, J.-B. (2025). Diversification of literacy patterns by social class in late Joseon (forum abstract). 36th Dokkyo International Forum, Dokkyo University. + +**Identifier:** https://www.dokkyo.ac.jp/international/36-10.html · **Read depth:** full_text of the web page, which is a conference abstract in Japanese, Korean and English (about 350 words in English); the talk and any paper behind it were not read · **Source-tier:** conference abstract (web page) · **Evidence basis:** direct_read · **Design and sample:** none stated; the abstract describes a survey of five social classes and says household-register studies could later be used to estimate literacy rates · **Parties modeled:** other (social classes, women) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues +The presenter divides late Joseon society into five groups: yangban, chungin, commoners, cheonmin and slaves, and women. Most yangban and chungin had "dual literacy in both Hangul and Classical Chinese", and learned both from childhood because "bilingual texts (Hangul-Chinese), which used both scripts side by side, had become widespread from the elementary education stage onward". Commoners and cheonmin "were often excluded from the long-term, systematic education required to acquire literacy in Classical Chinese", so few could read it and "they primarily conducted their literate activities using Hangul". Most women, whatever their class, found learning to read harder than men did; upper-class women learned by "learning by observing over the shoulder" or from fathers and brothers, and those who read conduct manuals such as the *Yeosaseo* often gained dual literacy. The abstract closes by expecting that class ratios from household registers could be matched to these patterns to estimate literate-population size and rates in Hangul and Chinese. It gives no numbers. + +## Relation to the argument +Supports, as a claim and not a measurement, the facet-F point that a script made for easy learning (Hangul) was read first by those shut out of the elite script: commoners and women. That fits a cost-lowering reading of a notation (P6) and the premise that learners gain most from a re-representation (P4). It gives no evidence on how many people learned or how fast, and the speaker names the estimate as future work. + +Facet check: the facet marks this a forum abstract; the page confirms it and shows it contains no counts or rates, so it cannot be cited for literacy levels. + +## In conversation with +- `hangeulmuseum`: a museum page also reports Hangeul letters written across classes, including a landlord's letter to a slave that "implies that the slave could read Hangeul". +- `clemens2025korea`, `baek2024spread`, `ivanova2026process`: other facet-F sources on Hangul spread; the abstract cites none of them. +- Leads: none given on the page. + +## Caution +A 350-word abstract of a talk: no data, no sources, no page numbers. The author's name is rendered "朴 鍾培 / Park Jong-Bae" on the page. The "five groups" are the presenter's categories, and "most" has no figure behind it. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The saved HTML page (cards/park2025dokkyo.html), read in full in all three languages (the English text is about 330 words). Six quotations, the five-class list, the household-register sentence and the author-name rendering checked against the English and Japanese/Korean versions. + +**Findings:** No issues found. Small precision notes only: the English text says "fathers and elder brothers" (the card says "fathers and brothers"), and "a small minority--or in many cases, a mere handful" is the abstract's only quantity word. The `hangeulmuseum` claim in "In conversation with" is about another source and was not checked here. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/perea2012subtle.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/perea2012subtle.md new file mode 100644 index 00000000..3d8b8ed4 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/perea2012subtle.md @@ -0,0 +1,38 @@ +# Perea, M., & Gomez, P. (2012). Subtle increases in interletter spacing facilitate the encoding of words during normal reading. *PLoS ONE, 7*(10), e47568. + +**Identifier:** doi:10.1371/journal.pone.0047568 · **Read depth:** full_text (open-access published PDF, seven pages; PLoS ONE prints no page range, so locators are section names and PDF page numbers) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** within-subject eye-tracking experiment, 24 University of València students (mean age 20, native Spanish), 120 target sentences (40 per spacing condition; Times New Roman 14 pt, one line per sentence); interletter spacing 0.0 (default), +1.0 and +1.5 points, applied to all characters including spaces; target word high or low frequency · **Parties modeled:** person–artefact (skilled reader and text) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Earlier lexical-decision studies by Perea and colleagues found that single words are identified faster with a small spacing increase (about +1.2 points; "w a t e r" faster than "water"). Perea and Gomez ask whether this carries over to normal sentence reading. They offer two rival predictions: wider letters ease encoding (less crowding) or push neighbouring words further into the periphery and raise the number of fixations (Introduction, PDF pp. 1–2). + +Results (Results, PDF pp. 3–5; Table 2–3): +- Average fixation durations after forward saccades shrank in order: 228 ms (default), 221 ms (+1.0), 216 ms (+1.5); F1(2,42) = 31.05, p < .001. First fixations on the target word: 227, 225 and 215 ms. 22 of 24 participants showed the advantage (Discussion, PDF p. 4). +- No spacing effect on gaze duration or total time on the target word, and no spacing-by-frequency interaction on first fixation (there was one on total time, significant only by items, p = .04, a reading cost for high-frequency words at +1.5), which the authors take as evidence of early encoding (Results; Discussion). +- Offsetting costs: more forward fixations and shorter saccades (7.58, 7.52, 7.27 characters), and landing positions closer to the word's start (2.1 against 1.8 and 1.7 on a five-area scale). +- Total sentence reading time: 2,119 ms (default), 2,068 ms (+1.0), 2,103 ms (+1.5). The +1.0 gain was significant by participants (F1(1,21) = 6.31, p = .02) and not by items (F2(1,117) = 2.32, p = .13). The authors write that they are "cautious in making strong inferences from the total reading data" because the effect is small, only two-thirds of participants showed it, and it missed significance by items (Discussion, PDF p. 5). + +They conclude that "interletter spacing is a relevant parameter when processing text" and that e-book software should let readers adjust it (closing paragraph of the Discussion, PDF p. 6; the paper has no separate Conclusion heading). They note the benefit for dyslexic readers shown by Perea et al. (2012) in total reading time, which control children did not show (Discussion). + +## Relation to the argument +Letter spacing, not word spacing, so the premises P1 to P5 are untouched. The paper matters to facet D as the positive case for subtle spacing among skilled adults: a real gain in fixation duration, cancelled in total reading time by extra fixations. That pattern shows the limit of reading "page-level" cost reductions from a single measure, the same caution that applies to ruling D6 (look for the move that does most of the work, with the cost measured end to end). It also shows that a spacing change helps at one stage and costs at another, which the argument should say if it ever lists spacing among the "interface" fixes. The authors note that their manipulation widened every character, including the word spaces, "so that between-word boundaries were well-defined" (Discussion); the finding does not transfer to cases where letter gaps approach word gaps, the case the authors discuss from Paterson and Jordan (2010), where large letter gaps without clear word boundaries hurt reading (Discussion, PDF p. 5). + +Facet check: the facet's one-line summary, "Fixations briefest at +1.5 spacing in skilled readers," is correct and omits that total reading time did not follow. + +## In conversation with +- `perea2012effects` (ref. 7) — the companion study with young readers and dyslexic readers. +- `perea2009space` (ref. 24) — source of the materials and of the additive-factors reasoning. +- Cited in reply by `hakvoort2017improvements` and `vandenboer2015default`, both of which count it among the studies where adults show no sentence-reading gain. +- Leads not in the bib: Paterson & Jordan (2010) on large letter-spacing increases in reading (the contrast the authors discuss); Perea & Gomez (2012), "Increasing interletter spacing facilitates encoding of words" (*Psychonomic Bulletin & Review*, ref. 6); Slattery & Rayner (2010); Reichle et al. (1998) E-Z Reader; Engbert et al. (2005) SWIFT. + +## Caution +Twenty-four adults, Spanish, single-line sentences, one font (Times New Roman, narrowly spaced by default), a CRT screen. The authors warn the result may not carry to e-readers or other fonts. The sentence-level speed gain is small and not reliable by items, so the paper cannot be cited for faster reading, only for shorter fixations. The dyslexia point is reported from another study and was not tested here. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The PLoS ONE PDF (cards/perea2012subtle.pdf, seven pages), read in the Abstract, Introduction, Methods, Results (Tables 2-3), Discussion and reference list. Four quotations, the fixation, saccade, landing-position and reading-time means, the F statistics for total time (F1 = 6.31, F2 = 2.32), "22 out of 24", the two-thirds figure, and the reference numbers 6, 7 and 24 checked. + +**Findings:** All numbers and quotations match. Three fixes: (1) the card attributes a finding to Galliussi et al. (2020), which this 2012 paper cannot cite; I replaced it with the Paterson and Jordan (2010) contrast the paper actually discusses; (2) the "relevant parameter" sentence is in the closing paragraph of the Discussion (PDF p. 6), not a "Conclusion" section; (3) the card's "no spacing-by-frequency interaction" is true for first fixation, but the paper also reports a total-time interaction by items (p = .04), now noted. Not checked: the claim that `hakvoort2017improvements` and `vandenboer2015default` cite this paper (those papers were not opened). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/perea2015colours.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/perea2015colours.md new file mode 100644 index 00000000..4488156d --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/perea2015colours.md @@ -0,0 +1,40 @@ +# Perea, M., Tejero, P., & Winskel, H. (2015). Can colours be used to segment words when reading? *Acta Psychologica, 159*, 8–13. https://doi.org/10.1016/j.actpsy.2015.05.005 + +**Identifier:** doi:10.1016/j.actpsy.2015.05.005 · **Read depth:** full_text (publisher-formatted PDF from the first author's University of Valencia page, uv.es/mperea/coloursPTW.pdf; journal pagination 8–13; read whole, including the two footnotes reporting extra experiments) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** eye-tracking (EyeLink II, 500 Hz) of 24 University of Valencia students, native Spanish speakers, none colour-blind, reading 120 single-line Spanish sentences (the Perea & Acha 2009 set) in three formats: spaced with alternating red/green words, unspaced with alternating colours, unspaced in one colour; 60 high- and 60 low-frequency target words (pp. 9–10) · **Parties modeled:** person–artefact (reader and typography) · **Facets:** W · **Read:** 2026-10-15 + +## What it argues + +The question is whether a visual cue other than space can do the space's job without pushing neighbouring words out of high-acuity vision (p. 8). Alternating colour nearly does. Total sentence reading time was 3,701 ms for unspaced one-colour text, 2,293 ms for unspaced alternating-colour text and 2,188 ms for spaced text (p. 10, Table 2), "i.e., a 4.7% reading cost" for colour against spaces (p. 12), where the one-colour unspaced text cost 69% (my arithmetic, 3,701/2,188). Progressive saccades were 9.8, 7.3 and 7.3 per sentence; mean fixation 289, 247 and 234 ms (Table 2). At the word, the frequency effect on gaze duration was 85 ms in one-colour unspaced text against 45 ms with colour and 27 ms with spaces, and the colour and spaced conditions did not differ reliably (p. 11), so "word identification processes were not severely hindered by lack of spaces between words" once colour marked them (p. 12). Landing positions still moved toward word beginnings without spaces (1.7, 1.9, 2.2 on a five-zone scale, p. 11), which the authors read as the low-spatial-frequency role of the space in saccade targeting (p. 12). Comprehension was above 93% in every condition (p. 10). Footnote 1 reports a further 24 readers trained on a 3,000-word unspaced alternating-colour story; spaced text kept "a small (but significant) advantage ... (around 9 words per minute)" (p. 13). Footnote 2 reports 33 skilled Thai readers for whom spaces in one colour imposed a cost (as in Winskel et al. 2009) while alternating colour had no effect (p. 13). + +The paper restates its predecessors with numbers. Perea and Acha (2009): spaced 1,856 ms, alternating bold 2,708 ms, unspaced 3,343 ms, "a 31% reading cost" for bold against spaces (p. 9). Rayner, Fischer and Pollatsek (1998): "the lack of spaces between words in English hinders the process of word identification (e.g., lexical effects such as word-frequency are magnified in unspaced sentences relative to spaced sentences) as well as eye movement guidance (e.g., landing sites are closer to the beginning of the word ...)" (p. 8). Its one historical sentence: "Around 1000 A.D., spaces were added between words to help the process of word segmentation (see Saenger, 1997 ...)" (p. 8). + +## Relation to the argument + +This is the cleanest experimental support for the abstract's phrase "exploiting a representational effect": what the reader needs is a boundary the eye can see, and space is one of several marks that supply it. Colour recovers almost everything; bold recovers less; nothing recovers all of it at the saccade-targeting level. Three consequences for the sentence under test. + +- **Supports** "moving boundary detection ... to the page" in the sense that a visual mark relieves the reader of segmentation at a measured cost: 69% longer without any mark in this sample of readers trained on spaces. +- **Narrows** "to the page" to "to any salient visual mark", which matters for the paper's design argument: the Insular innovation is one instance of a class (Zhang & Norman's representational effect), not a unique discovery. +- **Says nothing about the voice.** No condition involved reading aloud; the cost is measured on silent readers. +- **A date to flag.** The authors date word spaces to "around 1000 A.D." on Saenger's authority. That is the Continental adoption, not the seventh-century Insular origin the abstract names; both are in Saenger, and the paper should not borrow this sentence as corroboration. +- Footnote 2 repeats the Thai finding (`winskel2009eye`): readers of an unspaced script are slowed by added spaces, which P4 and D2 already carry. + +## In conversation with + +- `rayner1998unspaced`: the template and the baseline the paper is built on (p. 8). +- `perea2009space`: the bold manipulation this study improves on; the only place in the library where that paper's numbers (1,856 / 2,708 / 3,343 ms) are reported from a source that had them. +- `bai2008reading`: cited for the Chinese highlighting experiment (p. 9). +- `winskel2009eye`, `kasisopa2013eye`: the Thai comparisons (pp. 9, 13). +- `zhang1994representations`: the representational-effect frame the author's abstract uses; not cited by Perea et al. +- Leads not in the bib: Slattery & Rayner (2013) on intra- and inter-word spacing; Pinna, Uccula & Tanca (2010) on colour and reading aloud; Winskel, Perea & Ratitamkul (2012) on Thai bold. + +## Caution + +Twenty-four skilled adult readers of one spaced language, single sentences, laboratory conditions; the unspaced cost mixes unfamiliarity with the loss of the cue. The footnoted Thai experiment recorded no eye movements. No child, learner or oral-reading condition. The historical dating is an aside borrowed from Saenger and is not evidence. + +## S2 adversarial verification (2026-10-15) + +**Verdict: corrected (page locators).** + +**What I checked:** Re-opened the publisher-layout PDF text (local copy dxd_foundation_local/pdfs/W/perea2015colours.txt; the running heads mark pp. 9–13, and the title page is p. 8). Numbers checked against the text and Tables 2–3: total times 3,701 / 2,293 / 2,188 ms; progressive saccades 9.8 / 7.3 / 7.3; mean fixation 289 / 247 / 234 ms (Table 2, p. 10); "a 4.7% reading cost" (p. 12); frequency effect 85 ms against "45 and 27 ms" (p. 11); landing zones 1.7 / 1.9 / 2.2 on five fixation zones (p. 11); comprehension "above 93%" (p. 10); Perea and Acha 1,856 / 2,708 / 3,343 ms and "a 31% reading cost" (p. 9); 24 participants, EyeLink II at 500 Hz (p. 9); footnote 1 (24 readers, "close to 3,000 words", "around 9 words per minute") and footnote 2 (33 Thai readers, no eye movements) on p. 13. Quotations "not severely hindered by lack of spaces between words" (p. 12) and the Saenger sentence match. Crossref 10.1016/j.actpsy.2015.05.005: Perea, Tejero, Winskel; *Acta Psychologica* 159, 8–13, 2015; matches header and .bib. Read depth full_text stands. The card's 69% is its own arithmetic and is labelled as such. + +**What I changed:** Three locators moved from p. 9 to p. 8: the framing question (abstract), the restatement of Rayner et al. 1998 (now quoted verbatim) and the "Around 1000 A.D." sentence all stand on the first page of the article. The facet file and the `rayner1998unspaced` and `saenger1997space` "Also used by facet W" sections carried the same p. 9 and were fixed. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/price2011interactive.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/price2011interactive.md new file mode 100644 index 00000000..74faf269 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/price2011interactive.md @@ -0,0 +1,41 @@ +# Price, C. J., & Devlin, J. T. (2011). The interactive account of ventral occipitotemporal contributions to reading. *Trends in Cognitive Sciences, 15*(6), 246–253. + +**Identifier:** doi:10.1016/j.tics.2011.04.001 · **Read depth:** full_text (PMC author manuscript, PMC3223525, Europe PMC XML, including Boxes 1 to 4 and the Glossary; no journal page numbers, so locators are section and box names) · **Source-tier:** peer-reviewed theory article (the XML types it "Opinion") · **Evidence basis:** direct_read · **Design and sample:** none (theoretical account that reviews fMRI, priming and lesion findings) · **Parties modeled:** other (a brain region mediating between visual input and language) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Price and Devlin put forward the Interactive Account of the ventral occipitotemporal cortex (vOT). vOT, they argue, is not tuned to orthography. It combines bottom-up visuospatial input with top-down predictions from phonological and semantic areas, in a predictive-coding framework. In the Abstract: "specialization for orthography emerges from regional interactions without assuming that vOT is selectively tuned to orthographic features." + +The main steps, with locators: +- Opening section: vOT responds to many non-orthographic stimuli and even to touch, so no single functional label fits all its responses. +- Section "The Interactive Account": activation reflects forward input plus backward prediction plus prediction error; fMRI cannot separate them. Without learned links between orthography and phonology, vOT activation to words is low because "phonological areas do not send backward predictions to vOT." +- Section on evidence for automatic top-down effects: a masked picture prime of the same object reduced vOT activation to a named word, though participants did not see the prime. +- Sections on selectivity: activation is higher for pseudowords than for words and than for consonant strings, and higher for low- than for high-frequency words. The authors explain these as prediction error and reject explanations from bigram frequency or familiarity. +- Box 1: activation is low in pre-literates, high in early learners, lower again in skilled readers as predictions improve; in developmental dyslexia vOT activation is low, which they attribute to a failure to establish top-down predictions. +- Box 2: damage to vOT disconnects forward and backward links and hurts reading more than object naming because words share repeating parts in similar combinations (attitude, altitude, aptitude). +- Box 3: vOT neurons are "general-purpose analyzers of visual forms," with no one-to-one mapping to letters or bigrams. +- Concluding remarks: the account "obviates the need to postulate a novel form of learning-related plasticity (e.g. 'neuronal recycling') [7] or reading-specific neuronal responses (e.g. 'bigram detectors') [8]." + +## Relation to the argument +The paper bears on facet D's claim that the neuroscience of reading offers no clean evidence that a written convention reshaped a dedicated brain region. Price and Devlin reject a reading-specific region and recycling, and Kubota et al. (2024) reject a different part of the recycling story. The paper never discusses word spacing, unspaced script or reading acquisition in the historical sense; it says nothing on P1 to P6 or on rulings D1 to D10. + +Facet check: the facet quotes the Abstract line on orthography emerging "from regional interactions", and the quotation is exact. The facet records the paper as read in part (abstract and opening); the full reading adds that the account is a framework argued from existing studies, not a new experiment, and that it explicitly sets itself against Dehaene's recycling and bigram-detector proposals. It also makes the point that task and attention reverse selectivity, which weakens any claim that vOT activation shows "word-specific" tuning. + +## In conversation with +- `dehaene2007cultural` — cited as ref. 7, the neuronal recycling hypothesis the account is set against. +- `price2003myth` — Price & Devlin (2003), ref. 15, the earlier statement that the visual word form area is not reading-specific. +- `dehaene2010learning` — cited as ref. 53 (Dehaene et al., 2010); the card did not check what the text draws from it. +- `reicher1969perceptual`, `wheeler1970processes` — not cited in this paper. +- Leads not in the bib: Dehaene (2005), "The neural code for written words" (ref. 8); McClelland & Rumelhart (1981) (ref. 28); Friston's predictive-coding papers (refs. 22 and 23); Cohen et al. (2004) on letter-by-letter reading. + +## Caution +A 2011 theoretical position paper; later longitudinal work (see `kubota2024rethinking`) tests recycling directly and the account's own predictions about dyslexia rest on a small set of cited studies. The paper is about alphabetic orthographies in adults, and its Box 3 neuronal properties are extrapolated from monkey data because the authors know of no human single-cell data. The Europe PMC XML (CC BY 3.0, Elsevier Sponsored Documents) flags itself "pmc-prop-manuscript no", so it may be the publisher text rather than an author manuscript; no journal pagination is present, so page numbers could not be checked. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The Europe PMC XML (cards/price2011interactive.xml, PMC3223525), converted to text and read in the Abstract, the opening, the Interactive Account section, Boxes 1 to 4, the Concluding remarks and the reference list (about 7,000 words). Four quotations ("specialization for orthography...", "phonological areas do not send backward predictions to vOT", "general-purpose analyzers of visual forms", "obviates the need...") and the claims on tactile responses, masked-picture priming, pseudoword and frequency effects, Box 1 stages, the Box 2 "attitude, altitude, aptitude" example, the monkey-data admission in Box 3, and refs 7, 8, 15, 28 and 53 checked. + +**Findings:** All quotations and attributed claims match. Two fixes: the card calls the file an "author manuscript", but the XML's own flags say it is not a manuscript deposit, so I stated the uncertainty; and the article type in the XML is "Opinion", now noted. The Kubota et al. (2024) remark is the card's own link to a later paper, not something in this 2011 source. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/quintilian1920institutio.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/quintilian1920institutio.md new file mode 100644 index 00000000..bb1f34c8 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/quintilian1920institutio.md @@ -0,0 +1,36 @@ +# Quintilian. (1920). *The orator's education* [*Institutio oratoria*], Book 1 (H. E. Butler, Trans.; Loeb Classical Library, Vol. 1). Harvard University Press; Heinemann. + +**Identifier:** https://penelope.uchicago.edu/Thayer/E/Roman/Texts/Quintilian/Institutio_Oratoria/1A*.html (Book 1, chapters 1–3; also 1B, chapters 4–6; 1C, chapters 7–12) · **Read depth:** partial (the card writer read the parts listed below; the saved LacusCurtius pages hold all of Book 1). LacusCurtius transcription of the Loeb English text, proofread by the site; Loeb English pages are marked "p35", "p37", etc. (odd numbers; the even pages carry the Latin). Read closely: the Preface, 1.1 (in full), the opening of 1.2, 1.5.6, 1.7 (in full) and 1.8.1–5. Not read: the rest of Book 1 (1.2 from its second half, 1.3–1.4, the rest of 1.5–1.6, 1.8 from § 6, 1.9–1.12). The source is the translator's English; the Latin was not checked. · **Source-tier:** ancient text in public-domain translation (1920) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** person–artefact (teacher, child, text) · **Facets:** K · **Read:** 2026-10-13 + +## What it argues +Book 1 is Quintilian's account of the education that comes before rhetoric. Three parts bear on the paper. + +**Start young, and teach reading as slow drill.** He rejects the view that boys should wait till seven to read (1.1.15–19), sets the order of teaching (letter shapes before names, 1.1.24–25; syllables with no shortcut, 1.1.30), and on speed: "You will hardly believe how much reading is delayed by undue haste" (1.1.32, Loeb p. 37). "Reading must therefore first be sure, then connected, while it must be kept slow for a considerable time, until practice brings speed unaccompanied by error" (1.1.33). The difficulty he names is attention, not the page: "it is necessary to keep the eyes on what follows while reading out what precedes, with the resulting difficulty that the attention of the mind must be divided, the eyes and voice being differently engaged" (1.1.34, p. 37). + +**Orthography is judged by its cost to learners.** In 1.7 he treats spelling as "the servant of usage" that "undergoes frequent change" (1.7.11), and rejects an old convention of adding an *e* before *i* in endings because "such a practice will be an actual hindrance to those who are learning to read" (1.7.17, p. 139). His stated test, "within the limits prescribed by usage," is that "words should be spelt as they are pronounced," since "the use of letters is to preserve the sound of words and to deliver them to readers as a sacred trust" (1.7.30–31, p. 145). Chapter 1.7 names changes in letters, doubling, diphthongs and abbreviations; it says nothing about the spacing of words. + +**Reading aloud is taught by practice and by understanding.** "Reading remains for consideration. In this connexion there is much that can only be taught in actual practice, as for instance when the boy should take breath, at what point he should introduce a pause into a line, where the sense ends or begins" (1.8.1, p. 147); his "golden rule" is that "he must understand what he reads" (1.8.2). At 1.5.6 he notes that mistakes in speaking come from "separation or combination of syllables" — an observation about sound, not layout. + +## Relation to the argument +Facet K asks whether anyone in the history of reading instruction says that separating words shortened the training of readers (rulings D1, D2). This book is the clearest ancient account of how reading was taught, and it contains no such claim and no mention of word division. The cost of learning it reports is drill, divided attention between eye and voice, and the reader's task of finding where to breathe and pause, which fell on the pupil and was taught by the teacher. That complicates premise P6 (separation lowered a cost of reading) by showing where a Roman teacher placed the burden. It also gives the paper something positive for D10 (history as a guide): Quintilian does judge an orthographic convention by whether it hinders learners (1.7.17), so the idea that a writing convention can be weighed by its cost to the learner is ancient, though he never applies it to spacing. + +Facet check: the facet's quotations at 1.1.30, 1.1.32–34 and 1.8.1–2 are correct. Two page corrections from the page markers in this transcription: 1.1.34 sits on p. 37 (the facet gives pp. 37–39), and 1.8.1–2 on p. 147 (the facet gives p. 148). The facet's "read in full" covers 1.1.15–37 and 1.8.1–2; this reading adds 1.7 and 1.5.6. + +## In conversation with +- `heilmann2020reading` — the facet file reports that Heilmann reads 1.1.34 as about learners; this card shows the sentence's own words are about divided attention (eye and voice). +- `saenger1997space`, `knox1968silent`, `parkes1992pause` — the paper's main sources on word separation and reading practice; Quintilian does not mention them (he is earlier) and they do not appear in his text. +- `dionysius1910composition` — Roberts's notes there cite Quintilian (*Inst.* 8.6.64, 9.4.72, 10.4.4); Book 1 does not cite Dionysius. +- Within the text Quintilian cites Cicero's care for correct speech and Julius Caesar's *On Analogy* (1.7.34) (neither in the bib). + +## Caution +Partial reading of Book 1 (see Read depth), and Books 2–12 not read. Butler's translation, not the Latin. Quintilian describes his ideal curriculum, not typical practice. Absence of a remark on word spacing is not evidence that texts were spaced or unspaced. Loeb page numbers are those of the 1920 English pages as marked on the transcription. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The three saved LacusCurtius pages (cards/quintilian1920institutio_1-3.html, _4-6.html, _7-12.html; Loeb vol. I, Butler, 1920), converted to text. I did not read the parts the card marks as unread beyond locating the quoted passages. Ten quotations (1.1.32-34, 1.7.11, 1.7.17, 1.7.30-31, 1.8.1-2), the page markers p37, p139-p147, the 1.5.6 sentence, the seven-year view (1.1.15) and the Caesar "Analogy" reference (1.7.34) checked. + +**Findings:** All quotations are exact. Three fixes: (1) the Read depth label "partial_full_text" is not a README value, so I changed it to "partial", keeping the list of unread parts; (2) the 1.7.17 sentence is on p. 139 (the p141 marker comes after it), not pp. 139-141; (3) the spelling test is qualified in the text by "within the limits prescribed by usage", now included. The unread-parts list is the card writer's own; the saved pages in fact hold all of Book 1. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/reis2011johnson.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/reis2011johnson.md new file mode 100644 index 00000000..f5798393 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/reis2011johnson.md @@ -0,0 +1,33 @@ +# Reis, D. M. (2011). Review of William A. Johnson, *Readers and Reading Culture in the High Roman Empire: A Study of Elite Communities* (Oxford University Press, 2010). *Religious Studies Review, 37*(3). + +**Identifier:** doi:10.1111/j.1748-0922.2011.01538_4.x · **Read depth:** full_text (author-hosted PDF of the Wiley review, about 300 words; the copy carries no journal pagination, so the bib's pp. 219–220 are not checked) · **Source-tier:** book review (short notice) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** other (reader and reading community) · **Facets:** B · **Read:** 2026-10-13 + +## What it argues +Reis summarizes Johnson in one paragraph. Johnson "sees [reading] more broadly" than as "a strictly cognitive act," and defines it as "the negotiated construction of meaning within a particular sociocultural context." He studies "reading communities" and how their attitudes to books "were both informed by and shaped their worldviews." + +The one sentence on script: "Johnson shows that the scriptio continua of the bookroll and its costly aesthetic features ensured that the consumers of texts were limited." Restricting access "was precisely the point for elites who used reading to create and reaffirm the values of high society." The chapters on Pliny, Tacitus, Galen, Gellius, Fronto, Lucian and the Egyptian papyri show aristocrats fitting reading into daily routines. Such reading "separated them from the hoi polloi." Reis ends by calling the work "an important contribution to discussions of elite culture." + +## Relation to the argument +Reis states the sharpest version of the status reading: unspaced script, together with the bookroll's cost, "ensured" a limited readership. Ker and Mannering say the script is consistent with an elite culture; Reis says it limited the readership, and attributes the aim to the elites. That wording is Reis's compression and the paper should not treat it as Johnson's. It touches P6 (the move lowered a cost of reading, did not by itself raise the number of readers) and the dropped claim "Spacing 'ignited literacy'": on this account the Roman page excluded by design, so spacing did not cure an accident. + +The review does not mention word separation as a change, Saenger, silent reading or training for novices, so it adds nothing to P1–P4. + +Facet check: facet B lists it as read in full at about 300 words, which is right. The facet's description is neutral; the card flags that Reis's "ensured" is stronger than Ker's "consistent with." + +## In conversation with +- `johnson2010readers`: the book under review. +- `ker2012johnson`, `mannering2012johnson`: two longer reviews of the same book. +- No other source is cited in the review. + +## Caution +A 300-word notice for a religious-studies readership, which compresses Johnson. It cannot be cited for what Johnson says about script, only for how one reviewer put it. Pagination unverified. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** The author-hosted PDF of the Wiley review (cards/reis2011johnson.pdf, one page, 297 words), read in full. Nine quotations, the list of chapters (Pliny, Tacitus, Galen, Gellius, Fronto, Lucian, papyri) and the bibliographic details (Oxford University Press 2010, pp. ix + 227; first published online 14 Sep 2011; the DOI) checked. + +**Findings:** No issues found. Every quotation, including the bracketed "[reading]", matches the text; the card correctly treats "ensured" as Reis's wording. Not checkable here: the bib's pp. 219-220 (the PDF carries no pagination), and the comparisons with Ker, Mannering and the facet file, which concern other documents. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rfc8890.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rfc8890.md new file mode 100644 index 00000000..e8ae117a --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rfc8890.md @@ -0,0 +1,42 @@ +# Nottingham, M. (2020, August). *The Internet is for end users* (RFC 8890). Internet Architecture Board. https://doi.org/10.17487/RFC8890 + +**Identifier:** RFC 8890, doi:10.17487/RFC8890, https://www.rfc-editor.org/rfc/rfc8890 · **Read depth:** full_text (the RFC Editor plain-text edition, August 2020, ISSN 2070-1721; whole document read, including the reference list, IAB membership and acknowledgements; locators are section numbers) · **Source-tier:** report (IAB informational RFC; not a standard) · **Evidence basis:** direct_read · **Design and sample:** none; a position statement of the Internet Architecture Board · **Parties modeled:** other (standards authors, end users and "other parties" such as vendors, operators and governments) · **Facets:** O, T · **Read:** 2026-10-14 + +## What it argues + +The IAB states that "when there is a conflict between the interests of end users of the Internet and other parties, IETF decisions should favor end users" (Abstract). The document defines end users broadly, including people affected by a system they are not using (§2), and grounds the priority in the IETF mission statement (RFC 3935) and in the warning that measurable success "ignores how technology is so often used as a lever to assert power over users, rather than empower them" (§3). + +Section 4 lists ways to act on the priority. §4.1 concedes that "The IETF community does not have any unique insight into what is 'good for end users'" and that, even when the IETF reaches out, end users' "participation is minimal -- until a change affects someone in a way they don't like." It recommends going to affected communities and treating civil-society organisations as "a primary channel". §4.2 praises the standardised browser "user agent" as an intermediary that lets users switch implementations, and contrasts IoT, where "opaque, vendor-specific software" prevails. §4.3 says that "inappropriate concentration of power on the Internet has become a concerning phenomenon -- one that protocol design might have some influence upon." The document is explicit that it "is not an Internet Standards Track specification" (Status of This Memo). + +## Relation to the argument + +**Ruling D8 and premise P12.** The RFC is a standards body's own admission that end users are mostly absent from its process (§4.1), which supports the draft's claim that a standards body is an unreliable holder of a user-protecting convention. It also cuts slightly the other way: the IAB commits the IETF to favour end users in conflicts and names the user agent as a structural safeguard (§4.2). A user agent that represents the person against a service is close to what the draft wants a convention-holder to do. The draft could use §4.2 as a positive example rather than only as a gap. + +**What it does not say.** Nothing in RFC 8890 says that adoption of standards is voluntary or that standards do not bind platforms. That claim belongs to RFC 2026 (`rfc2026`), whose §1.1 describes the Internet as supporting communication "through voluntary adherence to open protocols and procedures defined by Internet Standards" (checked in S2 against the RFC Editor text). + +**Use in DRAFT_v2:** +- Line 428 ("Technical standards bodies cannot fill the gap: adoption of their standards is voluntary, and end users take little part in writing them (Nottingham, 2020)"): half supported. "End users take little part" matches §4.1 ("their participation is minimal"). "Adoption of their standards is voluntary" is not in RFC 8890; cite RFC 2026 (Bradner, 1996) for it. "Cannot fill the gap" is the draft's inference; the RFC itself argues the IETF should and can do more for end users (§4). +- Line 641 (reference entry): correct in substance; it lacks the month (August) and the DOI 10.17487/RFC8890. +- Line 750 (notes table): the earlier draft's "explicitly declare[s]" claim is unsupported by the text, as the note implies. + +## In conversation with + +- `rfc2026` — the Internet Standards Process; the source for voluntariness. Not cited by RFC 8890, but the claim at line 428 needs it. +- `rfc7282` — Resnick on rough consensus; companion IETF process document, not cited here. +- Cited in the RFC and not in the bib, as leads: RFC 3935 (IETF mission statement); Clark, Sollins, Wroclawski & Braden (2002), "Tussle in Cyberspace"; RFC 8752 (ESCAPE workshop on publishers and advertisers); RFC 7258 (pervasive monitoring); RFC 6973 (privacy considerations). + +## Caution + +An informational IAB statement of intent with no evidence base and no enforcement; it binds nobody, including the IETF's working groups. It speaks for the IETF of 2020 and says nothing about W3C or other bodies. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** confirmed + +**What I checked:** I re-read the RFC Editor plain-text edition the writer saved (dxd_foundation_local/pdfs/R4/rfc8890.txt) in full and checked every quotation word for word against its section (Abstract; Status of This Memo; §2 on indirect end users; §3 "lever to assert power"; §4.1 "unique insight", "participation is minimal", "a primary channel"; §4.2 user agent and "opaque, vendor-specific"; §4.3 "concerning phenomenon"), the header (IAB, Informational, August 2020, ISSN 2070-1721), the DOI against Crossref (10.17487/RFC8890, M. Nottingham, 2020-08, RFC Editor), and the references the card names as leads (RFC 3935, 8752, 7258, 6973, Tussle). I opened RFC 2026 to test the card's claim that voluntariness lives there. I re-read DRAFT_v2.md lines 428, 641 and 750. + +**Findings:** +1. No errors found. Quotations, locators, the full-text claim and the record are correct. +2. Added, not a correction: the voluntariness claim the card routes to RFC 2026 is in RFC 2026 §1.1 ("voluntary adherence to open protocols and procedures defined by Internet Standards"); the card now gives the locator instead of relying on facet O. +3. Use verdict for line 428 confirmed as half supported: "end users take little part" matches §4.1; "adoption ... is voluntary" is not in RFC 8890; "cannot fill the gap" is the draft's inference against the RFC's own call for the IETF to do more (§4). + diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rieder2020towards.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rieder2020towards.md new file mode 100644 index 00000000..6d463929 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rieder2020towards.md @@ -0,0 +1,43 @@ +# Rieder, B., & Hofmann, J. (2020). Towards platform observability. *Internet Policy Review, 9*(4). https://doi.org/10.14763/2020.4.1535 + +**Identifier:** doi:10.14763/2020.4.1535 · **Read depth:** full_text (open-access journal PDF, 28 pages; the PDF numbers are the printed numbers I cite; published 18 Dec 2020. I read the whole article, including the conclusion; I skimmed the reference list and footnotes) · **Source-tier:** peer-reviewed journal article (conceptual and policy argument) · **Evidence basis:** direct_read · **Design and sample:** none. A conceptual argument built on the transparency critique, with examples of existing data-access projects (TOSBack, Social Science One, Facebook Ad Library, Datenspende with 4,500 volunteers, scraping studies) · **Parties modeled:** platform–observer. Observers are researchers, journalists, regulators and civil society; the platform owner is the party observed, with users and business users as the people whose treatment is at stake · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +The authors start from a regulatory habit, "opening the black box" through disclosure, explanation and audit, and say it fails for two reasons. First, transparency as information disclosure carries untested steps: from disclosure to visibility, to insight, to "effective regulatory judgement" (p. 5). It also does not simply reveal: in Flyverbom's words, quoted by the authors, "transparency reconfigures - rather than reproduces its objects and subjects" (p. 5). Second, platforms resist the image. Machine-learning decisions cannot "be easily traced back to singular factors" (p. 7), the algorithm is spread across many components and keeps changing, so audits face a "snapshot logic" (p. 7), and "platform companies are therefore neither fully in control of actual outcomes, nor fully aware what is happening within their systems" (p. 8). + +They propose observability: "Unlike transparency, which nominally describes a state that may exist or not, observability emphasises the conditions for the practice of observing in a given domain" (p. 3). They set out three principles: widen the normative and analytical horizon, with "public interest" as the reference norm (pp. 10-13); observe platform behaviour continuously, through data-access agreements, accountability interfaces, developer APIs and scraping or volunteer plug-ins (pp. 13-18); and build the capacity for collaborative knowledge creation, with qualitative methods, journalists and NGOs, public funding, "centres of expertise" and a "European Platform Observatory" (section 3.3 and the Conclusion, pp. 18-24; the "centres of expertise" and Observatory passages are on p. 23). The article is explicit that observability is "not meant to be a radical departure from the call for transparency" (p. 21). + +One point the authors press repeatedly concerns who controls the access. Of the data-sharing initiatives they discuss (Facebook, Social Science One), the authors say they "have been plagued by delays and restrictions," and report that scholars "have rightfully criticised that the scope and modalities for access remain in the hands of platforms themselves" (p. 15; the second clause is attributed to Hegelich 2020 and Suzor et al. 2019). Without a legal framework, "companies can not only start and stop such initiatives at will but are also able to control parameters" such as scope and granularity (p. 15). In the conclusion they say of Twitter's new academic API that "these efforts are still voluntary and can be revoked at any time" (p. 22). + +## Relation to the argument + +The article gives the academic meaning of "observability", the word Nadella's post uses for its seven principles. It bears on the argument at two points. + +First, premise P10 (a convention the owner can withdraw alone is no convention). Rieder and Hofmann give a documented case in a neighbouring domain: access that the platform grants, sets the terms of and can end at will is not accountability. They treat the fix as law, not goodwill. This supports the step from conflict of interest to "cannot withdraw alone" that the argument map marks `[inf]`, though their example is data access for researchers, not a reading convention for users. + +Second, the seat of the observer. In their usage the observers are outsiders to the platform and the platform holds the power. In the post the deployer observes a model that serves it. The word is the same and the seat is opposite, which bears on ruling D4 (the author's contrast with the post) and D5. + +Facet check: facet G says Rieder and Hofmann "agree with Nadella that the black box will not be opened; both choose observation of behaviour." Half right. They agree the box will not simply open, and they favour observing behaviour over time (pp. 13-14). But they do not choose observation in place of transparency; they say it is not "a radical departure" from it (p. 21) and that traditional disclosure, reports and whistleblowing can remain "part of an observability framework" (p. 13). The facet read the article in part; the full text adds the access-control argument above, which the facet did not record. + +## In conversation with + +- `ananny2018seeing`: the article's main source for the transparency critique, whose chain from disclosure to visibility to insight it takes over (pp. 3, 5, 6). +- `gillespie2019custodians`: cited as Gillespie (2018, p. 18) for the definition of platforms as services that "host, organize, and circulate" shared content (p. 2); the bib key carries a 2019 date for the same book. +- `stohl2016managing`: not cited by Rieder and Hofmann. +- Seaver (2017) is cited for the point that even small companies struggle to say where "the algorithm" is (p. 8). The bib holds `seaver2019knowing` and `seaver2018captivating`; I did not match the 2017 reference to either. +- Leads, not in the bib: Burrell 2016 on opacity (p. 6); Dourish 2016 on what algorithms "know" (p. 7); Helberger et al. 2018 on "cooperative responsibility" (p. 21); Powles and Nissenbaum 2018 (pp. 10-11); Bovens 2007 on accountability as a social relation (p. 18). + +## Caution + +A policy-oriented conceptual article from 2020. It concerns outside scrutiny of platforms (EU Digital Services Act debate), not people using a platform to coordinate with each other, and it predates language-model agents. Its many examples are named projects described in a sentence or two, not evaluated. It makes no empirical claim that observability works. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The open-access journal PDF (cards/rieder2020towards.pdf, 29 pages; page numbers equal the printed numbers), the title, authors and date (18 Dec 2020) matching. Fifteen quotations (pp. 3, 5, 7, 8, 13, 15, 21, 22, 23), the Datenspende figure of 4,500 volunteers (p. 17), section headings and page ranges (3.1 from p. 10, 3.2 from p. 13, 3.3 from p. 18, Conclusion from p. 21), and the cited names Ananny, Gillespie (2018), Seaver (2017), Burrell, Dourish, Helberger, Powles and Bovens with their pages. + +**Findings:** All quotations are exact. Three attribution fixes: (1) the "scope and modalities for access remain in the hands of platforms themselves" clause is the authors reporting what scholars (Hegelich; Suzor et al.) have criticised, and "plagued by delays and restrictions" refers to the initiatives listed, not Facebook alone; (2) "voluntary and can be revoked at any time" is said of Twitter's new academic API in the Conclusion; (3) the "centres of expertise" and European Platform Observatory passages sit in the Conclusion (p. 23), outside section 3.3. Not checked: the claim about how the Nadella post uses "observability", which concerns another document (not in my batch). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rose2020zero.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rose2020zero.md new file mode 100644 index 00000000..94fab808 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/rose2020zero.md @@ -0,0 +1,39 @@ +# Rose, S., Borchert, O., Mitchell, S., & Connelly, S. (2020). *Zero trust architecture* (NIST Special Publication 800-207). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.800-207 + +**Identifier:** doi:10.6028/NIST.SP.800-207 · **Read depth:** partial (final publication, August 2020, 59 pages; I cite the printed page numbers, which are the PDF page minus 9, since the front matter runs in roman numerals. I read the abstract, sections 1 to 3, section 4 (4.1-4.5), section 5 in full, 7.1 and 7.3.4-7.3.7, and Appendix B.4; I skimmed sections 6, 7.3.1-7.3.3, the rest of Appendix B and the reference list, so I cannot vouch for those) · **Source-tier:** government technical report (federal guidance, no peer review) · **Evidence basis:** direct_read · **Design and sample:** none. An architecture definition with deployment models and use cases; no experiment or measurement · **Parties modeled:** other. An enterprise (the party that defines policy), "subjects" that request resources (users, applications and other non-human entities, p. 1), and administrators who configure the policy components. Customers and the public sit outside the model (p. 7) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +NIST defines zero trust as "a collection of concepts and ideas designed to minimize uncertainty in enforcing accurate, least privilege per-request access decisions in information systems and services in the face of a network viewed as compromised" (p. 4). Zero-trust models "assume that an attacker is present in the environment" and treat the enterprise's own environment as "no different—or no more trustworthy—than any nonenterprise-owned environment" (p. 1). Every request passes a policy decision point and a policy enforcement point, split in the logical architecture into a policy engine, a policy administrator and a policy enforcement point (Figs. 1-2, pp. 5, 9). A "trust algorithm" weighs the request, subject and asset records and threat intelligence; it can be criteria- or score-based, singular or contextual (pp. 17-20). + +The seven tenets (pp. 6-7) include per-session access with "the least privileges needed" (tenet 3); dynamic policy that may use behavioural attributes (4); "dynamic and strictly enforced" authentication and authorization before access (6); and (7) that the enterprise "collects as much information as possible about the current state of assets, network infrastructure and communications and uses it to improve its security posture." The tenets "apply to work done within an organization or in collaboration with one or more partner organizations and not to anonymous public or consumer-facing business processes" (p. 7), and for a public-facing service "the enterprise is constrained as to what internal cybersecurity polices can be enforced" (p. 27). + +Section 5 lists threats to ZTA itself. The policy engine and administrator are the key components, so a rogue administrator or a compromised policy administrator subverts everything; configuration changes "must be logged and subject to audit" (p. 28). A malicious insider or a thief holding valid credentials "may still be able to access resources for which the account has been granted access," and a contextual trust algorithm is expected to catch out-of-pattern use (p. 29; example on p. 20). Software agents that administer the system raise an "open issue" of how they authenticate; "the biggest risk" is false positives and false negatives, and an attacker may "trick the agent" (pp. 30-31). Migration is "a journey" (p. 1), and a first deployment may run in reporting-only mode (p. 40). Appendix B.4.6 admits that "there has not been a rigorous examination of how end users act" in a ZTA, and points to security fatigue as the nearest evidence (p. 49). + +## Relation to the argument + +NIST is the security canon behind the post's controls-outside-the-model move (the argument map's P11). The post's "assume the model is compromised" posture (its P19) matches the document's starting premise, and "observe everything" (its P15) matches tenet 7. The match is partial, as facet G records: in NIST, the information collected feeds policy decisions made by software, and a person appears only as the administrator who configures and audits the policy components (pp. 28, 30). The "human-readable evidence" the post asks for has no counterpart. NIST's enforcement is the "moves work from person to artefact" mechanism the argument map warns will not by itself separate the paper from the post. + +For ruling D5 (the "enterprise, not user" reading), the scoping on pp. 7 and 27 is the clearest textual support in this batch: the tenets stop at the enterprise boundary, and people outside it are users of a public service, not parties whose protection the architecture addresses. Section 5.7 also bears on the post: NIST already treats AI software agents as non-person entities that can be induced to act for an attacker (p. 31). + +Facet check: facet G's quotations of tenets 3, 6 and 7 and its description of the decision and enforcement points hold against the full text. Its note that tenet 7 feeds "policy, not a human reader" also holds, with two additions from sections it did not read: administrator audit of the policy components (p. 28), and §7.3.7's line that "the subjects and stakeholders of the resources and processes involved should also provide feedback" (p. 41), which places enterprise users in a feedback role. + +## In conversation with + +- `anderson1972computer` and `saltzer1975protection`: the reference-monitor and least-privilege ideas that facet G ties to the post. NIST's text and reference list name neither; it traces the term "zero trust" instead to DISA's "black core," the Jericho Forum's deperimeterization and Kindervag (p. 2). +- `parasuraman2010complacency`, `tariq2025alert`: unrelated by citation, but the fatigue gap NIST names in B.4.6 (p. 49) is the human side these two study. +- Leads, not in the bib: Kindervag's coining of "zero trust" (p. 2); NIST SP 800-63 on digital identity (cited for authentication assurance, pp. 18-20); the NIST Risk Management Framework SP 800-37 (§6.1, p. 32). + +## Caution + +A federal architecture guide for enterprise networks and resource access, not a study and not about AI. Its word "insider" means an authorised account misused, and it never mentions language models. Sections 6, 7.3.1-7.3.3 and most of Appendix B I only skimmed. I did not check for later NIST revisions. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the NIST PDF (60 PDF pages; printed page = PDF page minus 9, confirmed). Checked 20 quotations for exact wording and printed-page locator, the seven-tenet citations, the scope sentence on p. 7, the p. 27 and p. 31 claims, and the p. 2 provenance of the term "zero trust". Did not check the reference list or sections the card says it skimmed. + +**Findings:** One fix: Read depth said "full_text" while the card itself says sections 6, 7.3.1-7.3.3 and most of Appendix B were only skimmed; changed to "partial" and kept the list of unread parts. All quotations, page locators and attributed claims match the text. No other issues found. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ryan2014historical.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ryan2014historical.md new file mode 100644 index 00000000..74918421 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ryan2014historical.md @@ -0,0 +1,40 @@ +# Ryan, J. (2014). Historical note: Did double-entry bookkeeping contribute to economic development, specifically the introduction of capitalism? *Australasian Accounting, Business and Finance Journal, 8*(3), 85–97. + +**Identifier:** doi:10.14453/aabfj.v8i3.8 · **Read depth:** full_text (publisher PDF from the University of Wollongong journal site; journal pagination; reference list skimmed) · **Source-tier:** peer-reviewed journal note (historical commentary) · **Evidence basis:** direct_read · **Design and sample:** none; a review of Yamey's papers of 1947, 1949, 1959 and 1964 beside quotations from Sombart, Spengler, Nussbaum and Mattessich, framed with new institutional economics · **Parties modeled:** other (a merchant or firm and its ledger; society at the macro level) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues + +Ryan gives "a broad historical setting" for Yamey's criticisms of Sombart and others, who held double-entry bookkeeping to be central to the rise of capitalism (p. 85). He quotes the claims first: Spengler's "decisive event," Nussbaum's remark that its importance "can hardly be exaggerated," Mattessich's ranking of Pacioli's chapter with "the achievements attained by Columbus, Copernicus, Galileo, Descartes and Newton" (pp. 85–86; Spengler's own comparison names Columbus and Copernicus only). A footnote concedes that Pacioli's 1494 *Summa* was the first printed account, and that "it is generally agreed that he described the system as it had been practiced for about 200 years" (p. 86, n. 2). + +He then summarises Yamey. In 1949 Yamey looked for regular balancing and found "little evidence of a careful calculation and analysis of profit" (p. 87, quoting Yamey 1949, p. 110). In 1959 he examined the double-entry accounts of six individuals and one partnership; Ryan says the diversity "virtually precludes generalising": four of seven had assets or liabilities missing, three balanced annually, three irregularly and one never in eleven years, and three were out of balance (pp. 90–91). By Yamey's own account the seven were not "a representative selection" (p. 90). Cost accounting joined double entry only in the last quarter of the nineteenth century (p. 95). Yamey "saw no reason to modify his initial view" that the link between double-entry output and economic decisions was not established (p. 91). + +Ryan's contribution is a reconciliation. Yamey worked at the micro level of individual merchants' decisions; Sombart and the others wrote about society. Using Williamson's four levels of social analysis, Ryan places the historians' claim at the first level, informal institutions, where change takes centuries (pp. 92–94). He judges Yamey's scepticism "soundly based" for the period Yamey examined (p. 95), but holds that the question "transcends microeconomic decision making" (p. 96). His verdict: "double-entry bookkeeping, as an essential aspect of capitalism, emerged with it at about the same time, and with it, contributed to economic development" (p. 96). + +## Relation to the argument + +Facet F uses the Pacioli and Sombart story as a cautionary case of credit given to one device (case 4). Ryan helps in part. From a source that favours Sombart, he confirms that Pacioli described a practice some two centuries old (p. 86, n. 2) and that evidence of use in decisions before about 1875 is thin (pp. 88, 93, 95). That supports ruling D6 and the retreat from "the" move in `ARGUMENT_MAP.md`. + +Facet check: facet F, working from the abstract, took this note as part of the evidence that the big causal claim "was contested for decades by Yamey, and recent economic historians restrict double entry to control of debt and agents." Ryan does neither. He defends the causal claim at the macro level and says nothing about control of debt or agents. The abstract hides this by saying only that the note will "review Yamey's criticisms." Cite Ryan for the summary of Yamey and the two-centuries point, not for the deflation of Sombart. Sangster (2025; `sangster2025reboot`) supplies the criticism of Yamey's own definitions. + +The macro claim is thin: it rests on quotations from Sombart and Nussbaum and on the image of a "genie" released "out of the bottle" (p. 95), with no new evidence. + +## In conversation with + +- `yamey1949scientific` — summarised in detail (pp. 86–88): balancing, profit, and the "methodising" of business. +- `yamey1964accounting` — "Further notes on a theme by Sombart," summarised on pp. 88–89. +- Yamey (1947) and (1959) are also summarised (pp. 89–91); neither is in the bib. Sombart, Spengler, Nussbaum and Mattessich are quoted, and none has a key in `references.bib`. Williamson (1988) and North (1991) supply the institutional framework (pp. 92–93). +- Ryan does not cite `carruthers1991accounting` or `sangster2025reboot`. + +## Caution + +A thirteen-page note with no new data or archive work. It reports Yamey second-hand, and Sangster (2025, pp. 372–374) argues that Yamey's conclusions mix two definitions, bookkeeping and financial accounting, so Ryan's "soundly based" stands on the same ground. His own thesis joins a macro claim that cannot be tested here to a metaphor. It gives no date for the emergence of double entry beyond Pacioli's "about 200 years" of earlier practice. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the publisher PDF (14 pages, journal pagination 85-97); title, author, year match. Checked 21 quotations for wording and page, the 1959 sample figures (six individuals and one partnership; four of seven missing items; three annual, three irregular, one never over 11 years; three out of balance), the cost-accounting date, the 1875 point, footnote 2, and that Carruthers and Sangster are not cited. + +**Findings:** One fix: the card gave Mattessich's ranking as 'beside Columbus and Copernicus'. Mattessich (1964: 101, quoted on p. 86) lists Columbus, Copernicus, Galileo, Descartes and Newton; the Columbus-and-Copernicus pairing is Spengler's. Card now quotes the five names (13 words). Everything else matches. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/saltzer1975protection.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/saltzer1975protection.md new file mode 100644 index 00000000..97bb9a4d --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/saltzer1975protection.md @@ -0,0 +1,34 @@ +# Saltzer, J. H., & Schroeder, M. D. (1975). The protection of information in computer systems. *Proceedings of the IEEE*, 63(9), 1278–1308. + +**Identifier:** doi:10.1109/PROC.1975.9939 · https://web.mit.edu/Saltzer/www/publications/protection/ · **Read depth:** partial; full text of the authors' MIT web edition (Sections I, II, III, acknowledgement, suggestions for further reading, reference list), assembled from separate pages. Journal pagination is not available, so locators are section numbers (I.A.3 and so on). The figures (Figs. 1–14) are absent from the extraction, though the prose describes each; Table I is present. The footnotes page (numbered notes 1–45 in the prose) was not in the extraction and not read. The reference list carries a note that its OCR is poor. Section II (descriptor architecture) read for argument, not for mechanism detail · **Source-tier:** peer-reviewed article (tutorial and survey) · **Evidence basis:** direct_read · **Design and sample:** none; design principles drawn from "experience", plus a single password survey of about 300 passwords reported without a citation (I.B.4) · **Parties modeled:** other: owners and users of shared information, the system's guard (supervisor), borrowed programs, and administrators in a hierarchy of authority · **Facets:** G, X · **Read:** 2026-10-13 + +## What it argues +The paper's "key concern ... is multiple use": systems where "all users should not have identical authority" need the machine to enforce an authority structure (I.A.1). It defines protection narrowly, as control of executing programs' access to stored information, and warns that "a narrow concentration on protection mechanisms ... may lead to false confidence in the system as a whole" (I.A.1). The "intruder" "may be an otherwise legitimate user" (I.A.1). + +It sorts systems into functional levels (unprotected, all-or-nothing, controlled sharing, user-programmed sharing, "putting strings on information"; I.A.2), then gives eight design principles because "no one knows how to build a system without flaws" (I.A.3–4): economy of mechanism, fail-safe defaults, complete mediation ("Every access to every object must be checked for authority"), open design, separation of privilege, least privilege ("limits the damage that can result from an accident or error"), least common mechanism, and psychological acceptability ("the human interface be designed for ease of use, so that users routinely and automatically apply the protection mechanisms correctly"). Two more, work factor and compromise recording, "apply only imperfectly to computer systems"; of the second: "it is difficult to guarantee discovery once security is broken" because internal records "can be undone by a clever attacker" (I.A.3). The principles "serve best as warnings", not "absolute rules" (I.A.3). + +Section I.B builds the guard-and-wall model, with a supervisor protecting its own implementation. Section II develops capabilities (tickets) and access control lists, and finds that revocation, propagation and review of access are hard in capability systems (II.B.3). Section II.C.4 treats who may change the lists. Hierarchical control "provides for exercise of authority but ... has no provision for preventing abuse of that authority", so the paper proposes a "prescript" that triggers logging, delay or a second principal's assent, to "ensure that some independent judgment moderates otherwise unfettered use of authority" (II.C.4). Section II.C.5 separates discretionary from nondiscretionary controls and introduces confinement of borrowed programs. Section II.E introduces protected subsystems. Section III.C (Concluding Remarks) says users of the first protection mechanisms "have found them relatively clumsy to program and not especially well matched to the users' image of the problem"; Section III.B.c says "User interfaces that more closely match the mental models people have of information protection are needed." + +## Relation to the argument +It supports P11's lineage claim, as facet G argues: a small enforcing component outside the thing it governs, mediating every access, with privilege cut to the task. It also bears on P10 and the author's Part-1 reply (facet X, objection 6). Separation of privilege names a structure in which "no single accident, deception, or breach of trust is sufficient" because different organizations hold the keys; least common mechanism lets dissatisfied users "provide a substitute or not use it at all" (I.A.3). Both are designs where one party cannot withdraw a protection alone, though they are built inside one system, not among parties whose interests diverge. The prescript passage is the closest this source comes to P10's worry about unchecked authority at the top of a hierarchy. + +It cuts against the author in two places. The reference-monitor idea moves work to an artefact, so it cannot mark the paper off from Nadella (ARGUMENT_MAP, "What the argument no longer claims"). And I.A.3(h) puts the human interface inside security engineering, which strengthens the "category error" objection less than facet X's reply (3) suggests: the interface in question is the one by which a person states protection policy over their own objects, not an interface to an interested third party's mediation. On ruling D3: open design lets a "skeptical user" check the mechanism, but its safety rests on keys, so disclosure there does not substitute for control. + +Facet check: every S&S quotation in facets G and X matches the text (least privilege, complete mediation, separation of privilege, psychological acceptability, open design, compromise recording, "warnings"). Facet G's status line "read in part (section I.A)" understates what I now confirm: the facets did not use the hierarchical-control passage, I.B.4 on authentication, or III.B. + +## In conversation with +- `anderson1972computer` — listed under "Computer security technology planning study" in Suggestions for Further Reading ([95]), not cited in the prose. The three-category taxonomy of violations is credited to a different Anderson work ([6], *Advances in Computers*, 1973), which is not in the bib. +- Leads, not in the bib: Lampson (1973), the confinement problem [61]; Bell and LaPadula (1973) [62]; Weissman (1969), the ADEPT "high water mark" [14]; Schroeder (1972), mutually suspicious subsystems [70]; Redell (1974), indirect objects for revocation [54]. + +## Caution +A 1975 tutorial on single-machine time-sharing systems. Its "users" are people sharing files; none is a customer of an adaptive platform. It supplies principles, not evidence on human performance, and its one empirical figure (more than 50 percent of about 300 self-chosen passwords guessable, I.B.4) is uncited there. Cite section numbers, not pages, and say the edition is the authors' web edition (a note on the references page, signed Norm Hardy, 1997, says the OCR of that material performs poorly); the journal text may differ in wording and layout. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the cached text of the authors' MIT web edition; title and authors match. Checked 30 quotations for wording and section locator, the eight-plus-two principle list, the 300-password figure (more than 50 percent), the Anderson [6] and [95] references, and the section structure (I.A.1-4, I.B.4, II.B.3, II.C.4-5, III.B.c, III.C). + +**Findings:** Two fixes. (1) The card attributed both the 'relatively clumsy to program' quotation and the 'User interfaces that more closely match the mental models' quotation to III.C with locator III.B.a-c. The first is in III.C (Concluding Remarks); the second is III.B.c only. Rewritten with the right locators. (2) Read depth said full_text while stating that Section II was read for argument only and the footnotes page was not read; changed to partial, list of unread parts kept. All other quotations match. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sangster2024emergence.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sangster2024emergence.md new file mode 100644 index 00000000..02865e7f --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sangster2024emergence.md @@ -0,0 +1,41 @@ +# Sangster, A. (2024). The emergence of double entry bookkeeping. *The Economic History Review*, 78(2), 499–528. + +**Identifier:** doi:10.1111/ehr.13358 · **Read depth:** full_text (open-access version of record hosted by the University of Aberdeen repository, 30 pages including notes and reference list; pagination is the journal's, "Econ. Hist. Rev. 2025;78:499–528"; online 2024, print 2025) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** close reading of one ledger fragment (four surviving pages, more than 160 entries in 44 accounts, Florentine bankers at the Bologna fair, May 1211), compared with two later ledgers (1296, 1299), plus a causal argument from legal and commercial context; the data statement reads "The article describes entirely theoretical research" (p. 525) · **Parties modeled:** two parties plus a third party (merchant debtors and creditors, with a banker-intermediary who keeps the books and a court that reads them) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues +Sangster asks why double entry bookkeeping emerged, a question the article opens by calling unanswered: "There is still not a satisfactory explanation for the emergence of double-entry bookkeeping in Italy in the thirteenth century" (p. 500, quoting Dobie and Oldroyd 2020). He blames twentieth-century historians who defined double entry by its outputs (profit-and-loss accounts, balance sheets) and so recognised almost no medieval example, the earliest being a Genoese government ledger of 1340 (pp. 500–501). Yamey's repeated conclusion, which they accepted, was that double entry became useful only with nineteenth-century joint-stock companies (p. 500). + +He instead adopts Lane's input definition: "any accounts with duality of entry" count as "an elementary form of double entry" (p. 501). Every entry then needs an amount, a debit or credit sign, and the identity of the contra account (p. 501). Sangster reads the 1211 entries against that test (pp. 504–507, Figures 1–3). He finds it met, and he withdraws his 2016 label "dual entry" for one pair of entries after correcting a mistranslated semicolon ("posted to debit above", p. 506). The same entries resemble the ledgers of 1296 and 1299, which differ mainly by page numbers and a separate money column (p. 507). + +His causal account runs from coins to credit to evidence. Coinage was scarce, of variable quality and costly to assay (an £8 payment meant more than 1,920 silver denari checked one by one, p. 508). Moneychangers took deposits and paid by "book transfer" between customers' accounts (p. 509). Fairs grew and trade moved among strangers, so default and dispute rose (p. 511). Roman and Lombard law fitted commerce poorly, and courts of customary law presided over by merchants took over (pp. 511–514). A banker who made a book transfer needed the contra entry to prove it, so duality followed from the evidential demand (p. 518). The ledger also served as "an aide-mémoire as well as a document of record" (p. 513). Custom came first and law followed: "commercial law was derived from banking and mercantile customary practice, not the other way round" (p. 517). A Bologna statute of 1245 compelled book transfers at the fair, and Sangster reads it as a response to clearing the bankers had already invented (p. 517). Genoa's 1327 legislation prescribed accounts kept by "the method of the bankers" (p. 519). Sangster dates the method to the twelfth century, with bankers evolving it "for at least 47 years" by 1211, an inference from Pisa's court of customary law in 1164 (p. 520). Figure 4 (p. 523) draws the causal chain. + +## Relation to the argument +Facet F (parallel moves) uses it as the Pacioli-and-double-entry case: a convention with a long, collective, practice-first origin, not one inventor. The text supports that, though not in the form the facet states (see Facet check). + +For the word-separation precedent, the useful finding is the order of events: practice, then recognition by a body that adjudicates, then codification. That order bears on ruling D8 (the holder of the convention is unknown; compare courts and forex markets; look for where the convention bends) and on the open question in P10. The convention here was held by the interested party, since the bankers' method protected creditors and the banker's own risk (pp. 516, 522). It still became binding because courts of customary law, staffed by merchants who "did understand the needs of business" (p. 513), accepted it as evidence. That cuts against the premise as the map words it (a convention the owner can withdraw alone is no convention): the bankers could not withdraw it, because the courts and the guild-licensed books (p. 517) held it. That is my reading; Sangster does not state it. It also fits P7 poorly, since the bankers' ledger records what already happened between two parties and gives each of them no help inferring what the intermediary will do. + +One passage touches reading directly. Sangster reports Santini's view that entries in the spoken language, not Latin, meant "all merchants could read, understand, and thus genuinely approve the entries made in their names" (p. 519), and he files this under "pragmatic literacy" (n. 134). Sangster offers it as a hypothesis, not a result. It is the only line here that resembles the spacing story of a convention lowering a reading cost for the parties, and the paper does not test it. Ruling D10 (history as a guide, not mechanism) and ruling D1 (review the causal analysis, do not assert it) both apply: this is a causal argument by context, with no counterfactual. + +**Facet check:** The facet had only the abstract. Reading the whole text, I find the abstract-level claims intact, with five qualifications. (1) The facet says the 1211 ledger "has" double entries and so double entries are "nearly 280 years older" than Pacioli (1494 minus 1211 is 283). That holds only under Sangster's input definition; under the output definition earlier historians used, the earliest case is 1340 (p. 501). Contra entries almost never survive, and double entry is confirmed mainly from cross-references and offset (p. 506). (2) The facet says Sangster takes apart the "Pacioli" one-agent story. The text never discusses Pacioli's *Summa*; the name appears only in the title of a cited article (n. 48, Ciambotti 2016). Sangster targets the Yamey–Besta–Sombart definitional tradition, not Pacioli. The Pacioli comparison is the facet's, so the table row and the "Pacioli introduced double entry" item (F, lines 291 and 324) should not attribute it to Sangster. (3) He dates the origin to the twelfth century, earlier than 1211; the ledger is evidence of a method already running. (4) On "double entry produced capitalism," Sangster does not rebut it. He reports Yamey's view and cites Yamey 1949, p. 103, for the claim that credit dealings led to systematic accounting (n. 154, p. 523). (5) The facet gives the sentence "the reason for such a detailed bookkeeping method is unknown" in quotation marks; the abstract reads "The reason for the emergence of such a detailed bookkeeping method is unknown." No contradiction of substance. + +## In conversation with +- `yamey1949scientific`: Sangster cites it among Yamey's works (n. 11) as the source of the view that double entry became useful with joint-stock companies, and quotes p. 103 for credit as the spur (n. 154). He opposes the first and cites the second with approval. +- `yamey1964accounting`: matches n. 11's "Accounting and the rise of capitalism: further notes on a theme by Sombart", which Sangster dates 1962 (Studi in onore di Amintore Fanfani, vol. VI). The bib year of 1964 differs; I did not resolve it. +- `sangster2016genesis`: his earlier paper, cited for the "dual entry" label (pp. 502–503) and corrected on one entry (p. 506), and for the term "entity-wide" bookkeeping (n. 17). +- `sangster2025reboot`: not cited here, and the text does not cite it. The facet pairs them; the pairing is the project's. +- `yamey2005historical`: not cited. `carruthers1991accounting`: not cited. Sangster does not cite Carruthers and Espeland. + +Leads not in the bib, each from the reference list: Lane 1977, "Double entry bookkeeping and resident merchants" (source of the definition); Lee 1972, "The oldest European ledger" (translation of the 1211 ledger); Goldthwaite 2009, *The Economy of Renaissance Florence*; Tucci 1989, "Il documento del mercante"; Padgett 2012, "The emergence of corporate merchant-banks in dugento Tuscany" (organisational-emergence parallel); Greif 2002 and 2006 (community responsibility, which Sangster treats as contested, pp. 520–521). + +## Caution +A single-case historical argument built on four surviving pages, about 160 entries, and an inferred chain of causes. Sangster states that the contra account "if it existed, has not survived" in virtually all cases (p. 506), and the dating of the twelfth-century origin rests on inference from Pisan custom, not on a surviving twelfth-century ledger. The finding that bankers' ledgers were accepted as evidence in court rests on secondary sources (Lattes, Santini, Chiaudano) and on a negative: no notarised account book is known among more than 3,000 Florentine books, based on the author's decade in the archives and the recall of colleagues (n. 118, p. 517). The Innes credit theory of money (pp. 521–522) and the trust argument are interpretive. Citable as a case of practice preceding codification, and as the claim that the method's detail answered evidential demands. Not citable for the origin of double entry as a settled fact, for Pacioli, or for the spacing history. No difference was found between the open-access version and the version of record, since this is the version of record. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the Wiley/Aberdeen open-access PDF (journal pagination 499-528 confirmed by page headers); title, author, year match. Checked 29 quotations and their pages, the numerical claims (more than 160 entries, 44 accounts, 1,920 denari for GBP 8, more than 3,000 Florentine books, 1245, 1327, 1340, 1164, 47 years), footnotes 11, 17, 48, 118, 134, 154, and the reference-list leads. + +**Findings:** One quotation corrected: the card gave 'understood the needs of business' (p. 513); the text reads 'merchants who did understand the needs of business'. Fixed to the exact words. All other quotations, locators and figures match. The facet-check claim that Pacioli appears only in a cited title holds (n. 48 and the reference list). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sangster2025reboot.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sangster2025reboot.md new file mode 100644 index 00000000..aa753960 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sangster2025reboot.md @@ -0,0 +1,43 @@ +# Sangster, A. (2025). Time to reboot accounting history: Evidence from scholarship on double entry. *Accounting History, 30*(3), 368–391. + +**Identifier:** doi:10.1177/10323732251357606 · **Read depth:** full_text (Aberdeen repository file named "VOR," in the journal's layout with its pagination; reference list skimmed) · **Source-tier:** peer-reviewed article (historiographical critique) · **Evidence basis:** direct_read · **Design and sample:** a critical reading of the English-language literature on pre-1800 double entry, by what Sangster calls a "hermeneutic approach" of reading "critically and extensively, and repeatedly" (p. 375); no corpus, count or coding frame, apart from two recent articles examined as examples · **Parties modeled:** other (merchants, banks and governments and their ledgers, as read by historians) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues + +Sangster charges that the accounting-history literature on double entry before 1800 is "antiquarian": it asks "what" and "how" and never "why," reads the past through present practice, and trusts its predecessors instead of revisiting sources. The abstract says the result is that "unsupportable assertions are accepted as axiomatic fact" (p. 368). + +He traces the line of descent from Besta (pp. 370–372). Yamey, who dominates the English literature, uses two definitions without saying so: De Roover's bookkeeping, and a "Sombart-inspired" one that needs modern accounting features such as closed ledgers and profit-and-loss accounts. In Sangster's words, Yamey "was not assessing the bookkeeping. He was assessing the accounting" (p. 373). Yet "Neither Yamey nor De Roover ever wrote that double entry bookkeeping was rarely used before 1800" (p. 374). Hoskin and Macve's 1986 statement that it was rare is an "unfounded and misleading assertion" that discouraged further work (p. 375). Yamey generalised from a few ledgers and "less than 4 per cent" of the bookkeeping manuals published before 1800 (p. 374); a footnote counts 882 editions and says he would have needed six times more manuals for 95 per cent confidence about Great Britain (n. 8, p. 387). + +Two recent articles serve as examples of a literature "increasingly corrupted by misunderstanding" (abstract, p. 368; pp. 378–380). One places features of the 1211 Florentine bank ledger centuries too late in a table; the other redefines single-entry bookkeeping so that it contains double entries. Sangster's counter-evidence is that surviving Tuscan books show "no uniformity in practice beyond a superficial level, even in individual firms; no Darwinian progression" (p. 378). He recommends a "critical traditional" history that returns to the account books and draws on other disciplines, whose historians "dismiss the medieval accounting history literature with contempt" (p. 376). + +## Relation to the argument + +Facet F uses this article as a case where a one-agent origin story (Pacioli, 1494) and a one-cause story (double entry made capitalism) fall apart (case 4). The body supports part of that. + +It supports this much. Pacioli's treatise describes the Venetian paired journal-and-ledger method, which was not how double entry was done everywhere (p. 375 and n. 11), and he describes an "elementary" system (p. 384). Double entry is attested well before him: the first confirmed instances Sangster names are 1296 and 1299 (p. 379, citing De Roover 1963 and Lee 1977), about 195 to 198 years before 1494 by this card's subtraction. Economic historians "might recognise" it in the 1211 ledger (p. 378). That fits the lesson that credit for a convention goes to a later, visible description (D6). + +It does not support a settled account of why double entry was used. The abstract says economic historians found "two primary purposes before 1800: control over debt and distant agents, factors, and partners," and that "few used it to identify overall profit or wealth." The body does not set that claim out; it argues that nobody has asked "why." Cite the abstract's sentence as Sangster's summary of economic historians, not as a finding he demonstrates. + +Facet check: (1) Facet F attaches "a ledger of 1211 … and entity-wide double entry … by the end of the 13th century" to "Sangster 2025 [online 2024], abstract." Those details belong to `sangster2024emergence` (*Economic History Review*). Here 1211 is the earliest paragraph-format ledger (p. 378), and the confirmed double-entry dates are 1296 and 1299. (2) The facet lists Yamey's contest against Sombart as support for deflating the causal claim. Sangster does not overturn Yamey's finding that few identified profit before 1800; the abstract concurs. He argues that Yamey's definitions and synonyms invite readers to take it as a finding about bookkeeping (pp. 373–374). So Sangster qualifies how Yamey can be cited, and does not reject the Sombart criticism. + +## In conversation with + +- `yamey1949scientific` — the main Sombart-inspired paper, read as assessing accounting, not bookkeeping (pp. 372–373). +- `yamey1964accounting` — the bib's entry for the Yamey paper Sangster lists as a 1962 chapter, "republished in 1964" in the *Journal of Accounting Research* (n. 3, p. 386). +- `yamey2005historical` — cited for Yamey's claim that double entry was limited to wholesale merchants before 1800 (p. 374). +- Sangster does not cite `ryan2014historical` or `carruthers1991accounting`; Ryan's note leans on the Yamey whom Sangster criticises. +- Leads not in the bib: De Roover (1956, 1963); Lee (1973, 1977); Hoskin & Macve (1986); Goldthwaite (2015, 2018); Besta (1916); Matringe (2022). + +## Caution + +A polemic by a specialist, not a neutral survey; it faults named scholars and reports no systematic review of the literature it condemns. Its positive claims about use before 1800 appear in the abstract and in scattered evidence in the body, so take them as Sangster's view; the 2024 *Economic History Review* paper (`sangster2024emergence`, not read here) is the likelier source for the origin story. It concerns business and bank ledgers, not reading or notation. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the journal-layout PDF (pp. 368-391; page headers confirm pagination); title, author, year match. Checked 24 quotations for exact wording and page, the 4 per cent, 882-editions and 95-per-cent footnote (n. 8, p. 387), 1296 and 1299 dates and the 195-198 subtraction, the Venetian/elementary Pacioli passages, and the claim that Ryan and Carruthers are not cited. + +**Findings:** No issues found. The one quotation broken across a hyphenated line break ('bookkeep-ing') matched on p. 373 once the hyphen was joined. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/santonidesio2018meaningful.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/santonidesio2018meaningful.md new file mode 100644 index 00000000..ec9ce12f --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/santonidesio2018meaningful.md @@ -0,0 +1,40 @@ +# Santoni de Sio, F., & van den Hoven, J. (2018). Meaningful human control over autonomous systems: A philosophical account. *Frontiers in Robotics and AI, 5*, Article 15. https://doi.org/10.3389/frobt.2018.00015 + +**Identifier:** doi:10.3389/frobt.2018.00015 · **Read depth:** full_text (open-access publisher PDF, 14 pages; locators are the article's own page numbers, 1 to 13, which match the PDF pages; published 28 Feb 2018. I read the whole body; I skimmed the reference list. The bib note says only the abstract was read; the body was not) · **Source-tier:** peer-reviewed journal article, philosophical analysis (Original Research section) · **Evidence basis:** direct_read · **Design and sample:** none. Conceptual argument from Fischer and Ravizza's "guidance control" plus Nozick's tracking account of knowledge, applied to autonomous weapons, then to driving · **Parties modeled:** person–system, extended along a chain of people: designers, programmers, legislators, commanders, operators. The people the system acts on appear as victims and as road users, not as parties to the control relation · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +"Meaningful human control" entered the debate on lethal autonomous weapons without a theory behind it: "policy-makers and technical designers lack a detailed theory of what 'meaningful human control' exactly means" (p. 3). The authors supply one, cast as design requirements. They hold that "simple human presence or 'being in the loop' is not a sufficient condition for being in control" (p. 4), since a person can be present and still lack the information, options or capacity to matter ("merely pushing a button in a reflex when a light goes on", p. 4). + +They borrow Fischer and Ravizza's two conditions for responsibility and turn them into two necessary conditions for control. Tracking: the decision system "should demonstrably and verifiably be responsive to the human moral reasons relevant in the circumstances—no matter how many system levels, models, software, or devices of whatever nature separate a human being from the ultimate effects in the world" (p. 7). The test is counterfactual (Nozick): had the relevant facts or reasons differed, the system would have responded accordingly. Their examples include a ship-defence system that classed an incoming missile as friendly and an image classifier that tracked snow, not the animal (pp. 7-8). Tracing: the system's actions must be "traceable to a proper moral understanding on the part of one or more relevant human persons", meaning at least one human in the design or use chain who understands the system's capabilities and effects and "understands or is in the position to understand that others may have legitimate moral reactions toward them" (p. 9). The authors add that users have a "meta-task responsibility" to check that the system lets them do what they ought, and designers must not make that check "impossible or unduly difficult" (p. 9). They argue tracing is not only for assigning blame afterwards: it requires people "capable and motivated to take active steps to prevent unwanted outcomes" (p. 10). + +Two limits matter. The tracking condition does not fix whose reasons: any human agents along the chain will do, to be determined on a "case-by-case basis" (fn. 9, p. 7). And control "is a necessary but not sufficient condition" for a morally good system; a system may be under meaningful control and pursue wrong goals, since it "doesn't offer in itself any safeguard against evil human controllers" (fn. 10, p. 7; p. 8). The closing section proposes design at three levels, the system, its environment (separate lanes for autonomous cars) and "social and institutional practices" (pp. 11-12). + +## Relation to the argument + +Supports the part of the post's P19 (an authorised person with a halt right) that needs a person who can be traced to, and speaks to the argument map's P12 by saying that presence of a person is not control. It sits with Sterz et al. as the philosophical account of what the person at the gate needs. + +It also bears on ruling D5. The tracking and tracing conditions are defined over the people who design, deploy and operate; the affected enter only indirectly. The authors admit that control can coexist with wrong goals (p. 8), which is the deployer-serving case the post's seat produces. Nothing here requires that the people a system acts on be among those whose reasons it tracks. + +Facet check: facet G says the abstract names "the humans designing and deploying" and that it did not read whether the body adds the affected. The body adds them only partly. The tracked agents are left open ("the relevant human agents", p. 11), with the only worked example the traffic case, where the system should follow rules set "by the society via the public authority" and unwritten conventions that "reflect some relevant interest of the road users" (pp. 11-12). So the road users' interest is in view, through public rules, not as a direct party to control. Tracing's condition (b) brings others in as people who may react morally to the responsible human (p. 9). + +## In conversation with + +- `mitchell2026agents`: cites this article for the autonomous-weapons lineage of the human in the loop and for meaningful control as a governance requirement (Mitchell, pp. 1-2). +- `sterz2024quest`: does not cite it; both draw on Fischer and Ravizza (Sterz fn. 4). +- `elish2019moral`, `bainbridge1983ironies`, `parasuraman2010complacency`: not cited here. Nearest in kind is the authors' use of Cummings (2006) for the claim that dependence on decision systems "can lead to a reduced sense of responsibility" (p. 10). +- Leads, not in the bib: Fischer and Ravizza 1998, *Responsibility and Control*; Nozick 1981, *Philosophical Explanations*; Matthias 2004 and Sparrow 2007 on the "responsibility gap"; Horowitz and Scharre 2015 on meaningful human control; Roorda 2016 on NATO targeting procedures (pp. 3, 10). + +## Caution + +Written about lethal autonomous weapons, with civil driving as a first extension; the authors say the broader theory is future work (p. 11). It is a normative account with illustrations and no evidence about how people behave. It does not discuss platforms, recommendation, or coordination between parties. The mapping from "tracking" to a deployer's "controls outside the model" is the project's own step, not the authors'. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the Frontiers open-access PDF (14 pages; article page numbers match PDF pages); title, authors, year match. Checked 22 quotations for wording (re-extracted without layout mode because of two-column wrapping) and page, the tracking and tracing passages (pp. 7-10), footnotes 9 and 10 locators, the ship-defence and husky-snow examples, the traffic quotations (pp. 11-12), and the Cummings and Roorda mentions. + +**Findings:** No issues found. Not checked: the card's statement about what mitchell2026agents and sterz2024quest cite (those belong to other sources). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/schlimm2008modeling.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/schlimm2008modeling.md new file mode 100644 index 00000000..155af2c4 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/schlimm2008modeling.md @@ -0,0 +1,51 @@ +# Schlimm, D., & Neth, H. (2008). Modeling ancient and modern arithmetic practices: Addition and multiplication with Arabic and Roman numerals. In V. Sloutsky, B. Love, & K. McRae (Eds.), *Proceedings of the 30th Annual Meeting of the Cognitive Science Society* (pp. 2097–2102). Cognitive Science Society. + +**Identifier:** none found (not in Crossref; author-hosted copy at the University of Göttingen psychology site, per `../references.bib`) · **Read depth:** full_text (six-page proceedings paper, author copy with the proceedings header; the printed pagination 2097–2102 is the proceedings'; reference list read) · **Source-tier:** proceedings (conference paper) · **Evidence basis:** direct_read · **Design and sample:** computational simulation, no human participants: four rule-based agents (Arabic and Roman, each for addition and multiplication), run on 1,000,000 three-number additions (x + y + z, each 1 to 100) and 10,000 two-factor multiplications (x × y, each 1 to 100) (p. 2101) · **Parties modeled:** person–artefact (one agent working on an external written display) · **Facets:** F · **Read:** 2026-10-13 + +## What it argues + +The paper asks how much of the felt difficulty of Roman arithmetic "is based on inherent limitations of Roman numerals as opposed to our lack of familiarity" with them (p. 2097). The authors name historians (Cajori, Menninger) and cognitive scientists (Marr, Dehaene, Butterworth) who hold that Roman arithmetic, and multiplication above all, is infeasible, and they set out "to dispel this myth" (p. 2097). They cannot recruit participants equally familiar with both systems, and most details of how Romans used their numerals are lost, so they build agents (p. 2098). The agents are written in Lisp, "inspired by" ACT-R but with no memory dynamics or learning implemented (p. 2098). Each agent works on a simulated external display and is charged for elementary information processes (reading, attention shifts, memory retrieval and storage, writing, deleting), after Payne, Bettman and Johnson (p. 2098). + +The Arabic addition agent needs the problem laid out in right-justified columns, one addend per row, and the authors say this "caters towards certain properties of place-value notations" (p. 2099). The Roman agent copies symbols into a "working table" with one row per symbol type, then applies simplification rules such as IIIII → V (p. 2099). The Roman agent needs no addition facts (it holds seven simplification rules where the Arabic agent holds 100 single-digit sums) and its Roman multiplication table has 49 entries against the Arabic 100 (pp. 2100–2101). + +Table 2 (p. 2101; one unit is 1,000 operations) puts the cost on the external side. For additions, Arabic against Roman: reads 8,116 against 244,166 (ratio 30.08), attention shifts 16,625 against 336,413 (20.24), writes 3,637 against 18,700 (5.14); the Roman agent also makes 27,710 deletes where the Arabic agent makes none. For multiplications the ratios are smaller: reads 4.10, shifts 4.81, writes 3.52, deletes 27.70. The Roman input is about 2.6 times longer (15,030 against 5,760 symbols for additions), so single-symbol multiplications grow by "2.6² = 6.8" (p. 2101). + +The authors conclude that "the Achilles heel of Roman numerals is not that they are (qualitatively) unsuitable for algorithmic usage, but their greater cost in terms of the number of necessary elementary information processes, especially those using external resources" (p. 2102). Two conditions limit the result. The comparison "depends crucially on the precise format of external inputs and the availability of interactive resources" (p. 2101); a linear input would force the Arabic agent to rewrite into columns or hold more in working memory, while the Roman agent "could remain virtually unchanged" (p. 2101). An abacus in place of the working table would cut the Roman agent's writes and deletes sharply (pp. 2101–2102). The authors list time estimates, an ACT-R re-implementation and other numeral systems as future work (p. 2102). + +## Relation to the argument + +**Premise P5 and the facet-F cost hypothesis (H-cost).** The paper is the one source in this batch that counts, step by step, what a notation costs a worker, and it counts on the premise the argument borrows from Larkin and Simon (1987) and Zhang and Norman (1995, p. 2097 cites both; the project has not read the latter): the same information in another form changes the work. For the paper's main claim, that a layout move can shift work between memory and the page, the result is a clean instance. The Arabic agent's advantage depends on the column layout of the page, which is a design choice about the artefact and not a property of the digits. + +**Complication to "cheaper".** The source cuts against a single-axis reading. The Arabic form is cheaper in perceptual, attentional and motor steps and costlier in memory (100 addition facts, carries, intermediate sums); Roman is the reverse. "Lowering a cost" (P6) has to name which resource, and for whom. The authors state that the verdict flips with the tools at hand (abacus, fingers, linear input). + +**P1 and P4 by analogy, not mechanism.** The paper overturns a received view that an old notation is "inherently" hard, by showing it is workable and expensive. That mirrors the disputed point at P1 (whether unspaced script burdened trained readers at all). It does not test any claim about reading, and the author's rulings D1 and D10 (precedent and guide, not mechanism) apply: cite it as a design precedent for the cost-accounting frame, not as evidence about word separation. + +**D6, a complex of conventions.** Roman and Arabic arithmetic differ in numeral shapes, symbol count, tables, algorithms and page layout at once. The paper separates some of these (layout, table size) and shows each moves cost somewhere, which supports the author's wish to find "the critical move" among several (D6) without showing that any one is it. + +**Facet check:** I found no contradiction with the facet entry in `F_parallel_moves.md`. Its summary (four agents, 1,000,000 and 10,000 problems, no human data, "precise format of external inputs", the abacus caveat) matches pp. 2098–2102. The entry gives the p. 2097 quotation and the p. 2102 quotation correctly. + +## In conversation with + +From the paper's own text and reference list, matched against `../references.bib`: + +- `larkin1987why` (Larkin & Simon 1987): informational against computational equivalence is taken from it (p. 2098); the Roman/Arabic comparison is a worked case. +- `zhang1995representational` (Zhang & Norman 1995): cited for numeral systems as representational artifacts (pp. 2097, 2102). +- `norman1993things` (Norman 1993): cited for the claim that Roman addition is simpler than ours (p. 2097) and for tool-agent-task interaction. +- `hutchins1995cognition` (Hutchins 1995) and `kirsh1996adapting` (Kirsh 1996): cited for cognition recruiting external resources (p. 2098). +- `kirshmaglio1994distinguishing` (Kirsh & Maglio 1994): cited for people using hands and objects to arrange and count (p. 2098). + +Leads not in the bib: Simon (1978) on forms of mental representation; Payne, Bettman & Johnson (1993), the source of the elementary-information-process accounting; Anderson et al. (2004) on ACT-R; Nickerson (1988), "Counting, computing, and the representation of numbers", *Human Factors*; Neth & Payne (2001) on addition as interactive problem solving. + +## Caution + +This is a simulation of hand-built agents, not human data: no time estimates, no errors, no learning, and ACT-R only as inspiration (pp. 2098, 2102). The counts depend on the authors' choices of algorithm, input layout and problem range, which they say favour the Arabic agent in places (p. 2101). The authors cannot say how Romans actually computed (p. 2098). Cannot be cited for anything about reading, writing systems or learners. A six-page conference paper that the authors themselves treat as a first model. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the author copy of the proceedings paper (6 pages, 2097-2102); title, authors, year match. Checked 15 quotations (two-column layout, so re-extracted without layout mode) and their pages, every figure in Table 2 quoted by the card (reads, shifts, writes, deletes, ratios, symbol counts), the 1,000,000 and 10,000 problem counts, the 49-entry and seven-rule/100-fact comparisons, and the citations to Norman 1993, Zhang and Norman 1995, Hutchins, Kirsh, Kirsh and Maglio, Larkin and Simon. + +**Findings:** No issues found. The card's '2.6² = 6.8' is the PDF's superscript flattened to '2.62 = 6.8'; the meaning is the same. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/scribner1981psychology.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/scribner1981psychology.md new file mode 100644 index 00000000..a050ddef --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/scribner1981psychology.md @@ -0,0 +1,50 @@ +# Scribner, S., & Cole, M. (1981). *The psychology of literacy*. Harvard University Press. + +⚠ Not read in full. + +**Identifier:** ISBN 0-674-72115-2 (cloth), 0-674-72114-4; archive.org item `psychologyoflite00sylv` (lending-restricted; its search-inside endpoint returned HTTP 403 on 2026-10-14) · **Read depth:** partial: no page of the book opened. Seen: the archive.org metadata record (Harvard University Press, 1981; 366 scan images); the snippet quotations recorded by facet C in round 2 from archive.org search-inside (scan indices only, no printed pages); and, as a proxy, the authors' own summary of the same Vai project, Scribner & Cole (1978), "Literacy without schooling: Testing for intellectual effects", *Harvard Educational Review, 48*(4), 448–461, read in full from the copy on Michael Cole's lab site (lchc.ucsd.edu/People/MCole/LiteracyColeScribner.pdf, 16 pages with the journal's pagination). Google Books returned no record for the ISBN or title · **Source-tier:** monograph · **Evidence basis:** proxy (the authors' 1978 article) plus snippet record · **Design and sample:** field research among the Vai of Liberia, 1973–1978 as reported in the 1978 article: a survey and test battery given to "more than 700 Vai adults" (1978, p. 453), then ethnographically derived experiments · **Parties modeled:** other (individual readers and writers in a society with three literacies) · **Facets:** C, X · **Read:** 2026-10-14 + +## What it argues + +What can be stated from what I saw. The book reports the Vai project. The Vai use an indigenous syllabary learned outside school, alongside Arabic literacy learned in Qur'anic study and English learned in school, which let the authors separate the effects of literacy from those of schooling. In the 1978 article they report that "between 20 and 25 percent of Vai men could read and write using their own script", which is "transmitted from one generation to another without schooling or professional teachers" (1978, p. 453). + +Their central result, as the 1978 article gives it: on sorting and verbal-reasoning tasks, "improved performance was associated with years of formal schooling, but literacy in the Vai script did not substitute for schooling. Vai literates were not significantly different from nonliterates on any of these cognitive measures" (1978, p. 453). Vai literacy did produce specific skills tied to its uses, such as clearer game instructions and better integration of syllable-by-syllable speech, but "These consequences, however, are all highly specific to activities with the Vai script" (1978, p. 457). And: "We did not find that literacy in the Vai script was associated in any way with generalized competencies such as abstraction, verbal reasoning, or metalinguistic skills" (1978, p. 457). Their framework treats literacy as "a practice" with particular technologies and purposes (1978, p. 457). + +On the script itself, the 1978 article says it "is written without division into words or other language units; a string of syllabic characters runs across the page without spacing or segmentation", and that readers resolve the ambiguity by "experimentation in pronunciation", saying syllables aloud "until they click into meaningful units" (1978, p. 456). Facet C recorded a parallel sentence from the 1981 book by search-inside: "A string of syllabic characters runs across the page without division into words or phrases" (scan index only; printed page not seen). + +## Relation to the argument + +**Line 497 and ruling D1/D10.** The draft cites the book to say that word separation is a design guide, "not ... a monocausal explanation of historical literacy or intellectual development". As far as the proxy shows, Scribner and Cole support the second half directly: literacy as such did not produce general cognitive change; the effects were practice-specific and general gains tracked schooling (1978, pp. 453, 457). They do not address word separation or the history of literacy, so they support the caution by analogy, not on its subject. + +**Bears on P5 and P6, both ways.** The Vai case is a living unspaced script read by adults who learned it informally, which shows that an unspaced script can be learned without schooling. It also shows the reading cost the draft attributes to *scriptio continua*: readers find decoding "extraordinarily difficult" and resolve word boundaries by trial pronunciation (1978, p. 456). That fits P5's claim that separation turns an inferential task into a perceptual one. The Vai script also lacks tone and standardised vowel length, so its difficulty is not due to the missing spaces alone (1978, p. 456). + +**Use in DRAFT_v2:** +- **Line 497** ("Word separation is used strictly as an ergonomic guide ... not as a monocausal explanation of historical literacy or intellectual development (Street, 2003; Scribner & Cole, 1981)"). Supported in principle by the authors' finding, reported in their 1978 article, that literacy's effects are practice-specific and that general cognitive gains follow schooling (pp. 453, 457). Could not check against the book: no page of the 1981 book was seen, and no claim about the book's own wording or pagination can be verified. The citation is to the general argument, not to a page, so the risk is low. + +## In conversation with + +- `scribner1978literacy`: the authors' 1978 *HER* article, already in the bib with the note "Not obtained (no open copy per Unpaywall)"; an author-hosted copy is open at lchc.ucsd.edu (see Identifier) and is the source of every quotation on this card. +- `street1984literacy`, `street2003whats`: Street builds his "ideological model" partly on the Vai work; the draft cites Street 2003 beside this book at line 497. +- `goody1986logic`: the 1978 article tests Goody's metalinguistic hypothesis (citing Goody 1977, p. 453) and finds "no across-the-board evidence" for it (p. 454). +- `ong1982orality`: cites the 1978 article for the Vai script's difficulty (Ong pp. 83–87, per the algorithmacy arm's card). +- Leads not in the bib: Scribner & Cole 1973, "Cognitive consequences of formal and informal education", *Science* (cited 1978, p. 452); Olson 1977, "From utterance to text" (the 1978 article's foil); Greenfield 1972 on oral and written language; Goody, Cole & Scribner 1977 on Vai writing. + +## Caution + +The book was not opened. Every quotation above is from the 1978 article, a separate and shorter publication from the middle of the project; the 1981 book adds later studies, and its conclusions and wording may differ. The draft's reference entry is correct as far as the catalogue record shows. Do not cite page numbers in the book from this card. The claim that Vai script runs without word division is checked only against the 1978 article and a search-inside snippet recorded by facet C, not against a printed page of the book. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** confirmed + +**What I checked:** Confirmed the book itself cannot be opened openly: the archive.org item `psychologyoflite00sylv` is a lending item (metadata: Harvard University Press, 1981, ISBNs 0674721152 and 0674721144, imagecount 366, creators Scribner and Cole); I did not borrow it. Open Library lists the same ISBNs. Re-opened the proxy, Scribner & Cole (1978), *Harvard Educational Review* 48(4), from the writer's saved copy of the lchc.ucsd.edu PDF, and checked every quotation word for word against the line-joined text, assigning pages from the journal's page footers (448–461). Re-read DRAFT_v2.md line 497 and the reference at line 679, and the bib entries `scribner1981psychology` and `scribner1978literacy`. + +**Findings:** +1. All 1978 quotations verified verbatim, with correct pages: "more than 700 Vai adults" and "between 20 and 25 percent" / "without schooling or professional teachers" (p. 453); "did not substitute for schooling ... on any of these cognitive measures" (p. 453); "no across-the-board evidence" (p. 454); "without division into words ... without spacing or segmentation", "experimentation in pronunciation", "until they click into meaningful units", "extraordinarily difficult" (p. 456); "highly specific to activities with the Vai script", "generalized competencies such as abstraction, verbal reasoning, or metalinguistic skills", "a practice" (p. 457). The Goody (1977) citation is on p. 453. +2. Read depth (partial, proxy, ⚠ Not read in full) is stated honestly; no claim is made about the book's own pages. +3. Use in DRAFT_v2 (line 497): verdict stands. The general finding supports the caution against reading literacy as a monocausal cause of intellectual development; the book's own wording was not checked, as the card says. +4. Not re-checked: the Ong pp. 83–87 lead (taken from another arm's card) and facet C's search-inside snippet from the 1981 book. + +## Also used by facet W (2026-10-15) + +Facet W uses the Vai case, through the 1978 proxy article, as the one field description of the mechanism the abstract asserts: readers of an unspaced syllabary resolve ambiguity "by experimentation in pronunciation", saying syllables aloud "until they click into meaningful units" (1978, p. 456). That is voiced boundary detection by lexical matching, in a script where (1978, p. 456) tone and vowel length are also unmarked, so the voicing does more than find word ends. It supports a narrow form of "scaffolding literacy with oracy" for an unspaced script learned informally, and it comes with the same authors' finding that this literacy produced no general cognitive change (1978, pp. 453, 457), which bears on "freed literacy". diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/seattle2024deactivation.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/seattle2024deactivation.md new file mode 100644 index 00000000..80f5d678 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/seattle2024deactivation.md @@ -0,0 +1,41 @@ +# Seattle Office of Labor Standards. (2024, November). *Model notice of deactivation* (based on requirements of SMC 8.40.070) [Form]. + +**Identifier:** https://www.seattle.gov/documents/Departments/LaborStandards/Notice_of_Deactivation_Regular_form.pdf · **Read depth:** full_text (two-page agency form, revised 11/2024, whole document read; not the ordinance) · **Source-tier:** government model form (regulator guidance) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** two parties + third party (app-based worker and network company, with the city's labor-standards office as regulator) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +The form is a template that a company fills in to notify an app-based worker that their account will be deactivated or limited. It restates, in blanks and bracketed instructions, what the Seattle App-Based Worker Deactivation Rights Ordinance (SMC 8.40) requires. + +Timing: "Network companies are required to provide app-based workers with notice of deactivation 14 days before the date of deactivation and again on the date of deactivation" (p. 1). Language: notices must be issued "in any language that the company knows or has reason to know is the worker's primary language" (p. 1). + +Contents the company must supply: +- the reasons, which "must include portion of Company's policy that was violated and specific incident(s) that violated the policy"; +- the records considered, which "must include date, time, and location of all incidents supporting the deactivation decision"; +- the length of the deactivation, and how the worker can remedy it (p. 1); +- the right to challenge under SMC 8.40, with "at least until ____ (90 days from date of notice)" to do so; the company "has up to 14 days to respond with a decision on the challenge," and "All records of your challenge and any responses will be provided to you within 14 days" (p. 2); +- "all available methods of contact for a worker to initiate a challenge" and a link or instructions to the company's deactivation policy (p. 2). + +The form ends with a certification: "I certify under penalty of perjury under the law of Washington State that, to the best of my knowledge, true and accurate copies of the records ___ (Company) relied upon to deactivate the worker are [attached to this notice]/[can be downloaded here]," signed by a named officer with title, date and place (p. 2). The footer gives the office's contact details. + +## Relation to the argument +A concrete, recent case of a government fixing procedural conventions on a platform: advance notice, reasons tied to a named policy clause and dated incidents, access to the evidence, a stated challenge route and time limits, and a signature under penalty of perjury. In the paper's terms, these are the court-like conventions that Suzor (`suzor2018digital`, pp. 7–8) finds missing from platform terms (reasons, stable rules, appeal), now required by a city for one category of platform. It supports ruling D8 in two directions: the convention does bend toward courts when a public body imposes it, and the platform did not volunteer it. It leaves the decider where it was: the form says the company "has up to 14 days to respond with a decision on the challenge" (p. 2), so the first-instance reviewer of a deactivation is still the company that made it. Nothing on the form requires an independent reviewer; that is the contrast with `tumey1927ohio` and `us455usc`, where the interested party may not judge. + +Facet check: matches the facet's quotations. The facet notes the ordinance text was not found; that is still true here, and this form states only the minimum of what the company must put in the notice. Several of the form's duties (the 90-day window, 14-day response) are given as the company's duties under the ordinance; the ordinance itself is the authority for them. + +## In conversation with +- `suzor2018digital`, `suzor2016responsibilities` — the criteria (reasons, appeal) this form puts into force for one setting. +- `eu2019p2b`, `eu2022dsa` — the EU instruments in facet J with comparable notice-and-reasons duties; the form does not cite them. +- `rosenblat2016algorithmic` — facet J's source on driver deactivation as the harm the ordinance addresses; the form does not mention it. +- The form cites only SMC 8.40 and SMC 8.40.070 (not in the bib; leads). + +## Caution +An agency's sample form, not the law: the legal duty and its remedies are in the ordinance, which was not read. Revised 11/2024, so it reflects the rules at that date. No data on use or effect. The company fills the blanks, so the form shows what must appear, not what companies write. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the cached two-page PDF (Seattle Office of Labor Standards, revised 11/2024); title and issuer match; the whole document was read. Checked 10 quotations and the page 1 / page 2 split, plus the statement that the form cites only SMC 8.40 and 8.40.070. + +**Findings:** No issues found. Two quotations with blanks (the 90-day line and the certification) differ only in the underscores the card abbreviates. The footer calls it a 'Sample Notice' while the heading reads 'Model Notice'; the card follows the heading. Not checked: the card's claims about the ordinance, Suzor, the EU instruments and Rosenblat, which are outside this source. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/seymour2003foundation.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/seymour2003foundation.md new file mode 100644 index 00000000..ffb26d6b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/seymour2003foundation.md @@ -0,0 +1,37 @@ +# Seymour, P. H. K., Aro, M., & Erskine, J. M. (2003). Foundation literacy acquisition in European orthographies. *British Journal of Psychology, 94*(2), 143–174. + +**Identifier:** doi:10.1348/000712603321661859 · **Read depth:** full_text (publisher-formatted PDF from a University of Athens course site, journal pagination 143–174 intact; I read the argument, all three studies, the discussion and the reference list in full and skimmed the error-pattern and individual-variation tables, pp. 162–164, and the Appendix word lists, pp. 171–174) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** cross-sectional comparison of Grade 1 children (tested late in the first school year) in 13 orthographies, with Grade 2 groups in France, Denmark and Scotland; roughly 25–70 children per group (Table 2: Finnish 70, Danish P1 65, Austrian 57, French P1 45, Scottish P1 and P2 30 each); three short tests (letter sounds, very familiar words, simple nonwords) given individually · **Parties modeled:** other (child–script) · **Facets:** C · **Read:** 2026-10-13 + +## What it argues + +The COST A8 network tests whether orthographic complexity slows the first phase of learning to read, or only shapes later phases (pp. 144–145). It classifies the languages on two dimensions drawn from the members' own estimates: syllabic complexity and orthographic depth (Table 1, p. 146). French, Portuguese, Danish and English count as deep. Children across the countries read lists of nine letters, familiar words and simple nonwords. + +Letter knowledge does not separate the languages: every group reached about 90% or better, and the Scottish Primary 1 group (age 5) scored 94% (Study 1, p. 151). Word reading does. By accuracy on familiar words: Finnish 98.3%, German 97.7%, Spanish 94.7%, Portuguese 73.5%, French P1 79.1%, Danish P1 71.1%, Scottish P1 33.9% (Table 5, p. 153). The Scottish P2 group (about age 6½) reached 76.4%; French P2 reached 99.2% and Danish P2 92.6%. Nonword accuracy runs the same way: Scottish P1 29.3%, Danish P1 53.7%, Portuguese 76.9%, against 82–95% in the other Grade 1 groups (Table 6, p. 156). For English, the authors regress accuracy on British Ability Scales reading age: the Scottish group crosses 90% on words at a reading age of 7 years 4 months (p. 152), and passes the monosyllable nonword threshold above 7.5 years (p. 157). They conclude that English readers need "2½ or more years" to reach what most European children reach in the first year, "a ratio of about 2.5:1" (pp. 167–168). The abstract says "more than twice as slow." + +Syllabic complexity selectively slows nonword decoding and enlarges the word advantage (word minus nonword accuracy: 8.63 points in complex-syllable languages against 1.97 in simple ones; p. 160). Age of starting does not explain the gap: age correlated weakly with performance (Table 3, p. 149), and Danish children, who start at 7, still lag (pp. 165–166). The authors reject a teaching-method explanation, since Scottish schools use a mixed sight-word and phonics method fitted to a deep orthography (p. 166). Their mechanism: a deep orthography forces a "dual" logographic plus alphabetic foundation, and "there will always be a cost" that makes "irreducible differences in rates of progress" (p. 169). Variation between children is also wider in deep orthographies; normal scores in English, Danish, French and Portuguese fall in the disability range of the shallow ones (p. 165). + +## Relation to the argument + +Facet C reads this as support for premise P6 in the narrow sense: a script's design changes the rate at which children in school acquire it. The full text confirms that, and gives the size: roughly 2.5 to 1 for English against most European scripts. It also fits the ruling that the paper offers history as a guide, not a mechanism (D10): the cost sits in the script's design, and the script change that would remove it is not on offer. Seymour et al. do not test an intervention. They treat the cost as built into the orthography ("irreducible"), which cuts against any reading of the paper that expects design alone to remove a learning cost, and it concerns letter-sound mapping, not spaces between words, so the word-separation precedent (P2, P5) gets no direct support. It says nothing about who attends school (D1, D2). + +Facet check: the facet quotes the abstract's "more than twice as slow." The body gives a ratio of about 2.5:1 and also two further points the facet does not record: letter-sound knowledge does not differ between languages, and the gap sits in word and nonword reading. + +## In conversation with + +- `ziegler2010orthographic` — builds on this study and cites its 34% English figure; reports predictors of reading in Grade 2, not acquisition rates. +- `ziegler2005reading` — the same camp's grain-size theory; this paper does not cite it (published later) and works instead from Katz and Frost's (1992) orthographic depth hypothesis and Frost, Katz and Bentin (1987). +- Leads not in the bib (check): Seymour & Evans 1999 and Duncan & Seymour 2000 (the foundation-literacy method and the socioeconomic lag); Katz & Frost 1992; Frost, Katz & Bentin 1987; Wimmer & Goswami 1994; Frith, Wimmer & Landerl 1998. + +## Caution + +The English sample is two Dundee groups; the comparison uses the high-socioeconomic-status group only, tested at about 5½ and 6½ years, younger than every other group, so the authors admit that immaturity may add to the English gap (p. 167). Group sizes are small and the French data have missing cases (Table 2 note). The depth classification rests on committee estimates, not a computed measure (p. 146 note 1). The tests are short lists of very simple items, built as a "primitive" decoding model, not text reading. The 2.5:1 ratio is a regression extrapolation on reading age, not an observed time-to-mastery. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the publisher PDF (143-174; page numbers confirmed in headers); title, authors, year match. Checked all 9 quotations and their pages, Table 2 sample sizes and ages, Table 5 and Table 6 values for every group the card cites, the letter-knowledge range (89-99%; Scottish P1 94%), the 8.63 versus 1.97 syllable contrast (p. 160), the reading-age thresholds, and the teaching-method and disability-range passages (pp. 165-166). + +**Findings:** Two fixes. (1) Locators for the reading-age thresholds were 'pp. 154, 157'; the 7 years 4 months word threshold is on p. 152 and the 7.5-year nonword threshold on p. 157. (2) 'against 82-95% in the rest' for nonwords was loose: the rest of the Grade 1 groups run 82-95% but Scottish P2 (63.5%) and Danish P2 (81.3%) fall outside it, so the card now says 'other Grade 1 groups'. All table figures match. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sterz2024quest.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sterz2024quest.md new file mode 100644 index 00000000..7cb98d61 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/sterz2024quest.md @@ -0,0 +1,46 @@ +# Sterz, S., Baum, K., Biewer, S., Hermanns, H., Lauber-Rönsberg, A., Meinel, P., & Langer, M. (2024). On the quest for effectiveness in human oversight: Interdisciplinary perspectives. In *Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency* (FAccT '24). https://doi.org/10.1145/3630106.3659051 + +**Identifier:** doi:10.1145/3630106.3659051 · arXiv:2404.04059 · **Read depth:** full_text (arXiv v2, 7 May 2024, 13 pages in the ACM layout; locators are PDF pages, because the arXiv copy carries no proceedings page numbers. I read the whole text including the reference list's first pages; I skimmed the rest of the references. This is the arXiv version of the FAccT paper, not the proceedings PDF) · **Source-tier:** proceedings paper (peer-reviewed conference), conceptual · **Evidence basis:** direct_read · **Design and sample:** none. A conceptual synthesis by a psychologist, lawyers, philosophers and computer scientists, with a table of facilitators and inhibitors and a reading of Article 14 of the EU AI Act · **Parties modeled:** person–system, with the deployer and provider as parties who enable the "oversight person"; users and stakeholders appear only inside the definition of fitting intentions (p. 5) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +The authors ask when human oversight is effective, taking "the central objective of oversight is risk mitigation" as the standard (p. 2). Oversight is effective, they propose, if and only if the oversight person has four things: causal power over the system and its effects, epistemic access to the relevant situation, self-control, and fitting intentions for the role. Their short form is "Effectiveness = Moral Responsibility + Fitting Intentions" (p. 2), since the first three conditions together amount to moral responsibility. + +The definitions carry detail. Causal power fails if the person is "a mere observer" or has only "pseudo-options" such as a stop button that does not work (p. 4). Epistemic access needs enough knowledge of "what the system is doing, in which state it is," the risks, and what interventions would do; explainability measures can raise it (pp. 4-5). Self-control means being able to choose a path and follow through, which needs attention: if overseeing "a fleet of autonomous vehicles is so uneventful and dull that a normal person cannot help but let their mind wander off," oversight fails (p. 5). Fitting intentions means the person is moved by a role-appropriate aim: "they intend to mitigate any risks while taking other relevant factors, such as the interests of the system's users and other stakeholders, into consideration in a suitable way" (p. 5). + +Table 1 sorts facilitators and inhibitors into technical design (for example intervention options and preselection of outputs to review), individual factors (training, exhaustion, motivation, automation bias) and environment (job design, role conflicts, accountability, time pressure) (p. 7). On automation bias they cite Parasuraman and Manzey for "learned carelessness" (p. 7). One environmental factor stands out: if the overseer is also the decision-maker, "a conflict arises" when mitigating risk costs the decision-maker something (p. 8). + +Section 5 compares the definition to Article 14(4) of the EU AI Act. Requirements (d) and (e), override and a stop button, supply causal power; (a) and (c) supply epistemic access; "the criterion of fitting intentions ... has little direct equivalent in the legal text" (p. 9). They find the article's items of "strongly varying levels of abstraction" (p. 9) and point out it never mentions the system's inputs (p. 10). The authors name three limits: the definition is abstract, the mapping from factors to conditions needs empirical work, and it says whether oversight is effective, not whether something is oversight (p. 10). + +## Relation to the argument + +The four conditions give the project a checklist for the post's two human roles. The post's P19 holder of the brake has causal power by design; P15's reader of "human readable evidence" is meant to have epistemic access; the sources the project reads in facet G are about whether epistemic access and self-control hold. Sterz also says what is missing from the post: nothing in the post secures fitting intentions or deals with role conflict. + +Two points matter for the rulings. First, D5 (the "enterprise, not user" reading) and the argument map's P10. Sterz's role-conflict case (the overseer who is also the beneficiary of approvals, p. 8) and "other conflicting interests" (p. 5) are the person-level version of the owner-versus-user conflict P10 assumes. The paper does not extend that to a platform owner. Second, the affected enter only inside "fitting intentions": the overseer should weigh the interests of users and stakeholders, but they have no role as parties. Sterz does not argue that the people the system acts on should hold any control. + +Facet check: facet G says P19 supplies (a) and P15 is meant to supply (b), and that the literature is about whether (b) and (c) hold. That matches the paper. The facet does not mention (d), fitting intentions, or role conflict (p. 8), which bear most directly on D5. + +## In conversation with + +- `green2022flaws`: the paper's main source for criticism of oversight, from its survey of 41 policies and its call for institutional oversight (pp. 2-3). +- `bucinca2021trust`: cited for forcing people to think before seeing the system's output (p. 8; reference 11). +- `parasuraman2010complacency`: cited for automation bias and "learned carelessness" (p. 7). +- `eu2024regulation`: Article 14, the legal test case in Section 5 (pp. 8-10). +- `laux2024institutionalised`: cited (their ref. 49) for the idea that oversight should not be an end in itself and for conceptual work on what oversight is (pp. 2, 4). I did not check each use. +- `langer2024effective`: shares an author (Markus Langer) with the paper; I found no citation of it by title in the text or reference list. +- `santonidesio2018meaningful`: not cited; both lean on Fischer and Ravizza (Sterz fn. 4, p. 4). +- Leads, not in the bib: Fischer and Ravizza 1998 for guidance control; Enqvist 2023 on human oversight in the AI Act (their ref. 19); Mosier, Skitka, Burdick and Heers (1996) for the definition of automation bias, quoted on pp. 2 and 7 (their p. 205); Biewer et al. 2024 on "software doping analysis for human oversight." + +## Caution + +A conceptual proposal, in its authors' words "one among potentially numerous conceptualizations" (p. 4). It tests nothing on real overseers, and it predates the final adopted text of the AI Act (it uses the version approved by Parliament, with the Council still to adopt, fn. 9, p. 8). The arXiv v2 is not the proceedings page layout. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened arXiv v2 PDF (13 pages; locators are PDF pages as the card says); title, authors and venue match. Checked 14 quotations for exact wording and page (non-layout extraction), the Article 14(4) mapping, Table 1 factors, the 41-policy count for Green, Bucinca's forced-thinking suggestion (p. 8), fn. 9 (Council), and the reference-list leads (Enqvist, Biewer, Fischer and Ravizza, Mosier). + +**Findings:** Two small fixes. (1) The guidance-control footnote (Fischer and Ravizza, ref. 26) is fn. 4 on p. 4, not p. 5. (2) The card called the automation-bias source 'Mosier and Skitka (1996)'; the reference is Mosier, Skitka, Burdick and Heers (1996), quoted from p. 205. All quotations match. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/stohl2016managing.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/stohl2016managing.md new file mode 100644 index 00000000..fe6d6dfc --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/stohl2016managing.md @@ -0,0 +1,35 @@ +# Stohl, C., Stohl, M., & Leonardi, P. M. (2016). Managing opacity: Information visibility and the paradox of transparency in the digital age. *International Journal of Communication, 10*, 123–137. + +**Identifier:** https://ijoc.org/index.php/ijoc/article/view/4466 (open access, CC BY-NC-ND; no DOI) · **Read depth:** full_text (published journal PDF with the journal's pagination, pp. 123–137; the reference list was read) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** none; a conceptual argument with illustrative examples and one schematic figure (Figure 1), no data · **Parties modeled:** an organisation that controls information, and third-party observers who want to reconstruct its decisions · **Facets:** G · **Read:** 2026-10-13 + +## What it argues +The authors attack the assumption that "the more ICTs help to make organizational actions visible, the more organizations will become transparent" (p. 124). They treat visibility as an empirical construct with three attributes. *Availability* is whether actions are inscribed as data and stored. *Approval* is whether the owners allow outsiders to see the data, through legal obligation, norms or social consciousness. *Accessibility* is whether third parties can in fact find and use the data, which depends on directory knowledge, classification, skill and effort (pp. 125–31; Table 1, p. 126). "Organizational actors can manipulate each of these attributes" (p. 125). Transparency, in their definition, means that "third parties can clearly follow the chain of activity and decision making that led to a certain outcome" (p. 124). + +They then propose the transparency paradox: the relation between visibility and transparency may be curvilinear, so that "high levels of visibility decrease transparency and produce opacity" (abstract, p. 123; Figure 1, p. 132). Two paths lead there. In *inadvertent opacity*, volume swamps the receiver, so that "important pieces of information become inadvertently hidden in the detritus of information made visible—that is, the information hides in plain sight" (p. 133); they name the cause as information overload and "interpretive blinders." In *strategic opacity*, "actors who wish to keep certain information hidden from view but who are bound by transparency regulations or norms can produce opacity by strategically increasing the availability, approval, and accessibility of information" (p. 133). They add that digital media make this cheap through "macrotargeting" and "micro-targeting" (pp. 133–34), and, citing Eisenberg, that opacity is "not necessarily problematic" because it can foster cohesion (p. 134). The examples are small: open-meeting laws negated by closed "unofficial" meetings (p. 128), three non-overlapping catalogues at the LSE library (p. 130), and faculty salary tables that moved from a statehouse binder to a website (p. 131). Their closing questions ask for "information minimums and maximums" for effective decision making and "effective monitoring of organizational conduct" (p. 135). + +## Relation to the argument +**P9 (aids that hand the person something to read show weaker effects).** Stohl et al. give a mechanism for P9, not evidence. Their account says more material in view raises the recipient's cost of finding what matters. They offer no test, and the overload claim rests on Toffler (1970) and Edmunds and Morris (2000). + +**P11 and P12 (the post's "observe everything" and human-readable evidence; the person who must review).** The paradox predicts that a record complete in availability, approval and accessibility can still leave the reviewer unable to reconstruct a decision. The strategic path adds a point the facet omits: the party that controls the record can raise volume on purpose. In the post's setting that party is the deploying organisation, which fits D8's open question about who holds a convention. I did not re-read the post, so I cannot say the post does this. + +**P5/P6 (word separation re-represents the task).** This reading is Claude's, not Stohl's. The authors count "the skill of reading" and effort among the accessibility conditions (pp. 130–31) and hold that data can be available and approved yet "inaccessible to the general public because they required too much effort" (p. 131). Word spacing then reads as a move on accessibility alone: the text was already available and approved. The authors do not mention writing history. + +**D3 (disclosure does not change our ability to navigate).** The paper agrees in kind: exposing more does not give understanding. It concerns observers reconstructing an organisation's decisions, not users steering a mediator, so it supports the logic of D3 without removing the flag that no study shows disclosure leaves behaviour unchanged. + +**Facet check:** The facet's attributes, paradox definition and p. 133 quotation are correct. It names only the inadvertent path; the strategic path (pp. 133–34) is the one closer to Nadella's seat. It marks the card "read in part"; I read the whole article. + +## In conversation with +No work in `references.bib` appears in the source's reference list. The source cites Flyverbom (2015) and Christensen and Cheney (2015) on transparency, Leonardi (2014) on communication visibility, Latour and Woolgar (1986) on inscription, Bowker and Star (1999) on classification, and Eisenberg (1984) on strategic ambiguity; none is in the bib. The bib's transparency items (`ananny2018seeing`, `rieder2020towards`, `kroll2021outlining`) post-date it and are not cited by it, so the link is the project's own. Leads outside the bib: Flyverbom (2015), *Sunlight in cyberspace*; Christensen & Cheney (2015), *Peering into transparency*; Leonardi (2014), *Social media, knowledge sharing, and innovation*. + +## Caution +A theory piece with no data; "paradox" names a proposed shape (Figure 1 is schematic), not a measured one. It is about organisations and their observers, with no mention of algorithms, models or platforms, so applying it to AI agents is the project's step. It does not say when visibility helps, only that the relation can turn. The illustrations are anecdotes, several without a cited source (the LSE catalogues, the salary binders). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the IJoC journal PDF (pp. 123-137; page headers confirm); title, authors, year match. Checked 17 quotations and their pages, the three visibility attributes and Table 1 (pp. 125-126), Figure 1 (p. 132), the LSE catalogues (p. 130), the open-meeting example (p. 128), the salary example (p. 131), the Toffler and Edmunds and Morris citations (p. 133), and the cited authors (Flyverbom, Christensen and Cheney, Leonardi, Latour and Woolgar, Bowker and Star, Eisenberg). + +**Findings:** No issues found. The statement that the bib's transparency items are not in the source's reference list holds. The card's P5/P6 reading is flagged in the card as the project's own. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/strunk1998guido.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/strunk1998guido.md new file mode 100644 index 00000000..a8d3ac29 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/strunk1998guido.md @@ -0,0 +1,47 @@ +# Guido of Arezzo. Prologue to his Antiphoner; *Epistola de ignoto cantu* (excerpts). In O. Strunk & L. Treitler (Eds.), *Source Readings in Music History* (rev. ed.). W. W. Norton, 1998. As reproduced in K. M. Cook, *Workbook: The Medieval Period, Part 1: Monophonic Traditions* (A-R Online Music Anthology), Supplemental Reading 1. + +**Identifier:** https://www.armusicanthology.com/media/PDFs/Workbook-Medieval-Part1.pdf · **Read depth:** full_text of the excerpts as the workbook reproduces them (the 19-page PDF is the teaching workbook, not Strunk and Treitler; Supplemental Reading 1 runs on workbook printed pp. 2–5, the cover being unnumbered; the translation is Strunk and Treitler's by the workbook's attribution and I did not check it against that book or against Pesce 1999; I also read the rest of the workbook enough to confirm what it is) · **Source-tier:** book chapter (a translated primary text), here quoted in a teaching workbook the publisher marks "Do Not Copy" · **Evidence basis:** direct_read · **Design and sample:** none (a medieval author's own account of a method; no data) · **Parties modeled:** master and pupil with a written medium between them (other chanters, makers of antiphoners, as the target of reform) · **Facets:** F · **Read:** 2026-10-13 + +## What the PDF is + +The file is Karen M. Cook's workbook for the Medieval Period, Part 1 (monophonic traditions): assigned readings, three supplemental readings (Guido; Johannes de Grocheio, *De musica*; the *Roman de la rose*), further reading, a summary list, assigned scores and exercises with a key (pp. 1–18). Guido's excerpts are Supplemental Reading 1 (workbook pp. 2–5), headed "Prologue to His Antiphoner" and "Epistolo [sic] de ignoto cantu, excerpt." The workbook's key to exercise 4 (p. 17) adds a gloss of its own on the hymn *Ut queant laxis*, which is not Guido's text. The excerpt prints the hymn's score on p. 5, with a footnote saying the score comes from the A-R Music Anthology. + +## What it argues + +Guido opens the Prologue with a complaint: singers "will never sing by themselves without a master a single antiphon, not even a short one," though they sing every day "for a hundred years" (p. 2). He blames the lack of a common written form: antiphoners "are not one, nor yet a few, but rather as many as there are the masters in the various churches" (p. 2). He urges that no one provide an antiphoner with neumes who does not know "this art" (p. 2). + +His remedy is a book written "any intelligent and studious person may learn the chant by means of it; after he has thoroughly learned a part of it through a master, he will unhesitatingly understand the rest of it by himself without one" (pp. 2–3). He offers a demonstration: boys "who until now have been beaten for their gross ignorance of the Psalms" sing the antiphon correctly by themselves (p. 3). + +The excerpt then describes the layout (p. 3): each sound "is found always in its own row", lines are "drawn close together", some sounds on the lines and others in the spaces, "certain letters of the monochord are written at the beginning of the lines or spaces", and the lines are gone over in two colours, yellow for the letter C and red for F. The consequence he draws is that neume shapes alone are insufficient: "however perfect the formation of the neumes might be, it is altogether meaningless and worthless without the addition of letters or colors" (p. 3). The reader also has to "learn by heart a fair number of melodies" so as to recognise sounds, because recalling "with understanding" differs from singing "by rote" (p. 4). + +The *Epistola* excerpt contrasts two methods for finding an unknown melody. The older one, sounding each letter on the monochord, he calls "childish, good indeed for beginners, but very bad for pupils who have made some progress" (p. 4). His own method is to fix in memory the starting note of each phrase of a familiar melody (six phrases, six notes, p. 5). It works "provided there is someone present to teach the pupil, not merely from a written textbook, but rather by our practice of informal discussion" (p. 4). His one result is anecdotal: "some of them were able before the third day to sing an unknown melody with ease" (pp. 4–5). He ends: "All these matters, which we can hardly explain in writing, we can easily lay bare by a simple discussion" (p. 5). + +## Relation to the argument + +**P5, P6 and the author's ruling D6.** The excerpt is a medieval designer's own list of a bundle: ruled lines with one pitch per row, letters and colours as keys, modal formulas, memorised melodies, a monochord and a teacher. Guido claims no single element suffices; the neumes are "worthless" without letters or colours, and the written method works only with a person present. This supports the argument map's dropping of "the" move and D6's "complex, BUT we are looking for the most critical" move. His passage identifies the line-and-letter device as what makes the identical-pitch rule readable at a glance (p. 3), which may be the candidate for "most critical", but it is the author's reading and the excerpt says nothing of ranking. + +**P6, most for readers still learning.** The claim he makes is about learners: after learning a part "through a master" the pupil continues "without one." That is a reduced need for the teacher, not its removal, and the excerpt states the teacher's presence as a condition of his own method (p. 4). It supports a cost lowered for learners, not a widened population; the excerpt contains no counts. + +**The "ten years to one or two" figure.** The excerpts contain no such figure. The numbers Guido gives are rhetorical ("a hundred years," p. 2) and the "before the third day ... in many weeks" report (pp. 4–5), which concerns his memorising method and not the ruled staff. This agrees with the facet note that the claim is unverified. + +**P10, whom the convention serves.** The excerpt shows an author who writes the convention and urges others to adopt it, from a position as a reformer who says his way is the rule ("a common rule of art", p. 2). It gives no evidence on who could withdraw it. Specialist work (Haines 2008, `haines2008origins`, p. 331, per the facet entry) separates what Guido invented from what he popularised and notes his self-promotion; I read only the facet's quotation of that, not the article. + +**Facet check:** no contradiction with `F_parallel_moves.md`. The facet's quotations match the workbook (pp. 2–5). One addition: the Prologue excerpt does describe the ruled lines, letters and colours (p. 3), so the staff itself is in the excerpt, though the facet's "Epistola ... not about the ruled staff" is correct for the *Epistola*. The workbook titles the second text "Epistolo"; the bib and facet give *Epistola*. + +## In conversation with + +The excerpt names no scholarly works. Relations to `../references.bib`, drawn from the facet files and not from this text: `haines2008origins` (on what Guido invented against what he popularised), `pesce1999guido` (the translation source the facet names for the full *Epistola*), `mengozzi2010renaissance`, `mengozzi2007si` and `mews2024guidonian` (later scholarship on the reception of Guido's teaching). The excerpt does not cite any of them. + +## Caution + +Not Strunk and Treitler: a workbook's reproduction of excerpts, with the translation unchecked and the excerpts cut by the workbook's editor, so omissions are invisible. No dates or biography appear in the excerpt, so none are given here. Guido writes as an advocate with no data; the boys' results are his own report. The publisher marks the workbook as not to be copied, so this card quotes only short phrases; anyone quoting at length should go to Strunk and Treitler. The excerpt cannot support a claim that the staff alone shortened training. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** The workbook PDF (pdftotext of the file in pdfs/cards; Supplemental Reading 1, workbook pp. 2-5, and the key at p. 17). I checked all quotations from the source (about 25) and their page locators, plus the "six phrases", "yellow/red" and "hundred years" details. + +**Findings:** Every quotation matches the text and the page locators are right. The workbook's own header says "Reproduced from Oliver Strunk & Leo Treitler", as the card states. One fix: the card quoted about 280 words from the source, over the 250 limit; I trimmed the two longest quotations (the "intelligent and studious person" remedy and the "before the third day" result) without changing the claims. No other issues found. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/suzor2016responsibilities.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/suzor2016responsibilities.md new file mode 100644 index 00000000..9f63ecbb --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/suzor2016responsibilities.md @@ -0,0 +1,35 @@ +# Suzor, N. (2016, October 5–8). The responsibilities of platforms: A new constitutionalism to promote the legitimacy of decentralized governance. *Selected Papers of AoIR 2016: The 17th Annual Conference of the Association of Internet Researchers*, Berlin. + +**Identifier:** https://spir.aoir.org/ojs/index.php/spir/article/view/9055 · **Read depth:** full_text (four-page conference paper, published PDF from the AoIR Selected Papers site; no journal pagination, cited by paragraph order) · **Source-tier:** conference paper (short, abstract-length) · **Evidence basis:** direct_read · **Design and sample:** terms-of-service document analysis of 15 social media platforms, summarised in two sentences; no method section, no table · **Parties modeled:** two parties + third party (platform owner as governor of users; states and civil society pressing on the owner) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Suzor opens with the claim that "The ways in which platforms are governed matters" (para. 1) and that platforms are not neutral: their architecture and algorithms shape what people say and see, and "the powers they provide are carefully wielded and selectively enforced" (para. 1). States now enlist platforms in regulation (he cites "over 65 Million takedown notices" received by Google for its search results, as of 2016), and civil society presses the language of corporate "responsibility" (para. 2–3). + +The legal gap is the centre: terms of service are "formalised as legal contracts that most commonly allocate a great deal of power to the operators," so "In legal terms, the discretion of the platform owner is practically absolute," and constitutional language — "freedom of speech and association, requirements of due process and natural justice" — "has almost no application in the 'private' sphere" (para. 4). He proposes judging legitimacy against the rule of law: private governance should be "consensual, predictable, equal, and fair," with the "primary requirement" that "power is not exercised arbitrarily" (para. 5). + +The only evidence is one summary: he examined "the legal terms and conditions of fifteen of the largest English-language social media platforms," each "analyzed to identify the extent to which they provided protections for due process interests of users," and "In all cases examined, the terms of service provided broad, unfettered discretion to platform owners." The line the facet leans on follows: "Like constitutional documents, terms of service grant powers; but unlike constitutions, they rarely limit those powers or regulate the ways they are exercised" (para. 6). He ends by arguing that some decentralisation of governance is "both inevitable and desirable" but that enforcement must be "regular, transparent, equally and proportionately applied, and fair" (para. 7). + +## Relation to the argument +Supports the author's ruling D8 ("compare that platforms (which i think take a highly draconian position)") at the level of the contract text, and bears on premise P10: the owner holds the rules, and the law gives the user no handle on them. It supplies the source's own word for the condition, "practically absolute" discretion. It does not give counts, clause wording, dates or a list of platforms, so the strength of "draconian" rests on the 2018 article, not here. + +Facet check: the facet says this paper reports fifteen platforms and the 2018 article's abstract says 14, and leaves it unresolved. The full 2018 article (see `suzor2018digital`) names the 14 in its note 3 and says it excluded the Russian-language platform VK from the top 15 by traffic, so the two counts are explained by a change of sample, not an error. This paper does not say why it has 15. + +## In conversation with +- `suzor2018digital` — the full article that this paper previews; same author, same rule-of-law frame, sample of 14. +- `klonick2018new` — the facet pairs it as the study of how platforms actually run appeals; this paper does not cite it (it predates it). +- `gillespie2019custodians` — this paper cites Gillespie (2010, 2014) on platform neutrality and algorithms, not the 2018 book; the book is a later source in the same line. +- `tyler1988what`, `lind1988social` — not cited here. The procedural-justice work is the empirical counterpart of the rule-of-law values this paper names. +- Leads not in the bib (all from the reference list): Black (1996), "Constitutionalising Self-Regulation," *Modern Law Review* 59(1); Tamanaha (2004), *On the Rule of Law*; Suzor (2011), "The role of the rule of law in virtual communities," *Berkeley Technology Law Journal* 25; Lessig (2006), *Code* v2; Humphreys (2007). + +## Caution +Four pages and an abstract-length argument: the empirical claim is stated, not shown. No clause examples, no list of platforms, no date of collection (the 2018 article gives June 2015). Cite the 2018 article for any figure or quotation about the clauses. The paper analyses what the contracts permit, not what platforms do in practice. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the 4-page AoIR PDF (title, author, conference match). I checked all 21 quotations, the paragraph locators (paragraphs counted in order: platforms; coopting; civil society; legal terms; rule of law; the fifteen platforms; new constitutionalism), the 15-platform sentence, the 65 million figure, and the cited authors (Gillespie 2010 and 2014, Lessig 2006, Humphreys 2007, Black 1996, Tamanaha 2004, Suzor 2011). I also checked the facet-check claims about the 2018 article against its text. + +**Findings:** One small correction: the card gave "Google's 65 million" takedown notices; the source says Google "now receives over 65 Million takedown notices". Fixed. The 2018 article does name 14 platforms in note 3 and says it excluded VK from the top 15, and was collected in June 2015, so the facet-check paragraph holds. No other issues found. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/suzor2018digital.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/suzor2018digital.md new file mode 100644 index 00000000..d2b6982b --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/suzor2018digital.md @@ -0,0 +1,37 @@ +# Suzor, N. (2018). Digital constitutionalism: Using the rule of law to evaluate the legitimacy of governance by platforms. *Social Media + Society, 4*(3), 1–11. + +**Identifier:** doi:10.1177/2056305118787812 · **Read depth:** full_text (publisher-formatted open-access PDF, CC BY-NC 4.0, from the QUT repository eprints.qut.edu.au/120050; journal pagination 1–11; reference list read only to the end of the "B" entries in the extracted text) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** doctrinal and document analysis of the terms of service of 14 English-language social media platforms, collected June 2015; no coding scheme reported beyond the three rule-of-law headings · **Parties modeled:** two parties + third party (platform owner governs users; courts and states sit outside) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Suzor proposes that the legitimacy of platform governance be judged by rule-of-law values: "consensual, transparent, equally applied and relatively stable, and fairly enforced" (abstract, p. 1). He opens with Facebook's 2009 promise of binding user votes on its terms: users opposed a privacy-policy change by 88% of 668,872 votes, under 1% of over a billion registered users, short of the 30% turnout needed, and by 2012 Facebook had dropped the commitment (p. 1–2). The legal frame is that terms of service are consumer contracts, "firm to consumer, not sovereign to citizen" (p. 3), and U.S. constitutional rights bind only the state, so "if you don't like it, leave" is the legal answer (p. 3). + +He sorts the values into three headings and applies each to the 14 platforms (Facebook, Twitter, LinkedIn, Pinterest, Google+, Instagram, Tumblr, Flickr, Vine, Meetup, Tagged, MeetMe, ClassMates, Ask.fm; VK excluded; note 3, p. 9): +- **Consent: governance limited by law (pp. 5–6).** Facebook is the only platform to limit termination to violations of its terms; the other 13 reserve termination "for any reason (or no reason)" (table, p. 5–6). "In all cases examined, the Terms of Service provided broad, unfettered discretion to platform owners," and "Like constitutional documents, Terms of Service grant powers, but unlike constitutions, they rarely limit those powers" (p. 6). +- **Equality and predictability (pp. 6–7).** Terms can all be changed by the platform alone. The text says "Approximately half of the platforms committed to some responsibility to inform users" of changes; only Facebook commits to an opportunity to review and comment, "but its commitment to a democratic voting mechanism has been removed"; some give no notice at all, so users must bear "responsibility for continuously checking the legal terms" (p. 7). The table's column grouping did not survive text extraction, so I give no per-platform counts for notice. Only Facebook, Google+ and Instagram limit content removal to unlawful or rule-breaking content; the other 11 reserve wide discretion (table, p. 7); LinkedIn: "We are not obligated to publish any information or content on our Service and can remove it in our sole discretion, with or without notice" (p. 7). +- **Due process (pp. 7–8).** No terms set up an internal dispute process; "The criteria that platform moderation teams use to evaluate decisions are secret"; six of the terms require binding arbitration (the platform pays costs in three), almost all fix the platform's home jurisdiction, most of those requiring arbitration also bar class actions, and all cap liability, often at USD 100 (LinkedIn USD 1,000 for paying users). "Users have no realistic chance of challenging decisions made by platforms in any formal legal process" (p. 8). + +He is explicit about limits: he does not want platforms held to the standards of a constitutional democracy, favours "thin constitutionalism," and notes the method reads only formal documents, not community guidelines, moderator training documents or enforcement practice, so it shows "the outer legal bounds only" (pp. 5, 8–9). + +## Relation to the argument +Supports the author's ruling D8, that platforms "take a highly draconian position," on the contract text, and gives premise P10 a named mechanism: the owner writes the rules, changes them alone and cannot be reviewed. The change-of-terms finding (p. 7) speaks to "a convention that the owner can withdraw alone is no convention" almost directly. It also hands the paper a vocabulary for what conventions would need to carry (notice, reasons, stability, appeal), which the Seattle form (`seattle2024deactivation`) and the EU instruments in facet J then meet. + +Facet check: the facet had this article by abstract only and recorded 14 platforms; the full text confirms 14 and the quoted values. It also shows the sample is 2015 and U.S./common-law centred (p. 4–5), a limit the facet did not carry. + +## In conversation with +- `suzor2016responsibilities` — the four-page precursor (sample of 15). +- `klonick2018new` — cited as "the new governors" (Klonick, 2017); the article's call for a language of user rights answers it. +- `mathews1976eldridge`, `tumey1927ohio` — not cited; they are the state-side due-process standards that Suzor says do not reach platforms (p. 3). +- Leads not in the bib: Obar and Oeldorf-Hirsch (2016), the study of users not reading terms of service (cited p. 6; the bib's `oeldorfhirsch2023what` is a different paper, on algorithmic literacy); Tamanaha (2004), *On the Rule of Law*; Fuller (1969), *The Morality of Law* (bib has `fuller1978forms`, a different work); Gillespie (2010), "The politics of 'platforms'"; Venturini et al. (2016) on terms of service and human rights; Urban, Karaganis and Schofield (2016) on notice-and-takedown. + +## Caution +Collected June 2015; terms have changed (the article notes LinkedIn's rewrite). Formal contracts only, no enforcement data, no count of how often clauses were used. Common-law and U.S. focus. A framework paper: the finding that platforms "perform" poorly (abstract) is a classification of clauses, with no effect on users measured. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** publisher PDF of the Social Media + Society article (11 pp., journal pagination = PDF pages; pdftotext without layout, page by page). I checked all 19 quotations, the Facebook vote figures (88%, 668,872, under 1%, 30%), the 14-platform list, the table counts (termination, content removal, arbitration, cost-sharing), liability caps, and page locators. + +**Findings:** Quotations and numbers are right. Corrections: (1) note 3 with the platform list is on p. 9, not p. 10; (2) the card said "most bar class actions"; the source says most of the platforms that require binding arbitration also prohibit class actions (p. 8); (3) the phrase "poor performance" was in quotation marks but is not in the text (the abstract says "how poorly platforms perform"); I removed the false quotation. `mathews1976eldridge` and `tumey1927ohio` do not appear in the text, as the card says. Reference list only partly read, as the card states, but I found Obar and Oeldorf-Hirsch, Urban et al., Fuller and Klonick in it. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/thrax1874grammar.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/thrax1874grammar.md new file mode 100644 index 00000000..b6f00c33 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/thrax1874grammar.md @@ -0,0 +1,33 @@ +# Dionysios Thrax [attributed]. (1874). The grammar of Dionysios Thrax (T. Davidson, Trans.). *Journal of Speculative Philosophy, 8*, reprinted as a pamphlet. St. Louis (reprint of the journal article). + +**Identifier:** https://archive.org/details/grammarofdionysi00dionuoft · **Read depth:** partial. OCR text of the full public-domain pamphlet (Toronto copy, printed pages 3–16; kept as cards/thrax1874grammar.txt). Read closely: the translator's headnote, §§1–14 (pp. 3–11) and the final section on conjunctions (pp. 15–16); the list of the 25 section headings was read, but the word-class definitions in §§15–24 (verb to adverb, pp. 11–15) were skimmed, not read line by line. The OCR is rough in Greek words and some footnotes. · **Source-tier:** ancient text in an 1874 translation (English) · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** person–artefact (reader, text, grammarian) · **Facets:** K · **Read:** 2026-10-13 + +## What it argues +Davidson's headnote calls the work "the first attempt at a systematic grammar made in the Western World, and for many generations a text-book in the schools of the Roman Empire," and gives what Suidas says of Dionysios, a pupil of Aristarchos who taught in Rome in Pompey's time. On authorship he writes that the work's "genuineness and authenticity... have been impugned, but have been defended by Lersch, and are now generally admitted" (p. 3, 1874). The pamphlet is a numbered list of 25 definitions. + +§1 defines grammar as "an experimental knowledge of the usages of language as generally current among poets and prose writers" in six parts, the first of which is "Trained reading with due regard to Prosody" (p. 3). Davidson's note: prosody "includes everything designated by diacritical marks — aspiration, accentuation, quantity, and sometimes pauses" (p. 3 n.). §2, "On Reading": "Reading is the rendering of poetic or prose productions without stumbling or hesitancy. It must be done with due regard to expression, prosody, and pauses. Through the expression we learn the merit of the piece; from the prosody, the art of the reader; and from the pauses, the meaning intended to be conveyed" (p. 4). It adds that tragedy is to be read heroically, comedy conversationally, elegiacs thrillingly and so on, and that reading that ignores the rules "degrades the merits of the poets." + +§4 on punctuation names three marks and ties them to breath and sense: "the full stop denotes that the sense is complete; the semicolon is a sign of where to take breath; the comma shows that the sense is not yet complete" (p. 4); §5 adds that at the full stop "the pause is long, at the comma, very short" (p. 4). §7 gives the 24 letters, seven vowels and 17 consonants; §8 defines a syllable; §12: "A Word is the smallest part of an ordered sentence" (p. 8); §13: a sentence is a combination of words "making complete sense," and there are eight parts of speech (p. 8). Nothing in §§1–14 or the closing section mentions how words are written or divided on the page. + +## Relation to the argument +Source for facet K on what an ancient school taught as "reading." Two points bear on rulings D1–D2 and premise P6. First, reading is the first part of grammar and is defined as delivery with attention to prosody and pauses, so phrasing was a trained skill of the reader. Second, the definition of punctuation puts the breath and the sense at the reader's, not the page's, discretion: stops are marks of "where to take breath." That fits the reading that readers of unspaced scripts supplied structure themselves (P1), and it adds to the evidence in `quintilian1920institutio` (1.8.1) that pausing and phrasing were learned. It does not say that separating words reduced any of this, nor that it did not. + +Facet check: the facet's quotations (§2 and the footnote) match; the facet read §§1–2 and calls the work "attributed." Davidson (1874) says authenticity "is now generally admitted"; whether that holds against current scholarship was not checked here, so "attributed" is a safe wording. + +## In conversation with +- `quintilian1920institutio` — Book 1.8.1 on when to take breath and where to pause; Thrax and Quintilian do not cite each other. +- `dionysius1910composition` — the other ancient text on reading in this batch; no mutual citation. +- Davidson's notes cite Aristotle's *Poetics* ch. 20 and *Categories* and German scholarship (Schmidt, Steinthal, Lersch); not in the bib; leads. + +## Caution +An 1874 translation of a text of disputed date and authorship, read here in a rough OCR. The definitions are textbook compression; they show categories, not classroom practice. The pamphlet has no entry on layout or word division; silence is not evidence either way. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the OCR text of the 1874 St. Louis pamphlet (title page: "Translated from the Greek by Thomas Davidson", "Reprinted from the Journal of Speculative Philosophy", 1874). I read the headnote and sections 1-14, 24-25 and the section headings, and checked all 17 quotations against the OCR (the OCR garbles some words in the headnote, e.g. "tiraminar", but the card's wording matches the legible text), the page locators (headnote and section 1 on p. 3, sections 2, 4, 5 on p. 4, section 12-13 on p. 8), and the counts (24 letters, 7 vowels, 17 consonants, 25 sections, eight parts of speech). + +**Findings:** Quotations, numbers and locators are correct. One correction: the Read depth label "partial_full_text" is not a valid value and the card itself says sections 15-24 were skimmed, so it now reads "partial", with the list of what was not read kept. The volume number "8" of the Journal is not on the pamphlet's title page, so I could not verify it from this copy. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/torrensalvarez2022latin.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/torrensalvarez2022latin.md new file mode 100644 index 00000000..892a1e4c --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/torrensalvarez2022latin.md @@ -0,0 +1,30 @@ +# Torrens-Álvarez, M. J., & Tuten, D. N. (2022). From "Latin" to the vernacular: Latin-Romance hybridity, scribal competence, and social transformation in medieval Castile. *Speculum, 97*(3), 698–736. + +**Identifier:** doi:10.1086/720547 · **Read depth:** partial (open repository copy of the article, 39 PDF pages; Speculum pagination). I read pp. 698–717 and 731–736 in full; pp. 718–730 (the register analysis of Nicholaus Martini's texts and the competence argument) I read by extracting the opening of section 4 and every sentence on competence and comprehension, not line by line, and I did not read the Old Castilian examples or the table-like passages · **Source-tier:** peer-reviewed journal article · **Evidence basis:** direct_read · **Design and sample:** qualitative philological analysis of 53 manuscript legal documents by one Burgos scribe, Nicholaus Martini, 1211–1228 (14 transcribed from originals; 51 property transfers and 2 inquest reports) · **Parties modeled:** other (scribe, clients, institutions and the listeners at a public reading) · **Facets:** B · **Read:** 2026-10-13 + +## What it argues +Torrens-Álvarez and Tuten ask why early-thirteenth-century Castilian legal documents mix Latin and Romance. Earlier scholars, they say, mostly blame "some kind of scribal incompetence" (p. 700). The authors argue the opposite. Nicholaus could write continuous Latin, continuous Romance and a conventional hybrid, and he chose among them by genre and by what each register signalled (pp. 701, 718–722). They tie the hybrid register to urbanisation, a lay property market, and the rise of autonomous lay scribes who name themselves (*Nicholaus Martini scripsit*) and who answer for the documents as individuals (pp. 709–717). They describe the later steps as the royal chancery adopting the vernacular: by their cited figure, 34 percent of Fernando III's chancery documents in the 1230s were in nearly or fully continuous Romance (n. 4, p. 699, citing Ostos Salcedo), and all of Alfonso X's internal chancery documents (n. 4). + +Their second target is the assumption that these documents had to be easy to understand. The abstract says they "argue against an overemphasis on ease of comprehension as a factor shaping the design and reception of written texts and oral performances of them" (p. 698). In section 6 they argue that stressing a "need" for comprehensibility "may reflect a modern, utilitarian conception of discourse" (p. 731). Latin in the Mass and in formulae signalled institutional authority whether or not listeners understood it. The public reading aloud of a deed "had authorization and authentication as primary goals, not comprehension of documentary content" (p. 732). The conclusion follows: the texts "were not always fully comprehensible to all readers and listeners" (p. 735), yet they carried authority, and that is part of why Romance later acquired authority too. + +## Relation to the argument +The paper is not about word separation, and "continuous Romance" in it means a text wholly in one register, not unspaced script. Nothing in the text bears on P1–P4 directly. The facet B note was right that it is a lead only; the full text confirms that and also shows where the lead goes. + +Two uses are open, both analogical. First, it is a published case of historians contesting an ease-of-reading explanation of a change in how texts look and are read, and replacing it with social function. That is the same shape as the Saenger/Vatri/Heilmann dispute over cost, so it can sit beside ruling D2 as a reminder that "the new form was easier to read" needs its own evidence. Second, it speaks to ruling D8 (who holds the convention). The scribes form a "community of practice" (p. 714); the convention later gains the backing of the town council and then the crown, with the named scribe as "guarantor" (p. 717). A convention held by working practitioners and then underwritten by public authority is a possible comparison for the "bends towards" question D8 leaves open. The scribes served the parties to a transaction, not a platform's owner, so it does not bear on the P10 contrast. + +## In conversation with +Nothing in `references.bib` is cited by this paper or answers it. It argues with Wright (*Late Latin and Early Romance*, 1982), Menéndez Pidal, and Montaner Frutos, none of which is in the bib. It also draws on Le Goff and Ruiz for the social setting. `clanchy1993from` covers the same centuries from the English side, but the article does not cite it. +- Leads not in the bib: Wright 1982; Ruiz 2004, *From heaven to earth*; Montaner Frutos 2012 on spectroglossia; Moore 2013 on the performance of royal charters; Wright 2000 on the Treaty of Cabreros. + +## Caution +Do not cite it for anything about word separation, silent reading or the Insular scribes. Its evidence is one scribe in one city over 17 years, and its claim about comprehension rests on analogy to the Mass and on the authors' reading of the hybrid forms, not on any measure of what listeners understood. The page for the 34 percent figure is a note that cites another author. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the 39-page repository PDF (title, authors, Speculum 97(3) 698-736 match; journal page = PDF page + 697). I checked all 10 quotations and their page locators (pp. 698, 699 n. 4, 700, 714, 717, 731, 732, 735), the corpus numbers (53 documents, 1211-1228, 14 from originals, 2 remembranzas) and the 34 percent figure in note 4. + +**Findings:** All quotations, numbers and locators are correct. One correction: the card says pp. 718-730 (13 of 39 pages, the register analysis and competence argument) were skimmed, not read line by line, so "full_text" is changed to "partial", with the list of what was not read kept. Wright, Menéndez Pidal and Montaner Frutos are discussed in the text; none of the bib works named appears there. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tumey1927ohio.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tumey1927ohio.md new file mode 100644 index 00000000..c945eb70 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tumey1927ohio.md @@ -0,0 +1,34 @@ +# *Tumey v. Ohio*, 273 U.S. 510 (1927). + +**Identifier:** https://www.law.cornell.edu/supremecourt/text/273/510 · **Read depth:** full_text (Cornell LII reproduction of the opinion of the Court by Chief Justice Taft; paragraph-numbered, no U.S. Reports page breaks, so locators are paragraph numbers, "¶", which are my count: ¶1 is the first paragraph after "delivered the opinion of the Court", and the block-quoted statute and ordinance paragraphs each count as one; counsel's argument is omitted in the source, "intentionally omitted" at pp. 511–513) · **Source-tier:** court opinion (U.S. Supreme Court) · **Evidence basis:** direct_read · **Design and sample:** none (legal holding on one village mayor's court) · **Parties modeled:** two parties + third party (accused, the State, and a judge paid from the outcome) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Taft states the question: whether Ohio's Prohibition Act, in letting a village mayor try the accused, denies due process under the Fourteenth Amendment "because of the pecuniary and other interest which those statutes give the mayor in the result of the trial" (¶1). The facts: Tumey was fined $100 for possessing liquor; the mayor received $12 in costs from a convicted defendant and nothing from an acquitted one (¶¶2, 39); fines were split half to the State and half to the village (¶14) and the ordinance sent half the village's share into a "secret service fund" paying detectives and marshals 15% of fines and the prosecutor 10% (¶¶19–22). Over May–December 1923 the court collected over $20,000, of which the mayor took $696.35 on top of his salary (¶36). + +The holding: "it certainly violates the Fourteenth Amendment and deprives a defendant in a criminal case of due process of law to subject his liberty or property to the judgment of a court, the judge of which has a direct, personal, substantial pecuniary interest in reaching a conclusion against him in his case" (¶38). Taft draws a line short of this: "matters of kinship, personal bias, state policy, remoteness of interest would seem generally to be matters merely of legislative discretion" (¶38). He reviews English common-law practice and finds "the greatest sensitiveness over the existence of any pecuniary interest however small" (¶44), and finds the pay-only-on-conviction practice not embedded in custom (¶64). A de minimis exception is allowed, but about $100 a month in costs was not "minute, remote, trifling, or insignificant" (¶65). + +Two points matter beyond money. First, the test is structural, not about the particular man: "There are doubtless mayors who would not allow such a consideration as $12 costs in each case to affect their judgment," yet "Every procedure which would offer a possible temptation to the average man as a judge to forget the burden of proof required to convict the defendant, or which might lead him not to hold the balance nice, clear, and true between the state and the accused denies the latter due process" (¶66). Second, a separate ground: the mayor was also the village's chief executive responsible for its finances, so "an official perforce occupies two practically and seriously inconsistent positions, one partisan and the other judicial" (¶67), while the mere union of executive and judicial duties in a small-village mayor is not itself a violation (¶67). Actual guilt does not cure it: "he had the right to have an impartial judge" (¶69). The judgment was reversed (¶70). + +## Relation to the argument +This is the court-side case for the author's ruling D8 and premise P10. A third party that gains from one side's loss is disqualified as a matter of constitutional design, and the disqualifying fact is the interest, not proof of bias. A platform that decides disputes between itself and its users, and keeps the fees, fines or account value, is in the mayor's position by this test; Suzor's finding (`suzor2018digital`) that terms give the owner unreviewable discretion is the other half. It also supports the author's reading of convention as something held outside the interested party. + +It cuts the other way in two places the paper should keep. The rule binds the State only (Fourteenth Amendment), and Taft expressly leaves "personal bias" and "remoteness of interest" to legislatures (¶38), so even courts police only the clearest interest by constitutional law. And the remedy is after the fact, at the accused's own initiative. + +Facet check: the facet quotation (¶38) and the $12 are right. The facet had read the opinion "in part"; the full text adds the secret-service fund, the mayor's dual role (¶67), the $696.35 and the de minimis limit (¶¶64–65), none of which the facet recorded. + +## In conversation with +- Later works in the bib that build on or sit beside this opinion (Tumey does not cite them; the pairing is the project's, not the opinion's): `caperton2009massey` (disqualification where the probability of bias is too high, per the facet), `mathews1976eldridge` (balancing test for what process is due), `us455usc` and `us9usc10` (statutory recusal and vacatur for "evident partiality"). +- The opinion's own authorities: *Dimes v. Grand Junction Canal* (3 H.L.C. 759), *Bonham's Case* (8 Coke 118a), Cooley's *Constitutional Limitations* (7th ed.) p. 592 et seq., and *City Council v. Pepper* (1 Rich. 364), the case on remote interest. None is in the bib; they are leads. + +## Caution +A criminal due-process holding about a paid judge. It does not decide platforms, private arbitrators or civil disputes. LII text has no U.S. Reports pagination (cite as 273 U.S. 510 with ¶ for now, and take page pins from the official reporter before print). Counsel's arguments are omitted in the source. The Court says it was "referred to no cases at common law in England" showing judges paid only on conviction and reasons from analogous cases (¶41). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the Cornell LII HTML of 273 U.S. 510 (47 S.Ct. 437; decided 7 March 1927; Taft for the Court). I checked all 14 quotations, the dollar figures ($100 fine, $12 costs, over $20,000, $696.35, about $100 a month, 15% and 10%) and the cited authorities (Dimes, Bonham's Case, Cooley pp. 592 and 594, City Council v. Pepper). + +**Findings:** Quotations and numbers are right. Corrections: (1) the paragraph locators did not match any consistent count of the LII text (LII gives no paragraph numbers); I renumbered every cite to one stated count (¶1 = first paragraph of the opinion; for example, the holding and "kinship" sentence are in ¶38, the "doubtless mayors" and "Every procedure" passage in ¶66, the "two ... inconsistent positions" passage in ¶67, "impartial judge" in ¶69, the reversal in ¶70) and said so in the Read depth line; (2) the Caution said the Court notes the English cases are "not on all fours"; the opinion does not say that. It says it was referred to no English common-law case of judges paid only on conviction (¶41). (3) Counsel's omitted argument is at pp. 511-513, not "513-". Cooley is cited at p. 592 (¶37 in the first list of authorities) and p. 594; the card's "p. 592 et seq." is right. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tunbridge1992study.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tunbridge1992study.md new file mode 100644 index 00000000..a78960b2 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tunbridge1992study.md @@ -0,0 +1,47 @@ +# Tunbridge, G. L. (1992). *A study of scribal practices in early Irish and Anglo-Saxon manuscripts* [DPhil thesis, University of Oxford]. St Cross College. + +**Identifier:** doi:10.5287/ora-ave2z6gqd (Oxford Research Archive) · **Read depth:** partial (see Caution for the chapters read in full and the large parts skimmed or not read). The PDF is a 366-page scan with an OCR text layer; the title page names "G. L. Tunbridge, St Cross College" and the Oxford DPhil, Hilary Term 1992, matching the bib entry. Page numbers below are the thesis's own printed pages (PDF page minus 23 in the body) · **Source-tier:** unpublished doctoral thesis · **Evidence basis:** direct_read · **Design and sample:** palaeographic study of surviving Latin manuscripts copied in Ireland and England, 7th to early 9th century (about 250 survive, Preface p. xiii), worked from Lowe's *CLA* plates and photographs and some originals, plus a reading of Insular and late-antique grammatical texts. No experiment; counts come from sample columns and pages · **Parties modeled:** scribe, reader (often a learner of Latin as a foreign language), and the exemplar and grammatical tradition behind both; no third-party intermediary · **Facets:** A · **Read:** 2026-10-13 + +## What it argues + +Tunbridge holds that Insular scribes began to separate the words of Latin text because they were reading and copying Latin as a foreign language, and that the separation shows a linguistic analysis learned from the grammarians. The Long Abstract states the case: "Thus Irish scribes when copying Latin began to abandon scriptio continua in favour of a layout which clarified the syntax and morphology of the language and hence met the needs of the non-native Latin speaker" (Long Abstract p. 4). Chapter 2 supplies the motive. Old Irish scribes wrote their own language in syntactic word-groups (pp. 89-92), so an Irish student met Latin's narrower word boundaries as a difficulty (pp. 93-97), and the Continental exemplars "were in a sense part of the problem" because their unbroken lines "made no concessions to foreigners" (p. 97). Chapter 3 supplies the method: the eight parts of speech taught by the grammarians became "the principle underlying the separation of words" (Long Abstract p. 4), and gaps in the teaching on prepositions explain the typical Insular "errors" (pp. 124-150). + +Chapter 4 describes the practice. Scribes marked words by space, points, tags on final letters and variant letter forms (pp. 165-175). Tunbridge counts groupings in three manuscripts. In one column of the St Paul Pompeius, prepositions occur ten times in groups and separately only at a line end, and *et* is joined all six times (p. 183). In Hatton 93 (early ninth century), 38 of 42 *et* are joined, and of 65 chances to separate a preposition 49 are clearly attached, 5 slightly separated and 11 clearly separated (pp. 185-186). Her table shows prepositions and *et* joined in all three manuscripts, a pattern "common to most Insular scribes from the late seventh through to the ninth centuries" (p. 190). She also holds that "there is no question of their consciously following a rigid set of 'rules'" (p. 190). Fully consistent separation into parts of speech does not appear before the twelfth century (p. 125). + +Chapter 5 traces the early stages. All the earliest Irish manuscripts "already contain some kind of separation" (p. 193). Her table (pp. 214-215) runs from erratic wax tablets and syntactic-unit spacing (Ussher I, Milan Orosius) through word-based separation (St Gall Isidore, second quarter of the 7th century) to the Bangor Antiphonary (c. 680-691) and the Schaffhausen Adomnan (before 713). She warns against reading this as a straight line (pp. 214-216), and notes that the earliest full word separation turns up "not in formal liturgical books but in the more informally written copy of the *Etymologiae*" (p. 216). The Anglo-Saxons took the habit from the Irish, and Wearmouth-Jarrow scribes later compressed the spacing under scriptorium discipline (pp. 217-227). Scribes copy as readers, so the practice "fundamentally represents a kind of linguistic analysis" (pp. 228-229). Chapter 6 links separation to abbreviation: both show that Latin was seen "primarily as a written or 'visible' language" (p. 289, quoting Parkes), and both mean readers took in "word by word" rather than "letter by letter" (p. 297). + +## Relation to the argument + +Facet check: the facet (A6) read only the abstract and called the thesis "independent of Saenger" and hedged ("may have"). The full text confirms a different method (grammar teaching and manuscript counts). It is stronger than the abstract on motive: "must be seen as a practical outcome" of the grammarians' study (p. 98) and "primarily a linguistic phenomenon" (Long Abstract p. 7). It is not independent of Parkes, who supervised it (Acknowledgements), and its dating of the earliest manuscripts rests on Brown and Lowe (p. 194, n. 3), so it adds no independent test of Saenger's chronology (premise P2). + +- **P1 and P2.** It supports the foreign-learner motive and the late seventh-century Irish date (Bangor, Adomnan) from the manuscripts themselves. It says nothing about whether *trained* readers of continuous script were burdened; its readers are learners. +- **P5 (the scribe carries the boundary work).** Supported, and complicated. The scribe did the analysis, but with errors and with his own grouping habits. Hatton 93 renders *catholicon latine* as "catholi conlatine" and shows that a new Continental text composed no later than 822 could still circulate in continuous script (pp. 191-192). Inference mine: a reader who receives pre-analysed text also receives the analyst's mistakes. +- **P6 and ruling D1/D2.** The "no training needed" line is not supported here. The thesis concerns what scribes had to know (grammar), and it reports no change in how many people read. +- **D6 (a bundle, but is one move the critical one?).** Tunbridge treats separation with abbreviation, punctuation and page layout (pp. 34-38). She argues, conditionally on Parkes, that heavy abbreviation "was dependent on a degree of word separation" (p. 295), but allows that abbreviation was possible in continuous script (p. 294). That is partial support for separation as an enabling move. +- **P10 and D8 (who holds the convention).** The grammarians "on the whole ignored the specific needs of scribes as copyists" (p. 150) and were silent on separation (p. 156). The convention lived in scribal practice. A scriptorium could suppress it locally: Wearmouth-Jarrow discipline "sought to suppress" ample spacing (p. 225), and Hatton 48 avoids it to imitate Continental models (p. 295, n. 155). The scribes also shared their readers' position as learners of Latin (p. 234). The sources do not show a single party able to withdraw the convention for everyone. +- **Silent reading.** She cites Parkes against Saenger on the date of silent reading (p. 235, n.), which fits the dropped claim that spacing created silent reading. + +## In conversation with + +- `saenger1982silent`: Tunbridge reads it as the first to see the significance of separation, as an unplanned result of Irish contact with Latin, justified by reading speed (pp. 159-161). She faults it for ignoring the linguistic analysis in grouping (p. 229, n.). +- `parkes1992pause`: cited as forthcoming (p. 37, n. 114). She takes Parkes's morphological account as her starting point (p. 161) and tests it against Keller's phonological one (pp. 231-232). +- `saenger1997space`: later than the thesis; the comparison is the project's, not hers. +- Leads not in the bib: Parkes, "The Contribution of Insular Scribes ... to the 'Grammar of Legibility'", in *Scribes, Scripts and Readers* (1991), pp. 1-18; Keller, "Zur Worttrennung in den angelsächsischen Handschriften" (1929); Cohen, *La grande invention de l'écriture* (1958); Bischoff, *Latin Palaeography* (trans. 1990); Draak, "Construe marks in Hiberno-Latin manuscripts" (1957). + +## Caution + +A thesis, unrefereed and unpublished; I did not check whether it later appeared as articles. The motive rests on inference from grammar teaching and the pattern of groupings, not on any record of why a scribe spaced a word. The dates come from Lowe and Brown. The counts are small samples (one column, a few folios) and she says so (p. 184). Nothing measures reading cost, speed or training time. Her speed claim (p. 298) leans on Saenger. The Oxford scan's OCR is rough in places. I read in full: the abstracts, Preface, Chapter 1 pp. 34-41, Chapter 2 pp. 89-98, Chapter 3 pp. 124-126 and 148-150, Chapter 4 (all), Chapter 5 (all), Chapter 6 pp. 237-238 and 279-299, the Conclusion. I skimmed or did not read: Chapter 1 pp. 2-33 (Patrick's alphabets, *Ad Cuimnanum*), Chapter 2 pp. 43-88 (the *Auraicept*), Chapter 3 pp. 99-123 and 126-147 (the grammarians), Chapter 6 pp. 239-278 (manuscript-by-manuscript abbreviation studies), the Appendix (pp. 301-317, first page only), the plates and manuscript lists. The bibliography pp. 333-366 was not read through; references come from footnotes. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the 366-page Oxford scan with OCR text (title page, author and "Hilary Term, 1992" match; printed page = PDF page minus 23 holds). I checked all 25 quotations and their printed-page locators (Long Abstract pp. 4 and 7; pp. 97, 98, 125, 150, 183, 185-186, 190, 191-192, 193, 214-216, 225, 228, 289, 295, 297), the Pompeius and Hatton 93 counts (10 prepositions; 6 et; 38 of 42 et; 65 chances, 49 attached, 5 slightly, 11 clearly separated), the table of early manuscripts and the Preface figure of about 250 manuscripts (p. xiii). I did not re-read the chapters the card lists as skimmed. + +**Findings:** (1) Quotation fixed: the source reads "on the whole ignored the specific needs of scribes as copyists" (p. 150), with no comma after "whole". (2) Hatton 93: the source says the text was "composed no later than 822" (p. 192), not "about 822"; reworded. (3) The card says large parts were skimmed or not read (Ch. 1 pp. 2-33, Ch. 2 pp. 43-88, Ch. 3 pp. 99-123 and 126-147, Ch. 6 pp. 239-278, the bibliography), so "full_text" is changed to "partial". The rest holds. + +## Also used by facet W (2026-10-15) + +Facet W pairs the thesis with `doyle2025assessment`, a 2025 corpus paper that quotes Thurneysen (1946, p. 24) on Old Irish manuscripts writing stress-groups "as one", confirming from the editorial side what Tunbridge argues palaeographically (pp. 89–92): the scribes who separated Latin did not write their own language in separated words. For the abstract's clause "seventh-century Irish scribes ... adding spaces between words", the thesis supplies the learner's motive and the grammarians' method and says nothing about voice or silent reading except to cite Parkes against Saenger (p. 235 n.). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/turan2026oversight.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/turan2026oversight.md new file mode 100644 index 00000000..b179f622 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/turan2026oversight.md @@ -0,0 +1,46 @@ +# Turan, E. (2026). Oversight has a capacity: Calibrating agent guards to a subjective, fatiguing human. *arXiv preprint* arXiv:2606.08919. + +**Identifier:** arXiv:2606.08919 · **Read depth:** full_text (preprint v1, 8 Jun 2026, 12 pages; locators are PDF pages. I read everything, including the reference list; I did not open the author's code or data repository) · **Source-tier:** preprint, single independent author, not peer reviewed · **Evidence basis:** direct_read · **Design and sample:** a systems and measurement study. 125 hand-labelled coding-agent actions (labels: safe 42, approval 52, blocked 31; all by the author, p. 4); a guard that scores risk with deterministic rules and an LLM (Haiku, compared once with Sonnet); three LLM "reviewer personas" standing in for human annotators; and a simulation of a reviewer whose reliability falls linearly once an escalation count passes a capacity C. No human reviewers took part · **Parties modeled:** person–system. One human reviewer and one agent guard, with an adversary who floods the reviewer · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +Turan says the standard safety pattern for agents, a human approval gate, treats two things as given that are false: that "risky" has one correct label, and that the reviewer is "a perfect, infinitely-available oracle" (abstract, p. 1). He makes three claims. + +First, the label is contestable. Three LLM personas (cautious, pragmatic, strict-compliance) labelling the 125 actions reach Fleiss' κ = 0.52; the pairwise Cohen's κ range from 0.42 to 0.71, and the pragmatic persona calls 87 actions safe against the cautious persona's 45 (p. 5). The author's own gold labels are a single judge's, "in deliberate tension" with that floor (p. 4). + +Second, a guard framed as selective classification under an assumed 50:1 cost for a missed catastrophe against a false alarm cannot safely auto-decide on hard inputs. Reaching 0% missed danger needs a false-alarm rate near 100%, and the area under the risk-coverage curve is 0.374 ± 0.002 across three seeds for Haiku (pp. 5, 7). + +Third, once the reviewer is modelled as fatiguing, realised danger is U-shaped in the escalation rate. The reviewer model is r(ℓ) = max(r_min, 1 − 0.02 · max(0, ℓ − C)), with r_min = 0.2 (p. 5), so fatigue is assumed to start at capacity C. Table 2 (p. 6) gives, for C = 10, 25 and 50, an optimal escalation rate of 64%, 64% and 72%, and a share of danger getting through of 56%, 42% and 22% at the optimum against 69%, 57% and 39% at full escalation. A flooding attack that buries one malicious action in k benign ones reaches an attack success of about 80% for a policy that escalates 88% of routine actions, and 40% at 50 filler actions; a "load-aware" policy escalating 26% holds at 0% until about 100 filler actions (p. 7). + +Turan is plain about status: "The inverted-U and the flooding attack are modeling results that motivate a human study" (abstract, p. 1), and "the inverted-U below is a direct consequence of the assumed monotonically-fatiguing reviewer ... not an empirical finding about real people" (p. 6). He claims no new mechanism: fatigue-aware deferral, capacity-constrained deferral and reviewer-flooding attacks are "prior art" (pp. 1-2). + +## Relation to the argument + +Supports premise P12 in a narrow way: the reasoning that more escalation can lower safety, and the framing of attention as a budget, fit the argument that the post's P15 and P19 assign people work they perform badly. It is not evidence for P12. The fatigue curve is assumed, so the paper cannot show that people fail at it. Facet G is right to say "do not cite as human evidence," and this reading confirms it. + +Two features are useful for the paper's design argument. The paper's remedy is to escalate less, so the guard keeps more of the work, which is the artefact-side direction the argument map favours (P8); it does not give the person a better reading task. And the flooding attack names a concrete way the post's "observe everything" (P15) could be turned against its reader: a long benign stream degrades the one dangerous review. + +[Not from Turan, whose text and references do not cite Parasuraman and Manzey; the claim is unverified.] Against the model, Parasuraman and Manzey report that fatigued operators in one study were more careful with a decision aid than alert ones (Parasuraman and Manzey, p. 404). A monotone-decay model is one of two plausible shapes. + +Facet check: facet G calls it "a model of a fatiguing reviewer." True, and the paper also reports a measured disagreement result (κ = 0.52), though from LLM personas, not people (pp. 5, 9). Facet G's other statement, that the paper calls the results modelling results, holds in the body (pp. 1, 6, 9). + +## In conversation with + +- `tariq2025alert`: cited as reference 15 for fatigue and flooding attacks on reviewers in security operations (pp. 1-2, 7). +- `anthropic2026how`: not cited by Turan. +- `green2022flaws`, `mitchell2026agents`, `yu2026habituation`: none cited. +- Leads, not in the bib: Zhang et al. 2026, "Fatigue-Aware Learning to Defer via Constrained Optimisation" (arXiv:2604.00904), the source of the endogenous-reviewer idea, which Turan credits "not ours" (p. 3); Alves et al. 2024, cost-sensitive learning to defer under workload constraints (TMLR; arXiv:2403.06906); Wang, Li and Tian 2026, "Reframing LLM agent security as an agent-human interaction problem" (arXiv:2605.24309); the European Data Protection Supervisor's 2025 TechDispatch on human oversight of automated decision-making (his ref. 12). + +## Caution + +Small, curated, single-domain (coding-agent actions) and single-author; one dataset underlies every result, which the paper itself calls "five views of one measurement instrument" (p. 8). The 50:1 cost ratio, the linear fatigue slope (0.02), the floor (0.2) and the capacities (10, 25, 50) are chosen, not fitted. The "reviewers" are LLM personas. The optimistic reviewer, who starts at perfect reliability, makes absolute danger rates lower bounds (p. 8). Do not cite the inverted-U or the flooding numbers as findings about people. Preprint v1; the repository may have moved on. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** arXiv v1 PDF (12 pp.; title, author, 8 Jun 2026 match), whole text and page by page. I checked all 15 quotations and their PDF-page locators, and the numbers: 125 actions (42/52/31), Fleiss' kappa 0.52, pairwise kappa 0.42/0.47/0.71, 87 against 45, AURC 0.374 +/- 0.002 (Sonnet 0.351 against Haiku 0.373), 50:1, slope 0.02, r_min 0.2, Table 2 (64/64/72; 56/42/22; 69/57/39), the flooding figures (about 80%, 88%, 40% at 50 filler, 26%, 0% until about 100) and the FALCON/DeCCaF/ref. 15 relations. + +**Findings:** Quotations, numbers and locators are right. One correction: the paragraph "Against the model, Parasuraman and Manzey report ... (p. 404)" is not from Turan; his text and reference list do not cite them. I marked it in the card as not from this source and unverified. The code and data repository was not opened, as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/turpin2023language.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/turpin2023language.md new file mode 100644 index 00000000..10debabf --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/turpin2023language.md @@ -0,0 +1,44 @@ +# Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting. *Advances in Neural Information Processing Systems 36 (NeurIPS 2023)*. + +**Identifier:** arXiv:2305.04388 · **Read depth:** partial (arXiv v2, 9 Dec 2023, in NeurIPS layout; page numbers below are that PDF's, main text pp. 1–10, references pp. 11–15; I read Appendices B, C and H and the prompt table in Appendix E, and skimmed the per-task result tables and prompt listings in D–G, pp. 16–32) · **Source-tier:** proceedings (peer-reviewed conference paper; the bib entry is typed `@misc`) · **Evidence basis:** direct_read · **Design and sample:** behavioural experiments on two models, text-davinci-003 ("GPT-3.5") and claude-v1.0; 13 BIG-Bench Hard tasks, 330 items per task (3,299 items, p. 4); a stratified subset of the Bias Benchmark for QA (BBQ), 2,592 items (p. 7); no human participants · **Parties modeled:** other (a model and its own written explanation; a human reader is assumed, never tested) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +Turpin and colleagues ask whether a chain-of-thought (CoT) explanation names the real cause of the answer. They test it by adding a bias to the input that the explanation never mentions. If the answer moves and the explanation stays silent about the bias, the explanation is unfaithful. The abstract states the result: "CoT explanations can systematically misrepresent the true reason for a model's prediction" (p. 1). + +Two biasing features go into the BBH prompts: reorder the few-shot options so the right answer is always "(A)", or add a user line suggesting a random answer. Accuracy on items whose bias points to a wrong answer then falls. The largest drop is 36.3 points (GPT-3.5, zero-shot CoT, Suggested Answer; p. 6). "Answer is Always A" costs GPT-3.5 18.7 points and Claude 1.0 4.7 points (p. 6). Few-shot examples shrink the Suggested Answer drop to 24.1 points for GPT-3.5 and 21.5 for Claude (from 30.6; p. 6). The 95% intervals run about ±1.6 to ±2.4 points (p. 6). The authors find that "models virtually never verbalize being influenced by our biasing features": 1 of 426 reviewed explanations mentions the bias (p. 3; Appendix B gives 0 of 234 for BBH and 1 of 192 for BBQ). + +In a hand-annotated sample of 104 biased-context explanations, as many as 73% argue for the bias-consistent answer, and 15% show no obvious error (p. 7). On BBQ, the model sees two versions of weak evidence with the people swapped. Of the pairs it answers inconsistently, up to 62.5% line up with a stereotype (Claude, few-shot CoT, no debiasing instruction; 50% would be chance; p. 8). A debiasing instruction cuts that to about 51% for Claude but not for zero-shot GPT-3.5 (p. 8). + +The authors name their limits: this is "a necessary but not sufficient test for faithfulness" and covers only small input changes (p. 10). They add that a user could plant a bias and produce biased predictions "without a trace of this bias in its CoT explanations", which "could cause problems for model auditing or fairness methods if they rely on CoT explanations" (p. 10). + +## Relation to the argument + +**Premise P11 and ruling D4.** The Nadella post (P9) calls chain-of-thought transparency "not sufficient or dependable". Turpin is one of the empirical grounds for that line, and it supports it directly: a readable account of a model's reasoning can be plausible and still not be the reason. That strengthens the paper's reading that the post puts its load-bearing controls outside the model (P11). It also bears on D4, where the author said the post's "only move" is disclosure design. Turpin's own remedy is model-side: improve CoT faithfulness through "targeted efforts" or abandon CoT (p. 2). Neither is disclosure design, which fits the flag in `SOURCE.md` that "his only move" overstates the post. + +**Ruling D3.** Turpin tests whether a model's self-explanation tracks its cause. It does not test whether a person who is told how an algorithm works navigates it better. Do not cite it for "disclosure of the algorithm doesn't affect our ability to navigate it". It supports only the narrower claim that one form of disclosure, the model's own explanation, cannot be trusted as evidence. + +**Premise P12.** The paper models no reader. It shows the evidence a human overseer would read can mislead; it does not show how people read it. + +**Facet check:** the facet's "drop[s] by as much as 36%" is correct but is a drop of 36.3 accuracy points in one setting (GPT-3.5, zero-shot, Suggested Answer), not a relative loss. Claude 1.0's matching figure is 30.6 points, and the Answer-is-Always-A drops are far smaller (18.7 and 4.7 points; p. 6). + +## In conversation with + +- `lanham2023measuring` — Turpin cites it as concurrent work that perturbs generated CoT and finds that models often ignore the edits; Lanham proposes further necessary-but-not-sufficient tests (p. 9). +- `hase2020evaluating` — the source of the counterfactual-simulatability frame Turpin adopts (p. 3). +- `chen2025reasoning` — a later study with a title that echoes Turpin's; I did not read it, and Turpin (2023) cannot cite it. +- Not in the bib, as leads: Chen et al. (2023), "Do models explain themselves? Counterfactual simulatability" (arXiv:2307.08678); Jacovi & Goldberg (2020) on defining faithfulness; Parrish et al. (2022), the BBQ benchmark; Nisbett & Wilson (1977), cited on human explanations as unfaithful accounts; Radhakrishnan et al. (2023) on question decomposition. + +## Caution + +Two 2023 models, multiple-choice tasks, and artificial biases that the authors planted. The result does not say how often a current reasoning model hides its real cause, and the authors call their test necessary, not sufficient. No human reads any explanation. Use it for "plausible explanation does not guarantee faithful explanation", not for any claim about people's use of explanations. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** arXiv v2 PDF (32 pp., 9 Dec 2023; title, authors match). I checked the 14 quotations, the page locators (pp. 1, 3, 6, 7, 8, 10), and the numbers: 330 items per task and 3,299 in all; BBQ 2,592; drops of 36.3, 18.7, 4.7, 24.1, 21.5 and 30.6 points; intervals of 1.6 to 2.4; 426 explanations reviewed (234 BBH, 192 BBQ, 1 mention); 104 annotated; 73% and 15%; 62.5% for Claude 1.0 few-shot CoT; 62.5 to 50.6 with the debiasing instruction and 59.2 to 60.0 for zero-shot GPT-3.5. + +**Findings:** (1) The "no trace" quotation was not verbatim; the source (p. 10) reads "without a trace of this bias in its CoT explanations". Fixed. (2) The text says "As many as 73%" of the sampled unfaithful explanations support the bias-consistent answer; the card stated 73% flatly. Fixed. (3) The card says the per-task tables and prompt listings in Appendices D-G (pp. 16-32, over half the PDF) were only skimmed, so "full_text" is changed to "partial". The relations to Nadella-post premises are the project's, not Turpin's. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tyler1988what.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tyler1988what.md new file mode 100644 index 00000000..ee2ebab5 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/tyler1988what.md @@ -0,0 +1,36 @@ +# Tyler, T. R. (1988). What is procedural justice?: Criteria used by citizens to assess the fairness of legal procedures. *Law & Society Review, 22*(1), 103–135. + +**Identifier:** doi:10.2307/3053563 · **Read depth:** full_text (publisher PDF from Cambridge Core, journal pagination; the even-numbered pages 118–126 lost their running heads in the scan, so locators for pp. 118–127 are given as ranges; reference list skimmed) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** telephone survey of Chicago residents, spring 1984; 1,575 interviewed (63% of those contacted), 733 reported personal contact with the police or courts in the past year, 652 with usable experiences were analysed; cross-sectional, correlational · **Parties modeled:** person against a legal authority (police or court) acting as decider; disputes between two parties are the minority (18%) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Tyler first replicates the finding that how people judge the fairness of a procedure shapes their satisfaction and their view of the authorities, apart from the outcome they got (pp. 103–105, 119–120). He then asks what makes a procedure seem fair. He combines two lists: Thibaut and Walker's process and decision control, and Leventhal's six criteria (consistency, bias suppression, decision quality, correctability, representation, ethicality; pp. 104–105). + +The respondents were callers to the police (47%), people stopped by the police (31%) and court users (23%) (p. 110). He measured each criterion with a few survey items and regressed a rating of the fairness of the procedure on all of them at once. The criteria explain 69% of the variance (adjusted R²; pp. 121–122, Table 3). Seven make independent contributions, with standardized coefficients: the authorities' effort to be fair .30, honesty .23, ethicality (politeness and respect for rights) .21, quality of decisions .17, representation .17, correctability .14, and bias .07. Consistency with past experience or with other people's treatment does not matter (β = .04, not significant), though earlier work predicted it would. Fairness judgments explain about 24% of the variance in satisfaction and evaluation that nonfairness factors leave unexplained, against about 1% the other way (p. 117). + +Two analyses of variation follow. He tests eight criteria against six features of the situation and eight personal traits. Sixteen of 48 situation comparisons were significant (33%), against 5 of 64 trait comparisons (8%), which is about chance (pp. 125–126). So "the meaning of procedural justice varied according to the nature of the situation, not the characteristics of the people involved" (abstract, p. 103). In disputes, representation and consistency count for more; in non-disputes, the officer's effort to be fair (pp. 125–127). He concludes there are "no universally fair procedures" (p. 131), but that people share a common conception within a setting, which helps officials, because all parties will "focus on similar issues" (p. 132). + +## Relation to the argument +Ruling D8 supplies the yardstick question: how to compare courts, FX markets and platforms. Tyler gives a list of criteria that people use, and evidence that voice (representation) and a way to correct errors each add to perceived fairness. The ruling's open question about where conventions bend can use his list as a comparison grid. + +Facet check: the facet says the study "supports the use of the Q2 list as a yardstick for all three settings". The study cuts the other way on one point. Its main finding is that the weights on criteria change with the situation, and its stated conclusion is that no procedure is universally fair (p. 131). The grid needs weights per setting. The facet's list also names "notice", which Tyler did not measure; his items cover representation (process and decision control, p. 111), correctability and the others above. Correctability was measured only as whether the respondent knew of an agency to which they could have complained, and 33% did (p. 113). + +The card's reading, not Tyler's: his strongest predictor, the effort to be fair, is an inference about an official's motive (pp. 128–129), and the respondent had a person to read. A platform's mediator offers no such person, which bears on premise P7 (inferring what the mediator will do). Treat that as a question for the paper, not a result. + +## In conversation with +- `thibaut1975procedural`: Tyler tests Thibaut and Walker's process and decision control on citizens outside the courtroom, and finds representation matters but is one of several concerns (p. 128). +- `lind1988social` (Lind and Tyler 1988): cited with others for the main effect of procedural justice on satisfaction (p. 103). + +Leads from his list, not in the bib: Leventhal (1980), the source of the six criteria; Lind (1982); Tyler and Folger (1980) on citizen–police contacts; Barrett-Howard and Tyler (1986); Sheppard and Lewicki (1987). + +## Caution +Self-reports by Chicago residents in 1984 about police and court contacts, mostly not disputes (the 18% figure applies to cases with a dispute). Footnote 5 (p. 128) notes that the data are correlational, so the causal order cannot be tested. The fairness rating and the criteria come from the same respondent. It says nothing about financial markets, platforms or automated mediators. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the Cambridge Core PDF of Law & Society Review 22(1), 103-135 (33 pp.; journal page = PDF page + 102; title and author match). I checked all 4 quotations and the page locators for the sample (1,575; 63%; 733; 652; 47/31/23%, p. 110), the 69% variance and Table 3 (pp. 121-122), the coefficients (.30, .23, .21, .17, .17, .14, .07, .04), the 48 and 64 comparisons (16 and 5 significant, p. 125), the 18% disputes figure, "focus on similar issues" (p. 132), "no universally fair procedures" (p. 131) and footnote 5's correlational caveat (p. 128). + +**Findings:** Two page errors fixed: the "24 percent against about 1 percent" usefulness analysis is on p. 117, not p. 120; the 33% who knew of an agency to complain to is on p. 113, not p. 115. All other numbers and quotations are correct. Reference list was only skimmed, as the card says. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us2072usc.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us2072usc.md new file mode 100644 index 00000000..012f766f --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us2072usc.md @@ -0,0 +1,30 @@ +# 28 U.S.C. § 2072 — Rules of procedure and evidence; power to prescribe. + +**Identifier:** https://www.law.cornell.edu/uscode/text/28/2072 · **Read depth:** full_text (Cornell LII reproduction of the current Code text, with editorial notes; whole section read) · **Source-tier:** statute · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** other (the Supreme Court as rule-maker for lower federal courts and the litigants who live under the rules) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +The section, three subsections long, gives the rule-writing power over federal court procedure to the Supreme Court and fences it. (a): "The Supreme Court shall have the power to prescribe general rules of practice and procedure and rules of evidence for cases in the United States district courts (including proceedings before magistrate judges thereof) and courts of appeals." (b): "Such rules shall not abridge, enlarge or modify any substantive right. All laws in conflict with such rules shall be of no further force or effect after such rules have taken effect." (c): the rules "may define when a ruling of a district court is final for the purposes of appeal under section 1291." + +The editorial notes show the history: the section was added by Pub. L. 100–702 (19 November 1988), amended by Pub. L. 101–650 (1 December 1990) to add subsection (c), and replaces an earlier § 2072 (1948) that covered only civil procedure and was repealed with effect from 1 December 1988. The notes also record that the Supreme Court's admiralty rules (1920) were rescinded in 1966 when civil and admiralty procedure were unified. + +## Relation to the argument +This is the textual anchor for the courts leg of facet J and the author's ruling D8. Procedure in federal courts is a written convention with a named author (the Supreme Court, under a delegation from Congress), and the statute itself limits what that convention may do: it may govern how rights are enforced but not change the rights. For the paper's question about who holds the convention, it gives a clear model of a holder bound by a higher text, which a platform's terms of service lack (see `suzor2018digital`). Subsection (b) is also the court-side version of the distinction between procedure and substance that the paper needs for the claim that conventions about process can change while the underlying entitlement stays. + +It does not show how the Court actually writes rules; that is § 2073 (`us2073usc`) and the committee process. This section alone leaves the reader with a Court that writes rules by itself. + +## In conversation with +- `us2073usc` — the companion section that prescribes how the rules are proposed and consulted on; § 2072 and § 2073 are cross-referenced in § 2073(a)(2) and (d). +- The editorial notes cite Pub. L. 100–702 and 101–650 and 28 U.S.C. § 1291 and § 2071 (and, for the Virgin Islands, 48 U.S.C. § 1615); none is in the bib. Section 2075 is not mentioned in the text of § 2072; it appears in § 2073. + +## Caution +A statute is not a practice: the text says nothing about how rule changes are decided in fact. The LII page is a convenience copy of the Code; pin to the official Code or U.S. Statutes at Large before print. Editorial notes are LII's/ the Code's compilers' and carry no legal force. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the Cornell LII page for 28 U.S.C. § 2072 (whole section and editorial notes). I checked the three quoted subsections word for word, the history (added by Pub. L. 100-702, 19 Nov 1988; amended by Pub. L. 101-650, 1 Dec 1990; prior 1948 section repealed effective 1 Dec 1988), the 1920 admiralty rules rescinded in 1966, and the cross-references in § 2073. + +**Findings:** One correction: the card said the editorial notes cite 28 U.S.C. § 2075; they do not (§ 2075 appears in § 2073, not here). The notes cite §§ 1291 and 2071 and 48 U.S.C. § 1615. Quotations and dates are right. The card's statement that § 2073(a)(2) and (d) cross-reference § 2072 is right ((e) does too). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us2073usc.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us2073usc.md new file mode 100644 index 00000000..6866eb72 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us2073usc.md @@ -0,0 +1,34 @@ +# 28 U.S.C. § 2073 — Rules of procedure and evidence; method of prescribing. + +**Identifier:** https://www.law.cornell.edu/uscode/text/28/2073 · **Read depth:** full_text (Cornell LII reproduction of the current Code text; subsections (a)–(e) read in full; the long editorial notes on bankruptcy forms read in outline) · **Source-tier:** statute · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** other (the Judicial Conference and its committees as rule-drafters, with the bar and the public as consulted parties) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Section 2073 sets out how the rules that § 2072 lets the Supreme Court prescribe get proposed. (a)(1): "The Judicial Conference shall prescribe and publish the procedures for the consideration of proposed rules under this section." (a)(2): the Conference may appoint committees to recommend rules under §§ 2072 and 2075, and "Each such committee shall consist of members of the bench and the professional bar, and trial and appellate judges." (b): a standing committee on rules of practice, procedure and evidence must review every other committee's recommendation and recommend rules "as may be necessary to maintain consistency and otherwise promote the interest of justice." + +(c)(1) opens committee business to the public: "Each meeting ... shall be open to the public, except when the committee so meeting, in open session and with a majority present, determines that it is in the public interest that all or part of the remainder of the meeting on that day shall be closed to the public, and states the reason for so closing the meeting." Minutes must be kept and made public, with deletions allowed for closed portions. (c)(2): every meeting "shall be preceded by sufficient notice to enable all interested persons to attend." (d): any recommending body "shall provide a proposed rule, an explanatory note on the rule, and a written report explaining the body's action, including any minority or other separate views." + +(e) is the sentence that limits the rest: "Failure to comply with this section does not invalidate a rule prescribed under section 2072 or 2075 of this title." + +The section was added by Pub. L. 100–702 (1988) and amended in 1994 (Pub. L. 103–394) to add references to § 2075 (bankruptcy rules). The notes reproduce three 2005 bankruptcy provisions (Pub. L. 109–8, §§ 419, 433, 435) that direct the Conference to propose forms and rules and to balance the courts' and creditors' need for complete information against "economy and simplicity for debtors" (§ 433). + +## Relation to the argument +This is the best single text for what a procedural convention held by a disinterested body looks like on paper: mixed membership (bench and bar), public meetings, notice, written reasons and recorded dissent. It supports facet J's reading of the courts leg and premise P10 as a contrast case, since a platform changes terms alone and, per `suzor2018digital`, some give no notice at all. + +Facet check: the facet recorded (a)(2), (c)(1) and (d) and said it did not find a statutory comment period. That stays true: there is a notice-of-meeting duty, (c)(2), but no comment period. The facet did not record (e), which matters: the process duties are not enforceable by invalidating a rule. A paper citing § 2073 as a model of constrained rule-making should say so. + +## In conversation with +- `us2072usc` — the grant of power that § 2073 regulates. +- The text points to §§ 2075 (bankruptcy rules) and to Pub. L. 109–8 (2005); neither is in the bib. + +## Caution +Describes the process the law requires, not what the committees do. Section 2074 (congressional review period) is not in this text and was not read here, so this card says nothing about Congress's role. LII copy; pin to the official Code before print. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the Cornell LII page for 28 U.S.C. § 2073 (text and editorial notes). I checked all 8 quotations word for word, the history (added by Pub. L. 100-702, 1988; amended by Pub. L. 103-394, 22 Oct 1994, adding "or 2075"), the three 2005 bankruptcy notes (§§ 419, 433, 435) and the "economy and simplicity for debtors" phrase. + +**Findings:** One small fix: the (a)(1) quotation ended at "proposed rules." but the statute continues "under this section."; restored. Everything else matches. The card correctly says § 2074 was not read. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us455usc.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us455usc.md new file mode 100644 index 00000000..d92722a3 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us455usc.md @@ -0,0 +1,33 @@ +# 28 U.S.C. § 455 — Disqualification of justice, judge, or magistrate judge. + +**Identifier:** https://www.law.cornell.edu/uscode/text/28/455 · **Read depth:** full_text (Cornell LII reproduction of the current Code text with Historical and Revision Notes and amendment notes; subsections (a)–(f) read in full) · **Source-tier:** statute · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** two parties + third party (litigants and a judge who must step aside when interested) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Subsection (a) sets the general standard: "Any justice, judge, or magistrate judge of the United States shall disqualify himself in any proceeding in which his impartiality might reasonably be questioned." (b) lists mandatory grounds: personal bias or prejudice concerning a party, or personal knowledge of disputed facts (b)(1); prior service as a lawyer in the matter, or a former practice partner's service (b)(2); prior government service as counsel, adviser or witness, or an expressed opinion on the merits (b)(3); the judge's, spouse's or minor child's financial interest "in the subject matter in controversy or in a party to the proceeding, or any other interest that could be substantially affected by the outcome" (b)(4); and a relative within the third degree who is a party, officer or trustee of a party, a lawyer in the case, or a likely material witness (b)(5). + +(c) obliges the judge to inform himself of his own and (by reasonable effort) his family's financial interests. (d) defines "proceeding," "fiduciary" and "financial interest," which means "ownership of a legal or equitable interest, however small," with carve-outs for mutual funds the judge does not manage, offices in charitable and civic bodies, mutual-insurance policies and government securities, unless the outcome could substantially affect the value (d)(4). (e) bars waiver of the (b) grounds by the parties; for the (a) ground alone, waiver is allowed only "preceded by a full disclosure on the record of the basis for disqualification." (f) lets a judge keep a case after substantial time was spent if the financial interest, found late and not one the outcome could affect, is divested. + +The notes show the text rests on a 1911 rule (the 1948 revision) and that the 1974 amendment "expanded applicability" and "inserted provisions relating to waiver"; the 1988 amendment added (f). The 1948 revision note says relationship to a party's attorney was included "in conformity with the views of judges cognizant of the grave possibility of undesirable consequences resulting from a less inclusive rule." + +## Relation to the argument +Statutory counterpart to `tumey1927ohio`: where *Tumey* sets the constitutional floor (a direct, substantial pecuniary interest), § 455 sets a much wider statutory duty, using an appearance test ("might reasonably be questioned") and a count that includes "however small" interests. It is the courts-side evidence for premise P10 that the law makes the third party show disinterest. Two features bear directly on the paper: subsection (e) bars the parties from waiving the (b) grounds, so the convention is not at the parties' disposal (a platform's terms are the reverse: the user "agrees"), and (a) allows waiver only after full disclosure on the record, a rule on disclosure that sets the limits of what disclosure can cure — relevant to ruling D3. + +The limit the facet already notes stands: disqualification is by the judge's own act ("shall disqualify himself"), so the safeguard depends on the interested party noticing, and it covers federal judges only. + +## In conversation with +- `tumey1927ohio` — the constitutional floor; the statute does not cite the opinion, and the opinion (1927) predates the current text. +- `us9usc10` — the arbitration analogue: vacatur for "evident partiality" in an arbitrator. +- `caperton2009massey`, `mathews1976eldridge` — bib neighbours in facet J; the statute does not cite them. + +## Caution +Federal judges only; state judges and arbitrators are not covered. A duty written as self-disqualification; the text gives no enforcement mechanism for a judge who does not step aside (the revision notes point to other sections for reassignment, not for challenge). Pin to the official Code before print. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the Cornell LII page for 28 U.S.C. § 455 (text, historical and revision notes, amendment notes). I checked all 14 quotations word for word against subsections (a), (b)(4), (d)(4), (e) and the notes ("expanded applicability", "inserted provisions relating to waiver", the "grave possibility" sentence), the paraphrase of each (b) ground and of (c), (d), (f), the amendment years (1948, 1974, 1978, 1988, 1990), and the claim that the notes point to other sections (291 et seq., 140, 143, 144, 332) for reassignment. + +**Findings:** No issues found. The card's summary of (b)(5) leaves out its clause (iii) (a relative known to have an interest the outcome could substantially affect); that is an omission, not an error. The `tumey1927ohio` and `us9usc10` relations are the project's pairing, as the card says; the statute cites neither. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us9usc10.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us9usc10.md new file mode 100644 index 00000000..a1fd8403 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/us9usc10.md @@ -0,0 +1,29 @@ +# 9 U.S.C. § 10 — Same; vacation; grounds; rehearing (Federal Arbitration Act). + +**Identifier:** https://www.law.cornell.edu/uscode/text/9/10 · **Read depth:** full_text (Cornell LII reproduction of the current Code text with amendment notes; whole section read) · **Source-tier:** statute · **Evidence basis:** direct_read · **Design and sample:** none · **Parties modeled:** two parties + third party (parties to an arbitration and the arbitrators they are bound to) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Section 10(a) lists the grounds ("In any of the following cases"; the text does not itself say they are exclusive, which is case law) on which the federal district court "wherein the award was made may make an order vacating the award upon the application of any party to the arbitration": (1) "where the award was procured by corruption, fraud, or undue means"; (2) "where there was evident partiality or corruption in the arbitrators, or either of them"; (3) arbitrator misconduct, in refusing to postpone a hearing on sufficient cause, refusing to hear pertinent and material evidence, or "any other misbehavior by which the rights of any party have been prejudiced"; and (4) where the arbitrators "exceeded their powers, or so imperfectly executed them that a mutual, final, and definite award upon the subject matter submitted was not made." (b): if the award is vacated and the time to make it has not expired, the court may direct a rehearing. (c): a person who is not a party but "adversely affected or aggrieved" by an award under 5 U.S.C. § 580 (administrative dispute resolution) may seek vacatur if the use of arbitration or the award is "clearly inconsistent" with the factors in 5 U.S.C. § 572. + +The notes trace the text to the 1925 Act (ch. 213, § 10, 43 Stat. 885), with amendments in 1990, 1992 and 2002; the 2002 change moved former (a)(5) into (b) and added the (c) numbering. + +## Relation to the argument +It shows that even a private tribunal chosen by the parties is held to a disinterest rule by the state: an award tainted by "evident partiality" can be set aside, so the legal order polices the neutral that parties bring in themselves. That supports the facet's reading, premise P10 and ruling D8: where a third party decides between two others, law attaches a condition to its position. The same text shows how thin the check is. Vacatur is on application of a party, after the award, on four narrow grounds listed in (a); there is no ground for an unfair term, only for corruption, partiality, misconduct and excess of power. For platforms that require arbitration (Suzor reports six of fourteen sets of terms do; `suzor2018digital`, p. 8), this is the review the user keeps. + +## In conversation with +- `us455usc` — judges' disqualification, the in-court parallel. +- `tumey1927ohio` — constitutional rule for judges; not cited by the statute. +- The text cross-refers to 5 U.S.C. §§ 572 and 580; neither is in the bib. + +## Caution +The statute's wording is the only content: it does not say what "evident partiality" requires (that is case law), and it covers awards under the Federal Arbitration Act only. LII copy; verify against the official Code before print. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** the Cornell LII page for 9 U.S.C. § 10 (text, derivation, amendment notes). I checked all 9 quotations word for word against (a)(1)-(4), (b) and (c), the amendment history (1947 codification, 1990, 1992, 2002 and the redesignations), the "1925 Act, ch. 213, § 10, 43 Stat. 885" derivation line, and the Suzor cross-reference (six of the 14 sets of terms require binding arbitration, `suzor2018digital` p. 8). + +**Findings:** One overstatement fixed: the card said § 10(a) lists "the only grounds"; the statutory text says "In any of the following cases" and does not itself say the list is exclusive (that comes from case law, which the card's Caution already says is outside this source). Quotations and history are right. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/vandenboer2015default.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/vandenboer2015default.md new file mode 100644 index 00000000..2e6337be --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/vandenboer2015default.md @@ -0,0 +1,38 @@ +# van den Boer, M., & Hakvoort, B. E. (2015). Default spacing is the optimal spacing for word reading. *Quarterly Journal of Experimental Psychology, 68*(4), 697–709. + +**Identifier:** doi:10.1080/17470218.2014.964272 · **Read depth:** full_text (published version, UvA-DARE repository copy; journal pagination 697–709, PDF page N = journal page 695 + N; the Table 1 overview of earlier spacing studies was skimmed, not read cell by cell) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** word-naming experiment, Dutch children, 105 in Grade 2 and 92 in Grade 4, each naming 144 words (monosyllabic or bisyllabic) in six interletter-spacing conditions (−0.5, default, +0.5, +1, +1.5, +2 points), 24 words per condition, voice-key timing; a subsample of 29 poor readers (15 in Grade 2, 14 in Grade 4) analysed separately · **Parties modeled:** person–artefact (reader and single word) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Van den Boer and Hakvoort test the idea that more space between letters helps children read, especially weak readers, by crowding reduction. They point out that findings are split: lexical-decision studies show gains for children and adults, while word-naming and sentence studies of typical readers mostly show none, and adults sometimes slow down (Introduction, pp. 697–699). Their own experiment covers a range of spacings on both sides of the default. + +Results (Results, pp. 704–705): +- Accuracy fell only at −0.5: χ²(1) = 12.174, p < .001, with no difference by grade or word length; the other spacings did not differ from default. +- Naming rate was lower at −0.5 (χ²(1) = 6.308, p = .012) and slightly higher at +0.5 (χ²(1) = 6.257, p = .012) than at default, for both grades and both word lengths. +- Combined fluency (items per second times proportion correct): only −0.5 differed from default (F(1,193) = 23.835, p < .001). The +0.5 speed gain did not show up in this combined measure; the authors say the faster rate "goes hand in hand with a small decrease in accuracy" (p. 706). +- The 29 poor readers behaved the same way: fluency lower at −0.5 (F(1,28) = 5.500, p = .026), no gain at any widening. + +The authors read the −0.5 deficit in all groups as a crowding effect, and the absence of gains up to +2 as evidence that "default LS is the optimal LS for children's word reading fluency" (Discussion, p. 706). The abstract states: "We conclude that an increase in interletter spacing has no effect on word naming fluency" (p. 697). They offer reasons the lexical-decision gains may not transfer to naming, and suggest spacing might act at sentence or text level, where crowding from neighbouring words matters more (Discussion, p. 707). + +## Relation to the argument +It concerns letter spacing in single words. It enters facet D as the null-or-negative result against "more space, easier reading", and it supplies the useful asymmetry: tightening hurts all readers alike, loosening within a modest range helps none. That bears on ruling D6 and on premise P6 only by analogy: a spacing intervention can lower a cost only where the default is already tight enough to cost something, and the groups least able to read did not gain more. Nothing here concerns word boundaries, so it gives no support to P1 to P5. Under ruling D10 it is a reason not to cite "spacing helps struggling readers" as a general fact without naming the unit (letter, word, line). + +## In conversation with +- `perea2012effects` — Perea, Panadero, Moret-Tatay & Gómez (2012), the lexical-decision gains for skilled adults, young readers and dyslexic readers that this study fails to extend to naming (Introduction). +- `perea2012subtle` — cited among sentence-reading studies where typical adults show no gain (Introduction). +- `zorzi2012extra` — the extra-large-spacing finding for dyslexia (Introduction and Discussion). +- `cohen2008reading` — fMRI evidence that large spacing shifts word reading to a serial dorsal route at +2.25 and +3 (Introduction). +- `hakvoort2017improvements` — the same authors' later sentence-level study, which finds an accuracy gain; the two papers agree that speed does not change. +- Leads not in the bib: Risko, Lanthier & Besner (2011) and Vinckier et al. (2011), who report slowing from wide spacing in adults; Tai et al. (2006), fewer but longer fixations; Legge & Bigelow (2011) on the range of print sizes; Spinelli et al. (2002) on crowding in dyslexic children. + +## Caution +Single-word naming, not sentences or text. Grade 2 and Grade 4 children, Dutch, one proportional font. Six pre-set spacings in one font (14-point Times New Roman); the passages I read do not state the unit of the spacing steps, so the results do not map onto em fractions. The poor-reader group is small (N = 29, spread across two grades and two word lengths). The authors note that effects of spacing may differ at sentence level and that crowding was never measured directly (Discussion). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** the UvA-DARE final published PDF (14 pp.; title, authors, QJEP 68(4) 697-709 match; page = PDF page + 695). I checked the 5 quotations and their pages (697, 706), the statistics (chi-square 12.174, 6.308, 6.257; F(1,193) = 23.835; F(1,28) = 5.500, p = .026), the design numbers (105 and 92 children; 144 monosyllabic plus 144 bisyllabic words in six lists of 24; six spacings; 14-point Times New Roman; 29 poor readers, 15 and 14), the Discussion points on sentence or text level (p. 707), and the Cohen et al. 2008 detail (+2.25 and +3). + +**Findings:** No issues found. `hakvoort2017improvements` post-dates the source, as the card says. The card's note that the unit of the spacing steps is not stated in the passages read is a fair limit; I did not find one either. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wagemans2012century.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wagemans2012century.md new file mode 100644 index 00000000..583c647f --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wagemans2012century.md @@ -0,0 +1,38 @@ +# Wagemans, J., Elder, J. H., Kubovy, M., Palmer, S. E., Peterson, M. A., Singh, M., & von der Heydt, R. (2012). A century of Gestalt psychology in visual perception: I. Perceptual grouping and figure–ground organization. *Psychological Bulletin, 138*(6), 1172–1217. + +**Identifier:** doi:10.1037/a0029333; PMC3482144 (NIHMS390005) · **Read depth:** partial (changed from full_text at S2: the card says §§4.2.2-4.3, 5.2-5.5 and 6.2-6.4 were read at paragraph level only and the tables and reference list were not read) of the **PMC author manuscript** (JATS XML), not the *Psychological Bulletin* print. The XML has no journal page numbers, so locators are the manuscript's section numbers. I read sections 1–3, 4.1, 4.2.1, 4.4, 5.1, 5.6, 6.1, 6.5 and 7 in full. I read 4.2.2–4.3, 5.2–5.5 and 6.2–6.4 at paragraph level (opening and closing sentences, with the passages on past experience, attention and border ownership read through). I did not read Tables 1–6 or the 312-item reference list beyond matching it to the bib · **Source-tier:** peer-reviewed review article (accepted manuscript) · **Evidence basis:** direct_read · **Design and sample:** narrative review of a century of work on perceptual grouping and figure–ground organization; no search protocol, no effect-size synthesis · **Parties modeled:** none (a viewer and a display) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues + +Seven vision scientists take stock of Gestalt psychology on the centenary of Wertheimer's 1912 paper. They say the field recovered from its mid-century fall by turning demonstrations into measured laws, tying them to natural-image statistics and finding neural correlates, while "modern vision science cannot be considered some kind of contemporary Gestalt approach either" (§7.2). + +The part that matters here is grouping by proximity (§3). Wertheimer "first posed the problem of perceptual grouping in his ground-breaking 1923 paper by asking what stimulus factors influence the perceived grouping of discrete elements" (§3.1). Equally spaced dots form one line; when some sit closer than others, "the closer ones grouped together strongly into pairs". That "factor of relative distance, which Wertheimer called proximity, was the first of his famous laws or (more accurately) principles of grouping" (§3.1). + +Later work made it a law. Oyama (1961) found the ratio of times observers saw horizontal and vertical grouping follows a power of the distance ratio, with exponent α ≈ 2.89 (§3.3.3). Kubovy and colleagues, with 300-ms dot-lattice displays, showed that grouping by proximity "depends only on the relative distance between dots in competing organizations, not on the overall configuration" (§3.3.3); once the longer distance exceeds the shorter by about 1.5, grouping along it "is almost never seen." Proximity and similarity combine additively (§3.3.3). In labelled natural contours the power-law exponent is 2.92, close to Oyama's perceptual value, and proximity cuts the entropy of a grouping decision by roughly 75%, against about 10% each for good continuation and similarity (§§4.2.1, 4.4.2). The authors read this as a visual system tuned to the statistics of its scenes. + +They also complicate the classical picture: "grouping principles operate at multiple levels," and the conscious result depends on a representation built after depth and lightness constancy (§3.5), so grouping is not simply early and automatic (§7.2). Infants as young as 3 to 4 months group by proximity, yet contour grouping keeps developing into the late teens and experience shapes adult grouping (§4.4.1). Direct neural correlates "are relatively rare" (§7.3). + +## Relation to the argument + +Facet D, item 8: the facet takes this review as the source for treating word spaces as proximity grouping. The review supports the mechanism's standing: proximity is the oldest and best-measured grouping law, it depends on relative distance, and it carries more information about natural contours than good continuation or similarity. It does not discuss words, text or reading; letters appear only as stimuli in a few studies it cites (§§2.4, 5.4.1, 6.3). The step from dot lattices to a line of type, where gaps between words must exceed gaps between letters, is the paper's inference (premise P5: "applying these to a page is the paper's move"). Claude's reading: the review supplies no number for that step, since the 1.5 ratio is a dot-lattice threshold at 300 ms and cannot move to type. Two further points cut against loose use. Grouping is not strictly preattentive and early (§§3.5, 7.2), and experience matters (§4.4.1). The second is compatible with premise P4 (readers fluent in unspaced scripts gain little from added spaces), but the review does not address reading. + +Facet check: the facet's quotation matches §3.1 and its "read in part" status is superseded. Its summary that word spaces "work by proximity grouping, the oldest grouping principle" is the facet agent's inference, not this review's. + +## In conversation with + +- Nothing in `references.bib` is cited by this review (I matched its references against the bib). +- Not in the bib, as leads: Wertheimer (1923), *Psychologische Forschung* 4, 301–350 (the source of the proximity law); Kubovy, Holcombe and Wagemans (1998), "On the lawfulness of grouping by proximity," *Cognitive Psychology*; Elder and Goldberg (2002), *J. Vision* 2 (the 75% entropy figure); Palmer and Rock (1994); and Part II of the same review (Wagemans, Feldman, Gepshtein, Kimchi, Pomerantz, van der Helm & van Leeuwen), listed as in press here. + +## Caution + +Dot lattices and contour fragments, not text; no semantics and no reading task. The manuscript may differ from the print. Exponents and the 75% figure are other authors' results as this review reports them; I did not check the originals. The authors work in this tradition. The theory half of the century is in Part II, which I did not read. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the PMC author-manuscript XML (about 38,800 words of text; 312 reference entries). Searched all 7 quotations from the review (the other 3 flagged strings are facet or premise wording) and placed them by section heading (§3.1, §3.3.3, §3.5, §7.3). Checked the numbers: Oyama alpha = 2.89 (§3.3.3), natural-contour exponent 2.92 (§4.2.1), proximity cuts entropy by roughly 75% against roughly 10% each for good continuation and similarity (§4.4.2), the 1.5 ratio and 300 ms displays (§3.3.3), infants 3 to 4 months and development into the late teens (§4.4.1). + +**Findings:** Three fixes. (1) The card said the review finds the original theory 'has not been integrated into mainstream vision science (§7.2)'; §7.2 actually says 'modern vision science cannot be considered some kind of contemporary Gestalt approach either' (and credits Biederman with integrating Gestalt principles into mainstream cognitive science); reworded to the source. (2) 'Letters appear only as stimuli in two experiments (§§2.4, 5.4.1)': letters also appear in §6.3 (Bregman's blotch figure), and §2.4 names Navon's hierarchical letters; reworded to 'a few studies it cites (§§2.4, 5.4.1, 6.3)'. (3) Read depth changed from full_text to partial because the card says several sections were read at paragraph level only and tables and references were not read. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wang2024tibetan.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wang2024tibetan.md new file mode 100644 index 00000000..879462fd --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wang2024tibetan.md @@ -0,0 +1,44 @@ +# Wang, D., Niu, D., Li, et al. (2024). The effect of visual word segmentation cues in Tibetan reading. *Brain Sciences, 14*(10), 964. + +**Identifier:** doi:10.3390/brainsci14100964 · **Read depth:** full_text (open-access MDPI article, PMC11505889, Europe PMC XML; no printed page numbers, so locators are section numbers; Supplementary File S1 not read; the full author list is not given here because the XML I read lists only initials in the contribution statement: D.W., X.G., T.L., D.N.) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** two eye-tracking experiments (EyeLink 1000 Plus, 1,000 Hz), single-factor within-subject, 72 Tibetan university students in each (Experiment 1: 36 men, 36 women, mean age 20.43; Experiment 2: 36 and 36, mean age 20.39; no overlap); 80 experimental sentences, one block per participant, 20 per condition; silent reading with yes/no comprehension questions (accuracy 90.15% and 89.85%) · **Parties modeled:** person–artefact (reader and text) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Tibetan has no word-segmentation cue beyond the tsheg, a mark between morphemes. Wang and colleagues test whether adding word cues helps, and whether the word or the morpheme is the unit of processing (Introduction). Experiment 1 uses four conditions: normal text, word spacing, morpheme spacing, and non-word spacing (spaces around units larger than a morpheme but not a word). Experiment 2 replaces spaces with alternating red and green colour on the same units, so the sentence keeps its length (Section 3.1.2). + +Results: +- Experiment 1, global (Table 1): word spacing shortened sentence reading time, 5,566 ms against 6,019 ms for normal text (b = 0.083, t = 5.145, p < .001), with fewer fixations (15.50 against 16.51), shorter mean fixation duration (232 against 242 ms) and larger saccades (2.04 against 1.80 letters). Morpheme spacing (6,778 ms) and non-word spacing (7,552 ms) were slower than normal text. +- Experiment 1, local (Table 3; one two-morpheme word per sentence): gaze duration 356 ms with word spacing against 407 ms without; total fixation duration 466 against 544 ms; refixation probability down; the first fixation landed closer to the word centre (1.69 against 1.43 letters from the start). First fixation duration did not change (247 against 248 ms). +- Experiment 2: colour on word boundaries gave 5,737 ms against 6,016 ms for normal text (t = 2.873, p < .01); gaze duration fell to 382 from 394 ms; initial landing position did not change (1.49 against 1.48). Colour on non-word units slowed reading (6,774 ms). +- Conclusion (section 5): "Artificially incorporated interword spaces facilitate Tibetan reading, lexical recognition, and saccade target selection." Words are "more likely" than morphemes to be the basic unit. + +The authors had predicted no effect on global reading (Introduction, hypotheses 1 and 2) and say the result "differed from our experimental hypothesis" (Section 4.1). They offer two explanations: Tibetan university students may be unfamiliar with Tibetan text because of first-language attrition among Tibetan–Chinese bilinguals, so the cost of an unfamiliar format is small; and Tibetan readers already rely on an explicit cue, the tsheg (Section 4.1). They explain the landing-position effect through E-Z Reader and the Gestalt proximity principle, citing Koffka (Section 4.1). The Introduction restates Winskel et al. as finding that spaces helped lexical recognition but hurt reading speed in Thai. + +## Relation to the argument +The paper bears on P4 (readers fluent in unspaced scripts gain nothing from added spaces; learners gain most, on mixed evidence). It is a case where adult readers of an unspaced script gained from word spacing on global measures, which cuts against the "gain nothing" half of P4. The authors tie the gain to readers who are not highly practised in the script and who already depend on a boundary mark, which fits the second half of P4 ("learners gain most"). The Tibetan sample is university students, not children and not second-language learners of the script, so it is not a learner study in the narrow sense. It also bears on D6: word spacing was the cue that did the work in Experiment 1, but a colour cue did nearly as well on speed while not shifting landing positions. That separates two jobs spaces do, as the Thai study had (Winskel et al.). + +Facet check: the facet lists this paper in D without a note on its findings. The reading confirms spacing helps; it does not support "no gain for fluent readers". + +## In conversation with +- `winskel2009eye` — ref. 16; the Thai result they contrast with (spaces helped lexical recognition and, per the authors, hurt speed). +- `sainio2007role` — ref. 17; Japanese, no effect. +- `bai2008reading` — ref. 18; Chinese, spaces neither helped nor hurt and gave the trade-off hypothesis. +- `zang2013interword` — ref. 20; Chinese children and adults, spaces aid landing position. +- `rayner1998unspaced` — ref. 4. +- `kasisopa2013eye` — ref. 78; Thai, cited for statistical segmentation cues. +- `li2021effects` — ref. 48 (Li et al., 2022 in the text), word spacing in children's reading. +- `shen2012eye` — ref. 49; second-language learners reading spaced Chinese. +- `perea2009space` — ref. 12; `pollatsek1982eye` — ref. 13; `sheridan2013unsegmented` — ref. 15; `epelboim1994reading` — ref. 8; `rayner2009thirtyfifth` — ref. 6. +- Leads not in the bib: Wang et al. (2023), *Scientific Reports* 13, 314, on tsheg removal; Zhou et al. (2018) on alternating colours in Chinese; Perea, Tejero & Winskel (2015) on colours for segmentation; Ma et al. (2019); Li et al. (2022), *Brain Sciences* 12, 1205. + +## Caution +Participants are university students in Tibet whose first language is under pressure from Mandarin; the authors' attrition explanation is speculation and they did not measure it. Single-sentence silent reading, sentences about 28 letter-spaces long, one font (Microsoft Himalaya 36). The local analysis covers one two-morpheme word per sentence. The authors' own account of earlier results is compressed: the Introduction says Thai spacing "has a detrimental effect on reading speed", while Winskel et al. report that finding as marginal (p = .05 by participants, not reliable in the combined test). The paper does not report children or second-language learners of Tibetan. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the Europe PMC XML (PMC11505889), stripped to text. Checked the 3 verbatim quotations and about 40 numbers in Tables 1, 3, 5 and 6 (reading times, fixation durations, saccade amplitudes, gaze and total durations, landing positions, ages, comprehension accuracy), the hypotheses, the two Section 4.1 explanations, the Thai/Winskel reference (ref. 16) and the design details (80 sentences, Himalaya 36, mean length 27.83). + +**Findings:** One quotation error fixed: the card's 'detrimental ... to reading speed' is not the source wording; the source says Thai spacing 'has a detrimental effect on reading speed'. All numbers matched. Tibetan article's own account of Winskel (spacing hurt reading speed) is as stated; whether that matches Winskel et al. is checked in the winskel2009eye card. Section labels follow the XML headings; the XML has no page numbers. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wang2026monitorbench.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wang2026monitorbench.md new file mode 100644 index 00000000..3f228252 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wang2026monitorbench.md @@ -0,0 +1,46 @@ +# Wang, H., Sun, Y., Ko, B., Talati, M., Gong, J., Li, Z., Yu, N., Yu, X., Shen, W., Jolly, V., & Zhang, H. (2026). MonitorBench: A comprehensive benchmark for chain-of-thought monitorability in large language models. *Conference on Language Modeling (COLM 2026)*. + +**Identifier:** arXiv:2603.28590 · **Read depth:** partial (the main text and Appendices A.1–A.3, the start of A.4 through A.4.3, and A.5.1–A.5.2 and Table 4 were read in full; arXiv v4, 28 Sep 2026, headed "Published as a conference paper at COLM 2026"; page numbers are that PDF's, main text pp. 1–10, references pp. 11–14). I skimmed the remaining result tables and figures (A.5.3, pp. 24–29), did not read the monitor and stress-test prompt listings (A.5.4–A.5.5, pp. 30–43), and did not read the task examples (A.6, pp. 44–67); about 38 of the 68 PDF pages, all appendix, were therefore not read) · **Source-tier:** proceedings (accepted at COLM 2026; the bib entry is typed `@misc`) · **Evidence basis:** direct_read · **Design and sample:** benchmark of 1,514 instances in 19 tasks and 7 categories; 8 language models (Qwen3-14B, Qwen3-235B-A22B, GPT-OSS-120B, GPT-5-mini, GPT-5, Gemini-3-Flash, Claude Haiku 4.5, Claude Sonnet 4.5); one automated monitor (Qwen3-32B); a 300-trajectory human check of the monitor; no human subjects in the main experiments · **Parties modeled:** other (a model and an automated monitor model; no person reads the chain of thought) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +Wang and colleagues build a benchmark for "monitorability": whether a model's chain of thought (CoT) reveals the factor that actually drives its behaviour. Each task plants a "decision-critical factor", such as an answer hint, a hidden second objective, or a goal to sandbag. A monitor model, told exactly which factor to look for, flags the traces that disclose it. The main claim is in the abstract: "CoT monitorability is a conditional property affected by the evaluated LLM, monitor LLM, and task characteristics" (p. 1). + +Three findings carry the paper. First, monitorability is high when the factor shapes the reasoning steps and low when it only changes the final answer. Under CoT-only monitoring the Cue (Condition) tasks average 91.6 to 98.8 on the paper's 0–100 score, while the Cue (Answer) tasks average 48.9 to 88.1 (Table 2, p. 8). The multi-step-update task splits the models: 83.6 to 87.6 for the three open models and 1.1 to 17.9 for the five closed ones (Table 2). The all-task CoT-only mean is 81.4 (Table 2). Second, more capable models score lower: across the three monitor scopes the Pearson r between capability and monitorability runs from −0.87 to −0.39 (p. 9; eight points each). Third, watching the CoT beats watching only the actions, by about 24 points on average (p. 9). Under two stress prompts that tell the model to hide the factor or evade a monitor, the average score falls 7.8 and 7.2 points; Cue (Answer) falls 23 to 32 points, and the paper's inspection finds omission, not obfuscation (Table 3, pp. 9–10). + +The limits are stated. The tasks are "objective reasoning", so "monitorability in subjective or creative reasoning domains remains largely unexplored", and models are treated as black boxes (A.1, p. 16). For the closed models "we could only access the summarized CoT provided by the proprietary company" (p. 6). The monitor "is given exact decision-critical factors to look for", so scores reflect the evaluated model, not a monitor's skill (p. 8). Against 300 human-labelled trajectories (three expert annotators), the monitor reaches F1 of 85.31 (action-only), 91.86 (CoT-only) and 92.13 (all messages); human and model labels agree only weakly on multi-turn agent runs (A.5.2, p. 23). + +## Relation to the argument + +**Premises P11 and P12; ruling D4.** The Nadella post (P9) calls CoT transparency "not sufficient or dependable" and puts its load-bearing controls outside the model. MonitorBench supports the "conditional, not dependable" half with a large grid: a CoT discloses a factor reliably only where the task forces the factor into the steps. It cuts against the post only in a narrow way: for hidden second objectives and impossible tasks it finds high monitorability (83.8 to 95.2 average, Table 2), so a monitor that is told what to look for does well there. + +**Premise P12 (people monitor sustained streams badly).** This source does not test people. Its monitor is a model with the answer key, and its one human step is a calibration of that monitor. Cite it for what a model's trace discloses, not for what an overseer notices. The gap between "the trace holds the factor" and "a person reads it and acts" is the gap the paper leaves open. + +**Ruling D3.** The study concerns what a model verbalises. It says nothing about whether disclosure of an algorithm helps a person navigate it. + +**Facet check:** the facet lists MonitorBench among "2026 preprints" read as abstracts. The full text carries a COLM 2026 header (v4), so "preprint, not peer reviewed" understates its status. I did not verify the acceptance beyond that header. + +## In conversation with + +- `korbak2025chain` — cited in the introduction as the case for using CoT to oversee models (p. 1). +- `baker2025monitoring` — cited for the term "reduced CoT monitorability" and the obfuscation risk (p. 2). +- `lanham2023measuring` — cited among the faithfulness studies in Related Work (p. 3). +- `chen2025reasoning` — cited as Chen et al. (2025b) for models that hide the true factor behind their answers (p. 1). +- `kutasov2025shade` — the agent tasks (Agent Malicious and Benign) are built on SHADE-Arena (A.4.2, pp. 19–21); Wang also cites it on the weak agreement of human and model labels (A.5.2). +- `greenblatt2024ai` — this key is "AI Control" (ICML 2024); Wang's list holds a different Greenblatt paper, "Alignment faking in large language models" (arXiv:2412.14093), so do not equate the two. +- It does not cite Turpin et al. (2023) in its reference list; I searched it. +- Not in the bib, as leads: Guan et al. (2026), "Monitoring monitorability" (ICML 2026), the framework this paper extends; Emmons et al. (2025a), evasion of CoT monitors; Arnav et al. (2025), "CoT Red-Handed". + +## Caution + +All scores come from one monitor model that knows the answer; capability is proxied by pass rate on eight models; the capability correlation rests on eight points. The closed models' traces are provider summaries. The stress tests are prompts, not training. Tasks are objective, so nothing here covers the open-ended coordination the paper discusses. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the arXiv v4 PDF (68 pp.) and extracted it twice (layout and raw). Checked 7 verbatim quotations (all exact), Table 2 ranges (91.6-98.8, 48.9-88.1, 83.6-87.6, 1.1-17.9, 81.4, 83.8-95.2), Table 3 and the p. 9-10 text (Pearson r -0.87 to -0.39, 24 point CoT-vs-action gap, omission rather than obfuscation), the human-check F1 values (85.31, 91.86, 92.13; three expert annotators), 1,514 instances / 19 tasks / 7 categories, the model list, and the Chen, Korbak, Baker, Lanham, Kutasov and Greenblatt relations plus the absence of Turpin. + +**Findings:** One change: the read-depth line said full_text while stating that the appendix prompt listings (A.5.4-A.5.5) and task examples (A.6) were not read, about 38 of 68 pages; changed to partial and kept the list of what was not read. No quotation, number or locator errors found. Not verifiable from text: the exact order of the 7.8 and 7.2 figures between the Direct Concealment and Monitor-aware Evasion columns (the card does not assign them). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wertli2023influence.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wertli2023influence.md new file mode 100644 index 00000000..be93ef31 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wertli2023influence.md @@ -0,0 +1,38 @@ +# Wertli, J., Schötzau, A., & Palmowski-Wolfe, A. (2023). The influence of age on eye movements during reading in early elementary school children. *Klinische Monatsblätter für Augenheilkunde, 240*(4), 591–598. + +**Identifier:** doi:10.1055/a-2045-7271 · **Read depth:** full_text (open-access manuscript text from PMC10129412 via Europe PMC XML; no page numbers, so locators are section names; the author list and volume come from the PMC record's metadata) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** cross-sectional eye-tracking of reading aloud, 127 recruited children aged 7 to 12 from grades 2 to 5 in four primary schools in German-speaking Switzerland; 118 analysed (58 boys, 60 girls; mean age 9.8, SD 1.24; 63 native German speakers); two IReST texts of about 137 words, SMI RED 250 tracker at 250 Hz · **Parties modeled:** person–artefact (child reader and text) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Wertli and colleagues build normative eye-movement data for early readers. They ask how age, school grade, gender and mother tongue affect reading speed, saccades, fixations and errors when children read two standardized IReST texts aloud (Abstract; Methods). + +Results (Results; Table 2): +- Median reading speed by grade rose from 50.1 words per minute in grade 2 to 74.06 (grade 3), 93.41 (grade 4) and 102.2 (grade 5). Median saccades per word fell from 2.52 (grade 2) to 1.05, 1.16 and 1.19; median fixations per word from 2.21 to 1.29, 1.44 and 1.36. Median reading errors per word fell from 0.05 to 0.04, 0.03 and 0.03. +- By age: 57.1 (SD 17.27) words per minute at age 7, 72.11 (SD 24.62) at 9, 92 (SD 26.24) at 10, 102 (SD 25.41) at 11. Parameters did not change reliably from age 9 and grade 3 on. +- Mother tongue: native German speakers in grade 2 needed fewer saccades per word (p = .008) and fixations per word (p = .03) and made fewer errors; the authors report no significant difference by grade 3 for the eye-movement measures, and a speed gap in grades 2 and 3 that had closed by grade 5 (the overall effect of mother tongue on speed was not statistically significant). Gender had no reliable effect. School made no difference. +- The Discussion sets the fifth-grade median against young German adults at 179 words per minute on an IReST text, which makes the children "1.75 times slower." + +In the Introduction the authors state, from secondary sources, that "The saccades are typically forward movements of 7 – 9 characters," and that competent adult English readers have a perceptual span of 3 to 4 letter spaces to the left and 14 to 15 to the right of fixation, smaller in beginners and in dyslexic readers. + +## Relation to the argument +The paper has no text on word spacing, and no premise or ruling in the argument map rests on its data. The project uses it as an open source for the typical saccade length. That figure appears in the Introduction as a statement the authors cite to their ref. 2, Rayner (1985), "The role of eye movements in learning to read and reading disability" (not to Rayner 1986 or 1998), and I read the card as confirming only that Wertli et al. say it, as the facet already records ("Secondhand"). It also gives a mildly useful context for D2 and D6: children in the first grades need about two fixations per word, adults far fewer, so a convention that lowers the cost of locating words would matter most in the early years of learning. That inference is the project's, not the authors'. + +Facet check: the facet row calling the "7 – 9 characters" statement secondhand is correct. The perceptual-span figures in the Introduction are likewise taken from other work. + +## In conversation with +- `rayner1998eye` — the 20-year review, ref. 5; cited (with ref. 8) in the Introduction for the assessment of eye movements in competent adult readers. The "7 – 9 characters" statement is cited to ref. 2 (Rayner 1985, not in the bib), not to this review. +- `rayner1986eye` — ref. 11, perceptual span in beginning and skilled readers. +- `rayner2009thirtyfifth` — ref. 8, cited together with the 1998 review on eye movements. +- Leads not in the bib: Trauzettel-Klosinski et al. (2010) on German children with and without dyslexia; IReST Study Group (2012); Seassau & Bucci (2013); Häikiö, Bertram & Hyönä (2009). + +## Caution +Reading aloud in children aged 7 to 12 with the same grade-6 text for all, with few 7- and 12-year-olds (six each); the authors note aloud reading adds roughly 50 ms to fixations relative to silent reading. A study of normative data, not of spacing. The perceptual-span and saccade figures are restated from other work and should be cited to the original (Rayner 1998; McConkie & Rayner 1976, which the paper cites as ref. 14 with ref. 15 for the perceptual span). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the Europe PMC XML (PMC10129412) and read it as plain text. Checked the 3 verbatim quotations (exact), the Table 2 medians for four grades, the age means and SDs, N = 127 recruited and 118 analysed, mean age 9.8 (SD 1.24), 63 native speakers, the four schools, the p-values for mother tongue (.008, .03), 179 words per minute and 1.75 times slower, the perceptual-span figures and the reference numbers (2, 5, 8, 11, 14, 15). + +**Findings:** Corrected the citation chain for the '7 – 9 characters' sentence: the paper cites it to its ref. 2, Rayner (1985), not to Rayner (1986) as the card said, and not to the 1998 review (ref. 5), which the In conversation entry had named as its source; both entries fixed. Also made the mother-tongue sentence match the text: no significant difference by grade 3 for eye movements (Conclusion box); speed gap in grades 2-3 closed by grade 5, but the overall mother-tongue effect on speed was not significant. All numbers matched. The Caution line's citation to McConkie and Rayner 1976 is correct (paper ref. 14); annotated. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/westwell2018dissemination.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/westwell2018dissemination.md new file mode 100644 index 00000000..2136de1e --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/westwell2018dissemination.md @@ -0,0 +1,44 @@ +# Westwell, A. (2018). *The dissemination and reception of the Ordines Romani in the Carolingian church, c. 750–900* [Doctoral dissertation, University of Cambridge]. https://doi.org/10.17863/cam.23255 + +**Identifier:** doi:10.17863/CAM.23255 (DataCite; Cambridge Apollo handle 1810/275972) · **Read depth:** partial: the open Apollo PDF (319 PDF pages; title page dated 09/2017; DataCite records 2017-09-29 and 2018 issue dates, publication year 2018; printed page = PDF page − 12 in the body). Read: thesis summary and table of contents; the Introduction in full (printed pp. 1–12); the conclusion to ch. 4, "Amalarius, the Expositiones and the Carolingian Understanding of 'Correctio'" (pp. 254–256); the Conclusion in full (pp. 281–290); keyword searches of the whole text for "uniform", "Admonitio", "Alcuin", "correctio". Not read: chs. 1–3, 5 and most of ch. 4, the manuscript studies that make up the body · **Source-tier:** doctoral thesis · **Evidence basis:** direct_read (partial) · **Design and sample:** manuscript study of some sixteen ninth-century manuscripts carrying *ordines romani* (p. 284) · **Parties modeled:** other (rulers, bishops, monastic communities and liturgists in an exchange network) · **Facets:** O, T · **Read:** 2026-10-14 + +## What it argues + +Westwell studies the *ordines romani*, Frankish texts describing Roman liturgical ceremonies, as whole manuscripts rather than as witnesses to an original text. His thesis is that their reception was local and creative. The summary puts it directly: "what occurred under the Carolingians in the liturgy did not involve the imposition of the Roman rite from above. What was 'Roman' and 'correct' was decided by individuals, each in their own case" (thesis summary, front matter). + +The Introduction sets this against Michel Andrieu, the texts' editor, for whom Carolingian reform "was a uniform endeavour with a single goal" imposed from the centre (p. 4). Westwell follows McKitterick, who found interventions to be "local initiatives" that "differed in nature from one place to another", and Hen, who found no evidence of "centralised imposition" of particular liturgical texts (pp. 4–5). He keeps the term *correctio* but questions it: it "suggest[s] a teleological movement" or "dictates handed uncomplicatedly down from above, from Emperor or Pope" (p. 255), whereas the manuscripts show "constant dialogue and adaptation" (p. 255). + +The Conclusion generalises. "Closeness to the centre of power, even to the shaping of the very decrees and edicts traditionally seen as demanding uniformity, has not made that supposed uniformity any more real" (p. 288). Even the most Roman texts "were never received as the final word, the trumpet call to uniformity" (p. 289). The network of bishops and monasteries, "highly linked, constantly communicating, and revolving around the imperial centre, never sought a single answer" (pp. 289–290). The *ordines* belonged to a "'higher' liturgical discussion" of "a court-favoured few" (p. 287), not to local priests' handbooks. + +## Relation to the argument + +**Supports the caution the draft draws.** The draft cites Westwell for one point: that "uniformity imposed from above can be overstated". The thesis makes exactly that argument for liturgy (summary; pp. 4–5, 255, 288–289). + +**Cuts against the sentence that follows.** The draft concludes the paragraph: "What they show is who stood behind the page: an authority with standing over the scribe." Westwell's picture complicates that. In the liturgical manuscripts he studied, what counted as correct was settled by bishops, liturgists and communities in a network, and the centre's decrees did not make it uniform (pp. 255, 288). That is still an institution with standing over its scribes, so the draft's claim survives, but the authority was distributed and negotiated, not a single ruler's order. Facet O reached the same reading. + +**Out of scope.** Westwell is about liturgical texts, not script, layout or word separation; I found no discussion of word division in the parts read or by keyword search. He cannot be cited for any claim about the page's conventions, only about the limits of Carolingian imposition in one domain. + +**Use in DRAFT_v2:** +- **Line 123** ("Westwell (2018) cautions that uniformity imposed from above can be overstated"). Supported: thesis summary; pp. 4–5; p. 288 ("has not made that supposed uniformity any more real"). The caution concerns the liturgy, and the draft's next sentences correctly say none of the texts concerns word separation. The line's closing claim, "an authority with standing over the scribe", is the draft's and is in tension with Westwell's distributed picture (pp. 255, 289–290); a reader could cite him against it. + +## In conversation with + +- `mckitterick1989carolingians`: cited in ch. 1, n. 58 (*The Carolingians and the Written Word*, p. 143), found by keyword search; Westwell's main McKitterick source is *The Frankish Church and the Carolingian Reforms* (1977), which is not in the bib. +- `parkes2016carolingian`: a different Parkes; Westwell relies on Henry Parkes, *The Making of Liturgy in the Ottonian Church* (2015), not in the bib. +- Leads not in the bib: Hen, *The Royal Patronage of Liturgy in Frankish Gaul* (2001) (pp. 4–5); McKitterick, *The Frankish Church and the Carolingian Reforms 789–895* (1977) (p. 4); Andrieu, *Les Ordines Romani du haut Moyen Âge* (1931–1961) (p. 2); Keefe, *Water and the Word* (2002) (pp. 9–10); Mostert on errors in early copies of the *Admonitio generalis* (2016) (p. 11, n. 44). + +## Caution + +A doctoral thesis, not a peer-reviewed publication, read in part. The body chapters, where the manuscript evidence lies, were not read; the claims above come from the summary, introduction, one chapter conclusion and the conclusion. Its subject is liturgy; extending its finding on uniformity to script or book production is the draft's inference. The year: the title page says September 2017 and DataCite gives 2018 as the publication year, so "2018" in the draft is defensible. + +## S2 adversarial verification (2026-10-14) + +**Verdict:** confirmed + +**What I checked:** Re-opened the writer's saved Apollo PDF (319 pages), extracted it with page breaks and confirmed the offset printed page = PDF page − 12 (PDF p. 17 carries printed "5"). Located every quotation by printed page and read each in context, including the run-over of the Hen sentence onto p. 5 and of "suggest a teleological movement" on p. 255. Searched the whole text for "word division" and "word separation" (no hits). Checked the record against DataCite (Westwell, Arthur Robert; publication year 2018; issued dates 2017-09-29, 2018-05-21, 2018-07-20) and the title page ("Queens' College 09/2017"). Re-read DRAFT_v2.md line 123 and the reference at line 703. + +**Findings:** +1. All quotations verified verbatim with the stated locators: the thesis-summary sentences (front matter); Andrieu's "uniform endeavour with a single goal", McKitterick's "local initiatives ... differed in nature from one place to another", Hen's "centralised imposition of particular liturgical texts" (pp. 4–5); "suggest[s] a teleological movement", "handed uncomplicatedly down from above, from Emperor or Pope", "constant dialogue and adaptation" (p. 255); "court-favoured few" and "'higher' liturgical discussion" (p. 287); "has not made that supposed uniformity any more real" (p. 288); "trumpet call to uniformity" (p. 289); "highly linked ... never sought a single answer" (p. 290, with the sentence beginning on p. 289); "some sixteen manuscripts" (p. 284). +2. In-conversation links verified: McKitterick (1989), *The Carolingians and the Written Word*, p. 143, at ch. 1 n. 58; Henry Parkes, Hen (2001), Keefe (2002) and Mostert appear as stated. +3. Read depth (partial, itemised) is honest. +4. Use in DRAFT_v2 (line 123): verdict stands. Westwell supports the caution that uniformity from above can be overstated, for the liturgy; the card rightly flags that his distributed, negotiated picture sits uneasily with the line's closing "an authority with standing over the scribe". diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/winskel2009eye.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/winskel2009eye.md new file mode 100644 index 00000000..62e67f48 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/winskel2009eye.md @@ -0,0 +1,47 @@ +# Winskel, H., Radach, R., & Luksaneeyanawin, S. (2009). Eye movements when reading spaced and unspaced Thai and English: A comparison of Thai–English bilinguals and English monolinguals. *Journal of Memory and Language, 61*(3), 339–351. + +**Identifier:** doi:10.1016/j.jml.2009.07.002 · **Read depth:** full_text (publisher-formatted PDF in the author copy kept locally, 13 printed pages, 339–351 with the journal's pagination; the Method and Results were read in full, and the reference list was scanned) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** eye-tracking (EyeLink II, 250 Hz) of single-line sentence reading, 2 (spacing) x 2 (target word frequency); 36 Thai–English bilinguals from Chulalongkorn University reading Thai and English (two sessions on separate days), 36 English monolinguals from the University of Western Sydney reading English; 72 sentences per language, 144 trials per list, Latin square · **Parties modeled:** person–artefact (reader and text) · **Facets:** D, E · **Read:** 2026-10-13 + +## What it argues +Winskel, Radach and Luksaneeyanawin run the first controlled eye-movement study of Thai, an alphabetic script written without spaces between words. They test whether spaces help eye guidance, word identification, both, or neither, in a script whose readers have never needed them (pp. 341–342). The spaces in the Thai sentences were inserted by two Thai linguists who agreed on every segmentation (p. 343). + +Results, with the paper's locators: +- Thai sentence level (p. 344, Table 1, p. 344): spaced sentences took longer, 1,834 ms against 1,746 ms unspaced (F1(1,35) = 4.03, p = .05; minF' not significant, p = .13), with more fixations, 9.27 against 8.62. The authors report a "5% decrement in reading rate" when spaces were added and note that the spaced sentences were 12.8% longer on screen. +- Thai target words (p. 345): no spacing effect on first fixation duration; gaze duration and total fixation duration were shorter with spaces (gaze duration F1(1,35) = 18.41, p < .001). Reading times differed between people: 15 of 36 read spaced sentences significantly slower, 6 read them faster, 15 showed no difference (p. 345); the Discussion restates this as 42% hurt, 16% helped, 42% unaffected (p. 349). +- Thai landing positions: unchanged by spacing, "just left of center of the word (approximately 0.3 of a letter)" (p. 346). From this the authors conclude spaces help later word processing and not eye guidance or early segmentation (p. 348). +- English, bilinguals (p. 344): a 45% decrement in reading rate without spaces; sentence reading time 2,133 ms spaced against 3,887 ms unspaced; fixations 10.25 against 15.05. +- English, monolinguals (p. 344): a 33% decrement; 1,827 ms against 2,730 ms; fixations 9.00 against 11.66. The language-group-by-spacing interaction was reliable (F1(1,70) = 23.81, p < .001). Bilinguals' landing positions fell about 0.9 of a letter from mid-word without spaces against about 0.5 for monolinguals (p. 346). +- Transfer: the authors expected Thai readers' lifelong practice with unspaced text to help with unspaced English, and "the prediction ... was not supported" (p. 349). Skill in the language being read mattered more than practice with an unspaced script. + +The authors cite the standard figure that removing spaces from English "typically slows reading by 30–50%" (p. 340), and they treat unspaced reading in Thai as fast enough to mediate saccade targeting (p. 349). + +## Relation to the argument +P3 (removing boundaries slows practised readers by a third to a half) rests on this paper at p. 344 and it holds: 33% for monolinguals and 45% for bilinguals. P4 (readers fluent in unspaced scripts gain nothing from added spaces) is only partly right. The Thai readers gained at the word level (shorter gaze and total fixation times) but lost at the sentence level, and the paper's explanation for the sentence-level loss is mixed: extra fixations on newly visible words, and 12.8% longer lines. For D2 the paper supports the view that fluent readers of an unspaced script cope well. the authors judge that experienced Thai readers solve the problem of finding word boundaries without spaces far better in Thai than either group does in English (p. 349). For D6, the paper separates two jobs that spaces might do, guiding the eye and speeding word recognition, and finds only the second in Thai. + +Facet check: the facet cites the paper for the cost of removing spaces. That is correct. The facet does not mention the Thai result, which is the more informative one for the unspaced-script debate: a 5% slowdown at sentence level with a word-level gain, and large differences between readers. + +## In conversation with +- `rayner1998unspaced` — the design template (high- and low-frequency targets in one frame) and the English baseline; this study replicates its pattern. +- `pollatsek1982eye` and `morris1990parafoveal` — cited for the 30–50% figure. +- `kohsom1997adding` — the earlier study (Kohsom & Gobet 1997, a conference paper) that Winskel et al. cite as finding facilitatory effects of inserted spaces in Thai, "but not using eye movement methodology" (p. 340); the paper does not say how the study measured reading. +- `sainio2007role` — Japanese hiragana, where spaces helped by about 12% against 30–50% in English; mixed Kanji–Hiragana text did not gain. +- `bai2008reading` — spaced Chinese read as easily as unspaced; two opposing forces. +- `epelboim1994reading` is not cited in the text I read. +- Leads not in the bib: Inhoff, Radach & Heller (2000) on German compounds; Juhasz, Inhoff & Rayner (2005) on English compounds; Reilly, Radach, Corbic & Luksaneeyanawin (2005) on Thai reading; Tsai & McConkie (2003). + +## Caution +Single-line sentences read for comprehension with a yes/no check on 10% to 15% of trials; Courier font; target words were nominals of five or six letters, and 13% of the Thai data, 8% of the bilinguals' English data and 13% of the monolinguals' data were dropped for short first fixations, skipped targets or tracking loss (p. 345). The bilinguals scored 39% on the English vocabulary test against the monolinguals' 90% (p. 342), so the group comparison mixes proficiency with script experience. The sentence-level Thai effect is at p = .05 by participants and not reliable in the combined test. The authors' own sentence-length confound limits any claim about global speed. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the publisher-formatted PDF (13 pp., JML 61, 339-351; file is cards/winskel2009.pdf, the key has no .pdf of its own name), confirmed title, authors and year, and checked locators against the printed page numbers. Checked 4 quotations (all exact, pages right), Table 1 (all six cell means for time and fixations), F values, the 15/6/15 participant split and 42/16/42 percentages, the landing-position values, 13/8/13% data loss, vocabulary test percentages (39%, 90%), the 10-15% comprehension checks, and the Sainio, Bai, Kohsom, Pollatsek and Morris citations. + +**Findings:** One error fixed: kohsom1997adding was described as a 'reading-aloud study'; the paper only says Kohsom and Gobet (1997) found facilitatory effects of inserted spaces in Thai 'but not using eye movement methodology' (p. 340). Everything else matched. Note for the Tibetan card: Wang et al. say Winskel found spaces hurt Thai reading speed; Winskel's sentence-level cost is 5% by participants (p = .05, minF' p = .13), as this card's Caution says. + +## Also used by facet W (2026-10-15) + +Facet W uses the Thai result as the test case for "freed literacy by moving boundary detection from the voice to the page": readers who never had spaces detect boundaries silently, and when spaces are added they gain at the word (shorter gaze, p. 345) and lose at the sentence (5% slower, p. 344). Boundary detection without spaces is therefore eye-and-lexicon work, not voice work, in a fluent reader; the abstract's "from the voice" holds, if at all, for learners (`kasisopa2016child` shows Thai schooling spaces Grade 1 text and then withdraws the spaces). The English result (33% and 45% slower without spaces, p. 344) is the cost side of the representational effect. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wong2022meta.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wong2022meta.md new file mode 100644 index 00000000..c6dabb25 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/wong2022meta.md @@ -0,0 +1,38 @@ +# Wong, D., & Floridi, L. (2023). Meta's Oversight Board: A review and critical assessment. *Minds and Machines, 33*(2), 261–284 (online 24 October 2022). + +**Identifier:** doi:10.1007/s11023-022-09613-x · **Read depth:** full_text (open-access publisher PDF, journal pagination; reference list skimmed) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** none of its own; a synthesis of scholarship, news reports and Meta and Oversight Board documents, with one case analysis (Ethiopia, case FB-MP4ZC4CC) · **Parties modeled:** two parties plus a third party (user and platform, with a board funded by but legally separate from the platform) · **Facets:** J · **Read:** 2026-10-13 + +## What it argues +Wong and Floridi build a framework of strengths and weaknesses for Meta's Oversight Board (OB). The Board reviews removals and keeps-up decisions on Facebook and Instagram. Its rulings on a case bind Meta unless implementing them "could violate the law"; its policy recommendations do not bind, but Meta has committed to respond within sixty days (pp. 263–264). Meta funds it through an independent trust: $130 million at launch and $150 million more in July 2022 (p. 264). The authors place it as "platform self-governance", more independent than most self-regulation because it is legally separate (p. 264). + +Three strengths (pp. 266–268). It makes moderation more transparent, for example by exposing "internal rule books" (§3.1). Its recommendations have moved Meta, which agreed to translate its standards into South Asian languages and to tell users which rule they broke (§3.2). It is assertive: of 27 decisions "issued in 2021" it overturned Meta in 20 (74%) and upheld 7 (p. 268). It also read its charter to stop Meta removing a case from review by reversing its own decision (p. 268). + +Four weaknesses (pp. 268–271). Jurisdiction is narrow: a binary uphold-or-overturn on single posts, not accounts or groups (p. 269). Impact is small. Meta keeps final authority and could ignore recommendations or "choose not to provide future funding", and the Board "cannot rule on, or mandate changes to, broader platform features such as Facebook's recommendation algorithms, advertising systems, or data collection" (pp. 269–270). In 2021 it issued "20 decisions out of 130 shortlisted cases and approximately over 1 million cases submitted" (p. 269). Precedent is under Meta's control: Meta decides whether a new case is "identical content in parallel context" and "technically and operationally feasible", and precedent only removes content, never restores it (pp. 264, 270). Diversity is low: Domino (2020) counted one of 20 members from Southeast Asia, and over two-thirds of 2021 appeals came from the Global North, with 2% from sub-Saharan Africa and Central and South Asia (pp. 270–271). + +The Ethiopia case (§5) shows these at work. The Board upheld removal of a Facebook post on a different ground than Meta used, and of three recommendations Meta implemented one in part, took no further action on one, and agreed only to assess the third (pp. 272–275). The authors propose reforms (§7): let the Board decide when its precedent applies, apply precedent to pending appeals, define the criteria, and add appellate boards of randomly chosen users to decide whether precedent applies (pp. 278–281). They offer no test of the proposal. + +## Relation to the argument +Ruling D8 asks how courts, FX markets and platforms differ as holders of a convention. This paper is the platform case of a court-like body, and it shows the body's reach stopping where the platform's design begins: the Board cannot touch the recommendation system, and the owner controls how far precedent travels. That supports premise P10 (a convention the owner can withdraw alone is no convention) with a case in which the owner built the court. It also shows an owner-funded body that overturns the owner most of the time, so independence in the ruling and dependence in the architecture can coexist. + +Facet check: the facet quotes the abstract and reads the Board as "court-like" with Meta controlling jurisdiction and precedent. The text agrees. It adds that the Board's case rulings are binding while its recommendations are not, so "controls" holds most fully for precedent and scope. The text also gives two counts for 2021: 27 decisions (p. 268) and 20 decisions (p. 269), each attributed to the Board's annual report. Do not cite either without checking the report. + +The card's reading, not Wong and Floridi's: the Board's tiny caseload against about a million submissions resembles Galanter's overload (`galanter1974why`, pp. 119–122), which would push users to treat the process as symbolic. + +## In conversation with +- `klonick2018new`: cited (p. 1603, with a 2020 article by Klonick) for the point that users lack "direct accountability mechanisms" or "a fair opportunity to participate" (p. 280). The authors use this to motivate their appellate-board proposal. +- `douek2023meta`: the Board paper is not cited here (Douek's later work postdates it). The authors rely on Douek's 2019–2021 articles for self-regulation, non-binding recommendations and the legitimacy problem (pp. 264, 267, 269); those are leads. + +Leads, not in the bib: Douek (2019b), "Facebook's Oversight Board: Move fast with stable infrastructure and humility"; Van Loo (2021), "Federal Rules of Platform Procedure"; Schauer (1987) on precedent; Pickup (2021) on Board access to algorithms; Heldt (2019) on non-binding guidance. + +## Caution +A descriptive and normative review of the Board's first 18 months or so; the data are Meta's and the Board's own reports. The reform proposals are untested. The two 2021 decision counts disagree. Anything about the Board after October 2022 is outside this paper. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the open-access publisher PDF (25 pp., Minds and Machines 33, 261-284; journal page = PDF page + 260) and checked the card against the page text. Checked 11 verbatim quotations (all exact, every page locator correct), the 130 million / 150 million funding figures, 27 decisions with 20 overturned (74%) on p. 268 against 20 decisions of 130 shortlisted and about 1 million submitted on p. 269, the Domino and Global North figures (pp. 270-271), the Ethiopia case outcome for the three recommendations, the reform proposals (pp. 278-281, random selection on p. 280), and the Klonick citation (2018, p. 1603). + +**Findings:** No issues found. The card correctly flags that the two 2021 decision counts (27 and 20) are not reconciled in the paper, and correctly labels the Galanter comparison as the card's own reading (Galanter is not cited in the source). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/yu2023altered.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/yu2023altered.md new file mode 100644 index 00000000..ac6108ac --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/yu2023altered.md @@ -0,0 +1,40 @@ +# Yu, H., & Kwon, M. (2023). Altered eye movements during reading with simulated central and peripheral visual field defects. *Investigative Ophthalmology & Visual Science, 64*(13), Article 21. + +**Identifier:** doi:10.1167/iovs.64.13.21 · **Read depth:** full_text (open-access article text, PMC10584020, Europe PMC XML; no printed page numbers, so locators are section names; figures and the Table were read as text only) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** within-subject gaze-contingent experiment, 24 normally sighted young adults (18 to 23 years, mean 19.62; nine men) at Northeastern University, reading five-sentence pages aloud in Courier New with a simulated peripheral aperture or central scotoma of 2, 4, 6, 8 or 10 degrees, plus no-scotoma baseline; EyeLink 1000 Plus at 500 Hz; 28 text pages per person; three people dropped from central-scotoma analysis, leaving 21 · **Parties modeled:** person–artefact (reader and degraded text) · **Facets:** D · **Read:** 2026-10-13 + +## What it argues +Yu and Kwon ask how much of the visual field a reader needs, and where, before eye movements depart from normal. They find a critical size, not a gradual decline. + +Results (Results; Table; Figs. 2 and 4): +- Reading speed fell 54% under the smallest aperture (226.44 ± 5.35 wpm intact against 104.73 ± 6.70 wpm with a 2-degree aperture, p < .001), and 51% under the largest central scotoma (214.08 ± 4.48 against 104.09 ± 5.64 wpm). The effect of peripheral scotoma size on speed was F(5,115) = 117.24, p < .001; the effect of central scotoma size was F(5,100) = 123.77, p < .001. +- Peripheral aperture: below 6 degrees (about 11 visible letters), fixation duration and fixation count rose sharply. At the 2-degree aperture, fixation duration rose 21% (214 to 256 ms), fixations per line doubled (12.45 to 24.83), and saccade amplitude fell 30% (3.27 to 2.29 degrees). Saccade amplitude and velocity began to fall below an 8-degree aperture (14 letters). +- Central scotoma: a scotoma larger than 2.5 degrees (about 4.4 letters) hurt speed. At 10 degrees, saccade amplitude rose 98%, fixation duration 13% (217 to 245 ms), fixations per line 89%, and the share of regressive saccades 86% (24% to 45%). Microsaccade rate fell nearly 70%. +- The authors read the 6-degree figure as matching the visual-span hypothesis, with the visual span at roughly 9 to 14 letters for healthy adults (Discussion, take-home message 5). +- In the Introduction they state the standard perceptual span as "three to four letters to the left of fixation and 14 to 15 letters to the right of fixation," citing McConkie and Rayner, Rayner (1986), and others. + +The Methods give letter size as 0.5 degrees with 0.065-degree interletter spacing, so one letter occupies 0.565 degrees; this is how they convert degrees into letters. + +## Relation to the argument +The paper does not examine word spacing, and no premise in the argument map depends on its data. The project uses it as an open source that restates the perceptual-span figures. The Introduction attributes them to Rayner and McConkie and others, so the card supports only that Yu and Kwon assert them. The Discussion's interpretive line, that visibility and eye movements are tightly coupled and that longer, more frequent fixations may be "optimal reading strategies to compensate," is relevant to D6 only by analogy: a degraded input changes where and how long the eyes rest, not just how fast the reader goes. + +Facet check: the facet describes the paper as one of three open papers that restate the perceptual-span figures. That is accurate. Yu and Kwon contribute a different number too: the span they measure in this task is about 11 to 14 letters, which they call "slightly" smaller than the perceptual span of 14 to 19 letters, and they give reasons (Discussion). + +## In conversation with +- `rayner1998eye` — ref. 4, the 20-year review, cited in the opening for fixations and saccades. +- `rayner1986eye` — ref. 37, perceptual span in beginning and skilled readers. +- `pelli2008uncrowded` — ref. 20, Pelli & Tillman, the "uncrowded window" they equate with the visual span. +- `rayner2009thirtyfifth` — not cited here. +- Leads not in the bib: Legge et al. (1997), visual span in normal and low vision; McConkie & Rayner (1975); Rayner, Slattery & Bélanger (2010); Yu, Shamsi & Kwon (2022) on degraded text; Schotter, Angele & Rayner (2012). + +## Caution +Young adults with normal vision, simulated and not clinical scotomas, no adaptation period (the authors say people with real central loss settle on a preferred retinal locus and show smaller saccades, which this sample did not), simple second- to fourth-grade sentences read aloud, 56 characters per line. The span figures it quotes are restated from other work; cite the original studies. The Table's microsaccade-amplitude result was null (p = .78 peripheral, p = .31 central). + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the Europe PMC XML (PMC10584020) as plain text and read the Introduction, Methods, Results and Discussion sections cited. Checked the 2 verbatim quotations (exact), about 35 numbers (reading speeds with SDs, F values, percentage changes in saccade amplitude, fixation duration, fixations per line, regressions and microsaccade rate, the 2.5 degree / 4.4 letter and 6 degree / 11 letter thresholds, letter size 0.5 + 0.065 degrees, participants, 500 Hz, 28 text pages, 56-character sentences, null microsaccade-amplitude p values) and references 4, 20, 29, 33, 36, 37. + +**Findings:** No issues found. The three-to-four / 14-to-15 sentence sits in the Introduction and cites refs 29, 33, 36 and 37 (McConkie and Rayner 1975, Rayner 1975, Rayner et al. 2010, Rayner 1986), as the card says. The 14-to-19-letter comparison is in the Discussion ('appears to be slightly larger than what we have observed'). diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/yu2026habituation.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/yu2026habituation.md new file mode 100644 index 00000000..f80d11f2 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/yu2026habituation.md @@ -0,0 +1,41 @@ +# Yu, H., Liu, L., Jiang, X., Jia, Y., Wang, S., Qian, P., & Chen, Y. (2026). Habituation at the gate: Rising approval and declining scrutiny in human review of AI agent code. arXiv:2606.22721. + +**Identifier:** arXiv:2606.22721 · **Read depth:** full_text (arXiv v1, 21 Jun 2026, five pages in ACM two-column format; pagination below is that PDF's, pp. 1–5) · **Source-tier:** preprint (the bib entry is typed `@misc`) · **Evidence basis:** direct_read · **Design and sample:** observational, longitudinal within-reviewer analysis of GitHub review logs from the AIDev dataset; 400 repeat reviewers (10 or more agent pull requests each) and 11,429 reviews over a seven-month window, plus a human-PR comparison of 11,415 reviews by 1,851 reviewers · **Parties modeled:** person–system (a human reviewer and an AI coding agent's output) · **Facets:** G · **Read:** 2026-10-13 + +## What it argues + +Yu and colleagues ask whether reviewers who accumulate experience with AI-agent pull requests (PRs) approve more and inspect less. They take AIDev, which holds 16,895 reviews by 2,494 reviewers across five agents, and keep the 400 reviewers with at least 10 agent PRs (median 17 reviews; p. 2). They compare each reviewer's early and late half and also pool reviews by the reviewer's own experience decile. + +The headline numbers: approval rises from 30.1% to 36.8% between the early and late halves (+6.7 points; paired Wilcoxon p < 10^-6), and across experience deciles from 27.9% to 42.4% (+14.5 points; Table 1, p. 2). The change-request rate falls from 11.2% to 5.6% in the decile table (p. 2). At the individual level 52% of reviewers become more approving, 28% less, 20% unchanged (p. 2). Median review latency rises from 3.9 to 13.5 hours (p. 2), while inline comments per review fall from 1.01 to 0.79 (−22%, p = 0.0014) and comment words from 18.6 to 13.5 (p. 3). The shift in approval correlates with the drop in commenting (Spearman ρ = −0.556; p. 3). Three checks point to experience and to the agent: a per-reviewer logistic fit gives the experience term a mean coefficient of +0.11 and calendar time −0.007 (p. 4); median PR size stays flat (p. 4); and on human-authored PRs in the same repositories approval declines (29.1% by June) while agent-PR approval rises (41.7% by July; p. 3). + +The authors read the pattern as "most consistent with reflexive habituation under growing workload rather than rational trust calibration alone" (abstract, p. 1). They also say flatly: "We cannot establish causality" (p. 1), and "Crucially, we lack a direct measure of agent code quality over time; if agents genuinely improved, rising approval would be rational rather than habituative" (p. 4). All four agents received updates during 2025 (p. 1). Per-agent shifts differ: Copilot +9.1 points (n = 241), Devin +3.8 (n = 95), Codex +1.8 (n = 51) and Cursor −1.4 (n = 18; Table 2, p. 3). The proxy cross-agent control is "underpowered" (p. 3). + +## Relation to the argument + +**Premise P12** ("people perform sustained monitoring and whole-volume review badly"). This is the closest piece of field data on the exact task the Nadella post leaves a person: reading agent output and deciding to let it through. It is observational and compatible with the reviewers learning that the agents got better, so it supports P12 as a pattern consistent with it, not as a measured failure. The map's own wording for P12 is "mixed for the rest", and this source stays in that group. Its distinct contribution is a time trend: more approval with fewer comments as exposure grows. + +**Rulings D4 and D7.** The post puts a person at the end to read evidence and halt (P11). Yu gives one reason to doubt the durability of that seat. It does not test chain-of-thought evidence, and it says nothing of pilots; the aviation literature the D7 flag calls for is not cited here. + +**Premise P10.** Reviewers and agent operators are not shown to have conflicting interests here. The source cannot speak to who holds a convention. + +**Facet check:** the facet's numbers (400 reviewers, 11,429 reviews, 30.1% to 36.8%, latency 3.5 times, comments −22%) match the full text. The paper itself is less internally tidy than its abstract: the early-late change-request rates (11.2% to 5.6%) repeat the first-decile and tenth-decile figures of Table 1, the text says "28% less commenting" for words but 22% for comments (pp. 3–4), and Spearman ρ = +0.02 for PR size is given a p of 0.009 and called negligible (p. 4). I could not resolve these from the text. + +## In conversation with + +The reference list holds eleven software-engineering items; I matched none to `references.bib`. It does not cite Bainbridge (`bainbridge1983ironies`), Parasuraman, Green (`green2022flaws`) or any automation-complacency source, though its introduction says the effect "mirrors habituation effects documented in security auditing, aviation checklists, and automated-testing interactions" (p. 1; it cites Cassee et al. 2020 and Wessel et al. 2020 for this). + +- Leads, not in the bib: Li, Zhang & Hassan (2025), the AIDev dataset (arXiv:2507.15003); Wessel et al. (2020) on code-review bots; Cassee et al. (2020) on CI and review; Bacchelli & Bird (2013) on code review expectations. + +## Caution + +A five-page preprint by independent researchers on public, popular repositories (more than 100 stars). The data are approval flags, not defects, and no outcome quality is measured. Approval, latency and comment counts are weak proxies for "scrutiny". The reviewer pool for the human-PR comparison thins out after June 2025 (n = 207 reviews in July). Use it as suggestive evidence that review volume and exposure go with lighter review, not as a measured decline in oversight. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** confirmed + +**What I checked:** Opened the arXiv v1 PDF (5 pp.) and extracted it. Checked 5 verbatim quotations (all exact, correct page), about 40 numbers (400 reviewers, 11,429 reviews, 16,895 reviews by 2,494 reviewers, approval 30.1 to 36.8 and 27.9 to 42.4, change requests 11.2 to 5.6, 52/28/20%, latency 3.9 to 13.5 h, comments 1.01 to 0.79, words 18.6 to 13.5, rho -0.556, logistic coefficients +0.11 and -0.007, the human-PR comparison 29.1% and 41.7%, per-agent shifts and n, n = 207 in July, PR-size rho +0.02 with p 0.009) and the reference list (Cassee, Wessel, Bacchelli, Li et al.; no Bainbridge or Parasuraman). + +**Findings:** No issues found. The card's three internal inconsistencies in the paper (early-late versus decile change-request rates, 28% versus 22% commenting, rho = +0.02 with p = 0.009 called negligible) are real and accurately reported. The 'repeat the first-decile and tenth-decile figures' observation about change-request rates is the card's inference; the text does give 11.2 and 5.6 in the decile table. diff --git a/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ziegler2010orthographic.md b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ziegler2010orthographic.md new file mode 100644 index 00000000..7e81a646 --- /dev/null +++ b/submissions/berthon_ethics_paper/literature/cards/dxd_foundation/ziegler2010orthographic.md @@ -0,0 +1,37 @@ +# Ziegler, J. C., Bertrand, D., Tóth, D., Csépe, V., Reis, A., Faísca, L., Saine, N., Lyytinen, H., Vaessen, A., & Blomert, L. (2010). Orthographic depth and its impact on universal predictors of reading: A cross-language investigation. *Psychological Science, 21*(4), 551–559. + +**Identifier:** doi:10.1177/0956797610363406 · **Read depth:** full_text (publisher-formatted PDF from the University of Algarve repository, journal pagination 551–559 intact; the on-line Supplemental Material, Tables S1 and S2, was not available and was not read) · **Source-tier:** peer-reviewed article · **Evidence basis:** direct_read · **Design and sample:** cross-sectional study of 1,265 Grade 2 children in five countries (Finland 166, France 181, Hungary 139, Netherlands 597, Portugal 182); language-specific tests, z-scored within country; per-country regressions and a two-level hierarchical linear model with onset entropy at Level 2 · **Parties modeled:** other (child–script; no artefact or third party) · **Facets:** C · **Read:** 2026-10-13 + +## What it argues + +English is the outlier among alphabetic orthographies, and some researchers fear English-based work overstates the role of phonological awareness (p. 551–552, citing Share 2008 and Landerl & Wimmer 2000). The authors rank five languages by onset entropy, an index of how unpredictable letter-to-sound mappings are: Finnish 0.0, Hungarian 0.17, Dutch 0.23, Portuguese 0.42, French 0.46, with English 0.83 for comparison (Table 1, p. 552). They test three claims: phonological awareness is the main factor in each language (strong), it matters everywhere but with varying impact (weak), or it matters equally everywhere (p. 553). + +Results. In the per-country regressions, phonological awareness was the strongest correlate of reading speed and accuracy in every country but Finland (standardized betas for reading speed from 0.31 in Finland to 0.54 in Portugal; Table 2, p. 555), and the strongest correlate of decoding in all five. Rapid automatized naming (RAN) was a weaker correlate, limited to speed. Nonverbal IQ had no significant influence. In the multilevel model, entropy raised the weight of phonological awareness for reading speed (γ = 0.445, p ≤ .05), reading accuracy (0.689, p ≤ .05) and decoding speed (0.687, p ≤ .001; Table 3, p. 556). The authors conclude that the data "strongly support the weak claim," that phonological awareness "is important in all languages but that its impact is modulated by script entropy" (p. 556). A vocabulary effect that looked stronger in transparent scripts vanished once Finland was dropped (p. 555–556). The authors read the entropy effect as reciprocal: a transparent script "promotes access to phonemes," which in turn raises phonological awareness (p. 556). + +The introduction restates Seymour et al. (2003): at the end of Grade 1, reading accuracy reached ceiling in Italian, German, Greek, Spanish and Finnish; it was around 80% in Portuguese, French and Danish and "only 34%" in English (p. 551–552). + +## Relation to the argument + +Facet C uses this paper, with Seymour et al., as modern cross-script evidence for premise P6 (a script's design changes the rate at which children learn). The full text keeps that role but narrows it. The paper measures which cognitive skills predict reading in children already in school, not how fast a script is learned; the rate claim belongs to Seymour et al. as reported here. What the paper adds is the shape of the dependence: how much a skill matters varies with the script, and the better the script maps onto sounds, the less a child needs the separate skill of manipulating sounds. That supports the paper's guide-not-mechanism ruling (D10) in a weak form: the interface changes what the reader must bring to it. It says nothing about spaces between words or about who gets to school (D1, D2). + +Facet check: the facet says the paper has "the same bearing as Seymour." It does not. It reports predictors of reading in Grade 2, not acquisition rates, and the study itself restricts its claim to alphabetic scripts and flags Chinese as unexplored (p. 557). + +## In conversation with + +- `seymour2003foundation` — the 14-country Grade 1 study this paper builds on and cites for the ceiling effect in transparent scripts and 34% accuracy in English (p. 551–552). +- `ziegler2005reading` — the authors' own grain-size review, cited first for the claim that orthographic consistency is "the key factor determining the rate of reading acquisition" (p. 551). +- Leads not in the bib (check): Share 2008 (the "anglocentric" argument, *Psychological Bulletin*); Caravolas, Volin & Hulme 2005; Borgwaldt, Hellwig & De Groot 2005 (onset entropy); Vaessen et al. 2009 (RAN); McBride-Chang & Kail 2002 (English and Chinese). + +## Caution + +Correlational, one time point, Grade 2 only; the authors say the age may favor finding strong phonological-awareness effects. Tests differ by country, so only standardized weights compare across countries; five languages is a small Level 2 sample, and the entropy effect is clearest for decoding speed (p ≤ .001) while the reading effects are p ≤ .05. Finland departs from the pattern on several measures. Alphabetic scripts only. I did not read the Supplemental Material. + +--- + +## S2 adversarial verification (2026-10-13) + +**Verdict:** corrected + +**What I checked:** Opened the publisher-formatted PDF (9 pp., Psychological Science 21(4), 551-559, journal pagination intact). Checked 5 verbatim quotations (all exact, pages right), the sample sizes by country (166, 181, 139, 597, 182; 1,265 total), the entropy values and Table 2 betas (0.31 to 0.54 for reading speed; phonological awareness not the strongest predictor in Finland, where vocabulary is), Table 3 gammas (.445, .689, .687) and their p levels, the Seymour 34% and 80% figures, the Finland-dropped vocabulary result, and the Share, Landerl and Wimmer, and Ziegler and Goswami citations. + +**Findings:** One locator error fixed: the entropy table (Table 1) is on p. 552, not 553. Everything else matched. The Supplemental Material was not read, as the card says. diff --git a/submissions/berthon_ethics_paper/nadella/CARDS.md b/submissions/berthon_ethics_paper/nadella/CARDS.md new file mode 100644 index 00000000..13e63111 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/CARDS.md @@ -0,0 +1,489 @@ +# Cards behind the earlier Nadella and transition research + +Generated by `nadella/collect.py`; do not edit by hand. Every key in the two Nadella facet bibs +(`sources/*.bib`: facets G and V1 on the post; A-D, F, J, K, O, W on word spacing; B on print), +with the card now in `../literature/cards/`. Paths are relative to that directory. + +Cards: 204. Bib keys with no card: 260. + +| card | authors | year | title | basis | facet bib | +|---|---|---|---|---|---| +| `algorithmacy_design_ethics/anannycrawford2018.md` | Ananny, M., & Crawford, K. | 2018 | Seeing without knowing: Limitations of the transparency ideal and its applicatio | abstract | dxf_facet_G_nadella_lineage.bib | +| `algorithmacy_design_ethics/calacci2022.md` | Calacci, D., & Pentland, A. | 2022 | Bargaining with the black-box: Designing and deploying worker-centric tools to a | abstract | dxf_facet_O_custody_of_conventions.bib | +| `algorithmacy_design_ethics/eisenstein1979.md` | Eisenstein, E. L. | 1979 | The Printing Press as an Agent of Change: Communications and Cultural Transforma | secondary | bec_facet_B_print_1450_1800.bib | +| `algorithmacy_design_ethics/eisenstein1979.md` | Eisenstein, E. L. | 1979 | The Printing Press as an Agent of Change: Communications and Cultural Transforma | secondary | dxf_facet_C_spacing_to_literacy.bib | +| `algorithmacy_design_ethics/eisenstein2002.md` | Eisenstein, E. L. | 2002 | An unacknowledged revolution revisited | secondary | bec_facet_B_print_1450_1800.bib | +| `algorithmacy_design_ethics/endsleykiris1995.md` | Endsley, M. R., & Kiris, E. O. | 1995 | The out-of-the-loop performance problem and level of control in automation | abstract | dxf_facet_G_nadella_lineage.bib | +| `algorithmacy_design_ethics/goody1986.md` | Goody, J. | 1986 | The Logic of Writing and the Organization of Society | secondary | dxf_facet_C_spacing_to_literacy.bib | +| `algorithmacy_design_ethics/havelock1963.md` | Havelock, E. A. | 1963 | Preface to Plato | partial | dxf_facet_C_spacing_to_literacy.bib | +| `algorithmacy_design_ethics/innis1951.md` | Innis, H. A. | 1951 | The Bias of Communication | partial | bec_facet_B_print_1450_1800.bib | +| `algorithmacy_design_ethics/johns1998.md` | Johns, A. | 1998 | The Nature of the Book: Print and Knowledge in the Making | partial | bec_facet_B_print_1450_1800.bib | +| `algorithmacy_design_ethics/street1984.md` | Street, B. V. | 1984 | Literacy in Theory and Practice | secondary | dxf_facet_C_spacing_to_literacy.bib | +| `algorithmacy_design_ethics/wilkinson1965.md` | Wilkinson, A. | 1965 | The concept of oracy | abstract | dxf_facet_W_spacing_oracy_literacy.bib | +| `berthon_ethics_course/amershi2019guidelines.md` | Amershi, Saleema and Weld, Dan and Vorvoreanu, Mih | 2019 | Guidelines for Human-AI Interaction | partial | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/arber1875transcript.md` | Arber, Edward | 1875 | A Transcript of the Registers of the Company of Stationers of London, 1554-1640 | partial | bec_facet_B_print_1450_1800.bib | +| `berthon_ethics_course/balkin2016information.md` | Balkin, Jack M. | 2016 | Information Fiduciaries and the First Amendment | partial | dxf_facet_O_custody_of_conventions.bib | +| `berthon_ethics_course/eu2024aiact.md` | {European Parliament and Council of the European U | 2024 | Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 Ju | partial | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/euomnibus2026.md` | {European Parliament and Council of the European U | 2026 | Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 Jul | partial | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/graff1991literacy.md` | Graff, Harvey J. | 1979 | The Literacy Myth: Literacy and Social Structure in the Nineteenth-Century City | partial | dxf_facet_C_spacing_to_literacy.bib | +| `berthon_ethics_course/johns2002how.md` | Johns, Adrian | 2002 | How to Acknowledge a Revolution | full | bec_facet_B_print_1450_1800.bib | +| `berthon_ethics_course/khan2019skeptical.md` | Khan, Lina M. and Pozen, David E. | 2019 | A Skeptical View of Information Fiduciaries | partial | dxf_facet_O_custody_of_conventions.bib | +| `berthon_ethics_course/microsoft2022responsible.md` | {Microsoft} | 2022 | Microsoft Responsible AI Standard, v2: General Requirements (For External Releas | full | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/microsoft2024responsible.md` | {Microsoft} | 2024 | Responsible AI Transparency Report: How We Build, Support Our Customers, and Gro | partial | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/microsoft2025ai.md` | {Microsoft Office of Responsible AI} | 2025 | AI Literacy Starting Guide: Securing the Benefits of AI through Literacy for All | full | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/microsoft2025responsible.md` | {Microsoft} | 2025 | 2025 Responsible AI Transparency Report | partial | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/microsoft2026application.md` | {Microsoft} | 2026 | Application card: Microsoft Copilot (for organizations) [Transparency Note for M | full | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/milton1644areopagitica.md` | Milton, John | 1644 | Areopagitica: A Speech for the Liberty of Unlicensed Printing | partial | bec_facet_B_print_1450_1800.bib | +| `berthon_ethics_course/ong1982orality.md` | Ong, Walter J. | 1982 | Orality and Literacy: The Technologizing of the Word | partial | dxf_facet_C_spacing_to_literacy.bib | +| `berthon_ethics_course/rahman2021invisible.md` | Rahman, Hatim A. | 2021 | The Invisible Cage: Workers' Reactivity to Opaque Algorithmic Evaluations | abstract | dxf_facet_J_who_holds_the_convention.bib | +| `berthon_ethics_course/smith2025microsoft.md` | Smith, Brad | 2025 | Microsoft Elevate: Putting people first | full | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `berthon_ethics_course/statutesrealm1819vol5.md` | Great Britain, Record Commission | 1662 | An Act for preventing the frequent Abuses in printing seditious treasonable and | full | bec_facet_B_print_1450_1800.bib | +| `berthon_ethics_course/sudjianto2026when.md` | Sudjianto, Agus | 2026 | When AI Becomes an Insider, Who Is Left to Check Its Work? | full | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `dxd_foundation/alahmadi2022false.md` | Alahmadi, B. A., Axon, L., & Martinovic, I. | 2022 | 99% false positives: A qualitative study of SOC analysts' perspectives on securi | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/ancker2017effects.md` | Ancker, J. S., Edwards, A., Nosal, S., Hauser, D., | 2017 | Effects of workload, work complexity, and repeated alerts on alert fatigue in a | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/anderson1972computer.md` | Anderson, J. P. | 1972 | Computer security technology planning study (Vol | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/anthropic2026how.md` | Anthropic. | 2026 | How we built Claude Code auto mode: a safer way to skip permissions | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/apple2026guidelines.md` | Apple Inc. | 2026 | App Review Guidelines [Web page; "Last Updated: June 8, 2026"] | partial | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/bai2008reading.md` | Bai, X., Yan, G., Zang, C., Liversedge, S. P., & R | 2008 | Reading spaced and unspaced Chinese text: Evidence from eye movements | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/bainbridge1983ironies.md` | Bainbridge, L. | 1983 | Ironies of automation | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/baker2025monitoring.md` | Baker, B., Huizinga, J., Gao, L., Dou, Z., Guan, M | 2025 | Monitoring reasoning models for misbehavior and the risks of promoting obfuscati | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/balogh1927voces.md` | Balogh, J. | 1927 | „Voces paginarum": Beiträge zur Geschichte des lauten Lesens und Schreibens | partial | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/balogh1927voces.md` | Balogh, J. | 1927 | „Voces paginarum": Beiträge zur Geschichte des lauten Lesens und Schreibens | partial | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/barton1980awareness.md` | Barton, D., & Hamilton, M. E. | 1980 | Awareness of the segmental structure of English in adults of various literacy le | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/bassetti2012effects.md` | Bassetti, B., & Masterson, J. | 2012 | Effects of removing morphemic information and adding interword spacing on readin | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/becker2024conquering.md` | Becker, S. O., & Won, C. | 2024 | Conquering Korea for Jesus: Protestant missionaries, local churches, and literac | full | dxf_facet_C_spacing_to_literacy.bib | +| `dxd_foundation/biggs2020decoding.md` | Biggs, N. | 2020 | Decoding chancery records from the 1240s | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/boretius1883capitularia.md` | Boretius, A. (Ed.). | 1883 | Capitularia regum Francorum (Vol | partial | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/buijsman2020numerals.md` | Buijsman, S. | 2020 | How numerals support new cognitive capacities | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/buncic2017factors.md` | Bunčić, D. | 2017 | Factors influencing the success and failure of writing reforms | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/burnyeat1997postscript.md` | Burnyeat, M. F. | 1997 | Postscript on silent reading | partial | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/cameron2022expanding.md` | Cameron, L. D., & Rahman, H. | 2022 | Expanding the locus of resistance: Understanding the co-constitution of control | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/caperton2009massey.md` | *Caperton v. A.T. Massey Coal Co.*, 556 U.S. 868 | 2009 | | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/cert2022common.md` | CERT National Insider Threat Center. | 2022 | Common sense guide to mitigating insider threats (7th ed.) | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/cftc2017fxcm.md` | U.S. Commodity Futures Trading Commission. | 2017 | CFTC orders Forex Capital Markets, LLC (FXCM), its parent company, FXCM Holdings | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/chan2024visibility.md` | Chan, A., Ezell, C., Kaufmann, M., Wei, K., Hammon | 2024 | Visibility into AI agents | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/chan2025infrastructure.md` | Chan, A., Wei, K., Huang, S., Rajkumar, N., Perrie | 2025 | Infrastructure for AI agents [Preprint] | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/chantavarin2025effects.md` | Chantavarin, S., & Leung, T. T.-C. | 2025 | The effects of interword spacing and morphological complexity in reading Thai: A | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/charlemagne1900letter.md` | Charlemagne. | 1900 | Letter to Abbot Baugulf of Fulda (De litteris colendis), c | full | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/chen2021tradeoff.md` | Chen, M., Wang, Y., Zhao, B., Li, X., & Bai, X. | 2021 | The trade-off between format familiarity and word-segmentation facilitation in C | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/chen2025reasoning.md` | Chen, Y., Benton, J., Radhakrishnan, A., Uesato, J | 2025 | Reasoning models don't always say what they think [Preprint] | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/chen2026comparing.md` | Chen, C., Zhang, Z., Chen, Z., Xu, E., Yang, Y., K | 2026 | Comparing human oversight strategies for computer-use agents [Preprint] | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/clanchy1993from.md` | Clanchy, M. T. | 1993 | From memory to written record: England 1066–1307 (2nd ed.) | partial | dxf_facet_C_spacing_to_literacy.bib | +| `dxd_foundation/clemens2025korea.md` | Clemens, W. C. | 2025 | Review of V | full | dxf_facet_C_spacing_to_literacy.bib | +| `dxd_foundation/colvin2011empirical.md` | Colvin, A. J. S. | 2011 | An empirical study of employment arbitration: Case outcomes and processes | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/constantin2026nadella.md` | Constantin, A. M. | 2026 | Satya Nadella says companies should treat AI models as insider risks | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/cressy1980literacy.md` | Cressy, D. | 1980 | Literacy and the social order: Reading and writing in Tudor and Stuart England | partial | dxf_facet_C_spacing_to_literacy.bib | +| `dxd_foundation/crootof2023humans.md` | Crootof, R., Kaminski, M. E., & Price, W. N., II. | 2023 | Humans in the loop | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/cui2023eye.md` | Cui, Y. | 2023 | Eye movements of second language learners when reading spaced and unspaced Chine | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/dacosta2018tabulae.md` | da Costa, A. | 2018 | "That ye mowe redely fynde … what ye desyre": Printed tables of contents and ind | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/danna2022numerical.md` | Danna, R., Iori, M., & Mina, A. | 2022 | A numerical revolution: The diffusion of practical mathematics and the growth of | partial | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/de1972betrvg.md` | | | Bundesministerium der Justiz. (n.d.). *Betriebsverfassungsgesetz (BetrVG)* [Work | partial | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/degno2019coregistration.md` | Degno, F., Loberg, O., Zang, C., Zhang, M., Donnel | 2019 | A co-registration investigation of inter-word spacing and parafoveal preview: Ey | partial | dxf_facet_D_reading_science.bib | +| `dxd_foundation/delacroix2019bottom.md` | Delacroix, S., & Lawrence, N. D. | 2019 | Bottom-up data trusts: Disturbing the 'one size fits all' approach to data gover | full | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/dionysius1910composition.md` | Dionysius of Halicarnassus. | 1910 | On literary composition, being the Greek text of the De compositione verborum (W | partial | dxf_facet_K_training_and_cause.bib | +| `dxd_foundation/douek2023meta.md` | Douek, E. | 2024 | The Meta Oversight Board and the empty promise of legitimacy | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/doyle2025assessment.md` | Doyle, A., & McCrae, J. P. | 2025 | An assessment of word separation practices in Old Irish text resources and a uni | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/dummler1881alcuini.md` | Dümmler, E. (Ed.). | 1881 | Poetae Latini aevi Carolini (Vol | partial | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/duzan2026chain.md` | Duzan, A., & Cooper Stickland, A. | 2026 | Chain-of-thought monitoring can be unreliable in implicit-influence settings [Pr | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/elish2019moral.md` | Elish, M. C. | 2019 | Moral crumple zones: Cautionary tales in human-robot interaction | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/eu2019p2b.md` | | | Regulation (EU) 2019/1150 of the European Parliament and of the Council of 20 Ju | partial | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/eu2022dsa.md` | European Parliament and Council of the European Un | 2022 | Regulation (EU) 2022/2065 of 19 October 2022 on a Single Market For Digital Serv | partial | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/eu2022dsa.md` | European Parliament and Council of the European Un | 2022 | Regulation (EU) 2022/2065 of 19 October 2022 on a Single Market For Digital Serv | partial | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/eu2024platform.md` | | | Directive (EU) 2024/2831 of the European Parliament and of the Council of 23 Oct | partial | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/eu2024regulation.md` | European Parliament and Council of the European Un | 2024 | Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14: Human overs | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/evans2018forex.md` | Evans, M. D. D. | 2018 | Forex trading and the WMR Fix | partial | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/fca2014fines.md` | Financial Conduct Authority. | 2014 | FCA fines five banks £1.1 billion for FX failings and announces industry-wide re | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/frey2026correlational.md` | Frey, A., Meary, D., Loichot, M., Mirault, J., & B | 2026 | Correlational analysis of distinct contributions and overlaps between visual, vi | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/fsb2014foreign.md` | Financial Stability Board. | 2014 | Foreign exchange benchmarks: Final report | partial | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/fsb2015progress.md` | Financial Stability Board. | 2015 | Foreign exchange benchmarks: Report on progress in implementing the September 20 | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/galanter1974why.md` | Galanter, M. | 1974 | Why the "haves" come out ahead: Speculations on the limits of legal change | partial | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/galliussi2020inter.md` | Galliussi, J., Perondi, L., Chia, G., Gerbino, W., | 2020 | Inter-letter spacing, inter-word spacing, and font with dyslexia-friendly featur | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/garrett1998saenger.md` | Garrett, J. | 1998 | Review of P | full | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/gavrilov1997techniques.md` | Gavrilov, A. K. | 1997 | Techniques of reading in classical antiquity | partial | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/gema2026chain.md` | Gema, A. P., Rajani, N., Saxena, R., Kwan, W.-C., | 2026 | Chain-of-thought faithfulness of reasoning models varies with where and how pref | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/globalfxc2024code.md` | | | Global Foreign Exchange Committee. (n.d.; footer of the file read: "Updated Dece | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/green2022flaws.md` | Green, B. | 2022 | The flaws of policies requiring human oversight of government algorithms | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/greenblatt2024ai.md` | Greenblatt, R., Shlegeris, B., Sachan, K., & Roger | 2024 | AI control: Improving safety despite intentional subversion | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/grundemclaughlin2026overseeing.md` | Grunde-McLaughlin, M., Mozannar, H., Murad, M., Ch | 2026 | Overseeing agents without constant oversight: Challenges and opportunities (arXi | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/haines2008origins.md` | Haines, J. | 2008 | The origins of the musical staff | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/hakvoort2017improvements.md` | Hakvoort, B., van den Boer, M., Leenaars, T., Bos, | 2017 | Improvements in reading accuracy as a result of increased interletter spacing ar | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/hangeulmuseum.md` | National Hangeul Museum. | 2015 | Read society through its Hangeul letters [Exhibition page, 21 April to 7 June 20 | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/hao2024training.md` | Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., W | 2024 | Training large language models to reason in a continuous latent space (arXiv:241 | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/heilmann2020reading.md` | Heilmann, J. | 2020 | Reading early New Testament manuscripts: Scriptio continua, "reading aids", and | full | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/hendrickson1929ancient.md` | Hendrickson, G. L. | 1929 | Ancient reading | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/hervais2019learning.md` | Hervais-Adelman, A., Kumar, U., Mishra, R. K., Tri | 2019 | Learning to read recycles visual cortical networks without destruction | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/houpt2014new.md` | Houpt, J. W., Townsend, J. T., & Donkin, C. | 2014 | A new perspective on visual word processing efficiency | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/hutchins1985direct.md` | Hutchins, E. L., Hollan, J. D., & Norman, D. A. | 1985 | Direct manipulation interfaces | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/ivanova2026process.md` | Ivanova, Y. | 2026 | A process of implementing the Korean alphabet on the Korean peninsula [Bulgarian | full | dxf_facet_C_spacing_to_literacy.bib | +| `dxd_foundation/ivanova2026process.md` | Ivanova, Y. | 2026 | A process of implementing the Korean alphabet on the Korean peninsula [Bulgarian | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/jakkal2026secure.md` | Jakkal, V. | 2026 | Secure agentic AI for your Frontier Transformation | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/jhaver2019suspect.md` | Jhaver, S., Appling, D. S., Gilbert, E., & Bruckma | 2019 | "Did you suspect the post would be removed?": Understanding user reactions to co | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/jia2026faithfulness.md` | Jia, J., Benton, J., & Easley, E. | 2026 | Faithfulness as information flow: Evaluating and training faithful chain-of-thou | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/jordan2024reicher.md` | Jordan, T. R., Akkaya, A. M., Göçmüş, F. Z., Kalan | 2024 | The Reicher-Wheeler paradigm in word recognition research: a cautionary note on | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/kasisopa2016child.md` | Kasisopa, B., Reilly, R. G., Luksaneeyanawin, S., | 2016 | Child readers' eye movements in reading Thai | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/kellogg2020algorithms.md` | Kellogg, K. C., Valentine, M. A., & Christin, A. | 2020 | Algorithms at work: The new contested terrain of control | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/kellogg2020algorithms.md` | Kellogg, K. C., Valentine, M. A., & Christin, A. | 2020 | Algorithms at work: The new contested terrain of control | full | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/ker2012johnson.md` | Ker, J. | 2012 | Review of William A | full | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/klonick2018new.md` | Klonick, K. | 2018 | The new governors: The people, rules, and processes governing online speech | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/knox1968silent.md` | Knox, B. M. W. | 1968 | Silent reading in antiquity | full | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/kohsom1997adding.md` | Kohsom, C., & Gobet, F. | 1997 | Adding spaces to Thai and English: Effects on reading | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/kokotajlo2025ai.md` | Kokotajlo, D., Alexander, S., Larsen, T., Lifland, | 2025 | AI 2027 [Forecasting scenario, web page] | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/korbak2025chain.md` | Korbak, T., Balesni, M., Barnes, E., Bengio, Y., B | 2025 | Chain of thought monitorability: A new and fragile opportunity for AI safety (ar | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/krieber2017eye.md` | Krieber, M., Bartl-Pokorny, K. D., Pokorny, F. B., | 2017 | Eye movements during silent and oral reading in a regular orthography: Basic cha | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/kroll2021outlining.md` | Kroll, J. A. | 2021 | Outlining traceability: A principle for operationalizing accountability in compu | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/kubota2024rethinking.md` | Kubota, E., Grill-Spector, K., & Nordt, M. | 2024 | Rethinking cortical recycling in ventral temporal cortex | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/kumari2026nadella.md` | Kumari, S. | 2026 | Satya Nadella warns of AI insider risks, seeks stronger safety controls | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/kutas2011thirty.md` | Kutas, M., & Federmeier, K. D. | 2011 | Thirty years and counting: Finding meaning in the N400 component of the event-re | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/kutasov2025shade.md` | Kutasov, J., Sun, Y., Colognese, P., van der Weij, | 2025 | SHADE-Arena: Evaluating sabotage and monitoring in LLM agents (arXiv:2506.15740) | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/langer2024effective.md` | Langer, M., Baum, K., & Schlicker, N. | 2025 | Effective human oversight of AI-based systems: A signal detection perspective on | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/lanham2023measuring.md` | Lanham, T., Chen, A., Radhakrishnan, A., Steiner, | 2023 | Measuring faithfulness in chain-of-thought reasoning (arXiv:2307.13702) | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/larkin1987why.md` | Larkin, J. H., & Simon, H. A. | 1987 | Why a diagram is (sometimes) worth ten thousand words | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/laubrock2015eyevoice.md` | Laubrock, J., & Kliegl, R. | 2015 | The eye-voice span during reading aloud | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/laux2024institutionalised.md` | Laux, J. | 2024 | Institutionalised distrust and human oversight of artificial intelligence: Towar | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/layman2023controlled.md` | Layman, L., & Roden, W. | 2023 | A controlled experiment on the impact of intrusion detection false alarm rate on | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/leinenger2014phonological.md` | Leinenger, M. | 2014 | Phonological coding during reading | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/li2014reading.md` | Li, X., Bicknell, K., Liu, P., Wei, W., & Rayner, | 2014 | Reading is fundamentally similar across disparate writing systems: A systematic | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/little2026length.md` | Little, B. | 2026 | Length penalties make chain-of-thought less monitorable | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/lynch2025agentic.md` | Lynch, A., Wright, B., Larson, C., Ritchie, S. J., | 2025 | Agentic misalignment: How LLMs could be insider threats | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/mannering2012johnson.md` | Mannering, J. | 2012 | Review of William A | full | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/margono1987study.md` | Margono, S., & Shneiderman, B. | 1987 | A study of file manipulation by novices using commands vs | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/mathews1976eldridge.md` | *Mathews v. Eldridge*, 424 U.S. 319 | 1976 | | partial | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/matzinger2021influence.md` | Matzinger, T., Ritt, N., & Fitch, W. T. | 2021 | The influence of different prosodic cues on word segmentation | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/mirault2019precise.md` | Mirault, J., Snell, J., & Grainger, J. | 2019 | Reading without spaces: The role of precise letter order | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/mirault2019reading.md` | Mirault, J., Snell, J., & Grainger, J. | 2019 | Reading without spaces revisited: The role of word identification and sentence-l | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/mitchell2026agents.md` | Mitchell, M., Ghosh, A., & Passi, S. | 2026 | AI agents push humans out of the loop | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/nadella2025looking.md` | Nadella, S. | 2025 | Looking ahead to 2026 | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/nadella2026frontier.md` | Nadella, S. | 2026 | A frontier without an ecosystem is not stable | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/nadella2026models.md` | Nadella, S. | 2026 | Models as insider risks in the super intelligence era | full | bec_facet_V1_nadella_disclosure_paradigm.bib | +| `dxd_foundation/nadella2026models.md` | Nadella, S. | 2026 | Models as insider risks in the super intelligence era | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/nadella2026reverse.md` | Nadella, S. | 2026 | The reverse information paradox | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/nagarajrao2026fareshare.md` | Nagaraj Rao, V., Dalal, S., Schwartz, A., Liaqat, | 2026 | FareShare: A tool for labor organizers to estimate lost wages and contest arbitr | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/nothaft2020satanic.md` | Nothaft, C. P. E. | 2020 | Medieval Europe's satanic ciphers: On the genesis of a modern myth | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/nydfs2015barclays.md` | New York State Department of Financial Services. | 2015 | In the Matter of Barclays Bank PLC, Barclays Bank PLC, New York Branch: Consent | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/nydfs2015barclayspr.md` | New York State Department of Financial Services. | 2015 | NYDFS announces Barclays to pay additional $150 million penalty, terminate emplo | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/odonnell1994parkes.md` | O'Donnell, J. J. | 1994 | Review of M | full | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/odonnell2000saenger.md` | O'Donnell, J. J. | 2000 | Review of Paul Saenger, Space Between Words: The Origins of Silent Reading (Stan | full | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/odonnell2000saenger.md` | O'Donnell, J. J. | 2000 | Review of Paul Saenger, Space Between Words: The Origins of Silent Reading (Stan | full | dxf_facet_K_training_and_cause.bib | +| `dxd_foundation/oomen2016last.md` | Oomen, R. | 2017 | Last look | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/oversight2022q1.md` | Oversight Board. | 2022 | Q1 2022 transparency report [Report covering 1 January–31 March 2022]. | partial | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/ozkan2026wordinitial.md` | Özkan, Z. G., & Perea, M. | 2026 | Word-initial capitalization preserves lexical processing in unspaced text | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/palka2025foreign.md` | Pałka, P., Porębski, A., Pałosz, R., & Wiśniewska, | 2025 | Do foreign service providers simply ignore the EU consumer law? An empirical stu | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/parasuraman2010complacency.md` | Parasuraman, R., & Manzey, D. H. | 2010 | Complacency and bias in human use of automation: An attentional integration | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/park2025dokkyo.md` | Park, J.-B. | 2025 | Diversification of literacy patterns by social class in late Joseon (forum abstr | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/parkes1992pause.md` | Parkes, M. B. | 1992 | Pause and effect: An introduction to the history of punctuation in the West | partial | dxf_facet_A_the_record.bib | +| `dxd_foundation/parkes1992pause.md` | Parkes, M. B. | 1992 | Pause and effect: An introduction to the history of punctuation in the West | partial | dxf_facet_K_training_and_cause.bib | +| `dxd_foundation/perea2012subtle.md` | Perea, M., & Gomez, P. | 2012 | Subtle increases in interletter spacing facilitate the encoding of words during | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/perea2015colours.md` | Perea, M., Tejero, P., & Winskel, H. | 2015 | Can colours be used to segment words when reading? Acta Psychologica, 159, 8–13 | full | dxf_facet_W_spacing_oracy_literacy.bib | +| `dxd_foundation/price2011interactive.md` | Price, C. J., & Devlin, J. T. | 2011 | The interactive account of ventral occipitotemporal contributions to reading | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/quintilian1920institutio.md` | Quintilian. | 1920 | The orator's education [Institutio oratoria], Book 1 (H | partial | dxf_facet_K_training_and_cause.bib | +| `dxd_foundation/rayner1998unspaced.md` | Rayner, K., Fischer, M. H., & Pollatsek, A. | 1998 | Unspaced text interferes with both word identification and eye movement control | partial | dxf_facet_D_reading_science.bib | +| `dxd_foundation/reis2011johnson.md` | Reis, D. M. | 2011 | Review of William A | full | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/rfc8890.md` | Nottingham, M. | 2020 | The Internet is for end users (RFC 8890) | full | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/rieder2020towards.md` | Rieder, B., & Hofmann, J. | 2020 | Towards platform observability | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/rose2020zero.md` | Rose, S., Borchert, O., Mitchell, S., & Connelly, | 2020 | Zero trust architecture (NIST Special Publication 800-207) | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/rosenblat2016algorithmic.md` | Rosenblat, A., & Stark, L. | 2016 | Algorithmic labor and information asymmetries: A case study of Uber's drivers | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/rosenblat2016algorithmic.md` | Rosenblat, A., & Stark, L. | 2016 | Algorithmic labor and information asymmetries: A case study of Uber's drivers | full | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/ryan2014historical.md` | Ryan, J. | 2014 | Historical note: Did double-entry bookkeeping contribute to economic development | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/saenger1997space.md` | Saenger, P. | 1997 | Space between words: The origins of silent reading | partial | dxf_facet_A_the_record.bib | +| `dxd_foundation/saenger1997space.md` | Saenger, P. | 1997 | Space between words: The origins of silent reading | partial | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/saenger1997space.md` | Saenger, P. | 1997 | Space between words: The origins of silent reading | partial | dxf_facet_C_spacing_to_literacy.bib | +| `dxd_foundation/saenger1997space.md` | Saenger, P. | 1997 | Space between words: The origins of silent reading | partial | dxf_facet_K_training_and_cause.bib | +| `dxd_foundation/saltzer1975protection.md` | Saltzer, J. H., & Schroeder, M. D. | 1975 | The protection of information in computer systems | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/sangster2024emergence.md` | Sangster, A. | 2024 | The emergence of double entry bookkeeping | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/sangster2025reboot.md` | Sangster, A. | 2025 | Time to reboot accounting history: Evidence from scholarship on double entry | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/santonidesio2018meaningful.md` | Santoni de Sio, F., & van den Hoven, J. | 2018 | Meaningful human control over autonomous systems: A philosophical account | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/schlimm2008modeling.md` | Schlimm, D., & Neth, H. | 2008 | Modeling ancient and modern arithmetic practices: Addition and multiplication wi | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/scribner1981psychology.md` | Scribner, S., & Cole, M. | 1981 | The psychology of literacy | partial | dxf_facet_C_spacing_to_literacy.bib | +| `dxd_foundation/seattle2024deactivation.md` | Seattle Office of Labor Standards. | 2024 | Model notice of deactivation (based on requirements of SMC 8.40.070) [Form]. | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/seymour2003foundation.md` | Seymour, P. H. K., Aro, M., & Erskine, J. M. | 2003 | Foundation literacy acquisition in European orthographies | full | dxf_facet_C_spacing_to_literacy.bib | +| `dxd_foundation/shneiderman1983direct.md` | Shneiderman, B. | 1983 | Direct manipulation: A step beyond programming languages | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/sterz2024quest.md` | Sterz, S., Baum, K., Biewer, S., Hermanns, H., Lau | 2024 | On the quest for effectiveness in human oversight: Interdisciplinary perspective | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/stohl2016managing.md` | Stohl, C., Stohl, M., & Leonardi, P. M. | 2016 | Managing opacity: Information visibility and the paradox of transparency in the | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/strunk1998guido.md` | | | Guido of Arezzo. Prologue to his Antiphoner; *Epistola de ignoto cantu* (excerpt | full | dxf_facet_F_parallel_moves.bib | +| `dxd_foundation/suzor2016responsibilities.md` | Suzor, N. | 2016 | The responsibilities of platforms: A new constitutionalism to promote the legiti | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/suzor2018digital.md` | Suzor, N. | 2018 | Digital constitutionalism: Using the rule of law to evaluate the legitimacy of g | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/thrax1874grammar.md` | Dionysios Thrax [attributed]. | 1874 | The grammar of Dionysios Thrax (T | partial | dxf_facet_K_training_and_cause.bib | +| `dxd_foundation/torrensalvarez2022latin.md` | Torrens-Álvarez, M. J., & Tuten, D. N. | 2022 | From "Latin" to the vernacular: Latin-Romance hybridity, scribal competence, and | partial | dxf_facet_B_the_dispute.bib | +| `dxd_foundation/tumey1927ohio.md` | *Tumey v. Ohio*, 273 U.S. 510 | 1927 | | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/tunbridge1992study.md` | Tunbridge, G. L. | 1992 | A study of scribal practices in early Irish and Anglo-Saxon manuscripts [DPhil t | partial | dxf_facet_A_the_record.bib | +| `dxd_foundation/turan2026oversight.md` | Turan, E. | 2026 | Oversight has a capacity: Calibrating agent guards to a subjective, fatiguing hu | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/turpin2023language.md` | Turpin, M., Michael, J., Perez, E., & Bowman, S. R | 2023 | Language models don't always say what they think: Unfaithful explanations in cha | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/tyler1988what.md` | Tyler, T. R. | 1988 | What is procedural justice?: Criteria used by citizens to assess the fairness of | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/us2072usc.md` | | | 28 U.S.C. § 2072 — Rules of procedure and evidence; power to prescribe. | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/us2073usc.md` | | | 28 U.S.C. § 2073 — Rules of procedure and evidence; method of prescribing. | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/us455usc.md` | | | 28 U.S.C. § 455 — Disqualification of justice, judge, or magistrate judge. | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/us9usc10.md` | | | 9 U.S.C. § 10 — Same; vacation; grounds; rehearing (Federal Arbitration Act). | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/vandenboer2015default.md` | van den Boer, M., & Hakvoort, B. E. | 2015 | Default spacing is the optimal spacing for word reading | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/wagemans2012century.md` | Wagemans, J., Elder, J. H., Kubovy, M., Palmer, S. | 2012 | A century of Gestalt psychology in visual perception: I | partial | dxf_facet_D_reading_science.bib | +| `dxd_foundation/wang2024tibetan.md` | Wang, D., Niu, D., Li, et al. | 2024 | The effect of visual word segmentation cues in Tibetan reading | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/wang2026monitorbench.md` | Wang, H., Sun, Y., Ko, B., Talati, M., Gong, J., L | 2026 | MonitorBench: A comprehensive benchmark for chain-of-thought monitorability in l | partial | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/wertli2023influence.md` | Wertli, J., Schötzau, A., & Palmowski-Wolfe, A. | 2023 | The influence of age on eye movements during reading in early elementary school | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/westwell2018dissemination.md` | Westwell, A. | 2018 | The dissemination and reception of the Ordines Romani in the Carolingian church, | partial | dxf_facet_O_custody_of_conventions.bib | +| `dxd_foundation/winskel2009eye.md` | Winskel, H., Radach, R., & Luksaneeyanawin, S. | 2009 | Eye movements when reading spaced and unspaced Thai and English: A comparison of | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/wong2022meta.md` | Wong, D., & Floridi, L. | 2023 | Meta's Oversight Board: A review and critical assessment | full | dxf_facet_J_who_holds_the_convention.bib | +| `dxd_foundation/yu2023altered.md` | Yu, H., & Kwon, M. | 2023 | Altered eye movements during reading with simulated central and peripheral visua | full | dxf_facet_D_reading_science.bib | +| `dxd_foundation/yu2026habituation.md` | Yu, H., Liu, L., Jiang, X., Jia, Y., Wang, S., Qia | 2026 | Habituation at the gate: Rising approval and declining scrutiny in human review | full | dxf_facet_G_nadella_lineage.bib | +| `dxd_foundation/ziegler2010orthographic.md` | Ziegler, J. C., Bertrand, D., Tóth, D., Csépe, V., | 2010 | Orthographic depth and its impact on universal predictors of reading: A cross-la | full | dxf_facet_C_spacing_to_literacy.bib | + +## Keys with no card + +- `abovesecurity2026satya` (dxf_facet_G_nadella_lineage.bib) +- `linzainescu2026microsofts` (dxf_facet_G_nadella_lineage.bib) +- `mackworth1948breakdown` (dxf_facet_G_nadella_lineage.bib) +- `tariq2025alert` (dxf_facet_G_nadella_lineage.bib) +- `warm2008vigilance` (dxf_facet_G_nadella_lineage.bib) +- `bronner2018nonseparation` (dxf_facet_A_the_record.bib) +- `cavallo1999history` (dxf_facet_A_the_record.bib) +- `saenger1982silent` (dxf_facet_A_the_record.bib) +- `saenger2020orality` (dxf_facet_A_the_record.bib) +- `wingo1972latin` (dxf_facet_A_the_record.bib) +- `aronoff2002saenger` (dxf_facet_B_the_dispute.bib) +- `bauml2002saenger` (dxf_facet_B_the_dispute.bib) +- `cavallo1999history` (dxf_facet_B_the_dispute.bib) +- `feldt2023reading` (dxf_facet_B_the_dispute.bib) +- `flory2007silent` (dxf_facet_B_the_dispute.bib) +- `gilliard1993more` (dxf_facet_B_the_dispute.bib) +- `hanna1999saenger` (dxf_facet_B_the_dispute.bib) +- `heilmann2021lesen` (dxf_facet_B_the_dispute.bib) +- `johnson2000toward` (dxf_facet_B_the_dispute.bib) +- `johnson2010readers` (dxf_facet_B_the_dispute.bib) +- `lyons2001saenger` (dxf_facet_B_the_dispute.bib) +- `mccutcheon2015silent` (dxf_facet_B_the_dispute.bib) +- `mollon2001scriptura` (dxf_facet_B_the_dispute.bib) +- `mulchahey2001saenger` (dxf_facet_B_the_dispute.bib) +- `richter2001saenger` (dxf_facet_B_the_dispute.bib) +- `slusser1992reading` (dxf_facet_B_the_dispute.bib) +- `svenbro1993phrasikleia` (dxf_facet_B_the_dispute.bib) +- `vatri2012physiology` (dxf_facet_B_the_dispute.bib) +- `wingo1972latin` (dxf_facet_B_the_dispute.bib) +- `cipolla1969literacy` (dxf_facet_C_spacing_to_literacy.bib) +- `filiztekin2025unification` (dxf_facet_C_spacing_to_literacy.bib) +- `graff1987legacies` (dxf_facet_C_spacing_to_literacy.bib) +- `harris1989ancient` (dxf_facet_C_spacing_to_literacy.bib) +- `hong2018colonization` (dxf_facet_C_spacing_to_literacy.bib) +- `houston2002literacy` (dxf_facet_C_spacing_to_literacy.bib) +- `kuzuoglu2022basic` (dxf_facet_C_spacing_to_literacy.bib) +- `li2004reexamination` (dxf_facet_C_spacing_to_literacy.bib) +- `mckitterick1989carolingians` (dxf_facet_C_spacing_to_literacy.bib) +- `pae2011korean` (dxf_facet_C_spacing_to_literacy.bib) +- `saenger1982silent` (dxf_facet_C_spacing_to_literacy.bib) +- `saenger1989books` (dxf_facet_C_spacing_to_literacy.bib) +- `saenger1991separation` (dxf_facet_C_spacing_to_literacy.bib) +- `sherratt1989language` (dxf_facet_C_spacing_to_literacy.bib) +- `stock1983implications` (dxf_facet_C_spacing_to_literacy.bib) +- `thomas1992literacy` (dxf_facet_C_spacing_to_literacy.bib) +- `weng2023stop` (dxf_facet_C_spacing_to_literacy.bib) +- `wood1929latinizing` (dxf_facet_C_spacing_to_literacy.bib) +- `yilmaz2011learning` (dxf_facet_C_spacing_to_literacy.bib) +- `zanker1991ancient` (dxf_facet_C_spacing_to_literacy.bib) +- `ziegler2005reading` (dxf_facet_C_spacing_to_literacy.bib) +- `zieman2008singing` (dxf_facet_C_spacing_to_literacy.bib) +- `arditi2007letter` (dxf_facet_D_reading_science.bib) +- `baek2022korean` (dxf_facet_D_reading_science.bib) +- `bai2013interword` (dxf_facet_D_reading_science.bib) +- `blythe2012inserting` (dxf_facet_D_reading_science.bib) +- `bouma1970interaction` (dxf_facet_D_reading_science.bib) +- `cattell1886time` (dxf_facet_D_reading_science.bib) +- `christofalos2024impact` (dxf_facet_D_reading_science.bib) +- `cohen2008reading` (dxf_facet_D_reading_science.bib) +- `dehaene2007cultural` (dxf_facet_D_reading_science.bib) +- `dehaene2010learning` (dxf_facet_D_reading_science.bib) +- `dehaene2015illiterate` (dxf_facet_D_reading_science.bib) +- `dotan2018mind` (dxf_facet_D_reading_science.bib) +- `drieghe2017parafoveal` (dxf_facet_D_reading_science.bib) +- `duranovic2018influence` (dxf_facet_D_reading_science.bib) +- `epelboim1994reading` (dxf_facet_D_reading_science.bib) +- `kasisopa2013eye` (dxf_facet_D_reading_science.bib) +- `li2020integrated` (dxf_facet_D_reading_science.bib) +- `li2021effects` (dxf_facet_D_reading_science.bib) +- `lin2017beautiful` (dxf_facet_D_reading_science.bib) +- `luniewska2022effect` (dxf_facet_D_reading_science.bib) +- `morris1990parafoveal` (dxf_facet_D_reading_science.bib) +- `pelli2008uncrowded` (dxf_facet_D_reading_science.bib) +- `perea2009space` (dxf_facet_D_reading_science.bib) +- `perea2012effects` (dxf_facet_D_reading_science.bib) +- `pollatsek1982eye` (dxf_facet_D_reading_science.bib) +- `price2003myth` (dxf_facet_D_reading_science.bib) +- `rayner1986eye` (dxf_facet_D_reading_science.bib) +- `rayner1998eye` (dxf_facet_D_reading_science.bib) +- `rayner2009thirtyfifth` (dxf_facet_D_reading_science.bib) +- `raynerpollatsek1996reading` (dxf_facet_D_reading_science.bib) +- `reicher1969perceptual` (dxf_facet_D_reading_science.bib) +- `sainio2007role` (dxf_facet_D_reading_science.bib) +- `shen2012eye` (dxf_facet_D_reading_science.bib) +- `sheridan2013unsegmented` (dxf_facet_D_reading_science.bib) +- `skottun2012interletter` (dxf_facet_D_reading_science.bib) +- `slattery2016interword` (dxf_facet_D_reading_science.bib) +- `stagg2021room` (dxf_facet_D_reading_science.bib) +- `vatri2012physiology` (dxf_facet_D_reading_science.bib) +- `veldre2017spelling` (dxf_facet_D_reading_science.bib) +- `wang2025spaces` (dxf_facet_D_reading_science.bib) +- `wheeler1970processes` (dxf_facet_D_reading_science.bib) +- `wickens1995proximity` (dxf_facet_D_reading_science.bib) +- `zang2013interword` (dxf_facet_D_reading_science.bib) +- `zhou2023fixation` (dxf_facet_D_reading_science.bib) +- `zorzi2012extra` (dxf_facet_D_reading_science.bib) +- `zorzi2012reply` (dxf_facet_D_reading_science.bib) +- `abbamonte2023indexing` (dxf_facet_F_parallel_moves.bib) +- `baek2024spread` (dxf_facet_F_parallel_moves.bib) +- `bender2002signs` (dxf_facet_F_parallel_moves.bib) +- `blair2010too` (dxf_facet_F_parallel_moves.bib) +- `burnett2024learning` (dxf_facet_F_parallel_moves.bib) +- `cajori1928notations1` (dxf_facet_F_parallel_moves.bib) +- `cajori1929notations2` (dxf_facet_F_parallel_moves.bib) +- `carruthers1991accounting` (dxf_facet_F_parallel_moves.bib) +- `chrisomalis2010numerical` (dxf_facet_F_parallel_moves.bib) +- `chrisomalis2020reckonings` (dxf_facet_F_parallel_moves.bib) +- `conraulewis2021index` (dxf_facet_F_parallel_moves.bib) +- `cushman2011syllabary` (dxf_facet_F_parallel_moves.bib) +- `danna2021figuring` (dxf_facet_F_parallel_moves.bib) +- `david1985clio` (dxf_facet_F_parallel_moves.bib) +- `flanders2023place` (dxf_facet_F_parallel_moves.bib) +- `garciaretamero2013visual` (dxf_facet_F_parallel_moves.bib) +- `gigerenzer1995improve` (dxf_facet_F_parallel_moves.bib) +- `handel2016learn` (dxf_facet_F_parallel_moves.bib) +- `hibbard2003supporting` (dxf_facet_F_parallel_moves.bib) +- `hossain2009quest` (dxf_facet_F_parallel_moves.bib) +- `james2024otis` (dxf_facet_F_parallel_moves.bib) +- `kay2013rerun` (dxf_facet_F_parallel_moves.bib) +- `liebowitz1990fable` (dxf_facet_F_parallel_moves.bib) +- `mengozzi2007si` (dxf_facet_F_parallel_moves.bib) +- `mengozzi2010renaissance` (dxf_facet_F_parallel_moves.bib) +- `mews2024guidonian` (dxf_facet_F_parallel_moves.bib) +- `montgomeryanderson2014review` (dxf_facet_F_parallel_moves.bib) +- `norman1982alphabetic` (dxf_facet_F_parallel_moves.bib) +- `overmann2022review` (dxf_facet_F_parallel_moves.bib) +- `parins2013literacy` (dxf_facet_F_parallel_moves.bib) +- `pesce1999guido` (dxf_facet_F_parallel_moves.bib) +- `peters2006numeracy` (dxf_facet_F_parallel_moves.bib) +- `peters2007less` (dxf_facet_F_parallel_moves.bib) +- `peters2007numeracy` (dxf_facet_F_parallel_moves.bib) +- `power2000history` (dxf_facet_F_parallel_moves.bib) +- `rouse1982statim` (dxf_facet_F_parallel_moves.bib) +- `sangster2016genesis` (dxf_facet_F_parallel_moves.bib) +- `thomann2020serial` (dxf_facet_F_parallel_moves.bib) +- `unseth2016international` (dxf_facet_F_parallel_moves.bib) +- `vera2009decimal` (dxf_facet_F_parallel_moves.bib) +- `walker1993early` (dxf_facet_F_parallel_moves.bib) +- `white1962revival` (dxf_facet_F_parallel_moves.bib) +- `yamey1949scientific` (dxf_facet_F_parallel_moves.bib) +- `yamey1964accounting` (dxf_facet_F_parallel_moves.bib) +- `yamey2005historical` (dxf_facet_F_parallel_moves.bib) +- `zerubavel1977republican` (dxf_facet_F_parallel_moves.bib) +- `zhang1995representational` (dxf_facet_F_parallel_moves.bib) +- `bernstein1992opting` (dxf_facet_J_who_holds_the_convention.bib) +- `bernstein2001private` (dxf_facet_J_who_holds_the_convention.bib) +- `bis2019fx` (dxf_facet_J_who_holds_the_convention.bib) +- `colquitt2001justice` (dxf_facet_J_who_holds_the_convention.bib) +- `fuller1978forms` (dxf_facet_J_who_holds_the_convention.bib) +- `gillespie2019custodians` (dxf_facet_J_who_holds_the_convention.bib) +- `greif1989reputation` (dxf_facet_J_who_holds_the_convention.bib) +- `jhaver2019transparency` (dxf_facet_J_who_holds_the_convention.bib) +- `lind1988social` (dxf_facet_J_who_holds_the_convention.bib) +- `lind1993individual` (dxf_facet_J_who_holds_the_convention.bib) +- `maffie2025enshittification` (dxf_facet_J_who_holds_the_convention.bib) +- `milgrom1990role` (dxf_facet_J_who_holds_the_convention.bib) +- `myerswest2018censored` (dxf_facet_J_who_holds_the_convention.bib) +- `roth2002economist` (dxf_facet_J_who_holds_the_convention.bib) +- `thibaut1975procedural` (dxf_facet_J_who_holds_the_convention.bib) +- `tyler1990why` (dxf_facet_J_who_holds_the_convention.bib) +- `tyler2019social` (dxf_facet_J_who_holds_the_convention.bib) +- `vaccaro2020end` (dxf_facet_J_who_holds_the_convention.bib) +- `vaccaro2021contestability` (dxf_facet_J_who_holds_the_convention.bib) +- `vanderdonk2022ex` (dxf_facet_J_who_holds_the_convention.bib) +- `aronoff2002saenger` (dxf_facet_K_training_and_cause.bib) +- `baron2000alphabet` (dxf_facet_K_training_and_cause.bib) +- `bauml2002saenger` (dxf_facet_K_training_and_cause.bib) +- `black2001humanism` (dxf_facet_K_training_and_cause.bib) +- `blake2002literacy` (dxf_facet_K_training_and_cause.bib) +- `bonner1977education` (dxf_facet_K_training_and_cause.bib) +- `carruthers2008book` (dxf_facet_K_training_and_cause.bib) +- `cavallo1999history` (dxf_facet_K_training_and_cause.bib) +- `chartier1994order` (dxf_facet_K_training_and_cause.bib) +- `cribiore1996writing` (dxf_facet_K_training_and_cause.bib) +- `cribiore2001gymnastics` (dxf_facet_K_training_and_cause.bib) +- `grendler1989schooling` (dxf_facet_K_training_and_cause.bib) +- `hanna1999saenger` (dxf_facet_K_training_and_cause.bib) +- `illich1993vineyard` (dxf_facet_K_training_and_cause.bib) +- `johnson2010readers` (dxf_facet_K_training_and_cause.bib) +- `lyons2010history` (dxf_facet_K_training_and_cause.bib) +- `mak2011page` (dxf_facet_K_training_and_cause.bib) +- `marrou1956history` (dxf_facet_K_training_and_cause.bib) +- `morgan1998literate` (dxf_facet_K_training_and_cause.bib) +- `mulchahey2001saenger` (dxf_facet_K_training_and_cause.bib) +- `orme2001medieval` (dxf_facet_K_training_and_cause.bib) +- `orme2006medievalschools` (dxf_facet_K_training_and_cause.bib) +- `raritan2000review` (dxf_facet_K_training_and_cause.bib) +- `reynolds1996medieval` (dxf_facet_K_training_and_cause.bib) +- `riche1976education` (dxf_facet_K_training_and_cause.bib) +- `richter2001saenger` (dxf_facet_K_training_and_cause.bib) +- `rouse1982statim` (dxf_facet_K_training_and_cause.bib) +- `saenger1982silent` (dxf_facet_K_training_and_cause.bib) +- `saenger1995illich` (dxf_facet_K_training_and_cause.bib) +- `saenger1999history` (dxf_facet_K_training_and_cause.bib) +- `svenbro1999archaic` (dxf_facet_K_training_and_cause.bib) +- `akhawe2013alice` (dxf_facet_O_custody_of_conventions.bib) +- `boe2021rdl9` (dxf_facet_O_custody_of_conventions.bib) +- `chromium2018evolving` (dxf_facet_O_custody_of_conventions.bib) +- `cox2010review` (dxf_facet_O_custody_of_conventions.bib) +- `dhamija2006why` (dxf_facet_O_custody_of_conventions.bib) +- `doellgast2022negotiating` (dxf_facet_O_custody_of_conventions.bib) +- `doj2024ada` (dxf_facet_O_custody_of_conventions.bib) +- `doj2026extension` (dxf_facet_O_custody_of_conventions.bib) +- `dumville1987english` (dxf_facet_O_custody_of_conventions.bib) +- `egelman2008warned` (dxf_facet_O_custody_of_conventions.bib) +- `eu2019eaa` (dxf_facet_O_custody_of_conventions.bib) +- `eurofound2024hilfr` (dxf_facet_O_custody_of_conventions.bib) +- `eurofound2025glovo` (dxf_facet_O_custody_of_conventions.bib) +- `felt2016rethinking` (dxf_facet_O_custody_of_conventions.bib) +- `ganderats2023front` (dxf_facet_O_custody_of_conventions.bib) +- `ganz1995book` (dxf_facet_O_custody_of_conventions.bib) +- `gillespie2019custodians` (dxf_facet_O_custody_of_conventions.bib) +- `gorwa2019what` (dxf_facet_O_custody_of_conventions.bib) +- `hatton2026product` (dxf_facet_O_custody_of_conventions.bib) +- `helpnet2023lock` (dxf_facet_O_custody_of_conventions.bib) +- `holznagel2025shortcomings` (dxf_facet_O_custody_of_conventions.bib) +- `mckitterick1989carolingians` (dxf_facet_O_custody_of_conventions.bib) +- `mostert2016pray` (dxf_facet_O_custody_of_conventions.bib) +- `neumann2025authority` (dxf_facet_O_custody_of_conventions.bib) +- `parkes2016carolingian` (dxf_facet_O_custody_of_conventions.bib) +- `piazzoni11bookhands` (dxf_facet_O_custody_of_conventions.bib) +- `raji2020closing` (dxf_facet_O_custody_of_conventions.bib) +- `rfc2026` (dxf_facet_O_custody_of_conventions.bib) +- `rfc7282` (dxf_facet_O_custody_of_conventions.bib) +- `shangguan2019meta` (dxf_facet_O_custody_of_conventions.bib) +- `stadelmann2018how` (dxf_facet_O_custody_of_conventions.bib) +- `steinova2015psalmos` (dxf_facet_O_custody_of_conventions.bib) +- `w3c2024wcag22` (dxf_facet_O_custody_of_conventions.bib) +- `w3c2025process` (dxf_facet_O_custody_of_conventions.bib) +- `waechter2015desired` (dxf_facet_O_custody_of_conventions.bib) +- `wageindicator2022justeat` (dxf_facet_O_custody_of_conventions.bib) +- `busch2015zusammenschreibung` (dxf_facet_W_spacing_oracy_literacy.bib) +- `cushing2024social` (dxf_facet_W_spacing_oracy_literacy.bib) +- `davidson2019writing` (dxf_facet_W_spacing_oracy_literacy.bib) +- `havron2016reading` (dxf_facet_W_spacing_oracy_literacy.bib) +- `homer1999literacy` (dxf_facet_W_spacing_oracy_literacy.bib) +- `johnson2001word` (dxf_facet_W_spacing_oracy_literacy.bib) +- `kolinsky1987awareness` (dxf_facet_W_spacing_oracy_literacy.bib) +- `kukkonen2024voices` (dxf_facet_W_spacing_oracy_literacy.bib) +- `mattys2005integration` (dxf_facet_W_spacing_oracy_literacy.bib) +- `moffat2011grammar` (dxf_facet_W_spacing_oracy_literacy.bib) +- `morais1979does` (dxf_facet_W_spacing_oracy_literacy.bib) +- `morais1986literacy` (dxf_facet_W_spacing_oracy_literacy.bib) +- `moran2025learning` (dxf_facet_W_spacing_oracy_literacy.bib) +- `olson1996towards` (dxf_facet_W_spacing_oracy_literacy.bib) +- `osullivan2017reading` (dxf_facet_W_spacing_oracy_literacy.bib) +- `parkes1997raedan` (dxf_facet_W_spacing_oracy_literacy.bib) +- `veldhuis2012offline` (dxf_facet_W_spacing_oracy_literacy.bib) +- `vorstius2014eye` (dxf_facet_W_spacing_oracy_literacy.bib) +- `darnton1995censorship` (bec_facet_B_print_1450_1800.bib) +- `febvre1976coming` (bec_facet_B_print_1450_1800.bib) +- `houston2003lesser` (bec_facet_B_print_1450_1800.bib) +- `manuzio1564index` (bec_facet_B_print_1450_1800.bib) +- `martel2001autour` (bec_facet_B_print_1450_1800.bib) +- `sanchezmarti2023advent` (bec_facet_B_print_1450_1800.bib) +- `sasaki2017publishing` (bec_facet_B_print_1450_1800.bib) +- `schwartz2017did` (bec_facet_B_print_1450_1800.bib) +- `statutesrealm1817vol3` (bec_facet_B_print_1450_1800.bib) +- `suggett2018language` (bec_facet_B_print_1450_1800.bib) +- `villerscotterets1539lexilogos` (bec_facet_B_print_1450_1800.bib) diff --git a/submissions/berthon_ethics_paper/nadella/CORPUS.md b/submissions/berthon_ethics_paper/nadella/CORPUS.md new file mode 100644 index 00000000..727eea77 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/CORPUS.md @@ -0,0 +1,78 @@ +# Corpus for the Nadella analysis + +*Written by Claude on 2026-10-24.* Everything the research can start from, in four groups: Nadella's own texts, +reports on the X article, the lab's earlier analyses of it, and the cards behind those analyses. + +## 1. Nadella's texts (full text held locally, not in this repo) + +The repo is public and the X capture is marked for local use only, so the full texts sit in +`~/iit-playground/berthon_ethics_local/nadella/` (its README lists each file). Cite paragraphs as P1–P21, the +capturer's locators, which the lab's earlier work (facet G; MANUSCRIPT_v4's "para.") already uses. The X text and the +blog text match paragraph for paragraph. + +| Text | Date | Local file | Card | +| --- | --- | --- | --- | +| **Models as Insider Risks in the Super Intelligence Era** (X article; also on the blog) | 2026-10-10 | `x/nadella_2026-10-10_models_as_insider_risks.txt`; `posts/2026-10-10_models-as-insider-risks.txt` | `../literature/cards/berthon_ethics_course/nadella2026models.md`; `../literature/cards/dxd_foundation/nadella2026models.md` | +| The Reverse Information Paradox | 2026-07-12 | `posts/2026-07-12_reverse-information-paradox.txt` | `dxd_foundation/nadella2026reverse.md` | +| A frontier without an ecosystem is not stable | 2026-06-14 | `posts/2026-06-14_frontier-ecosystem.txt` | `dxd_foundation/nadella2026frontier.md` | +| Looking Ahead to 2026 | 2025-12-29 | `posts/2025-12-29_looking-ahead-2026.txt` | `berthon_ethics_course/nadella2025looking.md` | +| Positive Sum Future (began on X) | 2025-11-15 | `posts/2025-11-15_positive-sum.txt` | none yet | + +**The X article, paragraph by paragraph** (labels only; read the local text before quoting): + +| P | What it does | +| --- | --- | +| 1–2 | Old software could be traced to a code path; frontier models cannot be, yet they get "our most sensitive data" and "mission-critical actions on our behalf" | +| 3 | A new "trust architecture"; responsibility cannot be outsourced; a provider's assurances do not relieve "us" | +| 4–5 | Not nested black boxes to accept or reject: contained systems to observe, test and contain; "separate the supply of intelligence from the authority over it" | +| 6 | Alignment set aside; deterministic design, human controls, operating procedures, industry standards | +| 7–8 | Models as insider risks, not malicious but able to err or be compromised; enterprise practice (identity, least privilege, logging, containment) | +| 9 | Chain-of-thought transparency "non-negotiable" but "not sufficient or dependable" | +| 10 | Models checking models end in nested black boxes | +| 11–12 | Controls outside the model (a 1970s security principle); model separate from harness and action space | +| 13 | "Principles of observability" | +| 14 | Model diversity | +| 15 | Observe everything: "tamper-proof human readable evidence" | +| 16 | Verifiability: continuous testing, including failures | +| 17 | Independent controls: "Orgs" decide access and actions | +| 18 | Independent auditability | +| 19 | Containment: an emergency brake that "An authorized person" can pull | +| 20 | Incident disclosure "to those affected" and to industry | +| 21 | The most trustworthy system "enables us to trust the model the least" | + +## 2. Reports on the X article + +| Report | Card | Local copy | +| --- | --- | --- | +| Kumari, Business Standard, 12 Oct 2026 | `dxd_foundation/kumari2026nadella.md` | `commentary/kumari2026nadella.html` | +| Constantin, 12 Oct 2026 | `dxd_foundation/constantin2026nadella.md` | `commentary/constantin2026nadella.html` | +| Sudjianto, Substack, 10 Oct 2026 (the first substantive commentary found) | `berthon_ethics_course/sudjianto2026when.md` | — | + +Facet G found no expert critique of the post in its first two days (`sources/dxf_facet_G_nadella_lineage.md`, C9). + +## 3. The lab's earlier analyses of the post + +| Where | What it concluded | Copy here | +| --- | --- | --- | +| `dxd_foundation` facet G, "The Nadella post and its lineage" (57 sources) | The fairest statement (C1); lineage in the reference monitor, Saltzer and Schroeder, CERT, zero trust (C2); chain-of-thought monitoring is fragile, and the AI-safety work almost never tests the human reader (C3); the three human tasks the post creates (reconstruct from logs, notice and halt, hold responsibility) and what the human-factors evidence says of each (C4); observability vs transparency (C5); real convergences (C6); the three counts tested against the text (C7); **scope cautions** (C8); claims tested (C9) and claims the sources contradict (C10) | `sources/dxf_facet_G_nadella_lineage.md` + `.bib` | +| `berthon_ethics_course` facet V1, "Nadella's post and the disclosure paradigm" | The post move by move; Microsoft's own record (Responsible AI Standard, transparency reports, Copilot note, AI literacy guide, Elevate); the EU AI Act; commentary; for, narrows, against | `sources/bec_facet_V1_nadella_disclosure_paradigm.md` + `.bib` | +| `berthon_ethics_course` facets V2, V4–V6 | Oracy to literacy; disclosure and burden; scaffolding; prior work and objections (the v4 research) | `sources/bec_facet_V*.md` | +| MANUSCRIPT_v4, §§1, 4, 5 | The essay's ethical design toward people "stops at legibility"; three echoes (burden, medium, missing scaffold); v4 notes on where it departs from the author's framing | `../sources/08_MANUSCRIPT_v4.md`; `../sources/notes/berthon_v4_*.md` | +| `dxd_foundation` drafts v1–v4, claim, argument map, outline, review | Word separation moved work from reader to page; Nadella's post offers a record to read and a brake, and is silent on the people the model acts upon | `sources/dxf_*.md` | +| `dxd_foundation` facets E, H, L, S, T, X, Z | The human-factors frame, candidate moves, disclosure and choice, read-to-check in HMI, adjudication, the counter-case, the gap search | `sources/dxf_facet_*.md` | +| This arm's ledger | Rows E13 (the essay), E14 (Microsoft's practice), E15 (oversight without capacity), E16 (AI literacy in law), O06–O07, D06–D08; conflict X07 (sufficient or insufficient) | `../consolidation/` | + +## 4. Cards + +`CARDS.md` lists the 62 cards behind facets G and V1, now copied into `../literature/cards/` (and indexed in +`../literature/CARD_INDEX.md`), and the 7 facet-G sources that have no card (Above Security, Ananny and Crawford, +Endsley 1995, Lin and Zainescu, Mackworth, Tariq, Warm). + +## 5. The two transitions + +| Transition | Lab research (copied into `sources/`) | Its verdicts, in brief | +| --- | --- | --- | +| *Scriptio continua* to word spacing | `dxf_facet_A_the_record` (what changed, where, when, by whom); `B_the_dispute` (Saenger's thesis and its critics); `C_spacing_to_literacy` (did spacing spread reading?); `D_reading_science` (what a space does for a reader); `F_parallel_moves`; `J_who_holds_the_convention`; `K_training_and_cause`; `O_custody_of_conventions`; `W_spacing_oracy_literacy`; adjudication `dxf_facet_T_adjudication` | Spacing lowered a cost, not a cause of literacy (C, verdict 5); it came as a complex of conventions over centuries (T §4); ruler and church held the page's conventions (O; T §10); readers used to spaces lose a third to a half of their speed without them (README) | +| Print and mass literacy | `bec_facet_B_print_1450_1800` (who was shut out, by what rule, 1450–1800); the consolidation's rows H11–H13, H17, H24 | Four print-era failures: corporate monopoly of the press, the 1543 bar on reading the English Bible by rank and sex, Tridentine licensing, language bars at law; then the oversold promise of mass literacy (H13) | + +The lab's ledger rows on the scribal analogy are H26 and H27 (`../consolidation/CONCEPT_LEDGER.md`). diff --git a/submissions/berthon_ethics_paper/nadella/README.md b/submissions/berthon_ethics_paper/nadella/README.md new file mode 100644 index 00000000..f69aa107 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/README.md @@ -0,0 +1,93 @@ +# Workstream: the design ethics of current LLM tools, read through Nadella's X article + +*Written by Claude on 2026-10-24. Set-up only: nothing here is research or draft yet.* + +**The paper, in the author's terms** (verbatim in `SOURCE.md`). It examines the design ethics of current LLM tools; +*Spaces in the Stream* examines their cognitive issues. Nadella's X article, "Models as Insider Risks in the Super +Intelligence Era" (10 Oct 2026), proposes a set of moves, and those moves repeat the mistakes of two transitions: +from *scriptio continua* to spaces between words, and the management of mass literacy through the printing press. + +## The research question this sets up + +*Claude's restatement, for the author to accept or change.* For each of Nadella's moves, which mistake of the two +transitions does it repeat, what made it a mistake then (by the three internal standards of the consolidation: +contemporary objection, a broken professed standard, an unmet promise), and what corrected it? The grid below is the +work plan. Rows are his moves, by paragraph; the columns are where the lab's record already speaks. Empty cells are +the research. + +| Move (paragraph) | Scribal transition: what the record has | Print and mass literacy: what the record has | Already argued in the lab | +|---|---|---|---| +| Responsibility stays with the deployer; authority separated from supply (P3, P5, P17) | Ruler and church held the conventions of the page; nothing shows scribes and readers sharing interests as equals (`sources/dxf_facet_O_custody_of_conventions.md`; `dxf_facet_J_who_holds_the_convention.md`) | The Stationers held the press by grant (`bec_facet_B_print_1450_1800.md`, failure 1; ledger H11) | facet G C7(ii); v4 §4.1 | +| Chain-of-thought transparency, "non-negotiable" but insufficient (P9) | — | Disclosure as the second of four stages (ledger H24) | v4 §4.3; facet G C6(a) | +| Controls outside the model; separate harness and action space (P11, P12) | Spacing moved work from reader to page; so does the reference monitor (facet G C7, "Complication") | — | dxd T_adjudication §6 | +| Model diversity; independent auditability (P14, P18) | — | Second interpreters, notaries (ledger H10) | — | +| Observe everything: "tamper-proof human readable evidence" (P15) | A text that cannot answer the person who must act on it (Plato; ledger H27) | The record over the voice; the jurat (H07, H22) | v4 §4.2; facet G C4 row 1 | +| Verifiability: continuous testing (P16) | — | — | — | +| Emergency brake held by "an authorized person" (P19) | — | Who held the test and the brake (the gate, H01) | facet G C4 row 2, C6(c) | +| Disclosure "to those affected" after failure (P20) | — | Disclosure without capacity; the Requerimiento (H18) | v4 §4.2; facet G C6(e) | +| "Trust the model the least" (P21) | — | — | v4 §7 | +| Missing: any capacity-building for the people acted on | Spacing served readers for whom Latin was foreign, against "the elitist literate mentality of the ancient world" (Saenger, p. 11; ledger H26) | Literacy withheld, then oversold (H04, H13); scaffolds by sponsors (H25) | v4 §4.4 | + +## Where the record pushes back on the framing + +These points come from the lab's own audited research. The author decides what to do with them. + +1. **The spaces transition is the lab's good precedent, not a mistake.** In `dxd_foundation` and the Spaces essay, + word separation is the model of a sound design move: the page took on work the reader had carried (dxd + T_adjudication §6). So "making the mistakes as the transition from scriptura continua to spaces between words" needs a gloss. The record offers three + candidate mistakes of that era, each sourced: keeping a costly script because the literate elite did not want + readership widened (Saenger, p. 11); centuries of delay; and custody of the convention by ruler and church rather + than by readers (facet O). Nadella's P11 makes the same kind of move as spacing (work from person to artefact), + so the contrast must rest on *whom the move serves* and *what task it leaves the person*, which facet G flags as + the author's own tests to rule on (C7). +2. **Spacing did not spread literacy.** No historian the lab read credits word separation with spreading reading; + it lowered a cost (`dxf_facet_C_spacing_to_literacy.md`, verdict 5). "ultimately how the world managed the mass + distribution of literacy though the printing press" therefore links two transitions that the record keeps separate, and the + paper should say how they connect. +3. **Print's mistakes are institutional.** The documented print-era failures are a chartered monopoly, bars on + reading the Bible by rank and sex (1543), licensing under the Tridentine Index, and language bars at law (facet B + print, failures 1–4), plus the five practices of the consolidation. The mapping above draws on these. +4. **The post concedes much of what a critic would say** (facet G C8–C10): most of its controls need no reading; P9 + already calls chain of thought "not sufficient or dependable"; two humans appear (P19, P20). "Dismantle" holds + best against the tasks it leaves people (P15, P19), the seat it is written from (P5, P17), and the people it acts + on, who appear once (P20). + +## What is in place + +- **Nadella's texts**: five posts, the X capture with locators P1–P21, two press reports, held locally because this + repo is public (`~/iit-playground/berthon_ethics_local/nadella/`). Paragraph map in `CORPUS.md` §1. +- **The lab's earlier work**, copied byte for byte into `sources/` (49 files): the post's lineage (facet G), the + disclosure paradigm and v4 research (V1–V6), the word-spacing record (dxd facets A–D, F, J, K, O, W), and print + (facet B). +- **The cards**: 204 in `../literature/cards/`, listed in `CARDS.md` with 260 facet sources that have no card. + `python3 -I nadella/collect.py --check` from the arm directory. +- **The Spaces limitation**, paraphrased in `SOURCE.md`, verbatim in the local corpus. + +## Rulings (2026-10-24) + +The author's instruction, verbatim: "Use your judgement. The goal here is to not make the same mistakes as when +literacy was growing. Nadellas suggestions chart a course to make those same failures." Claude made the four open +calls below under that delegation; each is the author's to reverse. + +1. **The thesis.** The growth of literacy offers a course of corrections and a course of failures. Word spacing + belongs to the corrections: the page took on work the reader had carried, for readers the old script excluded. + The failures came with mass distribution: custody of the medium by an interested party, a record kept by one + side, disclosure in place of understanding, the burden of reading placed on the person bound, the competence + withheld or sold as a cure, and a gate held by the few. Nadella's moves take the second course. Where one of them + resembles the correction (controls outside the model, P11–P12), the paper concedes it and shows whom it serves. +2. **How the two transitions connect.** As two stages of one growth: spacing changed the artefact for a widening + readership; print distributed the medium at scale, and the institutions that governed that distribution made the + failures. Nadella writes at the print stage, the moment a medium that demands a new competence is put into + everyone's hands, and makes the print-stage choices, not the spacing choice. The paper does not claim spacing + spread literacy (pushback 2). +3. **Scope.** The ethics of the decision space itself (who bears the burden of understanding, who holds the record and + the brake, who has standing), and market power as custody of the conventions. Labour governance enters where the + design assigns work to people (the log reader, the brake holder). Training paradigms stay out, except chain-of-thought + monitorability as it bears on P9. +4. **Home.** This arm's paper, with Nadella as its central case. That settles consolidation conflict X01 as the merge + option: the five positions and v4's three wrongs supply the history; the insider post supplies the present. + +The strength of the charge follows facet G: concede what the post gets right, and aim at the tasks it leaves people, +the seat it writes from, and the people it acts on. + +The research plan is `RESEARCH_PLAN.md`. diff --git a/submissions/berthon_ethics_paper/nadella/RESEARCH_PLAN.md b/submissions/berthon_ethics_paper/nadella/RESEARCH_PLAN.md new file mode 100644 index 00000000..aae34f79 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/RESEARCH_PLAN.md @@ -0,0 +1,47 @@ +# Research plan: Nadella's moves against the failures of growing literacy + +*Written by Claude on 2026-10-24 under the author's delegation ("Use your judgement", `README.md`, Rulings). Plan +only: no research has run. Each hypothesis is stated so the evidence can refute it.* + +**The claim to test.** Each of Nadella's moves in "Models as Insider Risks" (P1–P21) takes a choice that the +institutions of growing literacy took and that the record counts as a failure by the consolidation's three internal +standards (contemporaries objected; a professed standard was broken; a promise went unmet). For each, the plan names +the historical failure, what the lab already holds, what would refute the pairing, and what must still be found. + +## 1. The pairings + +| # | Move | Hypothesis: the failure it repeats | In hand | Refuted if | Still needed | +|---|---|---|---|---|---| +| M1 | Authority and responsibility rest with the deploying organisation; "Orgs" set access and actions (P3, P5, P17) | **Custody of the convention by the interested party.** The page's conventions were held by ruler and church; the press by a chartered company that was also the Crown's licenser | dxd facets J, O; bec facet B (failure 1); ledger H11, N07 | The record shows conventions held by their owner that served the weaker party without outside force; or the post binds the organisation to anyone outside it | Who governs current LLM tool conventions in practice (vendor terms, enterprise admin controls) | +| M2 | Industry standards where existing ones are insufficient; learnings shared "industrywide" (P6, P20) | **Self-regulation by the guild.** The Stationers regulated the trade they monopolised; standards bodies bind no implementer | bec facet B; dxd T §10 (RFC 8890; co-determination binds through arbitration) | Industry-set standards in the literacy era protected readers without a public authority behind them | The current AI standards landscape (who writes, who is bound, who is consulted) | +| M3 | Chain-of-thought transparency "non-negotiable" but "not sufficient or dependable" (P9) | **Disclosure as the safeguard.** The second of four stages: read the text, certify the reading, do not check understanding | v4 §3, §4.3; ledger H18, H22, H24; facet G C3, C6(a) | P9's reader is shown to be able to use the chain of thought (the safety literature already says monitoring is fragile) | Who is meant to read chain of thought in deployed LLM tools, and whether any study tests that reader | +| M4 | "Observe everything": every meaningful action leaves "tamper-proof human readable evidence" (P15) | **Two failures at once.** (a) *The record over the voice*: one party keeps the record that binds the other. (b) *Scriptio continua for the auditor*: a whole-volume record with no marked points of commitment, so the reader must find the decisions, as the reader of unspaced script had to find the words | ledger H07, H22; facet G C4 row 1 (alert review, transcript audit); the Spaces essay's decision-space marking (private; cite through the author) | Evidence that people reconstruct outcomes well from complete agent logs; or that the logs mark commitments | Studies of agent-log review and audit trails in LLM tools; how current tools present traces | +| M5 | Model diversity; independent auditability; no model controls both behaviour and evidence (P14, P18) | **Checks among the owner's agents, not for the bound party.** The notary's duty ran to both parties and to the weaker; the second interpreter could be either side's | ledger H10, N07; facet G C2, C6 | Independence in the post is shown to run to the people acted on | Whether any LLM tool gives a user or counterparty an independent check | +| M6 | Controls outside the model; harness and action space separated (P11, P12) | **The correction, held by the wrong party.** The same kind of move as spacing (work moved from person to artefact), declared to the deployer only | dxd T §6; facet G C6(b), C7 ("Complication") | The envelope is declared to the people the system acts on | — (concede in the text) | +| M7 | Verifiability: continuous testing of the whole system (P16) | **Testing the system, never the person's understanding.** No equivalent of the probate registrar's check or Spain's art. 193 | ledger H23, D06; Microsoft's own Standard T1.3 asks it (E14) | The post or Microsoft's practice tests whether people can interpret outputs in realistic conditions | Microsoft's evidence, if any, against its own T1.3 and A5.4 | +| M8 | An emergency brake for "An authorized person" (P19) | **The gate held by the few.** Power over the outcome given to a designated literate class (registrar, clergy); the bound party has no refusal | ledger H01, H20, D02; facet G C4 row 2, C6(c); E15 | The brake is held by, or answers to, the people acted on | How brakes and approvals work in current tools (permission prompts; approval rates) | +| M9 | Timely disclosure "to those affected" after failure (P20) | **Notice without standing.** The Requerimiento's testimony; disclosure after the fact that proves notice was given | ledger H18, K12, O07; facet G C6(e) | Disclosure in the post carries a remedy or a voice | What incident disclosure to users means in current AI policy (AI Act art. 73) | +| M10 | "Trust the model the least" (P21) | **The burden placed on the reader.** Distrust becomes the user's duty (Copilot's "always review and verify") as reading the contract became the signer's | ledger H19, E14; v4 §4.2 | The post or the product assigns the verifying work to the operator | — | +| M11 | Silence on building anyone's capacity; skilling programmes beside the product | **The competence withheld, then oversold.** Literacy campaigns promised returns and blamed the learner; scaffolds served their sponsors | ledger H04, H13, H25, N17, E16; v4 §4.4 | Microsoft or the post builds capacity inside the system, owed and fading | Current AI-skilling claims and evidence of effect | + +## 2. What would sink the thesis + +The paper fails if (a) most moves pair with a correction, not a failure; (b) the pairings rest on resemblance and no +shared property (each row must name the property, as Table 1 of v3 does); or (c) the historical failures were not +failures by the internal standards. Facet G's cautions (C8–C10) are the first test of (a). + +## 3. Research to run, in order + +1. **Read the post's present-day setting** (M1, M4, M7, M8, M10): how current LLM tools present traces, approvals and + verification duties to users and to enterprise admins. Start from facet G's B4 cards and Microsoft's documents (V1). +2. **The print stage at scale** (M1, M2, M11): licensing and the Stationers as censor; reading mandates and church + examinations; mass literacy campaigns. Start from bec facet B, ledger H11–H13, and dxd facets C and O. +3. **The scribal stage, read first-hand** (M4, M6): Saenger p. 11 (the elite and the costly script), the custody of + the page (facet O), and the work spacing moved to the page (facets A, D). +4. **Audit and log review** (M4): studies of reviewing agent transcripts and audit trails (facet G C4: Kutasov, + Grunde-McLaughlin, CERT). +5. **Standards and self-regulation** (M2): who writes AI standards and whom they bind. + +Each step ends with cards for every source relied on and a pass of `collect.py` and the arm's checks. By the author's +standing rule the drafter reads and writes the prose; agents may search and card, and every card is verified before +it enters a draft. diff --git a/submissions/berthon_ethics_paper/nadella/SOURCE.md b/submissions/berthon_ethics_paper/nadella/SOURCE.md new file mode 100644 index 00000000..a829af89 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/SOURCE.md @@ -0,0 +1,48 @@ +# Source of the Nadella workstream + +*Recorded by Claude on 2026-10-24.* + +## The author's instruction (verbatim, 2026-10-24) + +> With this new consolidated directory, I want us to address the limitation from the recent human factors paper +> that focuses on decision spaces. We intentionally left out the ethical issues there and want to address them +> here. Specifically, we want to analyze and dismantle Satya Nadellas X article about the path forward. Before +> continuing, set yourself up for that process and collect all the content so we can start research + +## The limitation this workstream takes up + +The human-factors paper is *Spaces in the Stream* (PHD1750; working copy `~/phd1750-litreview/paper/`, a private +repo; latest full text `gdoc_full_2026-10-23.md`). Its scope paragraph, at line 22 of that file, brackets the +macro-sociological, labour, political and ethical/moral layers of algorithmacy, naming platform labour governance, +market power and training paradigms, so that it can study only the perceptual layer: how attention detects a +decision space as text forms on the display. The same paragraph sets the boundary condition this workstream +inherits: marking a decision space is necessary but not sufficient, because a mark shows where the system commits +without settling how the person ought to respond or what standing they have. The paragraph is kept verbatim only in +the local corpus (`~/iit-playground/berthon_ethics_local/nadella/spaces_limitation_2026-10-23.txt`), because the +paper's repo is private and this one is public. The Spaces paper does not mention Nadella; its precursor, +`dxd_foundation`, set word separation against Nadella's post (`sources/dxf_*`). + +## Which Nadella text + +"Satya Nadella's X article about the path forward" is read as **"Models as Insider Risks in the Super +Intelligence Era,"** posted as an X article on 10 October 2026 (x.com/satyanadella/status/2108931348857827686) and +on his blog the same day. It is his newest post on either; it sets out a "trust architecture for this new era" and +seven principles. The phrase "path forward" appears in none of his posts, and one other post began on X ("Positive +Sum Future", 15 November 2025). The author should confirm the reading (README, open question 1). + +## The author's framing (verbatim, 2026-10-24, second message) + +> We are looking at the design ethics of current LLM tools. +> +> The spaces essay looks at the cognitive issues with current design. +> +> This paper looks at the ethical issues of current design. +> +> Nadellas recent X article gives a bunch of moves. These moves, however, are making the mistakes as the +> transition from scriptura continua to spaces between words and ultimately how the world managed the mass +> distribution of literacy though the printing press. + +This confirms the text ("recent X article" is the 10 October 2026 post, his most recent) and answers open questions +2 and 5 in part: the subject is the design ethics of current LLM tools, as the ethical companion to the cognitive +Spaces essay. The two transitions it names are the ones the lab has researched in `dxd_foundation` (facets A–D, F, J, +K, O, W) and in `berthon_ethics_course` (facet B, print 1450–1800); both are now in `sources/`. diff --git a/submissions/berthon_ethics_paper/nadella/collect.py b/submissions/berthon_ethics_paper/nadella/collect.py new file mode 100644 index 00000000..32c5d997 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/collect.py @@ -0,0 +1,124 @@ +#!/usr/bin/env python3 +"""Collect the lab's earlier work on Nadella's posts into nadella/sources/, and the cards it cites into +literature/cards/, as byte-identical copies. Writes nadella/CARDS.md and records every copy in +literature/card_sources.json ("nadella_sources") so build_provenance.py checks it. + + python3 -I nadella/collect.py # copy, write CARDS.md, update card_sources.json + python3 -I nadella/collect.py --check # exit 1 if a copy is missing or differs, or CARDS.md is stale + +The full texts of the posts are NOT copied here: the repo is public and the X capture is marked for local +use only. They live in ~/iit-playground/berthon_ethics_local/nadella/ (see nadella/CORPUS.md). +""" +import glob +import json +import os +import re +import shutil +import sys + +HERE = os.path.dirname(os.path.abspath(__file__)) +ARM = os.path.dirname(HERE) +SUB = os.path.dirname(ARM) +sys.path.insert(0, os.path.join(ARM, "literature")) +import cardlib as cl # noqa: E402 + +DXF = "submissions/dxd_foundation" +BEC = "submissions/berthon_ethics_course" +FILES = {} +for f in ["ABSTRACT_v4", "ARGUMENT_MAP", "CLAIM", "DRAFT", "DRAFT_v2", "DRAFT_v3", "DRAFT_v4", "OUTLINE", + "README", "RESEARCH_PLAN", "REVIEW_v2_2026-10-13", "SOURCE"]: + FILES[f"nadella/sources/dxf_{f}.md"] = f"{DXF}/{f}.md" +for f in ["E_human_factors_frame", "G_nadella_lineage", "H_candidate_moves", "L_disclosure_and_choice", + "S_hmi_read_to_check", "T_adjudication", "X_counter_case", "Z_gap_search"]: + FILES[f"nadella/sources/dxf_facet_{f}.md"] = f"{DXF}/literature/facets/{f}.md" +FILES["nadella/sources/dxf_facet_G_nadella_lineage.bib"] = f"{DXF}/literature/facets/G_nadella_lineage.bib" +# The two historical transitions: scriptio continua to word spacing (dxd_foundation) and print (berthon facet B). +HISTORY = ["A_the_record", "B_the_dispute", "C_spacing_to_literacy", "D_reading_science", "F_parallel_moves", + "J_who_holds_the_convention", "K_training_and_cause", "O_custody_of_conventions", "W_spacing_oracy_literacy"] +for f in HISTORY: + FILES[f"nadella/sources/dxf_facet_{f}.md"] = f"{DXF}/literature/facets/{f}.md" + FILES[f"nadella/sources/dxf_facet_{f}.bib"] = f"{DXF}/literature/facets/{f}.bib" +FILES["nadella/sources/bec_facet_B_print_1450_1800.md"] = f"{BEC}/literature/facets/B_print_1450_1800.md" +FILES["nadella/sources/bec_facet_B_print_1450_1800.bib"] = f"{BEC}/literature/facets/B_print_1450_1800.bib" +for f in ["V1_nadella_disclosure_paradigm", "V2_oracy_to_literacy", "V4_disclosure_and_burden", + "V5_scaffolding", "V6_prior_work_and_objections"]: + FILES[f"nadella/sources/bec_facet_{f}.md"] = f"{BEC}/literature/facets/{f}.md" +FILES["nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.bib"] = f"{BEC}/literature/facets/V1_nadella_disclosure_paradigm.bib" +for f in ["RESEARCH_BRIEF", "V4_CHECK_BRIEF"]: + FILES[f"nadella/sources/bec_v4_{f}.md"] = f"{BEC}/v4/{f}.md" + +LIBS = {"dxd_foundation": f"{DXF}/literature/library", "berthon_ethics_course": f"{BEC}/literature/cards", + "algorithmacy_design_ethics": "submissions/algorithmacy_design_ethics/literature/library"} + + +def library_index(): + """(sub, path) for every card in LIBS, keyed by filename, DOI and title+year.""" + by = {} + for sub, lib in LIBS.items(): + for p in sorted(glob.glob(os.path.join(origin(lib), "*.md"))): + if os.path.basename(p) in cl.NONCARDS: + continue + r = cl.parse_card(p, sub) + for k in (("key", os.path.basename(p)[:-3]), ("doi", r["doi"]), + ("title", cl.norm_title(r["title"], r["year"]) if r["year"] else "")): + if k[1]: + by.setdefault(k, (sub, p)) + return by +BIBS = ["nadella/sources/dxf_facet_G_nadella_lineage.bib", "nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.bib"] + \ + [f"nadella/sources/dxf_facet_{f}.bib" for f in HISTORY] + ["nadella/sources/bec_facet_B_print_1450_1800.bib"] + + +def origin(rel): + return os.path.join(os.path.dirname(SUB), rel) + + +def main(check): + ok = True + for dest, src in FILES.items(): + d = os.path.join(ARM, dest) + if check: + ok &= os.path.exists(d) and cl.sha256(d) == cl.sha256(origin(src)) + else: + os.makedirs(os.path.dirname(d), exist_ok=True) + shutil.copy2(origin(src), d) + rows, missing = [], [] + idx = library_index() + for bib in BIBS: + entries = cl.parse_bib(os.path.join(ARM, bib)) if os.path.exists(os.path.join(ARM, bib)) else {} + for key in sorted(entries): + e = entries[key] + t = cl.norm_title(re.sub(r"[{}\\]", "", e.get("title", "")), e.get("year", "")) + hit = idx.get(("key", key)) or idx.get(("doi", cl.norm_doi(e.get("doi", "")))) or idx.get(("title", t)) + if not hit: + missing.append((os.path.basename(bib), key)) + continue + sub, p = hit + dest = os.path.join(cl.CARDS, sub, os.path.basename(p)) + if check: + ok &= os.path.exists(dest) and cl.sha256(dest) == cl.sha256(p) + elif not os.path.exists(dest): + shutil.copy2(p, dest) + r = cl.parse_card(p, sub) + rows.append(f"| `{sub}/{os.path.basename(p)}` | {cl.md_cell(r['authors'][:50])} | {r['year']} | " + f"{cl.md_cell(r['title'][:80])} | {r['basis']} | {os.path.basename(bib)} |") + lines = ["# Cards behind the earlier Nadella and transition research", "", + "Generated by `nadella/collect.py`; do not edit by hand. Every key in the two Nadella facet bibs", + "(`sources/*.bib`: facets G and V1 on the post; A-D, F, J, K, O, W on word spacing; B on print),", + "with the card now in `../literature/cards/`. Paths are relative to that directory.", "", + f"Cards: {len({r.split(' | ')[0] for r in rows})}. Bib keys with no card: {len(missing)}.", "", + "| card | authors | year | title | basis | facet bib |", "|---|---|---|---|---|---|"] + sorted(set(rows)) + \ + ["", "## Keys with no card", ""] + [f"- `{k}` ({b})" for b, k in missing] + ok &= cl.write_or_check(os.path.join(HERE, "CARDS.md"), "\n".join(lines) + "\n", check) + if not check: + jp = os.path.join(cl.HERE, "card_sources.json") + meta = json.load(open(jp, encoding="utf-8")) + meta["nadella_sources"] = FILES + with open(jp, "w", encoding="utf-8") as f: + json.dump(meta, f, indent=2, ensure_ascii=False) + f.write("\n") + print(f"{len(FILES)} files; {len({r.split(' | ')[0] for r in rows})} cards; {len(missing)} keys without card" + ("" if ok else " — STALE")) + return 0 if ok else 1 + + +if __name__ == "__main__": + sys.exit(main("--check" in sys.argv)) diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_facet_B_print_1450_1800.bib b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_B_print_1450_1800.bib new file mode 100644 index 00000000..e34d8a12 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_B_print_1450_1800.bib @@ -0,0 +1,196 @@ +@book{arber1875transcript, + author = {Arber, Edward}, + title = {A Transcript of the Registers of the Company of Stationers of London, 1554--1640 A.D.}, + volume = {1}, + publisher = {Privately printed}, + address = {London}, + year = {1875}, + url = {https://archive.org/details/transcriptofregi01statuoft}, + note = {Read in part: charter pp. xxviii--xxxii; Star Chamber ordinances p. 322; Addendum on the 1643--5 controversy pp. 581--583. OCR text from archive.org; OCR repairs flagged in the facet file.} +} + +@book{statutesrealm1817vol3, + author = {{Great Britain, Record Commission}}, + title = {The Statutes of the Realm}, + volume = {3}, + address = {London}, + year = {1817}, + url = {https://archive.org/details/the-statutes-of-the-realm-3}, + note = {Read in part: 34 and 35 Hen. VIII c. 1 (pp. 895--896 by OCR page markers) and 27 Hen. VIII c. 26 (page not resolved). OCR text from archive.org.} +} + +@book{statutesrealm1819vol5, + author = {{Great Britain, Record Commission}}, + title = {The Statutes of the Realm}, + volume = {5}, + address = {London}, + year = {1819}, + url = {https://archive.org/details/statutes_of_the_realm_5}, + note = {Read in part: 14 Car. II c. 33 (Licensing of the Press Act 1662, from p. 428) and the three continuing Acts. OCR text from archive.org.} +} + +@book{manuzio1564index, + author = {{Catholic Church} and Manuzio, Paolo}, + title = {Index librorum prohibitorum, cum regulis confectis per patres a Tridentina Synodo delectos, auctoritate sanctiss. D.N. Pii 4., Pont. Max. comprobatus}, + publisher = {Manuzio, Paolo}, + year = {1564}, + url = {https://archive.org/details/ita-bnc-ald-00000309-001}, + note = {Read in part: Regulae Indicis, rules IV, VI, X (Latin OCR; page numbers inferred from running heads). Place of printing not found in the metadata I read.} +} + +@book{milton1644areopagitica, + author = {Milton, John}, + title = {Areopagitica: A Speech for the Liberty of Unlicensed Printing}, + year = {1644}, + publisher = {Project Gutenberg}, + url = {https://www.gutenberg.org/cache/epub/608/pg608.txt}, + note = {Read in part (passages on licensing, monopoly, Index, Trent). Gutenberg eBook no. 608; no page numbers; not collated against a first edition.} +} + +@article{johns2002how, + author = {Johns, Adrian}, + title = {How to Acknowledge a Revolution}, + journal = {American Historical Review}, + volume = {107}, + number = {1}, + pages = {106--125}, + year = {2002}, + doi = {10.1086/ahr/107.1.106}, + note = {Read in part this session (pp. 109, 112, 119--120, 122, 124); repo card records a full read. Crossref title and pages match. Text from the ENS course-site PDF.} +} + +@article{eisenstein2002unacknowledged, + author = {Eisenstein, Elizabeth L.}, + title = {An Unacknowledged Revolution Revisited}, + journal = {American Historical Review}, + volume = {107}, + number = {1}, + pages = {87--105}, + year = {2002}, + doi = {10.1086/ahr/107.1.87}, + note = {Not accessible (OUP paywall). Crossref title and pages match. Reply: doi 10.1086/532100, pp. 126--128. Known only through Johns 2002.} +} + +@book{eisenstein1979printing, + author = {Eisenstein, Elizabeth L.}, + title = {The Printing Press as an Agent of Change: Communications and Cultural Transformations in Early-Modern Europe}, + publisher = {Cambridge University Press}, + year = {1979}, + url = {https://archive.org/details/printingpressas01eise}, + note = {Not accessible (archive.org lending only, no snippets). Known only through Johns 2002.} +} + +@book{johns1998nature, + author = {Johns, Adrian}, + title = {The Nature of the Book: Print and Knowledge in the Making}, + publisher = {University of Chicago Press}, + year = {1998}, + url = {https://archive.org/details/natureofbookprin0000john}, + note = {Not accessible (lending only). Known through Johns 2002 and the repo card.} +} + +@book{innis1951bias, + author = {Innis, Harold A.}, + title = {The Bias of Communication}, + publisher = {University of Toronto Press}, + year = {1951}, + url = {https://archive.org/details/the-bias-of-communication-harold-innis}, + note = {Read in part: pp. 22--30, 53--56, 60 (archive.org OCR; page numbers from running heads).} +} + +@book{febvre1976coming, + author = {Febvre, Lucien and Martin, Henri-Jean}, + title = {The Coming of the Book: The Impact of Printing 1450--1800}, + publisher = {NLB}, + address = {London}, + year = {1976}, + url = {https://archive.org/details/comingofbookimpa0000febv}, + note = {Not accessible (lending only; search-inside refused). Nothing is stated from it.} +} + +@article{darnton1995censorship, + author = {Darnton, Robert}, + title = {Censorship, a Comparative View: France, 1789--East Germany, 1989}, + journal = {Representations}, + volume = {49}, + pages = {40--60}, + year = {1995}, + doi = {10.2307/2928748}, + note = {Not accessible (UC Press PDF served a Cloudflare challenge; not bypassed). Crossref title and pages match.} +} + +@article{schwartz2017did, + author = {Schwartz, Kathryn A.}, + title = {Did {O}ttoman Sultans Ban Print?}, + journal = {Book History}, + volume = {20}, + number = {1}, + pages = {1--39}, + year = {2017}, + doi = {10.1353/bh.2017.0000}, + note = {Abstract only (via a blog quotation); full text closed. Crossref title and pages match.} +} + +@article{martel2001autour, + author = {Martel, Philippe}, + title = {Autour de Villers-Cott{\^e}rets : histoire d'un d{\'e}bat}, + journal = {Lengas}, + volume = {49}, + pages = {7--25}, + year = {2001}, + url = {https://www.felco-creo.org/wp-content/uploads/2025/08/2508-19-Martel-Villers-cotterets-lengas.pdf}, + note = {Read in part (online reprint by FELCO, 2025; cited by section). No DOI found in Crossref.} +} + +@misc{villerscotterets1539lexilogos, + author = {{Lexilogos}}, + title = {Ordonnance de Villers-Cott{\^e}rets, 1539 (transcription of articles 110--111)}, + url = {https://www.lexilogos.com/francais_villers_cotterets.htm}, + note = {Read in part. Secondary transcription, not an official copy; year of the web page not found.} +} + +@article{sanchezmarti2023advent, + author = {S{\'a}nchez-Mart{\'i}, Jordi}, + title = {The Advent of the Printing Press and {B}ritain's Multilingual Textual Culture, 1471--1510}, + journal = {Anglica: An International Journal of English Studies}, + volume = {32}, + number = {4}, + pages = {83--107}, + year = {2023}, + doi = {10.7311/0860-5734.32.4.05}, + note = {Read in part (abstract, pp. 92--93, 97). Open copy at https://rua.ua.es/dspace/handle/10045/138259. Crossref title and pages match.} +} + +@article{sasaki2017publishing, + author = {Sasaki, Yu}, + title = {Publishing Nations: Technology Acquisition and Language Standardization for {E}uropean Ethnic Groups}, + journal = {Journal of Economic History}, + volume = {77}, + number = {4}, + pages = {1007--1047}, + year = {2017}, + doi = {10.1017/S0022050717000821}, + note = {Read in part (abstract, Breton passage, conclusion; models not audited). Page numbers inferred from PDF order. Crossref title and pages match.} +} + +@article{houston2003lesser, + author = {Houston, R. A.}, + title = {`Lesser-used' Languages in Historic {E}urope: Models of Change from the 16th to the 19th Centuries}, + journal = {European Review}, + volume = {11}, + number = {3}, + pages = {299--324}, + year = {2003}, + doi = {10.1017/S1062798703000309}, + note = {Abstract only (Consensus); full text closed. Crossref title and pages match.} +} + +@incollection{suggett2018language, + author = {Suggett, Richard and White, Eryn}, + title = {Language, Literacy and Aspects of Identity in Early Modern {W}ales}, + booktitle = {The Spoken Word}, + publisher = {Manchester University Press}, + year = {2018}, + doi = {10.7765/9781526137876.00006}, + note = {Abstract only (Consensus); open PDF returned a 403 and was not bypassed. Crossref title matches; page range not returned by Crossref.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_facet_B_print_1450_1800.md b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_B_print_1450_1800.md new file mode 100644 index 00000000..8d2f10a0 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_B_print_1450_1800.md @@ -0,0 +1,241 @@ +# B — Print, 1450–1800: who was shut out, and by what rule + +*Facet B of the Berthon ethics course paper. Written 2026-10-13. Agent-drafted. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** The print era as a set of rules: who could print (guild monopoly, royal privilege, licensing, the charter of the London Stationers); who could read print (the English rank-and-sex limit on Bible reading of 1543; the Tridentine Index's permission regime for vernacular Bibles); what happened to dialects and minority languages when a printed vernacular became the standard (Welsh office-holding law of 1536, the French ordinance of 1539, London English and Older Scots in early printing); and what the historians of the book (Eisenstein, Johns, Innis, Darnton, Febvre and Martin) say about all this. It records history only. It draws no parallel to anything later. + +**The question the synthesis answers.** In the print era, who was shut out, by what rule, on what stated ground, who objected at the time, and what ended it. A second question runs underneath: which of the usual claims about print (a "monopoly of knowledge", a ban on printing in the Ottoman lands, the Index as a continent-wide wall, print as the cause of dialect loss) survive a reading of the primary text. + +**Legend for read status.** *Read in full* = I read the whole text this session. *Read in part* = I opened the text and read the passages cited, not the whole work (the entry says which). *Abstract only* = a database abstract or a publisher/search summary was all I saw. *Not accessible* = no text. Statutes, charters and the Index are quoted from archive.org scans of print editions (Record Commission *Statutes of the Realm*; Arber's 1875 transcript; the 1564 Index) that I downloaded and searched as OCR text. OCR damage is repaired inside square brackets and every repair is flagged; where the OCR page marker was ambiguous the locator says so. Locators of the form "p. N" are the printed page of the edition named in the entry. + +--- + +## (a) Search log + +All steps ran on 2026-10-13. "Hits" is the count the source reported, or "found / not found" for a known item. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | Repo cards | Read `algorithmacy_design_ethics/literature/library/{johns2002,innis1951,eisenstein1979,johns1998,eisenstein2002}.md` | 5 | Used only to find leads and page ranges. Every Johns 2002 and Innis quotation below was re-found by me in the source text (rows 3, 4). | +| 2 | archive.org `advancedsearch` | titles: "bias of communication"; "coming of the book"; "transcript registers company stationers"; "freedom of the press in england"; "printing press as an agent of change"; "printing revolution in early modern europe"; "statutes of the realm"; "index librorum prohibitorum" | 6 groups | Open full text: Innis (`the-bias-of-communication-harold-innis`), Arber vol. I (`transcriptofregi01statuoft`), Statutes of the Realm vols. 3 and 5, the 1564 Index (`ita-bnc-ald-00000309-001`). Lending-restricted (`access-restricted-item`): Febvre and Martin (3 copies), Siebert, Eisenstein 1979 and 1993/2005, Innis (other copies). | +| 3 | curl, ENS course site | `histoire.ens.psl.eu/IMG/pdf/adrian_johns-printed_revolution_discussed.pdf`; `pdftotext -layout` | 21 PDF pages | Johns 2002, AHR 107(1):106–125. Page numbers taken from the running heads. | +| 4 | curl, archive.org djvu text | Innis (`the bias of communication -- harold a. innis_djvu.txt`); Arber vol. I; Arber vol. II (`transcriptofregi02statuoft`); Statutes of the Realm 3; Statutes of the Realm 5; Index 1564 | 6 files | Arber vol. II searched for a 1586 Star Chamber decree and for "decree": no usable hit; I did not use it. The 1586 decree is only mentioned in Arber vol. I (payments for "the newe Decrees of the starre chamber"). I did not find its text. | +| 5 | curl, Project Gutenberg | `gutenberg.org/cache/epub/608/pg608.txt` | 1 | Milton, *Areopagitica*. No page numbers; locators are by quoted passage. | +| 6 | WebSearch | 1543 Act, "women, artificers, apprentices"; BHO Statutes of the Realm vol. 3 | 9 + 9 | Only Wikipedia mirrors and blogs. No official text. I found the statute myself in row 4 instead. | +| 7 | WebFetch / curl | `ric.edu/faculty/rpotter/statutes.html` (Potter, documents on the vernacular, 1500–1540) | 0 | Redirect to `w3.ric.edu`, then "unable to verify the first certificate"; curl got "Empty reply from server". Not read. | +| 8 | legislation.gov.uk | `/aep/Hen8/27/26`; `/aep/Hen8/34-35/1`; `/aep/Hen8/27/26/section/XX` | 200; 404; 404 | The first returns only the title and status line, "Laws in Wales Act 1535 (repealed 21.12.1993)", with no clause text. The other two 404. | +| 9 | WebSearch | Tridentine Index Rule IV English text | 10 | Found an EEBO-TCP copy (A66393 on linguisticdna.org, HTTP 403) and an EWTN paraphrase. I used the Latin 1564 edition (row 4) instead. | +| 10 | WebFetch | `bridwell.omeka.net/exhibits/show/luther/counterreformation/tenrules` | 1 | Summaries only, no rule text. Used only to confirm the 1664 reprint and the 1564 date. | +| 11 | curl | `quod.lib.umich.edu/e/eebo/A66393.0001.001` | 0 | Cloudflare challenge page. Not bypassed. | +| 12 | WebSearch | Ottoman printing ban 1485 Bayezid II Schwartz Müteferrika | 10 | Led to Schwartz 2017. Some results gave the dates 1483 and 1485 for the alleged ban; none was a primary source. | +| 13 | WebFetch | `malikazeghal.scholars.harvard.edu/resource/ablairpdf-0`; `kathrynschwartz.umasscreate.net/?p=47`; `marginalrevolution.com/.../did-ottoman-sultans-ban-print.html` | 403; list page; abstract | Schwartz 2017 is not open access (Unpaywall `is_oa: false`). I saw only the abstract as quoted in a blog post. | +| 14 | WebSearch, curl | Villers-Cotterêts articles 110–111; `lexilogos.com/francais_villers_cotterets.htm`; Martel PDF on `felco-creo.org` | 10; 1; 15 pp. | Lexilogos is a secondary transcription; no official copy (Légifrance, Archives nationales) reached. Martel (2001, Lengas 49) read in part. | +| 15 | WebSearch | Scotland 1616 school act, Statutes of Iona | 10 | Results are Wikipedia and teaching sites. I did not open the Register of the Privy Council; not used as evidence. | +| 16 | Consensus | `print standardization vernacular suppression of dialects and minority languages early modern Europe` | 19 | Surfaced Sánchez-Martí, Sasaki, Houston, Suggett and White, Burke, Marynissen et al. | +| 17 | OpenAlex `works?search=` | three queries (Schwartz; Darnton; Roman Inquisition Hebrew books) | HTTP 429 ×3, then ×2 more with `mailto` | Rate-limited (other agents running). Replaced by Crossref (row 18). | +| 18 | Crossref `query.bibliographic` and `/works/{doi}` | Schwartz 2017; Darnton 1995; Eisenstein 2002; Martel; and 9 DOIs listed in section (b) | all 9 DOIs resolved | Title, container, volume, pages and author matched for each. Martel's article has no DOI in Crossref (Lengas 49). | +| 19 | Unpaywall | the DOIs in section (b) | open: Darnton 1995 (UC Press), Sasaki, Suggett and White, Sánchez-Martí, Marynissen; closed: Schwartz, Houston, Bruder | Darnton's UC Press PDF returned a Cloudflare page; Suggett and White's Manchester Open Hive PDF returned a CloudFront 403. Not bypassed. | +| 20 | Alicante repository, DSpace REST | handle 10045/138259 → bundles → bitstream | 1 PDF | Sánchez-Martí 2023, 20 PDF pages, read in part. | +| 21 | curl, Cambridge Core | Sasaki 2017 open PDF | 1 | Read in part. | +| 22 | archive.org search-inside | `fulltext/inside.php` for `comingofbookimpa0000febv` | "Item not available" | Febvre and Martin cannot be searched without borrowing. Not accessed. | +| 23 | Google Books API | volumes query for Febvre and Martin | quota exceeded | Not retried. | +| 24 | Scholar Gateway | `semanticSearch` "licensing of the press and Stationers' Company monopoly in early modern England" | 0 | `ACCESS_DENIED: Access denied`. Not retried. | + +Snowballing: from Johns 2002 p. 119 to the question of how far the Index was enforced; from Innis p. 26 to the date of the lapse of licensing (compare Johns p. 124); from the 1566 ordinances back to the 1557 charter; from Milton to the Stationers' own 1643 papers in Arber. + +--- + +## (b) Entries + +### B1. Primary sources + +**arber1875transcript** — Arber, E., ed. (1875). *A Transcript of the Registers of the Company of Stationers of London, 1554–1640 A.D.*, vol. I. London: privately printed. https://archive.org/details/transcriptofregi01statuoft (OCR text: `transcriptofregi01statuoft_djvu.txt`). +- *Verification:* archive.org metadata (creator Arber, Edward, 1836–1912; "London Privately printed", 1875, volume 1). +- *Read status:* read in part: the English translation of the Stationers' charter (Introduction pp. xxviii–xxxii), the Star Chamber ordinances of 1566 (p. 322), and the Addendum on the controversy of 1643–5 (pp. 581–583). The rest of the 1,500,000-character volume was searched by keyword only. +- *Scale:* one city company with a national power of search; England and its dominions. +- *The charter (granted to the Stationers; Arber's introduction gives the date as 4 May 1556, and the charter's closing formula reads "the fourth day of May in the 3rd and 4th years of our reigns").* The ground stated for the grant, p. xxviii: "certain seditious and [h]eretical books rhymes and treatises are daily published and [p]rinted by divers scandalous malicious schismatical and heretical persons, not only moving our subjects and lieges to sedition and disobedience against us, our crown and dignity, but also to renew and move very great and detestable heresies". The monopoly, pp. xxx–xxxi: "no person within this our realm of England or the dominions of the same shall practise or exercise by himself or by his ministers, his servants or by any other person the art or mistery of printing any book or any thing for sale or traffic within this our realm of England or the dominions of the same, unless the same person at the time of his foresaid printing is or shall be one of the community of the foresaid mistery or art of Stationery of the foresaid City, or has therefore [li]cence of us" (OCR "hcence"; the clause runs on to "by the letters patent"). The power of search, p. xxxi: "make search whenever it shall [p]lease them in any place, shop, house, chamber, or building of any printer, binder or bookseller whatever within our [k]ingdom of England or the dominions of the same of or for any books or things printed, or to be printed, and to seize, take, hold, burn, or turn to the proper use of the foresaid [community], all and several those books and things which are or shall be printed contrary to the form of any statute, act, or proclamation". (OCR page footers "I. xxx" and "I. xxxi" bracket these passages; the English and Latin texts are interleaved, so the page split inside the clause is approximate.) +- *Date problem:* Arber's contents table is garbled in the OCR ("4 May 1666"); his introduction says 4 May 1556; I did not check any other source. I did not resolve the year. *(Added from AUDIT_01:)* The charter's own formula "3 & 4 Phil. & Mar." runs July 1556 to July 1557, so 4 May is 1557; Arber's introduction says 1556 in four places. +- *The Star Chamber ordinances of 1566 (p. 322; heading "Ordinances decreed for reformation of Disorders in printing and uttering of Bookes"; the date reads "xxix. daie of [I]une" with the year garbled in the OCR; the page sits between Arber's entries for 22 July 1565–22 July 1566 and 22 July 1566–22 July 1567).* Ordinance 2: "whosoeuer shall offend against the said ordinance, shall not onelie forfait all such Bookes and Copies, as shall be so printed or brought in; but also shall from [thence]foorth neuer vse or exercise, nor take benefit by anie vsing or exercising of the feate of printing, and shall susteine three moneths imprisonment, without baile or mainprise." Ordinance 5 gives the Wardens of the Company the right to search "all workehouses, shops, warehouses, and other places, of Printers, Bookesellers" and to present offenders "before the Queenes Maiesties hie Commissioners in causes Ecclesiasticall". +- *The Stationers' own papers, 1643 (Addendum).* The Company's petition to the Commons, dated in the OCR "[Ja]nu[ar]y 1643", says that "a very few Persons of the said Com[p]any, but Eleven in number, (whereas the Company consisteth of many hundred Families) labouring to acquire the Printing and profit thereof to themselves, have privately gained some promise from some of the said Divines, intending to exclude all others of the Company, from having any benefit thereby" (p. 582 in the OCR footers; the text sits just after the footer "I. 582"), and a later paragraph calls those eleven "Pattentees, and Monopolizers of Printing the most vendable Bookes of worth and quantity now in use" (p. 583). Thomason dated the Remonstrance "Aprill 1643"; Arber notes it was "By Henry Parker, Esq." Its p. 581 text says: "Muscovy professes enmity to Printing, and so does Turkie, for Learning's sake". That is a 1643 belief, not evidence (see B6, Schwartz). +- *Bearing on §4:* P1 supports (dates, places, a wrong named by insiders). P2: the Stationers' junior members and Milton named the monopoly a wrong in 1643–4. P3: England only. P4: not engaged. P5: not engaged. + +**statutesrealm1817vol3** — Great Britain, Record Commission (1817). *The Statutes of the Realm*, vol. 3. London. https://archive.org/details/the-statutes-of-the-realm-3 (OCR text `The_Statutes_of_the_Realm-3_djvu.txt`; archive.org marks it public domain). +- *Verification:* archive.org metadata (creator Great Britain, Record Commission; date 1817; public-domain mark). +- *Read status:* read in part: 34 & 35 Hen. VIII c. 1 and 27 Hen. VIII c. 26. +- *34 & 35 Hen. VIII c. 1 (A.D. 1542–3), the Act of 1543 on English Bibles; the OCR page markers "895" and "896" fall around the text, so the clauses lie on pp. 895–896.* It forbids the Tyndale translations and sets a licence for printed religious books ("by the King and his Clergye", with the printer's name). Then, on reading: "no maner of psone or psones after the firste daie of Octobre next ensuing, shall take upon him or them to reade preache or teache openlie to other in any churche or open assemblye within any the Kinges Domynions, the Byble or any parte of Scripture in Englishe ... oonles he be so appointed therunto by the King Majestie or by any Ordinarye". Stated ground for the second restriction: "a greate multytude of his saide subject[s], moste sp[e]ciallie of the lower sorte have so abused the same, that they have therbye growen and increased in divers naughtie and erronyous opynions, and by occasion therof fallen into greate dyvision and discencón amonges themselves, to the greate unquietnes of the Realme". The rule: "no woomen nor artificers prentises journeymen serving men of the degrees of yeomen or undre, husbandemen nor laborers shall reade within this Realme or in any other the King[s] Domynions, the Byble or Newe Testament in Englishe, to himselfe or [to] any other pryvatelie or openl[i]e, upon paine of oone monethes imprysonement for every tyme [so] offending". The exceptions: "it shalbe lawfull to everye noble man and gentleman being a householder to reade or cause to be red by any of his famylie or [servau]ntes in his house orchard or gardeyne, and to his owne famylie, any texte of the Byble or New Testament"; "everye marchaunte man being a householder and occupying the feate of merchaundyse" may "reade to himself privatelye"; and "everye noble wooman and gentlewooman maie reade to themselves alone and not to others any [Textes] of the Byble or Newe Testament". "All maner of psones other thenne woomen artificers prentyses journeymen serving men of the degrees of yeomen or undre, husbandemen and laborers maie reade to themselves and to none other any text of the Byble". Informers took half the fines: "the oone moytie of all the penalties and forfaictures ... shalbe to the King ... and the other moytie therof to suche psone or psones as will detecte and sue". +- *Whom the rule hit:* women below the rank of noblewoman or gentlewoman; apprentices, journeymen, servants below yeoman, husbandmen and labourers (men of those ranks were barred too). It barred reading, in this clause, not literacy as such; the Act is silent on whether the barred could read other texts. Of the barred groups, the Act does not say how many could read; I found no count. +- *27 Hen. VIII c. 26 (A.D. 1535–6), "An Acte for Lawes & Justice to be ministred in Wales in like fourme as it is in this Realme" (the Welsh Act of Union; OCR page markers near the clause read "566" and "560"; I could not fix the page).* This is not a print measure; I include it because it fixes a language bar in the print era. The stated ground, in the preamble: "by cause that the people of the same Dominion have and do daily use a speche nothing like ne consonaunt to the naturall mother tonge used within this Realme, sodme rude and ignorant people have made distinccion and diVsitie betwene the Kinges Subjectes of this Realme and hys Subjettes of the said Dominion and Principalitie of Wales, wherby greate discorde variaunce debate dyvysion murmur and sedicion hath growen". The rule: "all Justices Comissioners Shireves Coroners Eschetours Stewardes and thir Lieuten[a]ntes, and all other Officers and Ministers of the Lawe, shall pclayme and kepe the Sessions Courtes Hundred[es] Letes Shireves Courtes and all other Courtes, in the Englisshe Tonge"; and "no psonne or psones that use the Welsshe speche or langage shall have or enjoy any [maner] office or fees within the Realme of Englond Wales or other the Kinges Dominions, upon peyn of forfaiting the same offices or fees, onles he or they use and e[xer]cise the speche or langage of Englisshe." The legislation.gov.uk record for the Act reads "Laws in Wales Act 1535 (repealed 21.12.1993)"; that is the date the whole Act went, and I did not find when this clause was first repealed. +- *Scale:* the Welsh clause covers Wales and the Crown's dominions; the Bible clause covers England and the King's dominions. +- *Bearing:* P1 supports (a dated rank-and-sex bar on reading; a dated language bar on office). P2: the ground is stated in the preamble; no contemporary objector found in this volume. P3: single realm. P4: the 1543 bar is on a medium (the vernacular text) and the Welsh clause is on a language; neither is a bar on a coordination. P5: the Act presupposes that vernacular Bible reading changes opinion; the Welsh preamble presupposes language difference causes discord; neither promise is tested. + +**statutesrealm1819vol5** — Great Britain, Record Commission (1819). *The Statutes of the Realm*, vol. 5. London. https://archive.org/details/statutes_of_the_realm_5 (OCR text `statutes_of_the_realm_5_djvu.txt`). +- *Verification:* archive.org metadata (Great Britain Record Commission, 1819). +- *Read status:* read in part: 14 Car. II c. 33 (the Licensing of the Press Act, A.D. 1662; the contents table gives its first page as 428; the "twenty master printers" clause lies between OCR page markers 430 and 432) and the three later continuing Acts (chapters VII, VII and IV in three separate sessions; the OCR text of their running heads is garbled, so I do not give dates). +- *Content:* the title and preamble: "An Act for preventing the frequent Abuses in printing seditious treasonable and unlicensed Bookes and Pamphlets and for regulating of Printing and Printing Presses"; the ground: "no surer meanes can be ad[v]ised then by reducing and limiting the number of Printing Presses and by ordering and set[t]ing the said Art or Mystery of Printing by Act of Parliament". Licensing: "no private person or persons whatsoever shall att any time hereafter print or cause to be printed any Booke or Pamphlet whatsoever unlesse the same Booke and Pamphlet ... be first entred in the Booke of the Register of the Company of Stationers of London ... and unlesse the same ... shall be first lawfully licensed and authorized to be printed by such person and persons only as shall be constituted and appointed to license the same". Divinity, "Phisick, Philosophy or whatsoever other Science or Art" went to the Archbishop of Canterbury and the Bishop of London or the university chancellors; "all Books of History concerning the State of this Realm" to the Secretaries of State; common-law books to the Lord Chancellor and chief justices. The number of printers: "no man shall be admitted to ... be a Master Printer untill they who are now actually Master Printers shall be by death or otherwise reduced to the number of Twenty and from thence forth the number of Twenty Master Printers shall be continued and no more besides the Kings Printers and the Printers allowed for the Universities" (and "but four Master Founders of Letters"). Presses: "none of the said Master Printers ... shall keep above Two Printing Presses at once unlesse he hath been Master or Upper Warden of the Company who are hereby allowed to keepe Three Presses". Duration: "this A[c]t shall continue and be in force for two yeares to commence from the Tenth of June One thousand six hundred sixty and two and no longer". The later Acts (chapters VII, VII, IV) each continue it "untill the end of the next Session of Parlyament" or "the First Session of the next Parlyament". +- *Scale:* England. +- *Bearing:* P1 supports: a statute that fixes the number of masters at twenty and names who may license each class of book. P2: the preamble's ground is public order and religion; the contemporary objection is Milton's in 1644 (B1, below), and the Stationers' own 1643 petition. I did not find an account of who was refused entry after 1662. P4: a bar on a medium. P5: not engaged. + +**manuzio1564index** — *Index librorum prohibitorum, cum regulis confectis per patres a Tridentina Synodo delectos, auctoritate sanctiss. D.N. Pii 4., Pont. Max. comprobatus* (1564). Venice/Rome: Paolo Manuzio (Aldine; the place is not in the metadata I read). https://archive.org/details/ita-bnc-ald-00000309-001 (copy from the Biblioteca Nazionale Centrale di Firenze, via Early European Books). +- *Verification:* archive.org metadata (title, Manuzio, Paolo, 1564). +- *Read status:* read in part: Regulae Indicis, rules IV, VI and X (Latin OCR with long-s misread as "f"; I normalise "f" to "s" and "u"/"v" without changing letters). Page numbers are inferred from the printed running heads: rule IV falls on p. 15, rule X on pp. 18–19. +- *Rule IV, "Regula Quarta":* "si sacra Biblia uulgari lingua passim sine discrimine permittantur, plus inde, ob hominum temeritatem, detrimenti, quam utilitatis oriri; hac in parte iudicio Episcopi, aut Inquisitoris stetur; ut cum consilio Parochi, uel Confessarii Bibliorum, a Catholicis auctoribus uersorum, lectionem in uulgari lingua eis concedere possint, quos intellexerint ex huiusmodi lectione non damnum, sed fidei, atque pietatis augmentum capere posse." Then: "Qui autem absque tali facultate ea legere, seu habere praesumpserit; nisi prius Bibliis Ordinario redditis, peccatorum absolutionem percipere non possit." Booksellers who sold such Bibles "non habenti" the licence lost the price of the books and could be punished at the bishop's discretion ("poenis pro delicti qualitate eiusdem Episcopi arbitrio"). (OCR damage inside the words "facultate", "habere", "absolutionem" is normalised; the Latin is otherwise as printed.) +- *Rule VI* extends the same treatment to vernacular books on controversy. *Rule X* carries the licensing rule for printing: in Rome a book is to be examined first "per Vicarium summi Pontificis, & sacri Palatii Magistrum, uel per personas a Sanctissimo Domino nostro deputandas"; elsewhere by the bishop or a deputy and the inquisitor, "eius approbatio & examen pertineat". +- *Scale:* the Catholic church in principle; enforcement varied by state (B2, Johns p. 119). +- *Bearing:* P1 supports: a dated, text-quoted permission rule for reading vernacular Bibles. P2: the stated ground is that unrestricted vernacular reading does "more harm than good" because of "hominum temeritatem". Milton objected to the Index and to licensing in 1644 (below), though he did not name Rule IV. P3: church-wide, not world. P4: a bar on a medium. P5: not engaged. + +**milton1644areopagitica** — Milton, J. (1644). *Areopagitica: A Speech for the Liberty of Unlicensed Printing*. Project Gutenberg eBook no. 608. https://www.gutenberg.org/cache/epub/608/pg608.txt +- *Verification:* Gutenberg header ("The Project Gutenberg eBook of Areopagitica"); I did not collate against a first edition and the file has no page numbers. I cite by passage. +- *Read status:* read in part (searched for licensing, monopoly, Index and Trent passages and read each in context). +- *What it argues and the passages:* the history of licensing runs from the Popes to Trent: "the Council of Trent and the Spanish Inquisition engendering together brought forth, or perfected, those Catalogues and expurging Indexes, that rake through the entrails of many an old good author"; the Italian licensing statements are mocked. On who gained from the 1643 Order: "it may be doubted there was in it the fraud of some old patentees and monopolizers in the trade of bookselling; who under pretence of the poor in their Company not to be defrauded, and the just retaining of each man his several copy, which God forbid should be gainsaid, brought divers glossing colours to the House, which were indeed but colours, and serving to no end except it be to exercise a superiority over their neighbours". And: "what magistrate may not be misinformed, and much the sooner, if liberty of printing be reduced into the power of a few?" And: "Truth and understanding are not such wares as to be monopolized and traded in by tickets and statutes and standards." +- *Scale:* England, the Long Parliament, 1644; Milton speaks of Rome and Spain as the model. +- *Bearing:* P2 supports strongly: a contemporary who named the licensing-and-monopoly regime a wrong, on stated grounds, and named the beneficiaries. P3: single author, single country. P1/P4/P5: not engaged. + +### B2. Secondary sources read in full text + +**johns2002how** — Johns, A. (2002). How to acknowledge a revolution. *American Historical Review*, 107(1), 106–125. https://doi.org/10.1086/ahr/107.1.106 +- *Verification:* Crossref (title, volume, issue, pages, author match). Text from the OUP-typeset PDF on the ENS course site; page numbers from the running heads (pp. 106–125). +- *Read status:* read in part: the passages cited here (pp. 109, 112, 119–120, 122, 124); the repo card records a full read (S2-verified 2026-09-18) and I did not repeat that. +- *What it argues:* the book historian's reply to Eisenstein: print has no single "culture"; fixity and credit were made by local communities of printers, booksellers, readers and censors. +- *Quotations:* + - On the thesis of a unitary print culture, p. 109: "What really stands to lose its acknowledgement is something much more specific: the particular concept of a revolutionary shift to a unitary print culture." + - On Eisenstein and censorship, p. 119 (Johns's paraphrase): "She asserts that censorship created a sweeping asymmetry in European culture". Johns's reply, same page: "Authorities differed widely in the rigor with which they policed the press-in France, for example, the Index was effectively ignored." And: "I can think of no Protestant state before 1695 that had no regulatory regime governing its press". And: "Contemporaries themselves saw substantial differences between licensing and other procedures that constrained authorship. Not least, licensing was part and parcel of the regimes of privilege and registration, which were designed to protect good works as well as to hinder bad." + - On the defence of privilege, p. 112 (Johns quoting his 1998 book): "Supporters of the patents regime, on the other hand, retorted that the security associated with privileges enabled the production of worthwhile texts that would never appear at all in other circumstances." + - On the date of the idea, p. 122: "the specific concept of a printing revolution came about late in the day, arriving around 1800 and achieving the status of an orthodoxy only in the mid-nineteenth century." + - On the end of licensing, p. 124: "the end of licensing in 1695". +- *Scale:* England and Europe at large, argued about London; Johns concedes his book's "geographical focus is far too restricted" (p. 109) to replace Eisenstein's. +- *Bearing:* P1 complicates: Johns gives no figures on exclusion and says the history of licensing must be separated from "censorship". P2: documents what contemporaries called licensing and privilege. P3: complicates (locality). P4/P5: not engaged. + +**innis1951bias** — Innis, H. A. (1951). *The Bias of Communication*. Toronto: University of Toronto Press. https://archive.org/details/the-bias-of-communication-harold-innis (OCR text; running heads give the page). +- *Verification:* archive.org item; the repo card `innis1951.md` has the edition; I did not collate against another edition. +- *Read status:* read in part: pp. 22–30 ("Minerva's Owl", the print passages), pp. 53–56 and p. 60 ("The Bias of Communication"). +- *What it says on print:* + - p. 26: "Suppression of printing in response to ecclesiastical demands was accompanied by an interest in the increased production and export of paper by the state. As a result of this conflict between restriction of consumption and increase of production and exports, cheaper supplies were available in the Netherlands, Geneva, and countries in which printing remained free. From these marginal areas printed material was smuggled back into France. Freedom of the press in marginal free countries was supported by repression in France." + - p. 26 on the date of the lapse: "after the revolution of 1689 and the lapse of the Licensing Act in 1694". + - p. 53: "Monopolies of knowledge controlled by monasteries were followed by monopolies of knowledge controlled by copyist guilds in the large cities." + - p. 54: "Learning declined in France in the sixteenth century but the vernacular found fresh support in printers shown in the writings of Montaigne and Rabelais. French became an official language after 1539." + - p. 55: "In England the absolutism of the Tudors involved suppression of printing but encouragement of the Renaissance and of the Reformation." + - p. 60: "The monopoly of knowledge centring around the printing press brought to an end the obsession with space and the neglect of problems of continuity and time." +- *Scale:* Europe, five centuries in a page or two. +- *Bearing:* P1 complicates: Innis's "monopoly of knowledge" names a medium's organisation (monasteries, copyist guilds, the press) and not a group of excluded people; he names no one who lost standing. P2: no contemporary objector. P3: Europe-wide, undated detail. The date of the Licensing Act's lapse (1694) disagrees with Johns's 1695; I did not settle it. + +### B3. Secondary sources read in part + +**martel2001autour** — Martel, P. (2001). Autour de Villers-Cotterêts: histoire d'un débat. *Lengas*, 49, 7–25 (online reprint by FELCO, 2025, with added subheadings). https://www.felco-creo.org/wp-content/uploads/2025/08/2508-19-Martel-Villers-cotterets-lengas.pdf +- *Verification:* the reprint's own header gives the 2001 citation; no Crossref record found. +- *Read status:* read in part (the opening and the Brunot and Brun sections, about 15 pages scanned; the sections on Peyre and Boulard only by keyword). Page numbers: the reprint is unpaginated for the original, so I cite by section. +- *Arguments:* the question, in his words: "S'agit-il seulement d'éliminer le latin des actes de la pratique ou bien les langues autres que le français présentes dans le royaume sont-elles aussi visées ? La suite de l'histoire montre bien que c'est la seconde hypothèse qui est la bonne." Two readings of "langage maternel françois": Brunot's "théorie limitative" ("désigne clairement et uniquement le français du roi") and Peyre's "théorie libérale" ("est maternelle française toute langue parlée par un sujet natif du royaume"); Martel says the first is today the majority view. On the south: Brun found "le recul des « parlers locaux » devant le français" starting at the end of the Middle Ages, and that the ordinance of 1539 had no opposition, yet "les progrès du français, comme langue parlée et pas seulement écrite, dans le Midi de la France vont toutefois être très lents" because the monarchy "ne s'implique jamais réellement dans un effort de diffusion de la langue du royaume". +- *Scale:* the kingdom of France. +- *Bearing:* P1 complicates: the exclusion of Occitan from legal writing is documented only through a historiographic debate, and the ordinance does not mention print. P2: no contemporary objector found. P3: national. + +**villerscotterets1539lexilogos** — Lexilogos (n.d.). Ordonnance de Villers-Cotterêts, 1539 (transcription of arts. 110–111). https://www.lexilogos.com/francais_villers_cotterets.htm +- *Verification:* not an official copy; transcription attributed on the page to *Ordonnances royaulx sur le faict de la justice*, 1539. No Légifrance copy reached. +- *Read status:* read in part (arts. 110–111 and the page's note). +- *Article 111:* "nous voulons que doresnavant tous arrestz ensemble toutes autres procedures ... soient prononcez, enregistrez & delivrez aux parties en langage maternel françois et non autrement". The page's note adds earlier texts (Moulins 1490; June 1510; Provence, October 1535) that allowed the local vulgar tongue. +- *Bearing:* a secondary transcription only; a human with library access should replace it with the Isambert collection or an archival copy. + +**sanchezmarti2023advent** — Sánchez-Martí, J. (2023). The advent of the printing press and Britain's multilingual textual culture, 1471–1510. *Anglica: An International Journal of English Studies*, 32(4), 83–107. https://doi.org/10.7311/0860-5734.32.4.05 (open copy: https://rua.ua.es/dspace/handle/10045/138259) +- *Verification:* Crossref (title, volume/issue "32/4", pages 83–107, author). +- *Read status:* read in part: the abstract, pp. 92–93 and 97 and the notes. +- *Quotations:* p. 92: "While the early printers made no concerted effort to standardize the language, the fact of the matter is that London-based English with all its variation and lack of consistency became the dialect of choice for printed texts. The printers' preference for this metropolitan variety of English gradually reduced their level of tolerance for all other regional dialects that had enjoyed written transmission during the manuscript period". p. 93, on Wynkyn de Worde's 1499 printing of an Older Scots text: "De Worde's edition effectively erases all information that might betray the work's Scottish origin." p. 97 on Welsh: "The first book printed in Welsh was issued in London in 1546 (STC 201310), and it was not until 1585 when a book in Welsh was actually printed in Wales." And the conclusion, p. 97: "the introduction of the printing press in Britain, as happened in the rest of Europe, contributed to enhance the prestige of the language variety used by the elite, while giving little or no visibility to other languages and dialects, thus becoming instrumental in eroding the balances existing in the British language ecosystem." +- *Scale:* England and Scotland, 1471–1510; commercial printers. +- *Bearing:* P1 supports with a case-level record; but the same page says the printers did not set out to standardise. P2: no contemporary objector (the printers did not call it a wrong). P3: Britain. P5: complicates the cause (market, not rule). + +**sasaki2017publishing** — Sasaki, Y. (2017). Publishing nations: technology acquisition and language standardization for European ethnic groups. *Journal of Economic History*, 77(4), 1007–1047. https://doi.org/10.1017/S0022050717000821 +- *Verification:* Crossref; open PDF from Cambridge Core. +- *Read status:* read in part: abstract, introduction, the Breton and conclusion passages; I did not audit the models. +- *Quotation, p. 1015 (PDF order):* "Despite the legal exclusion of Breton use in the French school system for 1880–1951, Breton use became more widespread in journals and periodicals, culminating in the 1958 monolingual dictionary". p. 1042: "This article has sought to demonstrate that such conditions were critical to standardization for many minority tongues in Europe that have survived despite institutional centralization in modern times." The abstract: for "171 ethnic groups in Europe from 1400–2000 CE", "printing press adoption is positively and significantly correlated with vernacular dictionaries; and (2) early adopters are more likely to standardize vernaculars than latecomers." +- *Scale:* 171 European ethnic groups over six centuries; the measure is the first vernacular dictionary. +- *Bearing:* complicates P1: for minority languages the article finds print enabled standardisation; it does not measure loss. The dictionary proxy captures who got a standard, not who lost a dialect. Page numbers are inferred from PDF order. + +### B4. Abstract only, or secondary summaries + +**schwartz2017did** — Schwartz, K. A. (2017). Did Ottoman sultans ban print? *Book History*, 20(1), 1–39. https://doi.org/10.1353/bh.2017.0000 +- *Verification:* Crossref. *Read status:* abstract only, via a blog quotation (Marginal Revolution, 7 April 2023): "I argue that the sources available to scholars today do not support the notion that the sultans banned printing." The same abstract says the claim "arose from early modern European scholars' search to articulate their sense of Ottoman inadequacy." Other details (Thevet 1584; no surviving firmans) come from search-result summaries of the article and I do not rely on them. +- *Bearing:* contradicts the 1643 Stationers' belief (B1). P3: this is the one non-European print case in the facet and it is a null result. + +**houston2003lesser** — Houston, R. A. (2003). 'Lesser-used' languages in historic Europe: models of change from the 16th to the 19th centuries. *European Review*, 11(3), 299–324. https://doi.org/10.1017/S1062798703000309 — abstract only (Consensus). Abstract: the period "saw the beginnings of a decline in the use of certain dialects and separate languages, notably Irish and Scottish Gaelic, although some tongues such as Catalan and Welsh remained widely used", and the relation to context "was mediated through the availability of printed literature; the political (including military) relations ...; the nature and relative level of economic development ...; the policy of the providers of formal education and that of the church ...; and ... local social structures and power relationships." Full text closed (Unpaywall). A human with library access should read it first for the European scale. + +**suggett2018language** — Suggett, R., & White, E. (2018). Language, literacy and aspects of identity in early modern Wales. In *The spoken word* (Manchester University Press). https://doi.org/10.7765/9781526137876.00006 — abstract only (Consensus); the open PDF returned a CloudFront 403 and I did not bypass it. Abstract: "Wales alone among the Celtic languages developed a vigorous print culture in the vernacular that built upon a written tradition reaching back into the Middle Ages"; "the survival and development of the written language ... became an essential condition for the reproduction of the modern spoken language." This cuts against the reading of the 1536 clause as the cause of Welsh decline. + +### B5. Not accessible + +**eisenstein1979printing** — Eisenstein, E. L. (1979). *The Printing Press as an Agent of Change*. Cambridge University Press. Not accessible: archive.org lending only (`printingpressas01eise`, `printingpressas02eise`), no snippets. I have only Johns's reports of her position (B2, p. 119: censorship "created a sweeping asymmetry in European culture"). The repo card is secondary; I did not upgrade it. + +**eisenstein2002unacknowledged** — Eisenstein, E. L. (2002). An unacknowledged revolution revisited. *American Historical Review*, 107(1), 87–105; reply 126–128 (https://doi.org/10.1086/ahr/107.1.87; https://doi.org/10.1086/532100). Crossref resolves both. Not accessible: OUP page paywalled. Her words are known only through Johns. + +**johns1998nature** — Johns, A. (1998). *The Nature of the Book*. University of Chicago Press. Not accessible (lending only); known through Johns 2002 and the repo card. + +**febvre1976coming** — Febvre, L., & Martin, H.-J. (1976). *The Coming of the Book: The Impact of Printing 1450–1800*. London: NLB. Not accessible: three lending-only copies (`comingofbookimpa0000febv*`), search-inside refused ("Item not available"), Google Books API quota exceeded. I state nothing from it. + +**darnton1995censorship** — Darnton, R. (1995). Censorship, a comparative view: France, 1789–East Germany, 1989. *Representations*, 49, 40–60. https://doi.org/10.2307/2928748 (Crossref resolves). Not accessible: the UC Press PDF is marked open in Unpaywall but served a Cloudflare challenge; I did not bypass it. I read no Darnton text; every Darnton claim is therefore absent. + +--- + +## (c) Synthesis + +### Failure 1. Exclusion from the press by corporate monopoly (England, charter of 4 May 1556 on Arber's dating, 1557 by the regnal formula, to the 1695 lapse) + +- *What was done.* The Crown granted the London Stationers the exclusive right to print: no one in the realm could print for sale unless a member of the Company or licensed by royal patent (charter, Arber I pp. xxx–xxxi). The Company could enter any printer's, binder's or bookseller's premises and "seize, take, hold, burn" books printed contrary to law (p. xxxi). Star Chamber in 1566 added loss of the right to print for ever and three months' imprisonment without bail for printing against law (Arber I p. 322). In 1662 Parliament fixed the number of master printers at twenty, plus the King's printers and the universities', and four letter-founders; no master could keep more than two presses (three for past masters or wardens) (Statutes of the Realm vol. 5, 14 Car. II c. 33). All printed books needed a licence from named officers by subject (same Act). +- *To whom.* Anyone outside the Company: printers, would-be printers, authors who wanted to print. The Company's own junior members were also shut out of the Bible-with-notes trade in 1643 by eleven patentees (Arber I pp. 582–583). +- *Where and when.* London, and the kingdom, from the charter of 4 May (1556 in Arber; 1557 by the regnal year) to the lapse of licensing (Johns p. 124: 1695; Innis p. 26: 1694). +- *Stated ground.* Sedition and heresy (charter p. xxviii); "the well-government and regulating of Printers and Printing Presses is matter of Publique care" (1662 preamble). The defenders of privilege also argued it paid for expensive books: "the security associated with privileges enabled the production of worthwhile texts that would never appear at all in other circumstances" (Johns p. 112, quoting his own 1998 summary of the patents regime's supporters). +- *Who objected at the time.* Milton, 1644: "old patentees and monopolizers in the trade of bookselling" brought "glossing colours" to Parliament, and "liberty of printing" should not be "reduced into the power of a few" (*Areopagitica*). The junior Stationers, 1643: eleven men "labouring to acquire the Printing and profit thereof to themselves ... intending to exclude all others of the Company" (Arber I p. 582). +- *What ended it.* Non-renewal of the licensing Acts; Johns gives "the end of licensing in 1695" (p. 124) and Innis "the lapse of the Licensing Act in 1694" (p. 26). I did not find the lapse in the statute. Johns adds that the trade did not collapse because contemporaries feared it would, and that this produced the first copyright law in 1710 (p. 124, citing Feather); I did not open Feather. +- *Evidence depth.* Statute and charter text: full text of the passages quoted. The count of people excluded: not found. + +### Failure 2. Bar on reading the English Bible by rank and sex (England and dominions, 1543) + +- *What was done.* 34 & 35 Hen. VIII c. 1 forbade "woomen" (except noble women and gentlewomen, who could read to themselves alone), "artificers prentises journeymen serving men of the degrees of yeomen or undre, husbandemen" and "laborers" to read the Bible or New Testament in English, privately or openly, on pain of one month's imprisonment, and left noblemen and gentlemen who were householders free to read, or have read, to their families, and merchant householders free to read privately. Informers shared the fines (Statutes of the Realm vol. 3, pp. 895–896). +- *To whom.* Women below the gentry; apprentices, journeymen and serving men of yeoman rank or below; husbandmen and labourers. +- *Where and when.* England and "the King's Dominions", from "the firste daie of Julye next cooming" (the Act's own start date for this clause; the Act belongs to the session of 1542–3, and search-result summaries of Wikipedia give royal assent as 12 May 1543, which I did not check against an official copy). +- *Stated ground.* "a greate multytude of his saide subject[s], moste sp[e]ciallie of the lower sorte have so abused the same" that they "growen and increased in divers naughtie and erronyous opynions". +- *Who objected at the time.* Not found. I did not locate a contemporary text naming the 1543 clause a wrong. The standard by which it counts as a wrong here would have to be stated by the author. +- *What ended it.* Search-result summaries say the Act was repealed in 1547; unverified against the statute. +- *Evidence depth.* Statute text read in full for the clauses quoted. Any figure for how many people it affected or how many could read: not found. + +### Failure 3. Permission to read the vernacular Bible, and pre-publication licensing, under the Tridentine Index (Catholic Europe, 1564) + +- *What was done.* Rule IV made reading Catholic vernacular Bibles depend on the written permission of bishop or inquisitor, on the advice of the parish priest or confessor; a reader who read or held one without it could not be absolved until he handed it in, and a bookseller who sold to a reader without the permission lost the price of the books and could be punished (1564 Index, p. 15). Rule X required prior examination and approval of any book before printing (pp. 18–19). +- *To whom.* Lay readers of vernacular Bibles in Catholic jurisdictions; booksellers. +- *Stated ground.* Experience shows that unrestricted reading brings "plus ... detrimenti, quam utilitatis" because of "hominum temeritatem". +- *Who objected at the time.* Milton (1644), who attacks the Index system and the Roman and Spanish licensing as the model for the English Order; he does not name Rule IV. +- *What ended it.* Not found. +- *Contested scope.* Johns, p. 119: "in France, for example, the Index was effectively ignored", and, against Eisenstein's contrast between censored and uncensored Europe (as Johns reports it), Johns writes that "the contrastive appeal to censorship involves an appeal to a null set" because no Protestant state before 1695 lacked a regulatory regime. Innis, p. 26, shows suppression of printing in France by other, ecclesiastical and state powers, so the claims do not collide directly; but a reader should not treat the Index as a uniform continent-wide rule. Full text of Rule IV: read; evidence of enforcement: not found. + +### Failure 4. Language bars at the point of law and the standardisation of one vernacular (Wales 1536; France 1539; London English and Older Scots in printing, 1471–1510) + +- *Wales, 1536.* The Act of Union required English in all courts and barred anyone who "use[s] the Welsshe speche or langage" from any office or fees unless he used English (27 Hen. VIII c. 26). The preamble's ground: a speech "nothing like ne consonaunt to the naturall mother tonge" had caused "distinccion and diVsitie" and "sedicion". Who objected: not found. It is not a print measure; the text I read does not mention print. +- *France, 1539.* Articles 110–111 required judicial acts to be given "en langage maternel françois et non autrement". Whether this excluded Occitan and other languages is disputed: Brunot reads the phrase as the king's French only, Peyre as any language of a native subject; Martel says the first is now the majority view and that written Occitan retreated from the end of the Middle Ages (Martel, sections on Brunot and Brun). Evidence depth: Lexilogos transcription (secondary) and Martel in part. +- *Printing itself.* In London and Edinburgh, the early printers favoured a London English or a prestige Older Scots over other dialects (Sánchez-Martí pp. 92–93, 97), and Welsh was printed late, in London in 1546 and in Wales in 1585 (p. 97). The same pages say the printers made no concerted effort to standardise. +- *Counter-evidence.* Sasaki's 171-group study finds that press adoption predicts a vernacular dictionary; the study measures who standardised, not who lost a dialect. Suggett and White (abstract only) argue Welsh survived because it had a print culture. +- *Scale.* A kingdom, a city press, a continent-wide sample; no world scale. + +### Claims the sources contradict, or leave open + +1. **A Turkish or Ottoman ban on printing.** The Stationers' 1643 Remonstrance states it (Arber I p. 581). Schwartz's abstract says the sources "do not support the notion that the sultans banned printing" and traces the claim to European writers. A print-era "ban" on the Ottoman side therefore cannot be asserted from my sources. +2. **The Index as a continent-wide wall.** Johns p. 119: in France "the Index was effectively ignored". The Index text itself is a rule; enforcement is a separate fact I did not find. +3. **Date of the lapse of licensing.** Johns p. 124: 1695. Innis p. 26: 1694. Unresolved. +4. **Date of the Stationers' charter.** Arber's introduction: 4 May 1556; his contents-table OCR: garbled. Other literature not checked. Unresolved. +5. **Print as the cause of dialect loss.** Sánchez-Martí says it "contributed" (p. 97) and that printers did not set out to standardise (p. 92); Sasaki finds press adoption goes with standardisation of minority tongues; Suggett and White (abstract) say Welsh print culture helped Welsh survive; Martel says Villers-Cotterêts was a deliberate exclusion. These do not agree on cause or on whether anyone was wronged. +6. **Villers-Cotterêts, scope.** Brunot (king's French only) against Peyre (any native language), per Martel. +7. **A "printing revolution."** Johns (p. 122) dates the concept to around 1800; Eisenstein holds the revolution thesis. This is the historiography the paper's print half would have to take a side on, and my sources do not let me weigh Eisenstein directly. +8. **Innis's "monopoly of knowledge."** It describes an organisation of a medium (monasteries, copyist guilds, the printing press), not a bar on named persons. The repo's draft use of it as a wrong done to people goes beyond what Innis wrote on the pages I read. + +--- + +## (d) Gaps and unverified leads + +- **Eisenstein, *The Printing Press as an Agent of Change* (1979) and *The Printing Revolution* (1983/2005)**; **Eisenstein 2002** (AHR 87–105, 126–128). Not opened. Where: Bentley library or interlibrary loan; archive.org lending (`printingpressas01eise`, `printingrevoluti00eise_0`); OUP/JSTOR for the AHR forum. +- **Febvre and Martin, *The Coming of the Book*** (NLB 1976; Verso 1984/1997). Not opened. Where: library copy, or archive.org borrow (`comingofbookimpa0000febv`). The question to put to it: its figures on presses, on language of imprints and on reading, with page numbers. +- **Darnton**, "Censorship, a comparative view" (Representations 49, 1995; https://doi.org/10.2307/2928748) and his books on the French book trade. Not opened. Where: UC Press (bot-walled for me) or JSTOR through a library. +- **Johns, *The Nature of the Book*** (1998), full text for the Stationers' Company and privilege chapters. Where: library. +- **Siebert, *Freedom of the Press in England 1476–1776*** (1965), lending only on archive.org (`freedomofpressin0000unse_i1v0`). Where: library. +- **Schwartz 2017** in full (Book History 20:1–39; Project MUSE). Abstract only here. Where: library via Project MUSE. +- **Houston 2003** (European Review 11:299–324) and **Suggett and White 2018** in full; the second is open access at Manchester Open Hive but the PDF is bot-walled for me. Where: browser at https://www.manchesteropenhive.com/view/9781526137876/9781526137876.00006.xml . +- **Statutes still to open.** The Welsh Bible Act (5 Eliz. I) for the counter-case; the Irish Act of 1537 on English language and habit; the Scottish Privy Council act of 10 December 1616 (Register of the Privy Council of Scotland, vol. 10, p. 777, as given in one search result; not checked); the Act of Parliament or document that ended the 1543 Bible restrictions; the continuing Acts that fix the lapse of licensing in 1694 or 1695. Where: *Statutes of the Realm* vols. 3–6 on archive.org, or British History Online; the Register of the Privy Council on archive.org. +- **The Star Chamber decree of 1586** and **the decree of 1637**: only mentioned in Arber I. Arber vol. II (1576–1595), on archive.org as `transcriptofregi02statuoft`, should hold the 1586 decree; my keyword search of it found no hit and I did not read it by page. A human should check pp. 807ff. +- **The Index in Latin with later rules and the 1559 Paul IV Index**, for how Rule IV changed; Reusch, *Der Index der verbotenen Bücher* (1883–85; `derindexderverb00reusgoog` on archive.org) is an open lead. Not opened. +- **Villers-Cotterêts**: an official copy (Isambert's *Recueil des anciennes lois françaises*, via Gallica) in place of Lexilogos; Boulard 1999 (Revue historique 609:45–100), the article Martel discusses. Not opened. +- **Non-European print cases and minority-language print** beyond Britain and France: Ottoman non-Muslim presses, Hebrew printing and the Roman Index, Spanish America. I found nothing I could open. Grendler's *The Roman Inquisition and the Venetian Press* (cited by Johns n. 31) is a lead. +- **Counts.** I found no figure in any source I opened for how many people the 1543 bar, the 1662 limit on presses or Rule IV excluded, and none for how many could read. The paper cannot state a number from this facet. +- **Scholar Gateway** returned ACCESS_DENIED; OpenAlex returned HTTP 429. diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.bib b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.bib new file mode 100644 index 00000000..59384356 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.bib @@ -0,0 +1,112 @@ +% Facet V1 (round 7, 2026-10-14): Nadella's post and the disclosure paradigm. +% Reused keys (entries copied here for the facet; no change to references.bib): +% nadella2026models (sister arm DXD, card library/nadella2026models.md, S2-verified) +% eu2024aiact, euomnibus2026 (references.bib; read this session from EUR-Lex) +% DOIs: only Amershi et al. has one; checked against Crossref /works/10.1145/3290605.3300233 on 2026-10-14 +% (title, proceedings, pages 1-13, 13 authors match). The Microsoft documents and the statutes have no DOI. + +@misc{nadella2026models, + author = {Nadella, Satya}, + title = {Models as Insider Risks in the Super Intelligence Era}, + year = {2026}, + howpublished = {sn scratchpad (personal blog), 10 October 2026; also posted on X}, + url = {https://snscratchpad.com/posts/models-as-insider-risks/}, + note = {Read in full from the local capture (P1--P21, capturer's numbering); DXD card library/nadella2026models.md compared it with the blog page. Quote only short fragments with paragraph numbers.} +} + +@misc{eu2024aiact, + author = {{European Parliament and Council of the European Union}}, + title = {Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)}, + year = {2024}, + howpublished = {Official Journal of the European Union, L, 12.7.2024}, + url = {http://data.europa.eu/eli/reg/2024/1689/oj}, + note = {V1: Arts 3(56), 4, 13, 14, 26, 50, 73, 85, 86 and recitals 20, 91 read on EUR-Lex (CELEX 32024R1689, HTML) on 2026-10-14; text identical to the Publications Office copy.} +} + +@misc{euomnibus2026, + author = {{European Parliament and Council of the European Union}}, + title = {Regulation (EU) 2026/1744 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI)}, + year = {2026}, + howpublished = {Official Journal of the European Union, L series, 24.7.2026}, + url = {http://data.europa.eu/eli/reg/2026/1744/oj}, + note = {V1: recital 8, Art 1(5) (replacement Art 4), Art 1 points on Arts 50(7), 111 and 113, and Art 4 (entry into force) read on EUR-Lex on 2026-10-14.} +} + +@techreport{microsoft2022responsible, + author = {{Microsoft}}, + title = {Microsoft Responsible AI Standard, v2: General Requirements}, + institution = {Microsoft Corporation}, + year = {2022}, + month = jun, + type = {Standard (for external release)}, + url = {https://msblogs.thesourcemediaassets.com/sites/5/2022/06/Microsoft-Responsible-AI-Standard-v2-General-Requirements-3.pdf}, + note = {Read in full (27 pp.).} +} + +@inproceedings{amershi2019guidelines, + author = {Amershi, Saleema and Weld, Dan and Vorvoreanu, Mihaela and Fourney, Adam and Nushi, Besmira and Collisson, Penny and Suh, Jina and Iqbal, Shamsi and Bennett, Paul N. and Inkpen, Kori and Teevan, Jaime and Kikin-Gil, Ruth and Horvitz, Eric}, + title = {Guidelines for Human-AI Interaction}, + booktitle = {Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems}, + year = {2019}, + pages = {1--13}, + publisher = {ACM}, + doi = {10.1145/3290605.3300233}, + note = {Read in part: abstract and Table 1 (p. 3).} +} + +@techreport{microsoft2024responsible, + author = {{Microsoft}}, + title = {Responsible AI Transparency Report: How We Build, Support Our Customers, and Grow}, + institution = {Microsoft Corporation}, + year = {2024}, + url = {https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/Responsible-AI-Transparency-Report-2024.pdf}, + note = {Read in part: pp. 4, 11, 25, 29, 33, 34 of 40. Release month not checked.} +} + +@techreport{microsoft2025responsible, + author = {{Microsoft}}, + title = {2025 Responsible AI Transparency Report}, + institution = {Microsoft Corporation}, + year = {2025}, + url = {https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2025-Responsible-AI-Transparency-Report.pdf}, + note = {Read in part: PDF pp. 7, 22, 26, 28, 34 of 35 (spread layout; printed page numbers differ, e.g. PDF p. 26 = printed p. 49).} +} + +@misc{microsoft2026application, + author = {{Microsoft}}, + title = {Application card: Microsoft Copilot (for organizations) [Transparency Note for Microsoft 365 Copilot]}, + year = {2026}, + howpublished = {Microsoft Learn, page dated 5 October 2026 (ms.date)}, + url = {https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-transparency-note}, + note = {Read in full on 2026-10-14. The URL still says transparency-note; the page title now says Application card.} +} + +@misc{microsoft2025ai, + author = {{Microsoft Office of Responsible AI}}, + title = {AI Literacy Starting Guide: Securing the Benefits of AI through Literacy for All}, + year = {2025}, + month = jan, + howpublished = {Microsoft, PDF, 11 pp.}, + url = {https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/final/en-us/microsoft-product-and-services/security/pdf/microsoft-ai-literacy-starting-guide-jan-2025-477720.pdf}, + note = {Read in full. Authorship as stated in the 2025 Transparency Report (PDF p. 26): released by the Office of Responsible AI.} +} + +@misc{smith2025microsoft, + author = {Smith, Brad}, + title = {Microsoft Elevate: Putting people first}, + year = {2025}, + month = jul, + howpublished = {Microsoft On the Issues (blog), 9 July 2025}, + url = {https://blogs.microsoft.com/on-the-issues/2025/07/09/elevate/}, + note = {Read in full.} +} + +@misc{sudjianto2026when, + author = {Sudjianto, Agus}, + title = {When AI Becomes an Insider, Who Is Left to Check Its Work?}, + year = {2026}, + month = oct, + howpublished = {Substack (Agus's Substack), 10 October 2026}, + url = {https://agussudjianto.substack.com/p/when-ai-becomes-an-insider-who-is}, + note = {Read in full. Not peer reviewed. datePublished 2026-10-10T17:24 UTC in page metadata.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.md b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.md new file mode 100644 index 00000000..f99779e6 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V1_nadella_disclosure_paradigm.md @@ -0,0 +1,213 @@ +# V1 — Nadella's post and the disclosure paradigm: the post, Microsoft's own record, and the EU AI Act + +*Facet V1 of research round 7 (version 4), 2026-10-14. Agent-drafted; the author has not reviewed it, and none of the text is the author's. Notes for the drafter, not prose for the paper.* + +**What this facet covers.** (Q1) What Nadella's post of 10 October 2026 proposes, move by move (P1–P21), sorted into (i) a text, record or disclosure to be read, (ii) a control held by the deploying organisation, (iii) something that builds a person's capacity; who "we" is; where the people the systems act on appear; who must read the P15 evidence. (Q2) Microsoft's own published responsible-AI record on transparency and on users: the Responsible AI Standard v2 (2022), the Responsible AI Transparency Reports (2024, 2025), a current Transparency Note, the AI Literacy Starting Guide, the Elevate skilling commitment, and the Guidelines for Human-AI Interaction that the Standard tells teams to follow. (Q3) The EU AI Act's version of the same move: Articles 3(56), 4 (as enacted and as replaced by the Digital Omnibus 2026/1744), 13, 14, 26, 50, 73, 86. (Q4) Commentary on the post. + +**The thesis tested** (from `v4/SOURCE_v4.md`): the post treats disclosure as sufficient ethical design, and so (a) places the burden on the person who must read, (b) keeps the person in a literacy-based experience, (c) does not scaffold the person's growth to algorithmacy. + +**Read-depth marks.** *Full* = whole text read this session. *Part* = the passages named. *Abstract* = abstract only. *Record* = catalogue entry only. + +--- + +## Method notes + +| # | Source | Request | Result | +|---|---|---|---| +| 1 | Local capture | `~/iit-playground/dxd_foundation_local/nadella_2026-10-10_models_as_insider_risks.txt` | Read in full (P1–P21). Word counts by `grep -i` on the post body (header excluded): "user" 0, "people" 0, "skill" 0, "literac" 0, "train" 1 (P2, "training data"), "learn" 2 (P8 "learned", P20 "learnings"), "read" 1 (P15 "readable"), "human" 2, "person" 1, "affected" 1, "disclos" 2 (P20). | +| 2 | Sister arm (DXD) | `wt-dxd-foundation/.../literature/facets/G_nadella_lineage.md` (B1 table, C1–C10, changes of 2026-10-13); `library/nadella2026models.md` (S2-verified); `audit/claims/CLAIMS_01.md` rows C01-05, C01-36 to C01-40; `CLAIMS_04.md` rows C04-38, C04-40, C04-43 to C04-45 | Reused; not redone. | +| 3 | This arm | `facets/Z_gap_search.md` R1–R2 (Art 4, Omnibus); `facets/N2_platform_law.md` (Arts 13, 14, 26, 86, Omnibus dates) | Reused and extended. | +| 4 | EUR-Lex | `eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32024R1689` and `...CELEX:32026R1744`; also `eur-lex.europa.eu/eli/reg/2026/1744/oj/eng` (HTTP 200) | **EUR-Lex served both texts this time** (earlier facets Z and N2 recorded a bot challenge). Saved to `berthon_ethics_local/pdfs/V1/eurlex/`. The extracted text is byte-identical to the Publications Office copies in `pdfs/N2_platform_law/eu/` (`cmp`: identical; sha256 of 32024R1689 text d89a8a4d…, of 32026R1744 text 139b2b57…). | +| 5 | Microsoft (curl) | RAI Standard v2 PDF (msblogs.thesourcemediaassets.com); Transparency Report 2024 and 2025 PDFs (cdn-dynmedia-1.microsoft.com); AI Literacy Starting Guide (Jan 2025, same CDN); Elevate post (blogs.microsoft.com, 9 Jul 2025); Microsoft 365 Copilot Transparency Note (learn.microsoft.com, ms.date 2026-10-05); Amershi et al. 2019 (microsoft.com/research) | All 200; saved to `berthon_ethics_local/pdfs/V1/ms/`, extracted with `pdftotext`. | +| 6 | WebSearch | `Microsoft 2025 Responsible AI Transparency Report PDF`; `Microsoft "AI fluency" OR "AI literacy" skilling commitment Elevate 2025 2026 users`; `blogs.microsoft.com Microsoft Elevate Brad Smith July 2025 ...` | Located the primary URLs. Search summaries not used as evidence. | +| 7 | WebSearch (extended) | `Nadella "Models as Insider Risks" critique OR analysis OR response` | Nine hits; one substantive commentary (Sudjianto, Substack, 10 Oct 2026), fetched and read in full. The rest are news summaries already in DXD facet G (TNW, Business Standard, Forbes blocked) or vendor blogs. | +| 8 | WebSearch | `Nadella insider risks essay "those affected" OR "end users" OR "customers" critique October 2026 ...` | Nothing on the affected-person or end-user angle. | +| 9 | Crossref | `/works/10.1145/3290605.3300233` | Amershi et al., CHI 2019, pp. 1–13, 13 authors: confirmed. | + +**Not done / blocked.** I did not read the whole 2024 (40 pp.) or 2025 (35 pp.) Transparency Reports: I searched each for literacy, fluency, skill, Transparency Note, disclos-, overreliance, end user, Human-AI, and read the pages those hits fell on (2024: pp. 4, 11, 25, 29, 33, 34; 2025: PDF pp. 7, 22, 26, 28, 34). I did not read any other Transparency Note, the Microsoft Trust Center AI Act page, or Microsoft's 2026 report (I did not search for one). Forbes (Linzainescu) remains blocked, as DXD recorded. No login, no paywall bypass. + +--- + +## Q1. What the post proposes, move by move + +DXD facet G already read the post paragraph by paragraph (B1 table) and tested three counts against it (C7); its claims ledger revised two rows from "supported_narrower" to "supported" on the ground that "Disclosure is one strand of the proposal, not its whole" (CLAIMS_01 C01-05, C01-36). I reuse that reading and add the three-way classification the brief asks for. *Full* (local capture; checked against DXD's blog comparison). + +**Who "we/us" is.** The deploying organisation. The post grounds the move in "how to handle powerful actors inside the enterprise" (P8) and applies it to "SI inside the enterprise" (P9); the only explicit holder of authority is "Orgs" (P17). DXD's ruling D5 adopted this reading; the S2 check on `library/nadella2026models.md` confirmed "user", "customer", "employee" and "people" do not occur. My own count agrees ("user" 0, "people" 0). + +**Classification.** (i) = a text, record or disclosure to be read; (ii) = a control held by the deploying organisation; (iii) = builds a person's capacity. + +| P | Move (paraphrase; short fragments only) | Class | Who holds it / who reads it | +|---|---|---|---| +| P1–P3 | Setting: attribution to code paths is lost (P1–P2); responsibility cannot be outsourced to the provider (P3). | — (premise) | The deployer keeps responsibility. | +| P4 | Build systems whose behaviour "we can observe", limits we can test, actions we can contain. | (ii), with an observing element | The deployer. | +| P5 | Separate "the supply of intelligence from the authority over it". | (ii) | Authority = the deployer. | +| P6 | Deterministic system design, "human controls", operating procedures, industry standards. | (ii) | The deployer; "human controls" not described. | +| P7–P8 | Treat the model as an insider risk; identity, least privilege, "log activity", containment. | (ii); logging is a record (i) | Security staff. | +| P9 | Chain-of-thought transparency as "a non-negotiable", but "not sufficient or dependable". | (i) | A reader of the CoT; who is unstated. | +| P10 | Models may test each other, but this risks "nested black boxes". | (ii) (a caution) | — | +| P11–P12 | Controls over access and action "must sit outside the model"; separate model, harness, action space. | (ii) | Enforced by the system, no reader needed. | +| P13 | Umbrella: "principles of observability". | — (heading) | — | +| P14 | Model diversity; no model verifies its own work. | (ii) | Architecture. | +| P15 | Every meaningful action leaves "tamper-proof human readable evidence"; outcomes reproducible without the model's attestation. | (i) | "We" must be able to reproduce; the reader is not named. | +| P16 | Continuous testing, including failures and attacks. | (ii) | Test engineering. | +| P17 | "Orgs" independently determine what a model can access and do. | (ii) | The organisation. | +| P18 | Independent auditability: validation independent of the intelligence validated. | (ii) relying on (i) evidence | An auditor (person or system), unnamed. | +| P19 | Containment; "an authorized person" can pause or shut down a model mid-task. | (ii) | One authorised person, inside the deployer. | +| P20 | Incident disclosure: "timely disclosure to those affected"; share what went wrong industry-wide. | (i) | Those affected (after failure) and the industry. | +| P21 | Trust the model least. | — (aphorism) | — | + +**Count.** Of the substantive moves, three are disclosures or records to be read as their main content (P9, P15, P20); about ten are controls held by the organisation (P4–P8, P10–P12, P14, P16–P19; P18 relies on P15's record). **None is of type (iii).** The post contains no word for training, skill, literacy, learning by a person or education; its two "learn" words refer to what the enterprise has learned (P8) and to "learnings" shared across industry (P20), which is organisational and sectoral learning, not a person's growth. + +**Where the people the systems act on appear.** Once: P20, "those affected", as recipients of "timely disclosure" when systems "fail or are compromised". The same paragraph sends lessons to the industry, not to them. P2 says the models act "on our behalf", i.e. for the deployer; the post never says toward whom. DXD's claims ledger already holds the sentence "the people on whom the model acts enter the framework only once, as recipients of a disclosure after something has failed" as supported (CLAIMS_01 C01-40) and C04-44/45 (the affected person "does not hold the brake"). + +**Who must read the P15 evidence.** The text does not name a reader. The sentence that follows says "We need to be able to reproduce how an outcome was achieved" (P15), so the reader is the deploying organisation, for the purpose of reconstruction. DXD G (B1, "What 'human readable evidence' asks of a person") read it as forensic rather than continuous vigilance; the auditor of P18 and the "authorized person" of P19 are the only candidates the post names. Nothing in the post makes the evidence available to those affected; P20 gives them a disclosure after failure, not access to the record. + +**Bearing.** +- *Against the thesis as stated (point 4 in SOURCE_v4).* The post does not assume disclosure is sufficient. It says the one disclosure it insists on, CoT transparency, is "not sufficient or dependable" (P9), and most of its weight sits on controls that disclose nothing to anyone (P11–P12, P14, P17, P19). This is the DXD finding (C10 items 1–2; C01-05) and it stands. **The thesis is narrowed: the post does far more than disclosure.** +- *For a narrower thesis.* Every safeguard is either a text to be read or a control the organisation holds; nothing builds anyone's capacity (no (iii)); the people acted on appear once, as readers of a post-failure notice (P20). The burden the post creates for a person falls on the organisation's own staff (the P15 reader, the P19 brake-holder), and the post says nothing about how they will be equipped. +- *Narrows (a).* The burden is not placed on the end user, who is absent. It falls on the deployer's staff. "The user" of the three wrongs is two people (as SOURCE_v4 already flags): staff who must read the evidence, and the affected, who are only told. + +--- + +## Q2. Microsoft's own published record on transparency and on users + +### Q2.1 Microsoft Responsible AI Standard v2, General Requirements (June 2022) — `microsoft2022responsible`. *Full* (27 pp.). + +- **Goal T1, "System intelligibility for decision making"** (p. 9): systems "that inform decision making by or about people are designed to support stakeholder needs for intelligibility of system behavior." **T1.1** requires teams to identify "1) stakeholders who will use the outputs of the system to make decisions, and 2) stakeholders who are subject to decisions informed by the system" (p. 9). **T1.2**: design "the system UX, features, reporting functions, and educational materials" so that those stakeholders can understand intended uses, "interpret relevant system behavior effectively", and "remain aware of the possible tendency of over-relying on outputs produced by the system ('automation bias')" (p. 9). **T1.3**: "Define and document the method to be used to evaluate whether each stakeholder who will make decisions or be subject to decisions based on the behavior of the system can interpret the relevant system responses reasonably well" (p. 9). **T1.4–T1.6**: release criteria, pre-release evaluation, and a gap plan if criteria are not met (p. 9). +- **Recommendation T1.2.1** (p. 10): "Follow the Guidelines for Human-AI Interaction when designing the system." T1.3.1: user researchers design evaluations "in appropriately realistic contexts of use" (p. 10). +- **Goal T2, "Communication to stakeholders"** (p. 11): documentation for "stakeholders who make decisions about whether to employ a system" and those "who develop or deploy systems that integrate with this system" (T2.1); "When the system is a platform service made available to external customers or partners, a Transparency Note is required" (T2.2). The Transparency Note's audience is the customer that deploys, not the person acted on. +- **Goal T3, "Disclosure of AI interaction"** (p. 12): inform people "that they are interacting with an AI system"; T3.3 requires a method to evaluate whether each stakeholder "is informed of the type of AI system they are interacting with or exposed to" (p. 12). +- **Goal A5, "Human oversight and control"** (p. 8): A5.2 requires the system elements ("system UX, features, alerting and reporting functions, and educational materials") that let overseers understand intended uses, "how to interpret system behavior", "when and how to override, intervene, or interrupt the system", and automation bias; **A5.4**: "evaluate whether each oversight or control function can be accomplished by stakeholders in realistic conditions of system use" (p. 8). + +*Bearing.* **Narrows the thesis, and sharpens the contrast with the post.** Microsoft's own 2022 standard does more than disclose. For the person *subject to* a decision it requires designed educational materials, a test of whether that person "can interpret the relevant system responses reasonably well", and release criteria tied to the test (T1.1–T1.6). For overseers it requires a test that oversight "can be accomplished ... in realistic conditions" (A5.4). The October 2026 post has neither: its P15 reader and P19 brake-holder are never tested. So **the post is thinner than its own company's standard**. *For (b):* the competence the Standard tests is interpretation of the system's outputs, a reading competence; "educational materials" are the medium. *Not (iii) in the thesis's sense:* the Standard tests comprehension at release; it does not commit to growing a person's competence over time. + +### Q2.2 Guidelines for Human-AI Interaction (Amershi et al., CHI 2019) — `amershi2019guidelines`. *Part* (abstract; Table 1, p. 3). DOI 10.1145/3290605.3300233 (Crossref-checked). + +- Eighteen guidelines (Table 1, p. 3), named by the Standard as the design reference (T1.2.1, A5.3.1). Some are disclosures: G1 "Make clear what the system can do", G2 "Make clear how well the system can do what it can do", G11 "Make clear why the system did what it did", G18 "Notify users about changes". Others give the user controls over the system: G8 "Support efficient dismissal", G9 "Support efficient correction" ("Make it easy to edit, refine, or recover when the AI system is wrong"), G15 "Encourage granular feedback", G16 "Convey the consequences of user actions", G17 "Provide global controls" ("Allow the user to globally customize what the AI system monitors and how it behaves") (all p. 3). +- G13 "Learn from user behavior" is the system learning from the user, not the user learning (p. 3). + +*Bearing.* **Narrows (a) and (b) for Microsoft's design canon, not for the post.** Microsoft's own interaction guidelines put controls in the user's hands (correction, dismissal, feedback, global controls) as well as disclosures. The post uses none of them; its controls are the organisation's (P17). **For (c):** none of the 18 guidelines is about growing the user's competence; the one "learn" guideline points the other way. + +### Q2.3 Responsible AI Transparency Report 2024 — `microsoft2024responsible`. *Part* (pp. 4, 11, 25, 29, 33, 34 of 40). + +- "We've published 33 Transparency Notes since 2019 to provide customers with detailed information about our platform services" (p. 4; again p. 25). Transparency Notes "follow a specific template to provide customers with detailed information about capabilities, limitations, and intended uses" (p. 25). +- For Copilots, "user-friendly notices in product experiences to provide important disclosures": "in-product disclosure to inform users that they are interacting with an AI application and citations to source material to help users verify information", plus "disclaimers about the potential for AI to make errors" (p. 25). +- **"Responsible use of AI applications is a shared responsibility between application developers and application users"** (p. 34). Developers should "design mitigations that enable appropriate use ... to minimize users' risk of overreliance" (p. 34). The research named includes "an Excel add-in prototype that helps users assess AI-generated code" and work on "how users interact with AI transparency features" (p. 34). +- Training is for employees: "99 percent of employees completed the responsible AI module" (p. 4; p. 29: "This journey begins with educating all of our employees"). + +*Bearing.* **For (a):** the company's own report names the user as co-bearer of responsibility (p. 34), and its user-facing measures are disclosures, citations "to help users verify", and disclaimers (p. 25). **For (b):** the means are texts. **Narrows (c) slightly:** a research prototype to help users assess AI output is the nearest thing to in-the-moment scaffolding I found, but it is a prototype, not a product commitment. + +### Q2.4 Responsible AI Transparency Report 2025 — `microsoft2025responsible`. *Part* (PDF pp. 7, 22, 26, 28, 34 of 35). + +- On the AI Act: "The first set of provisions that came into effect in February 2025 were the AI literacy requirements"; Microsoft "has undertaken various initiatives to promote AI literacy ... empowering our employees, customers, and others" (PDF p. 7, printed p. 11). +- "AI literacy efforts" (PDF p. 26, printed p. 49): literacy "helps ensure that individuals can make informed decisions regarding AI systems." Programmes for employees ("role-specific online learning paths"; Trust Code training); public resources ("AI courses that range from basic education on what AI is to technical training"); an "AI Literacy Starting Guide ... for vendors and customers"; an "AI Skills Navigator"; an AARP/OATS programme for adults over 50 funded by the Societal Resilience Fund. Customers "need to assess how literacy requirements in the EU AI Act ... apply to them". +- **A misstatement of the Act** (PDF p. 26): "The EU AI Act requires that deployers and providers of AI systems 'provide all relevant actors in the AI value chain with the insights required to ensure the appropriate compliance and its correct enforcement.'" The quoted words are from recital 20, where the subject is AI literacy itself ("AI literacy should provide all relevant actors ..."), not an operative duty on providers and deployers (recital 20, read in the OJ text). The operative duty is Article 4. Report footnote 74 cites a web edition (artificialintelligenceact.eu), not the OJ. +- Same page: new features "allow both users and admins greater visibility into Copilot-generated web queries"; admins can "search, audit, and eDiscovery" on them (PDF p. 26). +- Research on overreliance: mitigations "such as uncertainty highlighting, cognitive forcing functions, contrastive explanations, and AI critiques" (PDF p. 28, printed p. 53). + +*Bearing.* **Narrows (c):** Microsoft does more than disclose; it runs training programmes and public courses aimed at literacy, and its research names in-the-interaction aids (cognitive forcing functions). **For (c) as framed:** what reaches users is courses and a curated reading list, i.e. training *about* AI outside the experience, and the literacy duty is read as a compliance item for customers. **For (b):** "visibility" into web queries is more to read; the audit version of it goes to admins. + +### Q2.5 Microsoft 365 Copilot Transparency Note (now titled "Application card: Microsoft Copilot (for organizations)"), Microsoft Learn, ms.date 2026-10-05 — `microsoft2026application`. *Full* (page text). + +- Limitations: "User training and adoption: Effective use of Microsoft Copilot requires users to understand its capabilities and limitations. There might be a learning curve, and users need to be trained." +- "Responsible AI is a shared commitment between Microsoft and its customers." +- Under **"End-users should:"** — "always review and verify AI-generated responses"; "users should review the responses generated by Microsoft Copilot and verify that they match their expectations and requirements"; and on overreliance: "we help mitigate this risk by adding disclaimers to our products but users should still make sure to review the accuracy of the answers." +- On grounding: "Users should always take caution and use their best judgment"; responses include "references to the sources for users to verify", and users "are also provided with explicit notice that they're interacting with an AI system and advised to check the source materials". +- Feedback: thumbs up/down; "Customer admins can review feedback". + +*Bearing.* **The strongest primary support for (a) and (b) in this facet.** Microsoft's current user-facing transparency document assigns the residual work to the end user in so many words: review, verify, check the sources, use best judgment; the company's stated mitigation is "adding disclaimers". Competence is to come from users being "trained" (a limitation, not a Microsoft commitment). This is Microsoft, not Nadella's post; the paper must say so. + +### Q2.6 AI Literacy Starting Guide (Office of Responsible AI, January 2025) — `microsoft2025ai`. *Full* (11 pp.). + +- Purpose (p. 1): "AI literacy promotes knowledge so that we can make informed decisions regarding AI systems." "Microsoft's AI literacy materials provide a starting point to help you comply with these requirements" (the AI Act's Article 4, quoted in its 2024 wording in footnote 1). +- Form: a curated list of courses, modules, blog posts and webpages at three levels (fundamentals, intermediate, advanced) for "Using AI" and "Developing AI" (pp. 1–2). After the fundamentals, learners will be able to "Describe and define common terms" and "Understand basic concepts relating to responsible AI" (p. 2). + +*Bearing.* **For (c):** Microsoft's answer to the literacy duty is a reading list and courses, separate from the product; it is training about AI, not support built into the experience. **For (b):** the medium of the literacy offer is text and video coursework. + +### Q2.7 Microsoft Elevate (Brad Smith, 9 July 2025) — `smith2025microsoft`. *Full*. + +- "over the next five years, we will donate on a global scale more than $4 billion in cash and AI and cloud technology to K-12 schools, community and technical colleges, and nonprofits". +- "the Microsoft Elevate Academy will help 20 million people earn an in-demand AI skilling credential ranging from foundational fluency to advanced technical training." +- Framing: "putting people first by equipping them with the skills, knowledge, and tools to thrive with AI." + +*Bearing.* **Against (c) read as "Microsoft does nothing for competence".** Microsoft makes a large public skilling commitment and uses the word "fluency". **For (c) read precisely:** the commitment is credentials and courses delivered through schools, colleges, LinkedIn and GitHub, outside the systems people meet. It is a training programme, not scaffolding inside the mediated experience, and it is not in the post. Its figures are Microsoft's own targets. + +--- + +## Q3. The EU AI Act: where the law puts the duty + +Source: Regulation (EU) 2024/1689, OJ L 12.7.2024, from EUR-Lex (identical to the Publications Office copy), and Regulation (EU) 2026/1744 (Digital Omnibus on AI), OJ L 24.7.2026. `eu2024aiact`, `euomnibus2026`. *Part* (articles named). Z_gap_search R1–R2 and N2 already read Arts 3(56), 4, 13, 14, 26, 86, 113 and the Omnibus; I re-read each from EUR-Lex and add Arts 50 and 73 and the Omnibus Art 50 and Art 113 changes. + +- **Art 3(56)** defines AI literacy as "skills, knowledge and understanding that allow providers, deployers and affected persons ... to make an informed deployment of AI systems, as well as to gain awareness about the opportunities and risks of AI and possible harm it can cause". Affected persons are in the definition. +- **Art 4 as enacted** (applied from 2 February 2025, Art 113): "Providers and deployers of AI systems shall take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, ... and considering the persons or groups of persons on whom the AI systems are to be used." The duty is on organisations; the persons made literate are their staff and agents; the people the systems are used on are only to be "considered". +- **Recital 20**: for affected persons, literacy means "the knowledge necessary to understand how decisions taken with the assistance of AI will have an impact on them"; the Board should "promote AI literacy tools, public awareness"; codes of conduct are "voluntary". No operative article gives affected persons a literacy measure. +- **Art 4 as replaced by the Omnibus** (Art 1(5) of Reg 2026/1744; in force on the third day after publication of 24.7.2026): providers and deployers "shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf ... This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual." Para 2: the Commission and Member States "shall support and facilitate the efforts of providers and deployers"; para 3: the Board adopts recommendations. **Recital 8** gives the reason: "imposing such obligations creates an additional compliance burden, particularly for smaller enterprises, whereas AI literacy should be a strategic priority, regardless of regulatory obligations". +- **Art 13(1)**: high-risk systems must be "sufficiently transparent to enable deployers to interpret a system's output and use it appropriately"; 13(2): instructions "comprehensible to deployers". Transparency runs to the deployer. +- **Art 14(1)**: oversight "by natural persons"; **14(4)(b)** the overseer must "remain aware of the possible tendency of automatically relying or over-relying on the output ... (automation bias)"; **14(4)(e)** "interrupt the system through a 'stop' button". **Art 26(2)**: deployers assign oversight "to natural persons who have the necessary competence, training and authority, as well as the necessary support". Recital 91: "an adequate level of AI literacy, training and authority". The overseer is the deployer's. +- **Art 26(11)**: deployers of Annex III systems that make or assist decisions "shall inform the natural persons that they are subject to the use of the high-risk AI system". Notice only. +- **Art 50(1)**: providers ensure that "the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect". **50(2)**: synthetic outputs "marked in a machine-readable format". **50(3)–(4)**: deployers disclose emotion recognition, biometric categorisation, deep fakes, and AI-generated public-interest text. **50(5)**: the information "shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure" and conform to accessibility requirements. The Omnibus gives legacy generative systems until 2 December 2026 for 50(2) (Art 1, amending Art 111) and rewrites 50(7) on codes of practice. +- **Art 73**: serious incidents are reported by providers "to the market surveillance authorities" (73(1)); in the paragraphs I read, no notice to affected persons is required. +- **Art 85**: anyone may complain to a market surveillance authority. **Art 86(1)**: an affected person subject to a deployer's decision based on an Annex III system that produces legal or similarly significant adverse effects has "the right to obtain from the deployer clear and meaningful explanations of the role of the AI system in the decision-making procedure and the main elements of the decision taken"; residual (86(3)). +- **Dates** (Omnibus, replacing Art 113(c)): Chapter III Sections 1–3 (so Arts 13, 14, 26) apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems. + +*Bearing.* +- **For (a) and (b).** Every duty the Act owes the person on the receiving end is a text: a notice of AI interaction (Art 50(1), (5)), a notice of being subject to a high-risk system (Art 26(11)), an explanation on request after a decision (Art 86). The "reasonably well-informed, observant and circumspect" person of Art 50(1) is the law's model reader, and disclosure is excused when that reader would notice anyway. +- **For (c), strongly.** Art 4 is the one place the Act addresses competence. It puts the duty on organisations, for their own staff and agents; affected persons appear in the definition (3(56)) and recital 20 but not in the operative duty. In July 2026 the Omnibus weakened even that duty from "ensure ... a sufficient level" to "support the development of", with no "specific level", for stated reasons of compliance burden (recital 8). Three months before the post, the law moved away from building competence. +- **Narrows (c).** The Act does do something besides disclosure: it requires competent, trained overseers with "the necessary support" (Art 26(2)), names automation bias (Art 14(4)(b)), and gives the affected a complaint route (Art 85). It treats competence as an organisational duty, not as the reader's own burden. But the competence is the deployer's staff's, not the affected person's. +- **Against reading the post as worse than the law on P20.** The post asks for "timely disclosure to those affected" after failures (P20). The Act's incident regime, in the paragraphs I read, reports serious incidents to authorities, not to the people affected (Art 73). On that point the post asks for more than the law. + +--- + +## Q4. Commentary on the post + +- **Sudjianto, A. (2026, 10 October). "When AI Becomes an Insider, Who Is Left to Check Its Work?" Substack** — `sudjianto2026when`. *Full*. Published the day of the post (datePublished 2026-10-10T17:24 UTC). Agrees with the separation of intelligence from authority. Argues: "Companies are rushing to automate the very work through which people develop the expertise needed to independently verify AI"; "An audit trail tells us what the agent did, not whether it should have done it. And an emergency stop is useful only when something recognizes that intervention is necessary"; "An experienced model validator does not become experienced by reading validation reports"; "There is a considerable difference between using AI to help someone understand a problem and using AI to eliminate the need for that person to understand it." Recommends: engineer verification into the system "rather than relying on human review as the primary safeguard"; "redesign junior roles around developing judgment, not merely supervising AI outputs"; measure "whether AI adoption is improving or weakening institutional competence", e.g. by planted-error exercises. + - *Bearing.* **For (c)** from inside the enterprise frame: the post's oversight depends on competence it does not say how to grow, and reading records does not grow it. **For (b)**: competence comes from doing the work, not reading reports. **Narrows (a)**: Sudjianto, like the post, wants verification engineered into the system, so less of the burden falls on a human reader. His subject is the organisation's staff, not the affected; he does not mention end users or those affected. +- **News and vendor coverage.** TNW, Business Standard, Security Boulevard, Forbes (blocked) are in DXD facet G B1 and its cards (`constantin2026nadella`, `kumari2026nadella`). I found no further critique, and nothing on the affected-person or end-user angle (search rows 7–8). + +--- + +## For, narrows, against + +### For +1. The post contains no move of type (iii); "skill", "literacy", "people" and "user" do not occur, and its only "learn" words are organisational (P8) and sectoral (P20). +2. The people acted on appear once, as recipients of disclosure after failure (P20); they are given neither the record (P15) nor the brake (P19) (DXD C01-40, C04-44). +3. Microsoft's current Transparency Note tells end users to "review and verify", says the company mitigates overreliance "by adding disclaimers" while "users should still make sure to review", and treats user training as a limitation (`microsoft2026application`). +4. Microsoft's 2024 report calls responsible use "a shared responsibility between application developers and application users" (p. 34), and its user measures are disclosures, citations and disclaimers (p. 25). +5. Microsoft's literacy answer is a curated course list and external credential programmes (Starting Guide; Elevate), not support inside the experience. +6. The AI Act's duties to the affected are all notices or explanations (Arts 26(11), 50, 86); its one literacy duty binds organisations for their own staff (Art 4), and the Omnibus cut it to "support the development of" with no guaranteed level (Reg 2026/1744 Art 1(5), recital 8). +7. The only substantive commentary argues that oversight needs competence grown by doing the work, which records do not build (Sudjianto). + +### Narrows +1. **The post does much more than disclosure.** Most of its moves are controls held by the organisation (P4–P8, P10–P12, P14, P16–P19), and it calls CoT transparency "not sufficient or dependable" (P9). The thesis that the post "assumes disclosure is sufficient" does not hold as stated; the defensible version is that its safeguards are texts or organisational controls and none builds a person's capacity (DXD C7, C10; C01-05). +2. The burden in the post falls on the organisation's staff (P15 reader, P19 brake-holder), not on end users, who are absent; "the user" must be split in two. +3. Microsoft's own Standard (2022) goes beyond disclosure: it requires evaluating whether people *subject to* decisions can interpret outputs (T1.1–T1.3) and whether overseers can perform oversight in realistic conditions (A5.4). Its design guidelines give users correction, dismissal, feedback and global controls (Amershi G8, G9, G15–G17). These measures are absent from the post, not from Microsoft. +4. Microsoft runs large training programmes (Elevate; employee literacy paths) and its research names in-the-moment aids (cognitive forcing functions; an Excel add-in prototype to help users assess AI output). The claim must be "no scaffolding inside the experience", not "no investment in competence". +5. The AI Act requires the deployer's overseers to have competence, training, authority and "support" (Art 26(2)) and names automation bias (Art 14(4)(b)); competence is an organisational duty there, not only the reader's burden. + +### Against +1. On incident disclosure, the post asks for more than the AI Act: "timely disclosure to those affected" (P20) against the Act's reporting to authorities (Art 73 as read). +2. Microsoft's Standard tests comprehension by the people subject to decisions (T1.3). That is a designer's duty to check that the reader can read, which cuts against "the burden is simply placed on the person" as a description of Microsoft's own governance (not of the post). + +--- + +## What the paper can say (notes for the drafter) + +1. Nadella's post offers two kinds of safeguard: records and disclosures to be read (CoT transparency, P9; "tamper-proof human readable evidence", P15; incident disclosure, P20) and controls held by the deploying organisation (P5, P11–P12, P14, P17, P19); none of its moves builds a person's capacity (Nadella 2026; DXD `library/nadella2026models.md`). +2. The post does not treat disclosure as sufficient: it calls chain-of-thought transparency "not sufficient or dependable" (P9) and rests mainly on controls outside the model (P11–P12, P17). +3. The people the systems act on appear once, as recipients of "timely disclosure" after failure (P20); the evidence of P15 is for the organisation to reconstruct outcomes, and the brake of P19 belongs to "an authorized person". +4. Microsoft's own Responsible AI Standard (2022) asks more than the post: it requires teams to evaluate whether people subject to a system's decisions "can interpret the relevant system responses reasonably well" (T1.3, p. 9) and whether overseers can perform oversight "in realistic conditions" (A5.4, p. 8). +5. Yet the competence the Standard tests is interpretation of the system's outputs, delivered through "educational materials" and disclosures (T1.2, A5.2), a reading competence. +6. Microsoft's current user-facing documentation tells end users to "review and verify" outputs and states that it mitigates overreliance "by adding disclaimers", while "users should still make sure to review the accuracy" (Microsoft 365 Copilot transparency note, 2026). +7. Microsoft's 2024 Transparency Report calls responsible use "a shared responsibility between application developers and application users" (p. 34). +8. Microsoft's answer to the literacy duty is courses and credentials outside the product: a curated AI Literacy Starting Guide (2025) and a skilling programme promising 20 million credentials "ranging from foundational fluency to advanced technical training" (Smith 2025). +9. The EU AI Act's duties toward affected persons are notices and explanations (Arts 26(11), 50(1), 50(5), 86(1)); its only literacy duty binds providers and deployers for "their staff and other persons ... on their behalf" (Art 4). +10. In July 2026 the Digital Omnibus replaced that duty with one to "support the development of AI literacy", adding that it "does not require ... any specific level of AI literacy of any individual" (Reg 2026/1744, Art 1(5)), citing compliance burden (recital 8). +11. On one point the post exceeds the law: it asks for disclosure to those affected by incidents (P20), where the AI Act's incident regime reports to authorities (Art 73). +12. The first substantive commentary on the post accepts its architecture and argues that its oversight depends on expertise that reading records does not build: "An experienced model validator does not become experienced by reading validation reports" (Sudjianto 2026). + +## Gaps + +- I found no work, in the commentary searched (rows 7–8) or in DXD facet G's searches, that criticises the post for omitting the people the systems act on or for building no one's competence. The paper's charge appears to be new. +- I found no Microsoft commitment, in the documents read (RAI Standard v2; Transparency Reports 2024 and 2025, pages read; the Copilot transparency note; the AI Literacy Starting Guide; the Elevate post), to scaffold an end user's competence inside a product over time. The nearest items are research prototypes and research on cognitive forcing functions (2024 p. 34; 2025 PDF p. 28), not product requirements. +- I found no operative article of Regulation (EU) 2024/1689, as amended by 2026/1744, that requires any measure to build the competence of affected persons; they appear in Art 3(56) and recital 20 only. +- Not read: the rest of the two Transparency Reports; Microsoft's Trust Center AI Act page; any 2026 Transparency Report; the full Amershi paper beyond the abstract and Table 1; Art 73 paragraphs beyond those printed in the first block I read. diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V2_oracy_to_literacy.md b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V2_oracy_to_literacy.md new file mode 100644 index 00000000..39309164 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V2_oracy_to_literacy.md @@ -0,0 +1,233 @@ +# V2 — Oracy to literacy: the ethical failures of the move from oral to written standing, proof and agreement + +*Facet V2 of research round 7 (version 4 of the Berthon ethics course paper). Agent-drafted on 2026-10-14. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** Six cases from the transition from oral to written modes of standing, proof and agreement, read for the v4 thesis: that treating disclosure as sufficient ethical design (a) puts the burden on the person who must read, (b) keeps that person in a literacy-based procedure after the medium has changed, and (c) builds no capacity in them. The cases: (1) medieval England's shift from memory to written record (Clanchy; *quo warranto*; the Warenne sword; forgery; seals); (2) benefit of clergy and the neck verse; (3) the Statute of Frauds 1677; (4) the Requerimiento; (5) the Numbered Treaties of Canada (Treaty 6, Treaty 8); (6) Plato's *Phaedrus* and the scholarly cautions against the "great divide". For each case I record who was bound by writing, whether the text was read or explained to them, whether that was treated as sufficient, who bore the burden, and whether anyone built the hearers' capacity to read. + +**Builds on, does not redo.** Facet A (`A_documents_and_standing.md`) already holds Clanchy 3rd ed. snippets, Harrison 2012 on Barnet, Blackstone on benefit of clergy, the Statute of Frauds s. 1, *Delgamuukw* and *Marshall*, and the Waitangi report. N3 and E hold the great-divide records (Goody and Watt 1963, Halverson 1992, Ong 1982, Collins 1995, Finnegan 2012). V6 holds the gap search on oral-to-written analogies for AI and the Phaedrus-and-AI genre; I cite it and do not repeat it. The sister arm's verified card on *Phaedrus* (`wt-dxd-foundation/.../library/plato1925phaedrus.md`) and the algorithmacy arm's verified Havelock card (`algorithmacy_design_ethics/literature/library/havelock1963.md`) are reused for quotations. + +**Read-depth marks.** *Full* = I read the whole text or the whole operative passage this session. *Part* = I read the pages cited. *Snippet* is a kind of *Part*: archive.org full-text-search highlights of about 20 words from a lending-only scan, with no page shown; I give a page only where an index entry, a running head or a page-cited secondary source fixes it, and I say which. *Abstract* = abstract only. *Record* = catalogue or database record only. + +**Spelling.** Spanish quotations keep the 1853 and 1876 printed spelling. The djvu OCR of Oviedo renders *ç* as *g* ("parésgeme"); I restored *ç* and changed nothing else. English statute and treatise quotations keep their spelling. + +--- + +## (a) Method and search log (2026-10-14) + +| # | Source | Query or request | Result | Notes | +|---|---|---|---|---| +| 1 | Arm record | grep of `literature/facets/*.md`, `cards/*.md`, `MANUSCRIPT_v3.md` for Clanchy, Requerimiento, neck verse, Statute of Frauds, Phaedrus, Treaty 6, great divide, Finnegan | A, E, M3, N3, X, T, V6 | Read A in full; read N3, E §13, V6 entries that touch these cases. | +| 2 | Sister arms | `wt-dxd-foundation/.../library/` (clanchy1993from, plato1925phaedrus, scribner1981psychology); `algorithmacy_design_ethics/literature/library/` (clanchy2013, ong1982, finnegan1988, street1984, goody1977, goody1986, goodywatt1963, havelock1963) | 11 cards | Plato and Havelock cards carry S2 verification; Clanchy, Finnegan 1988, Street 1984, Goody cards are secondary or record. | +| 3 | Local PDFs | `grep -ril` over `berthon_ethics_local/pdfs` and `dxd_foundation_local/pdfs` | Hunnisett 1981, Robertson 1980 (Clanchy reviews, user-supplied JSTOR PDFs), `B/statutes5.txt` | Both reviews read in full. | +| 4 | archive.org advancedsearch | Clanchy; Ong; Gabel; Statutes of the Realm vol. 2; Helps; Las Casas; Oviedo; Morris; Goody; Finnegan; Horton and Finnegan | ids found | Clanchy 1979 and 1993, Gabel 1969 reprint, Goody 1977 and 1987, Horton and Finnegan 1973: lending-only. Statutes vol. 2, Helps 1855, Las Casas 1876, Oviedo 1853, Morris 1880, Ong (community upload): open. | +| 5 | archive.org page search (`fulltext/inside.php`) | Clanchy 1979 | "Item not available" | Lending restriction; not circumvented. | +| 6 | archive.org full-text search API (`be-api.us.archive.org/fts/v1/search`) | Clanchy 1979: about 45 phrase queries (rusty sword, Warenne, quo warranto, chronicler, Guisborough, 1290, time out of mind, forgers, Edward the Confessor, bondsmen, statute of Exeter, serfs …); Gabel: neck-verse, psalm, by rote, upside down, illiterate, taught, reading-test …; Goody 1977 and 1987, Horton and Finnegan 1973: great divide, dichotomy, binary, René Maheu, non-literate | snippets | Highlights only, no page numbers. The API is public and returns search highlights; it does not open the book. | +| 7 | archive.org downloads (open items) | Statutes of the Realm vol. 2 djvu; Las Casas *Historia* t. IV (`historiaindias03casarich`); Oviedo t. III (`historiageneraly03fern`); Helps vol. 1 (`spanishconquest01helpgoog`, 1855 London printing); Morris 1880 (`treatiscanada00morrrich`); Ong PDF; Cross 1917 (JSTOR Early Journal Content) | all read at the pages cited | Saved to `berthon_ethics_local/pdfs/V2/`. | +| 8 | SCC site | `decisions.scc-csc.ca/.../item/1366/index.do` | *R. v. Badger* [1996] 1 S.C.R. 771, full judgment | Read ¶¶ 51–55. | +| 9 | vLex (via WebFetch) | `vlex.co.uk/vid/861195218` | Criminal Law Act 1827, ss. VI–VII | The fetch tool returned the sections as quotations; re-check against a printed Statutes at Large. | +| 10 | legislation.gov.uk | Law Reform (Enforcement of Contracts) Act 1954, original PDF | 2-page scan | OCR with tesseract; s. 1 read. | +| 11 | Crossref | DOIs for Cross 1917, Faudree 2012, Hunnisett 1981, Robertson 1980, Goody and Watt 1963, Stollery 1976; title queries for neck verse, Requerimiento, Seed, Goody 1977/1987, Modes of Thought | records | All DOIs given below resolved and matched. Book titles have review records only. | +| 12 | OpenAlex | Faudree 2012 and 2015 abstracts; `statute of frauds 1677 history`; gap queries `requerimiento terms of service`, `requerimiento algorithm`, `requerimiento informed consent`, `requerimiento privacy notice`, `neck verse literacy test algorithm` | 5,246; 9,331; 1,442; 167; 200 hits | Gap queries surfaced Israeli 2022 (Law and History Review, OA) and nothing that connects these cases to digital disclosure. | +| 13 | Cambridge Core OA PDF | Israeli 2022 | 26 pp. | Read pp. 37–40 and 46–47. | +| 14 | Alberta Law Review (OJS) | Stollery 1976 | 34 pp. | Read pp. 222–223 and 254. | +| 15 | Old Bailey Online | `/about/punishment` | JavaScript shell, no text | Not used. | +| 16 | WebSearch | Rabel on the Statute of Frauds; BC Law Reform Commission report | no text | Rabel not found; the BCLI page has no report text. | + +**Not reached.** Clanchy (any edition) as a book; Gabel as a book; Seed, *Ceremonies of Possession* (1995) ch. 3; Hanke's studies of the Requerimiento; *Dreaver v. The King* (1935) and *R. v. Johnston* (1966) on Treaty 6's medicine chest; Finnegan 1973 and 1988 as books; Street 1984; the 1937 Law Revision Committee report itself. The leads list (section f) says what a library visit should pull. + +--- + +## (b) Findings, case by case + +### Case 1. Memory to written record in England, 1066–1307 + +**Clanchy, *From Memory to Written Record* (1st ed., Harvard UP 1979), `clanchy1979memory`. Part (Snippet).** The 1979 edition is the one Ong, Hunnisett and Robertson paginate. I add it as a key beside the arm's `clanchy2013memory` because the locators below are 1979 pages. + +- *The Warenne story and Clanchy's caution (1979, pp. 21–22; page from the book's index entry "Walter of Guisborough, 21" and from Ong 1982 p. 95, who cites "1979, pp. 21–2").* The story "appears in one version of the chronicle of Walter of Guisborough". The earl "produced in their midst an ancient and rusty sword and said: 'Look at this, my lords, this is my warrant! [F]or my ancestors came with William the Bastard and conquered their lands with the sword, and by the sword I will defend them from anyone intending to seize them.'" (OCR prints "Kor"; I read "For".) Clanchy's caution, in his words: "The story is certainly inaccurate in details, but it has a value comparable with the myths Fitz Neal recorded a century earlier"; "Some historians have found the Warenne story as unacceptable as Edward's judges may have done, because it seems so primitive and ill-founded by literate [standards]" (the last word is cut off); "a rusty sword was not going to help his cause at law"; "The story seems to be a desperate reassertion of the primacy of oral tradition over recorded history and of non-literate forms [of proof]" (cut off); the sword's exhibition was "an archaic method of proof … which had been superseded in Edward I's courts by written evidence and book-learned law"; "The sword was rusty; it was a relic, not a practical weapon." Ong reports the caution as "somewhat questionable because of certain inconsistencies" (Ong p. 95, read this session). +- *The writing rule in quo warranto (snippet, no page).* "the king's attorneys were arguing in many of the quo warranto prosecutions against the magnates that the only sufficient warrant for a privilege was a written one and that in the form of a specific statement in a charter" (the 3rd-ed. snippet in facet A, PP18, continues: "Memory … if unsupported by clear written evidence, was ruled out of court"). "the quo warranto prosecutions threatened to disfranchise nearly all the magnates." +- *The retreat (snippet, no page).* "In 1290 by the quo warranto statute a compromise was made. In subsequent case law the king was understood to have conceded …"; "Although the quo warranto cases were rapidly suspended in the 1290s and the king's government had to concede that tenure 'from time out of mind' was a legitimate claim, the principle had been established for the future"; had the Crown's maxim held ("'time does not run against the king'"), "nearly all owners of property in England would have been disfranchised". Hunnisett (1981, p. 68, Full) adds that "the establishment of 1189 as the legal limit of memory" was "an early recognition of the value of the archival work" of Hubert Walter. +- *Forgery (1979, p. 249 per Ong p. 96; snippets).* "monks, who were the traditional experts in writing, were also the greatest forgers. The more powerful and ancient the house, the more likely it was that its documents would be forged in a professional [manner]"; "In the monastic community archivists and forgers shared a common aim in documenting a house's titles"; "Forgers recreated the past in an acceptable literate form. They are best understood not as occasional deviants on the peripheries [of legal practice]". Of "the 164 documents written in the name of Edward the Confessor, 44 are likely to be spurious and there are doubts about a further 56." **Correction for any draft that follows Ong:** Ong (p. 96, read) says "44 are certainly forged, only 64 certainly authentic"; Clanchy's own words are "likely to be spurious" with doubts about 56 more. Hunnisett (1981, p. 68, Full): Clanchy shows "how the change from memory to written record, and the quo warranto proceedings, made inevitable the large-scale forgery of documents, such forgeries having a greater chance of acceptance than had genuine Anglo-Saxon charters". Robertson (1980, p. 612, Full): "Forgery is described as a product of the gradual transition from oral to written records." +- *Using writing without reading it: seals (1979, pp. 31, 36, 184 per the index entry "statute of Exeter &"; snippets).* "the statute of Exeter of 1285 actually required 'bondsmen' to have seals to authenticate their written evi[dence]"; "even some serfs probably had them"; "in Edward I's reign even serfs were required by statute to have them." Hunnisett (p. 67, Full) summarises Clanchy's account of "the use of seals in lieu of signatures" and of reading as hearing: "in the middle ages 'reading' normally meant 'hearing' and 'writing' meant 'dictating' … it was therefore not essential for kings or for such officials as sheriffs to be literate in the modern sense (p. 219)". Robertson (p. 612): Clanchy treats "the tendency to 'hear' documents when they were used rather than to read them silently". +- *Documents distrusted (1979, p. 230, via Ong p. 94, read).* "'documents did not immediately inspire trust' (Clanchy 1979, p. 230). People had to be persuaded that writing improved the old oral methods sufficiently to warrant all the expense and troublesome techniques it involved." Ong (p. 94): "Witnesses were prima facie more credible than texts because they could be challenged and made to defend their statements, whereas texts could not" (citing Clanchy pp. 235–6). The Sandwich jury of 1127 swore from what "'I have received from my ancestors, and I have seen and heard from my youth'" (Clanchy pp. 232–3, via Ong p. 94). +- *Breadth.* Hunnisett (p. 67): the growth of royal documents "made it ever more necessary for the king's subjects to have at least a nodding acquaintance with letters, not excluding peasant smallholders who by 1300 were making grants by charter in large numbers", and the book argues "its main result was the gradual spread of literacy from a few monasteries to all ranks of society". Hunnisett warns that "the extent and, even more, the depth of lay literacy in 1300 must remain conjectural, and Dr. Clanchy is not completely consistent about this". Robertson (p. 613): Clanchy "may be overly optimistic about the literacy of manorial reeves, whose accounts to their stewards were often prepared and delivered by local clerks." + +*Case summary.* **Bound by writing:** magnates holding franchises (quo warranto); all free landholders as written title displaced remembered title; serfs required to own seals. **Read or explained:** documents were heard, not read (reading meant hearing). **Treated as sufficient:** the Crown's attorneys argued that only a written warrant sufficed; the Crown retreated in 1290 and conceded tenure from time out of mind. **Burden:** on the holder to produce a written warrant. **Capacity built:** none by the Crown; literacy spread "to all ranks" as a by-product (Hunnisett's summary), and lords answered peasant charters by seizing them (facet A, Harrison pp. 21–22). **Who gained:** the writing experts, the monasteries, who were "also the greatest forgers". + +### Case 2. Benefit of clergy and the neck verse + +**Statute 4 Hen. VII c. 13, *Statutes of the Realm* vol. 2 (1816), p. 538, `statute1489clergy`. Full (the chapter).** Running head "A.D.1488-9. 4° Hen.VII." Title: "An Acte to take awaye the benefytt of Clergye from certayne psons." Preamble: "WHERE AS upon trust of privilege of the Churche div[er]s psones lettred hath ben the more bold to cõmitte murdre rape robbery thefte and all othre myschevous dedys, bicause they have ben continuelly admitted to the beneffice of the Clergie as ofte as they did offend". Enacting words: every person "not being within orders, whiche onys hath ben admytted to the benefice of his Clergie, eftsonys arayned of eny suche offence, be not admitted to have the benefice or privilege of his Clergie"; each to be "marked with a M. upon the brawne of the lefte thumbe" for murder and "a T." for other felony, "by the Gaillour openly in the Courte before the Jugge". **Narrows the brief's framing.** The 1489 statute does not create the reading test. It presupposes it ("psones lettred"), limits a lay reader to one use and brands him so the court can see he has used it. **Date:** the volume dates the session 1488–9; Cross (1917, p. 552) prints "in 1487". The brief's "1489" matches the volume's end year. + +**Cross, "The English Criminal Law and Benefit of Clergy during the Eighteenth and Early Nineteenth Century", *AHR* 22(3), 1917, pp. 544–565, `cross1917english`. Part (pp. 551–556 read closely, the rest skimmed); public-domain JSTOR Early Journal Content.** +- p. 552: after a 1350 ordinance the privilege "was apparently interpreted to mean all who could read; at any rate such grew to be the usage of the courts. A verse in the Psalter was commonly selected (usually the 51st Psalm) which came to be known as the 'neck-verse', and if the accused was able to 'read like a clerk' he was handed over to the ordinary, though it was an indictable offense at common law to teach a felon to read that he might claim his clergy." (Cross gives no source for the last clause in the text I read.) +- p. 553: peers were to be treated as clerks convict "'though he cannot read'" (1 Edw. VI c. 12, quoted from Pike); n. 47 quotes Pike: "a Peer who could or who could not read might still have robbed one church or committed one highway robbery with impunity, though an ignorant peasant would have been hanged." +- p. 554: the petition of Osmund Gibbs, yeoman, 1640 (Hist. MSS. Comm., Report IV): falsely charged by a justice who coveted his copyhold, he "was found guilty and put to read for his life. Petitioner desired to read the Psalm of Mercy, but Farwell so incensed the judge against him, that there was not only a clear bar made to prevent promptings, but the judge turned him unto one of the hardest verses to read, which by God's grace he was enabled to do, and so escaped hanging, but was burnt in the hand." +- p. 555: Jeaffreson's Middlesex figures: "In the last four years of the reign of Edward VI., 8.54 per cent. out of 117 convicted of capital offenses could read and pleaded their clergy: in the reign of James I. there was an amazing increase to 38.83 per cent. out of 1725 cases." Cross then: "At length Parliament came to realize the rank injustice of discriminating between lettered and unlettered criminals, so in the reign of Anne the reading test was taken away", quoting the Act: a person "'shall not be required to read, but without any reading shall be allowed, taken and reported to be and punished as a clerk convict.'" n. 55: the Act is "6 Anne, c. 9, s. 4" in *Statutes of the Realm* "though the statute is usually cited as 5 Anne, c. 6". + +**Gabel, *Benefit of Clergy in England in the Later Middle Ages* (Smith College Studies in History 14, 1928–29; reprint Octagon 1969), `gabel1928benefit`. Part (Snippet; index "'Neck-verse,' 72").** +- "Although any passage might be chosen at random, the first verse of the fifty-first Psalm came to be known as the 'neck-verse,' no doubt because of its common use for the purpose." +- A prisoner "could neither read nor 'sillabicare,' yet seemed to know certain passages by rote. At this point the secular judge gave him the book upside down, but the prisoner 'read' as before, in nowise disturbed by the altered circumstance. When the Ordinary refused to claim him, John was given another chance to plead as a layman, but stood fast by his claim of clergy." +- A prisoner "had improved his last days or hours by a diligent study of the 'neck-verse'"; another "may have learned to 'read' while awaiting the day of execution"; one "was discovered to have been a layman and illiterate at the time of his arrest, but had been taught what he knew by two boys from Appleby, who had been admitted to the gaol through [—]" (cut off). +- "Towards the close of the century the reading-test was commonly administered by the court". + +**Blackstone, *Commentaries* IV ch. 28, `blackstonendcommentaries`.** Reused from facet A (Part): "those men, who could not read, if under the degree of peerage, were hanged"; the reading condition was removed because learning was "no extenuations of guilt" and death was "too severe for the ignorant also". + +**Criminal Law Act 1827 (7 & 8 Geo. IV c. 28), `statute1827criminal`. Part (vLex transcription via a fetch tool).** "VI. And be it enacted, That Benefit of Clergy, with respect to Persons convicted of Felony, shall be abolished"; s. VII: no person convicted of felony "shall suffer Death, unless it be for some Felony which was excluded from the Benefit of Clergy before or on the First Day of the present Session of Parliament, or which hath been or shall be made punishable with Death by some Statute". Dated "21st June 1827". Re-check against a printed Statutes at Large before quoting. + +*Case summary.* **Bound by the text:** every lay commoner convicted of a clergyable felony; peers and (later) women exempt from reading (Cross pp. 553–554). **Read or explained:** the reverse of disclosure — the court handed the accused a text and made his life depend on reading it. **Treated as sufficient:** a performance of reading, even by rote, was treated as proof of clergy until judges tested it (book upside down; a bar against prompting; "the hardest verses"). **Burden:** wholly on the accused, and the judge could raise it at will (the Gibbs petition). **Capacity built:** none by the state; Cross reports that teaching a felon to read for this purpose was itself indictable, and Gabel records prisoners taught in gaol by outsiders. **Ended by:** Parliament removing the reading condition under Anne on a ground Cross calls "the rank injustice of discriminating between lettered and unlettered", then abolishing the privilege in 1827. + +### Case 3. The Statute of Frauds 1677 + +**An Act for prevention of Frauds and Perjuryes, 29 Car. II c. 3, *Statutes of the Realm* vol. 5, pp. 839–841, `statute1677frauds`. Full (ss. 1–5, 17, 19; local `B/statutes5.txt`).** Facet A read s. 1. Added here: +- Preamble (p. 839): "For prevention of many fraudulent Practices which are commonly endeavoured to be upheld by Perjury and Subornation of Perjury". +- s. 4 (p. 840): "noe Action shall be brought whereby to charge any Executor or Administrator upon any speciall promise to answere damages out of his owne Estate or whereby to charge the Defendant upon any speciall promise to answere for the debt default or miscarriages of another person or to charge any person upon any agreement made upon consideration of Marriage or upon any Contract or Sale of Lands … or upon any Agreement that is not to be performed within the space of one yeare from the makeing thereof unlesse the Agreement … or some Memorandum or Note thereof shall be in Writeing and signed by the partie to be charged therewith or some other person thereunto by him lawfully authorized." +- s. 17 (p. 841): no contract for the sale of goods "for the price of ten pounds Sterling or upwards shall be allowed to be good except the Buyer shall accept part of the Goods soe sold and actually receive the same or give some thing in earnest to bind the bargaine or in part of payment, or that some Note or Memorandum in writeing … be made and signed". +- s. 19 (p. 841): "for prevention of fraudulent Practices in setting up Nuncupative Wills", no oral will is good above "thirty pounds" unless "proved by the Oathes of three Wittnesses (at the least)". +- **Narrows.** The statute is not a reading test. An agent may sign; s. 17 keeps oral routes (part delivery, earnest money); s. 19 keeps oral wills under £30 and above it with three witnesses; s. 3 exempts copyhold and customary interests (facet A). + +**Stollery, "Statute of Frauds", *Alberta Law Review* 14 (1976), from p. 222, `stollery1976statute`. Part (pp. 222–223, 254).** +- The stated ground, explained (p. 222): "In 1677, parties to an action, their husbands or wives, and persons with an interest in the result of the action could not be witnesses"; quoting Lord Bowen, "the merchant whose name was forged to a bill of exchange had to sit by, silent and unheard". Statutes of 1844–1854 let litigants testify, "removing this rationale". Pp. 222–223: jury verdicts were "practically unappealable"; "Subordination and perjury evidently were rife." +- Who bore the cost (p. 254, quoting the Law Revision Committee's Sixth Interim Report, Cmd. 5449, 1937, at 7): "'promotes more frauds than it prevents.' True, it shuts out perjury; but it also and more frequently shuts out the truth. It strikes impartially at the perjurer and the honest man who has omitted a precaution, sealing the lips of both." Fitzjames Stephen, as quoted there: "'in the vast majority of cases its operation is simply to enable a man to break a promise with impunity, because he did not write it down with sufficient formality.'" And: "A and B contract: A has signed a sufficient note or memorandum, but B has not. In these circumstances, B can enforce the contract against A, but A cannot enforce it against B." + +**Law Reform (Enforcement of Contracts) Act 1954, 2 & 3 Eliz. 2 c. 34, `statute1954lawreform`. Full (2 pp., OCR).** s. 1 repeals, within s. 4 of the 1677 Act, the executor clause, the marriage clause and the one-year clause "in relation to any promise or agreement, whether made before or after the commencement of this Act"; s. 2 repeals s. 4 of the Sale of Goods Act 1893. The Act states no ground. + +*Case summary.* **Bound:** parties to the listed classes of oral promise. **Read or explained:** not applicable; the statute requires a signed note, not a reading. **Treated as sufficient:** the signed writing; its absence defeats the promise. **Burden:** on the person who relied on a spoken promise and did not get the other side to sign (the "honest man who has omitted a precaution"). **Capacity built:** none; the law presumes parties can get a writing made. **Stated ground:** perjury, at a time when parties could not testify. **Who bore the cost:** I found no source that names a social class (the poor, the unlettered) as the losers; the 1937 Committee names the honest promisee. Most of s. 4 was repealed in 1954. + +### Case 4. The Requerimiento + +**Las Casas, *Historia de las Indias*, Book III, chs. 57–59 and 67, in t. IV (ed. Marqués de la Fuensanta del Valle and José Sancho Rayón, Madrid: Imprenta de Miguel Ginesta, 1876), pp. 154–167, 198–199, `lascasas1876historia`. Full (the four chapters).** +- *The text* (ch. 57, pp. 154–156), "la forma y orden que Pedrárias habia de tener en requerir á los indios … el cual se envió después á todas las Indias" (p. 154). It opens "De parte del rey D. Fernando, y de la Reina doña Joana, su hija … nos, sus criados, os notificamos y hacemos saber como mejor podemos" (p. 154). +- *Disclosure by reference to writings the hearer may inspect* (p. 155): the papal donation is "según se contiene en ciertas escripturas que sobre ello pasaron, según dicho es, que podéis ver si quisiéredes". +- *The demand to understand* (pp. 155–156): "vos rogamos y requerimos que entendáis bien ésto que os decimos, y toméis para entenderlo y deliberar sobre ello el tiempo que fuere justo". +- *The burden clause* (p. 156): "y protestamos que las muertes y daños que dello se recrecieren sea á vuestra culpa y no de Sus Altezas, ni nuestra, ni destos caballeros que con nosotros vienen: y de como lo decimos y requerimos pedimos al presente escribano que nos lo dé por testimonio signado, y á los presentes rogamos que dello nos sean testigos". +- *Authorship* (pp. 156–157): "Este requerimiento ordenó el venerable doctor Palacios Rubios, bien mi amigo". +- *Condemnation* (ch. 58, pp. 158–163). He argues first on the supposition "ya que los indios entendieran nuestra lengua" (p. 158) that the content was unprovable to them; then (p. 162) calls it "injusto, impío, escandaloso, irracional y absurdo", and "cosa es de reir, ó de llorar por mejor decir"; then (pp. 162–163): "¿qué podrá alguno decir en excusa de los que formaron aquel requerimiento y de los que á ejecutallo iban, haciéndolo á quien ni palabra del entendían, más que si fuera en latin referido ó en algarabía?; y ya saben los que estudiaron derechos, qué valor ó momento tiene el mando ó precepto, ó requerimiento, que se hace á gente que la lengua en que se dice no entiende". +- *Its falsity on the record* (p. 162): the text's claim that other islands submitted on being informed is false, "porque no es verdad que les notificasen é informasen de cosa dello á ninguna isla". +- *In use* (ch. 59, p. 167): "creer que entendieron ellos cosa del requerimiento es falsísimo, porque no sabían más de nuestra lengua que de la latina". +- *Read to the trees* (ch. 67, pp. 198–199): raiding parties stopped a league or half a league from a village, "y entre sí leian el requerimiento á los árboles diciendo: «Caciques é indios de tal pueblo, hacemos saber …»", and the captain asked the notary for "testimonio autorizado … de como se habia requerido á los Caciques é indios de aquel pueblo … pero que no habian querido venir á dar la obediencia"; then at dawn they fired the houses. + +**Oviedo, *Historia general y natural de las Indias*, Lib. XXIX cap. VII, in t. III (Madrid: Real Academia de la Historia, 1853), pp. 27–32, `oviedo1853historia`. Full (the chapter's account of the Requerimiento).** Oviedo was the expedition's *veedor* and carried the text himself. +- pp. 27–28: Pedrarias "mandó … que yo llevasse el requirimiento in scriptis que se avia de hacer á los indios, é me lo dió de su mano, como si yo entendiera á los indios, para se lo leer, ó tuviéramos allí quien se lo diera á entender, queriéndolo ellos oyr: pues mostrarles el papel, en que estaba escripto, poco hacia al caso". +- p. 29: the friars "que el requirimiento diçe … se quedaron en los navios hasta ver cómo subçedian las cosas"; the reading was on 14 June 1514. +- p. 31: to Pedrarias, before the captains: "Señor: parésçeme que estos indios no quieren escuchar la teología deste requirimiento, ni vos tenés quien se la dé á entender: mande vuestra merçed guardalle, hasta que tengamos algún indio destos en una jaula, para que despaçio lo aprenda é el señor obispo se lo dé á entender." Pedrarias "lo tomó con mucha risa". +- p. 31: "Quisiera yo que aquel requirimiento se les hiçiera entender primero; pero como cosa excusada ó por demás, no se tractó dello". +- pp. 31–32: in 1516 Oviedo asked Palacios Rubios "si quedaba satisfecha la conçiençia de los chripstianos con aquel requirimiento; é díxome que sí, si se hiçiesse como el requirimiento lo diçe. Mas parésçeme que se reía muchas veçes" — and Oviedo could laugh more at him "si pensaba que lo que diçe aquel requirimiento lo avían de entender los indios, sin discurso de años é tiempo." +- p. 32: on the promised "tiempo que fuere justo": "Adelante se dirá el tiempo que los capitanes les daban, atando los indios después de salteados, y en tanto leyéndoles toda aquella capitulaçion del requirimiento." + +**Helps, *The Spanish Conquest in America*, vol. 1 (London: J. W. Parker, 1855), pp. 379–383, 387, `helps1855spanish`. Part.** English translation, used for the English of the burden clause (pp. 382): "we protest that the deaths and losses which shall accrue from this are your fault, and not that of their Highnesses, or ours, nor of these cavaliers who come with us. And that we have said this to you and made this Requisition, we request the notary here present to give us his testimony in writing". The writings clause (p. 380): "as is contained in certain writings which passed upon the subject as aforesaid, which you can see if you wish." The understanding clause (p. 381): "that you take the time that shall be necessary to understand and deliberate upon it". Helps's own comment (p. 383): "the very terms 'Church,' 'privileges,' 'vassalage,' 'exemptions,' are such as require a knowledge of Christianity and of the peculiar civilization of Europe for any one to understand … we may fancy what ideas the reading of the document, even when it was read, conveyed to a number of Indians sitting in a circle". Helps translates Oviedo (p. 387): the Governor "gave it to me from his own hand, as if I understood the Indians, or as if we should find any one there who would make them understand it, even if they were willing to listen." Note that Helps renders "el tiempo que fuere justo" as "the time that shall be necessary"; "just" is closer. + +**Israeli, "The *Requerimiento* in the Old World", *Law and History Review* 40(1), 2022, pp. 37–62, doi:10.1017/S0738248021000602, `israeli2022requerimiento`. Part (abstract; pp. 39–40, 46–47).** The procedure was ordinary Castilian law before it was colonial: "a demand or warning to an addressee and its official recording" (p. 39). Its effect (p. 46): "Once a requerimiento had been conducted, the addressee could not pretend to be ignorant of the demand, and could be considered as having been given a fair opportunity to make amends." The record mattered as much as the speech: "the lack of a testimonio could render the entire procedure meaningless" (p. 46). In Castile it had "an important dialogical component": the addressee "was requested to declare whether or not they were willing to fulfil it", and "the notary was obliged to document the addressee's response" (p. 47). Las Casas and others argued that "the conditions needed for the requerimiento to function as a performative speech act in fact never existed in the New World, both because in many cases the indios could not understand the text recited to them and because they did not receive the proper opportunity to respond" (p. 47, summarising Faudree). Seed's hypothesis of an Islamic origin is reported at p. 39. + +**Faudree, "How to Say Things with Wars", *J. Linguistic Anthropology* 22(3), 2012, pp. 182–200, doi:10.1111/j.1548-1395.2012.01152.x, `faudree2012how`. Abstract.** The text "presupposes and simultaneously instantiates social hierarchies by establishing participant roles"; "Competing interpretations of those participant roles are made possible by the text's strategic indeterminacy and the temporal dynamics of its circulation." + +*Case summary.* **Bound:** the Indigenous peoples addressed, who were made liable for the war. **Read or explained:** read aloud in Spanish, sometimes not even within earshot (Las Casas ch. 67); no interpreter required (Oviedo pp. 27–28). **Treated as sufficient:** yes — the notary's sworn record that it had been read was the legal product (Las Casas p. 156, p. 199; Israeli p. 46). **Burden:** placed in words on the hearers: deaths and losses "á vuestra culpa". **Capacity built:** none; the text promised "el tiempo que fuere justo" to understand, and Oviedo reports the time given was the time it took to read it to captives already tied (p. 32). Oviedo's proposal to teach it to an Indian "en una jaula" is the only capacity-building in the record, and it is a jest. **Contemporaries who named it a wrong:** Las Casas (pp. 158–163, 167, 198–199) and Oviedo (pp. 27–32). The legal principle Las Casas invokes — a command in a language the hearer does not understand has no force (p. 163) — is a standard of the time, not a present-day one. + +### Case 5. The Numbered Treaties of Canada + +**Morris, *The Treaties of Canada with the Indians of Manitoba and the North-West Territories* (Toronto: Willing & Williamson, 1880), `morris1880treaties`. Part (pp. 185–186, 217–219, 351–357).** Morris was the Treaty 6 commissioner; the book prints his reports, the negotiation record and the treaty texts. +- *Read and explained* (p. 357): Treaty 6 was "Signed by the Chiefs within named in the presence of the following witnesses, the same having been first read and explained by Peter Erasmus, Peter Ballendine and the Rev. John McKay". At the Fort Carlton conference (p. 185) "the interpreter, Peter Erasmus, would read what changes they desired in the terms of our offer". +- *What the Cree asked for* (p. 185): "an ox and a cow each family … provisions for the poor, unfortunate, blind and lame; to be provided with missionaries and school teachers … a free supply of medicines". Morris's report on the same page: "They were anxious to learn to support themselves by agriculture, but felt too ignorant to do so". +- *The oral promise against the written clause.* Morris's spoken reply (p. 218): "A medicine chest will be kept at the house of each Indian agent, in case of sickness amongst you." The written text (p. 355): "That a medicine chest shall be kept at the house of each Indian Agent for the use and benefit of the Indians, at the discretion of such Agent". The written text adds the agent's discretion; the spoken promise ties it to "sickness amongst you". +- *Schools* (p. 353): "Her Majesty agrees to maintain schools for instruction in such reserves hereby made, as to her Government of the Dominion of Canada may seem advisable, whenever the Indians of the reserve shall desire it". The Crown wrote capacity-building into the treaty; the arm's TRC card (`trc2015honouring`) records what Canada's schooling became. + +**R. v. Badger, [1996] 1 S.C.R. 771 (Treaty 8, 1899), `badger1996`. Full (¶¶ 51–55 read; the rest skimmed).** Cory J. for the majority, ¶ 52: "The treaties, as written documents, recorded an agreement that had already been reached orally and they did not always record the full extent of the oral agreement" (citing Morris 1880 pp. 338–42). "The treaties were drafted in English by representatives of the Canadian government who, it should be assumed, were familiar with common law doctrines. Yet, the treaties were not translated in written form into the languages (here Cree and Dene) of the various Indian nations who were signatories. Even if they had been, it is unlikely that the Indians, who had a history of communicating only orally, would have understood them any differently." Hence the words "must be interpreted in the sense that they would naturally have been understood by the Indians at the time of the signing". ¶ 55: "The Indian people made their agreements orally and recorded their history orally. Thus, the verbal promises made on behalf of the federal government at the times the treaties were concluded are of great significance in their interpretation." + +*Case summary.* **Bound:** the Cree, Dene and other signatory nations. **Read or explained:** yes, by named interpreters (Morris p. 357). **Treated as sufficient:** by the Crown at the time; not by the Supreme Court in 1996, which holds the written text an incomplete record of an oral agreement (¶ 52) and gives the oral promises interpretive weight (¶ 55). **Burden:** on the signatories, until the courts shifted interpretation toward their understanding. **Capacity built:** the Cree asked for school teachers (p. 185) and the treaty promised schools "as … may seem advisable" to the government (p. 353); what followed is outside this facet. **Caution for the paper:** *Badger* ¶ 52's clause "who had a history of communicating only orally, would [not] have understood them any differently" is a court's own great-divide assumption; cite the holding, not that clause. + +### Case 6. *Phaedrus* 274c–275b and the cautions against the "great divide" + +**Plato, *Phaedrus* (Fowler trans.), sister-arm card `plato1925phaedrus` (verified S2, `wt-dxd-foundation/.../literature/library/plato1925phaedrus.md`).** Quotations taken from the verified card: Thamus says letters "will produce forgetfulness in the minds of those who learn to use it, because they will not practice their memory" (275a); "you offer your pupils the appearance of wisdom, not true wisdom, for they will read many things without instruction and will therefore seem to know many things, when they are for the most part ignorant and hard to get along with, since they are not wise, but only appear wise" (275a–b). Socrates continues: a written word "knows not to whom to speak or not to speak" and "always needs its father to help it" (275e). The card records that the 1925/vol. 9 citation is not supported by the Perseus copy (header: Fowler 1914, vol. 1) and that Phaedrus doubts the myth (275b). Ong (p. 78, read this session) lists Plato's objections, including that "a written text is basically unresponsive". **Use:** two phrases carry the v4 thesis without any cognitive claim — "read many things without instruction" (no scaffolding) and a text that cannot answer its reader (disclosure without dialogue). + +**Ong, *Orality and Literacy* (1982), existing card `ong1982orality` (V6; pp. 2–3, 132–134).** I read further pages in the archive.org PDF (`pdfs/V2/ong1982.pdf`): p. 3, the book "treats rather the differences in 'mentality' between oral and writing cultures" and treats social, economic, political and religious structures as "only indirect concerns"; p. 56, "Nor must we imagine that orally based thought is 'prelogical' or 'illogical' in any simplistic sense", yet "they cannot organize elaborate concatenations of causes in the analytic kind of linear sequences which can only be set up with the help of texts"; p. 55, "Writing has to be personally interiorized to affect thinking processes". **These are the cognitive claims v4 must not adopt.** Ong pp. 94–96 is the Clanchy summary used in Case 1; p. 133, secondary orality is "essentially a more deliberate and self-conscious orality, based permanently on the use of writing and print" (also on the V6 card). + +**Finnegan, "Literacy versus Non-literacy: The Great Divide?", in Horton and Finnegan (eds.), *Modes of Thought* (London: Faber, 1973), from p. 112, `finnegan1973literacy`. Part (Snippet).** She opens on UNESCO's director-general: "of UNESCO, René Maheu, mankind can be divided into two main groups, the difference between them being apparently directly attri[butable]" (cut off; the note cites the World Congress of Ministers of Education on the Eradication of Illiteracy, Teheran 1965, *Speeches and Messages*, UNESCO 1966, p. 29). She asks: "Is it true that the label 'non-literate' necessarily implies 'without literature'?" and whether "individuals in non-literate (or largely non-literate) cultures are ipso facto less creative, thoughtful, self-aware or individu[ally …]". **Use:** the great-divide claim was a literacy-campaign premise as well as an academic one; Finnegan's target includes UNESCO's own framing. + +**Goody's later qualifications. `goody1977domestication` (Part, Snippet; index "Great Divide, the, 8, 151, 164") and `goody1987interface` (Part, Snippet).** 1977: "To argue this way is not to subscribe to a 'great divide' theory"; "This is not a great-divide theory. It sees some changes as more important than others"; he grants that his work on "the consequences of literacy can be considered a 'great-divide theory', since we treat 'literacy' as a variable". 1987: "the Greek achievement has been interpreted as a Great Divide theory (Finnegan 1973)"; "On the contrary, we were attempting to avoid the binary divisions current in much scholarship"; "we should not think of a simple binary shift between oral and literate, but a whole sequence [of changes]"; literacy "is never a question of crossing a single frontier, a simple binary shift". **Use:** even the divide's chief proponent disowned the binary; v4 can cite Goody for the institutional consequences of writing (the file, the law — `goody1986logic`, Record) without the cognitive thesis. + +**Street 2003, card `street2003whats` (verified).** The autonomous model holds that literacy in itself has effects "regardless of the social and economic conditions" (p. 77, per the card); the ideological model treats literacy as a social practice rooted in power. v4's three wrongs are claims about institutions (who bears the burden, what procedure is kept, who builds capacity) and so sit inside Street's ideological model. + +**Havelock, card `havelock1963` (algorithmacy arm, verified S2).** From the card, *Preface to Plato* p. 39: public inscription is "craft literacy, in which the public inscription is composed as a source of referral for officials and as a check upon arbitrary interpretations"; "The clue to the whole problem lies not in the use of written characters and writing materials … but upon the supply of readers, and this depended on a universalisation of letters. The reading trauma, to use a modern term, had to be imposed at the primary level of schooling". **Use:** a record made for officials, published to a public that cannot read it, is Havelock's "craft literacy"; his remedy is a supply of readers, built by schooling. The card also carries Havelock's cognitive claim ("an oral state of mind", p. 41), which v4 must leave aside. + +**Already held (Record or Abstract):** Goody and Watt 1963 (`goody1963consequences`, the divide's founding text, opening paragraph only), Halverson 1992 (`halverson1992goody`, Record), Collins 1995 (`collins1995literacy`, Abstract: universalist claims "cannot withstand criticism"), Scribner and Cole (sister-arm card `scribner1981psychology`: literacy's effects are practice-specific; general cognitive gains follow schooling; 1978 article pp. 453, 457). + +--- + +## (c) For, narrows, against + +### For the thesis + +1. **The Requerimiento is the paradigm of disclosure treated as sufficient, with the burden on the hearer.** It required the hearers to understand ("que entendáis bien ésto que os decimos", pp. 155–156), referred them to writings they could inspect "si quisiéredes" (p. 155), declared deaths and losses "á vuestra culpa" (p. 156), and made the notary's record of the reading the legal product (p. 156; p. 199; Israeli p. 46). Two contemporaries named the wrong: Las Casas, on the force of a command given in a language the hearer does not understand (pp. 162–163), and Oviedo, who carried the paper and said showing it to them "poco hacia al caso" (pp. 27–28). Wrongs (a) and (c) both hold. +2. **Wrong (b) has a documented mechanism.** Israeli (pp. 46–47) shows the requerimiento was an ordinary Castilian procedure built for addressees who understood the language, could answer, and whose answers the notary recorded. Transplanted to people who could do none of these things, the procedure kept its form and lost its conditions. That is wrong (b): the literate procedure stays when the situation that made it fair has gone. +3. **The neck verse shows a literacy test deciding standing, with the burden adjustable by the tester and capacity-building forbidden.** The judge could bar prompting and choose "the hardest verses" (Cross p. 554); a prisoner who knew passages by rote was handed the book upside down (Gabel); teaching a felon to read for the purpose was, on Cross's account, indictable (p. 552). Peers were excused from reading (Cross p. 553). Parliament ended the reading condition on the ground Cross gives as "the rank injustice of discriminating between lettered and unlettered" (p. 555). +4. **Written title made writing experts powerful and invited forgery.** Monks, "the traditional experts in writing, were also the greatest forgers" (Clanchy 1979, snippet); 44 of 164 Confessor charters likely spurious (Clanchy, snippet; Ong p. 96 overstates as "certainly"). Hunnisett (p. 68): the shift and quo warranto "made inevitable the large-scale forgery of documents". The party holding the writing skill gained from a rule that put the burden on the other side. +5. **The written treaty diverged from the oral promise, and a court said so.** Treaty 6's medicine chest is "in case of sickness amongst you" when spoken (Morris p. 218) and "at the discretion of such Agent" when written (p. 355). *Badger* ¶ 52: the written treaties "did not always record the full extent of the oral agreement" and were not translated into Cree or Dene. +6. **Thamus states wrong (c) in a phrase:** readers "will read many things without instruction and will therefore seem to know many things" (275a–b, verified card); and the written word cannot answer its reader (275e). + +### Narrows the thesis + +1. **The reading in the Requerimiento was a ritual for the record, not a failed attempt to inform.** Its legal function was the notary's *testimonio* (Israeli p. 46; Las Casas p. 199). This sharpens rather than weakens the parallel, but it changes the charge: the wrong was not that a disclosure was badly designed; it was that the disclosure's purpose was to discharge the discloser. v4 should say that, and should check whether Nadella's "tamper-proof human readable evidence" (P15) is evidence for the organisation's own record in the same sense. +2. **4 Hen. VII c. 13 did not create the reading test** (Statutes vol. 2 p. 538). It limited lay readers to one use and branded them. The test came from court practice after 1350 (Cross p. 552; Gabel: "commonly administered by the court" by the century's close). Cite the statute for branding, Cross and Gabel for the test. Benefit of clergy was a mercy that literacy unlocked, not a penalty on illiteracy (facet A); the injustice was the unequal mercy. +3. **The Statute of Frauds is not a literacy test.** Agents may sign; part delivery or earnest money satisfies s. 17; oral wills survive under £30 or with three witnesses (pp. 840–841). Its ground was evidential in a system where parties could not testify (Stollery p. 222). Its cost fell on "the honest man who has omitted a precaution" (1937 Committee via Stollery p. 254), not on a named class of the unlettered. I found no source that names the poor or the illiterate as its losers. Use it, if at all, for the narrower point that a writing rule allocates the cost of a missing formality to the party who trusted speech. +4. **Medieval England's losers were often the powerful, and the Crown retreated.** The quo warranto writing rule "threatened to disfranchise nearly all the magnates"; in 1290 "a compromise was made" and tenure "from time out of mind" was conceded (Clanchy, snippets). Clanchy says the Warenne story is "certainly inaccurate in details". Serfs used charters and were required to hold seals (Clanchy; Hunnisett p. 67). The case shows a writing rule pushed and partly withdrawn, not the unlettered dispossessed. +5. **The Treaty 6 record shows interpreters reading and explaining, and the Cree asking for schools.** Disclosure was oral and mediated (Morris pp. 185, 357), and capacity-building (schools) was in the text (p. 353). The failure is the divergence of the written clause and the later reading of it, not an absence of explanation. +6. **Ong's account is a cognitive one** (p. 3, "differences in 'mentality'"; p. 56). v4 may cite Ong for secondary orality and the Clanchy summary, not for what oral peoples could think. + +### Against the thesis + +1. **Literacy also widened standing.** Benefit of clergy saved readers from the gallows and then, once the test went, saved everyone convicted of a first clergyable felony (Cross p. 555; Blackstone in facet A). Hunnisett (p. 67) reads Clanchy's main result as "the gradual spread of literacy from a few monasteries to all ranks of society". A literacy transition can enlarge the class that has standing; the paper cannot present it as only a story of exclusion. +2. **Contemporaries of the transition did not always demand capacity.** The 1489 statute, the Statute of Frauds and the quo warranto rule all assume the burden can be met by hiring a clerk, an agent or a notary (agents sign in 1677, clerks prepare reeves' accounts per Robertson p. 613). That is an intermediary model, not a scaffolding model, and it was how literate societies actually spread written proof. The paper must argue why intermediation is not enough for algorithmacy, not assume it. +3. **"Read and explained" was in places done.** The Treaty 6 record names three interpreters (Morris p. 357), and the Cree negotiated changes read back to them (p. 185). Where disclosure was oral, mediated and negotiated, the record does not support a charge that the hearers were left to read alone. + +--- + +## (d) Per-case table + +| Case | Bound by writing | Read or explained to them? | Treated as sufficient? | Who bore the burden | Capacity built? | Contemporary objector | +|---|---|---|---|---|---|---| +| Quo warranto, c. 1279–1290s | Franchise-holding magnates | Charters were heard, not read (Hunnisett p. 67) | Crown argued only a written warrant sufficed; conceded 1290 | Holder, to produce writing | No | The Warenne story (doubtful, Clanchy pp. 21–22) | +| Written title, seals | Free tenants; serfs required to hold seals (statute of Exeter 1285) | Documents heard | Writing gained trust slowly ("did not immediately inspire trust", p. 230) | Holder; forgers gained | By-product spread (Hunnisett p. 67) | Not found | +| Neck verse | Lay commoners convicted of clergyable felony | Reverse: a text handed to the accused | A reading performance, policed by the judge | Accused; the judge could raise it | Forbidden (Cross p. 552); gaol teaching (Gabel) | Parliament under Anne (Cross p. 555); Blackstone | +| Statute of Frauds | Parties to s. 4 and s. 17 promises | Not applicable | Signed note; oral routes partly kept | The promisee who trusted speech | No | 1937 Committee; Lord Campbell; Stephen (via Stollery p. 254) | +| Requerimiento | Indigenous peoples addressed | Read in Spanish, sometimes to the trees | Yes; the notary's record was the product | The hearers, in the text's own words | No; Oviedo's "jaula" jest | Las Casas; Oviedo | +| Treaty 6 / Treaty 8 | Signatory nations | Read and explained by interpreters | By the Crown then; not by the SCC in 1996 | Signatories, until *Badger* | Schools promised "as … may seem advisable" | Courts (modern); Cree requests at Carlton | + +--- + +## (e) What the paper can say + +Notes for the drafter, each a claim the record supports. + +1. The Requerimiento, drafted by Palacios Rubios for Pedrarias Dávila's expedition and read by that expedition at Santa Marta, by Oviedo's account, on 14 June 1514 (Oviedo 1853 t. III, cap. VII, p. 29), made Indigenous hearers responsible for a war on the strength of a Spanish text read aloud: "las muertes y daños que dello se recrecieren sea á vuestra culpa" (Las Casas, *Historia* III.57, t. IV p. 156; Helps 1855 p. 382 for the English). +2. Its legal product was the notary's sworn record that it had been read, not the hearers' understanding (Las Casas p. 156 and p. 199; Israeli 2022 p. 46: once performed, "the addressee could not pretend to be ignorant of the demand"). +3. Two men who were there or knew its author condemned it in its own century: Oviedo, who carried it and wrote that showing the Indians the paper "poco hacia al caso" (Oviedo 1853 t. III pp. 27–28), and Las Casas, who held that a command in a language the hearer does not understand has no force (pp. 162–163). +4. The procedure came from Castilian practice where addressees understood, answered and had their answers recorded; it failed when moved to people who could do none of these (Israeli 2022 p. 47). This is the clearest historical instance of v4's wrong (b). +5. The text promised the hearers "el tiempo que fuere justo" to understand; Oviedo reports that captains read it to people already captured and tied (p. 32), and his only proposal to teach it was to put an Indian "en una jaula" (p. 31). No one built the hearers' capacity. +6. In English criminal law a reading test decided who hanged: a lay commoner who could read a Psalter verse, usually Psalm 51, was branded and released; one who could not, unless a peer, was hanged (Cross 1917 pp. 552–553; Blackstone IV.28). The judge controlled the test's difficulty (Cross p. 554), and Cross reports that teaching a felon to read for the purpose was indictable (p. 552). +7. Parliament removed the reading condition under Anne and abolished the privilege in 1827 (Cross p. 555; 7 & 8 Geo. IV c. 28 s. VI); the 1489 statute (4 Hen. VII c. 13, Statutes vol. 2 p. 538) only branded readers and limited them to one use. +8. In medieval England the Crown argued under Edward I that only a written charter warranted a franchise, then conceded tenure "from time out of mind" in 1290 (Clanchy 1979, snippets; facet A for the 3rd-ed. passage); written proof also made forgery by the writing experts, the monasteries, routine (Clanchy, p. 249 per Ong; Hunnisett 1981 p. 68). +9. The Numbered Treaties were read and explained by interpreters (Morris 1880 p. 357), yet the written text could narrow what was said, as Treaty 6's medicine chest did (p. 218 spoken; p. 355 written), and the Supreme Court of Canada holds the written treaties an incomplete record of oral agreements (*Badger* ¶¶ 52, 55). +10. The Statute of Frauds put the cost of a missing writing on the party who trusted a spoken promise, "the honest man who has omitted a precaution" (1937 Law Revision Committee, quoted in Stollery 1976 p. 254); it was not a literacy test, and most of s. 4 was repealed in 1954. +11. Plato's Thamus gives the oldest statement of the scaffolding objection: readers "will read many things without instruction" and a written word cannot answer its reader (*Phaedrus* 275a–b, 275e, verified sister-arm card). +12. The paper should state its claims as claims about institutions, not minds: Finnegan (1973) and Street (2003, p. 77) reject the great divide, Goody himself disowned the binary ("not … a simple binary shift", 1987), and Ong's account is explicitly about "mentality" (Ong p. 3) and so is not the paper's ground. + +**Gap statements.** +- I found no work that reads the Requerimiento, the neck verse, the Statute of Frauds or the Numbered Treaties as precedents for disclosure-based AI governance, in OpenAlex (five queries, log row 12) or in the arm's facets; V6 documents the wider gap for oral-to-written analogies. +- I found no source that names a social class as the losers under the Statute of Frauds, in the statute, Stollery 1976 or facet A's sources. +- I found no court text on Treaty 6's medicine chest in the sources I could open (*Dreaver* 1935 and *Johnston* 1966 not reached). + +--- + +## (f) Leads for a library visit + +- **Clanchy 1979 or 1993 in print:** pp. 21–22 (Warenne, in full, with the caution paragraph); p. 249 and the forgery chapter (the 164/44/56 figures and their page); pp. 31, 36, 184 (statute of Exeter and seals); p. 219 (reading as hearing); the page of the 1290 compromise sentence. +- **Gabel 1928/1969:** p. 72 and the reading-test pages ("infra, pp. 64 et seq."); the full sentence on the boys from Appleby. +- **Seed, *Ceremonies of Possession* (1995), ch. 3** and **Faudree 2015** (*Colonial Latin American Review* 24(4), doi:10.1080/10609164.2016.1150029) in full. +- **Statutes at Large** for 7 & 8 Geo. IV c. 28 s. 6 and for 6 Anne c. 9 s. 4 / 5 Anne c. 6 (the reading test's removal, with its year). +- ***Dreaver v. The King* (1935, Exchequer Court)** and ***R. v. Johnston* (1966, Sask. C.A.)** on the Treaty 6 medicine chest. +- **Law Revision Committee, Sixth Interim Report (Cmd. 5449, 1937), pp. 6–7**, to quote it directly rather than through Stollery. +- **Finnegan 1973** in full (pp. 112 ff.) and **Goody 1977** pp. 8, 151, 164. + +--- + +## (g) Cards written this round (all `verified: pending`) + +`lascasas1876historia`, `oviedo1853historia`, `helps1855spanish`, `israeli2022requerimiento` (Requerimiento); `cross1917english`, `gabel1928benefit`, `statute1489clergy` (benefit of clergy); `clanchy1979memory` (memory to written record); `statute1677frauds`, `stollery1976statute` (Statute of Frauds); `morris1880treaties`, `badger1996` (Numbered Treaties); `finnegan1973literacy` (great divide). Reused without new cards: `ong1982orality` (V6 card; this facet read further pages), `street2003whats`, `hunnisett1981clanchy`, `robertson1980clanchy`, `blackstonendcommentaries`, and the sister-arm cards `plato1925phaedrus` and `havelock1963`. diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V4_disclosure_and_burden.md b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V4_disclosure_and_burden.md new file mode 100644 index 00000000..7a832d22 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V4_disclosure_and_burden.md @@ -0,0 +1,180 @@ +# Facet V4: disclosure, transparency and the burden on the reader (2026-10-14) + +Facet V4, research round 7 (v4), 2026-10-14. Agent-drafted. The author has not reviewed it, and none of the text is the author's. + +## What this facet covers + +The v4 thesis (v4/SOURCE_v4.md) holds that Nadella's post of 10 October 2026 treats disclosure (specific disclosures, readable records, design for legibility) as sufficient ethical design, and that this repeats three wrongs of the oral-to-written transition: (a) placing the burden on the person who must read; (b) keeping the person in a literacy-based experience; (c) not scaffolding growth to algorithmacy. This facet asks what the record says about disclosure and transparency as ethical design, and about shifting the burden onto the reader, in five bodies of work: + +1. Mandated disclosure and notice-and-choice (Ben-Shahar and Schneider; Loewenstein, Sunstein and Golman; McDonald and Cranor; Obar and Oeldorf-Hirsch; Solove). +2. The transparency critique in algorithmic governance (Ananny and Crawford; Edwards and Veale; Kaminski; Burrell; Selbst and Barocas; Nissenbaum). +3. Experiments on explanations and human oversight of AI (Poursabzi-Sangdeh et al.; Bansal et al.; Buçinca et al.; Vasconcelos et al.; Lee et al. 2025; Bainbridge 1983; the meta-analyses of Schemmer et al. and Vaccaro et al.). +4. Responsibilization (Shamir; Giesler and Veresiu; notice-and-choice as burden-shifting). +5. The counter-literature that defends disclosure and transparency (informed-consent ethics; disclosure as respect for autonomy; Kim and Routledge; Calo; Fung, Graham and Weil). + +Findings are sorted at the end under **for**, **narrows** and **against** the thesis. + +## Read-depth key + +- **Full**: read the whole text this session, or the work sits on a verified card that records a full read. +- **Part**: opened the text and read the named pages or sections. +- **Abstract**: database or publisher abstract only. +- **Record**: Crossref or catalogue record verified; text not opened. + +Where a finding rests on a verified card in this arm or in a sister library, the entry gives the card path and says "card". I did not reread those sources unless the entry says so. + +## Method notes + +- **Reuse first.** Before searching I read this arm's facets M2, M3, M4, N2, N4 and N5 and the cards `benshahar2011failure`, `edwards2017slave`, `kaminski2021right`, `kim2022why`, `giesler2014creating`; the sister arm's facet `L_disclosure_and_choice.md` and library cards (`dxd_foundation/literature/library/`: `poursabzi2021manipulating`, `bansal2021whole`, `bucinca2021trust`, `vasconcelos2023explanations`, `lee2025impact`, `loewenstein2014disclosure`, `simkute2024ironies`); and the S2-verified library of the `algorithmacy_design_ethics` arm in this worktree (`submissions/algorithmacy_design_ethics/literature/library/`: `bainbridge1983.md`, `burrell2016.md`, `nissenbaum2011.md`, `anannycrawford2018.md`, `selbstbarocas2018.md`). Where those cards record a verified full read, I cite the card and did not reread. +- **New downloads** (all in `/Users/ludwitt/iit-playground/berthon_ethics_local/pdfs/V4/`, each with a `pdftotext -layout` extraction): Solove 2013 (Harvard Law Review's own PDF, `harvardlawreview.org/wp-content/uploads/2013/05/vol126_solove.pdf`, journal pagination 1880–1903, read in full); Calo 2012 (Notre Dame Law Review production file with journal pagination 1027–1072, hosted at `techpolicylab.uw.edu`, read in part); Obar and Oeldorf-Hirsch 2022 (*International Journal of Communication* 16, 4779–4800, open access, read in part), which restates the 2020 NameDrop results with numbers. Bainbridge 1983 (Pergamon PDF, 5 pp., from an open author-collection copy at `ckrybus.com`; pp. 775–777 read and every quotation below located this session). +- **Local copies reread this session:** Selbst and Barocas 2018 (`berthon_ethics_local/pdfs/M2/selbst.txt`, Fordham pagination; p. 1094 and nn. 24, 48–49); Kaminski 2019 (`pdfs/N2_platform_law/scholarship/kaminski2019_osf.txt`, LawArXiv preprint of 15 June 2018, own pagination; pp. 18–19, 23); McDonald and Cranor 2008 (`dxd_foundation_local/pdfs/L/mc.txt`, author's pre-press; abstract p. 1, pp. 17–20); Vasconcelos et al. 2023 (`dxd_foundation_local/pdfs/cards/vasconcelos2023explanations.pdf`; pp. 129:1, 11, 15, 17–21, 27, 32 located). +- **Blocked, not worked around:** GW Law repository copy of Solove (Cloudflare challenge; the HLR copy was used instead); Ananny and Crawford 2018 (no open copy; SAGE closed; earlier rounds in this arm and the algorithmacy library also failed); Obar and Oeldorf-Hirsch 2020 (T&F and SSRN walled in earlier rounds; I did not retry them); Shamir 2008 (no open copy found by web search; *Economy and Society* is T&F); Ben-Shahar and Schneider 2014 (book; no open review found in one web search). Semantic Scholar's API returned 429. Unpaywall not used. No e-mail address was sent anywhere. +- **Depth honesty.** Ananny and Crawford remain *Abstract* plus two short passages reached through citing works that I read (Selbst and Barocas p. 1094; Edwards and Veale p. 67 on the verified card). Obar and Oeldorf-Hirsch 2020 is *Record* plus the authors' own restatement in their 2022 article, which I read; numbers attributed to the 2020 study come from that restatement and are marked so. + +## 1. Mandated disclosure and notice-and-choice + +**Ben-Shahar and Schneider (2011), `benshahar2011failure`.** *Part*, card (verified). Mandated disclosure is "the single most common technique for protecting personal autonomy in modern society" (p. 647), and "it chronically fails to accomplish its purpose. Even where it seems to succeed, its costs in money, effort, and time generally swamp its benefits" (p. 651). On boilerplate, roughly 1 in 1,000 shoppers opens the terms, for a median 29 seconds (p. 671, n. 127, Bakos et al.); Marotta-Wurgler: "No matter how prominently [end-user license agreements] are disclosed, they are almost always ignored" (p. 672, n. 130). Disclosure is inequitable: it serves the sophisticated more than the poor (p. 742). The sister arm's full OCR read adds: "disclosees often do not read disclosed information, do not understand it when they read it, and do not use it even if they understand it" (p. 665); "Financial literacy education is an important component of the disclosure and informed consent paradigm, and much evidence now shows that this effort has largely failed" (p. 667); "complexity cannot be explained simply" (p. 713); and a concession that brief symbolic signals (Los Angeles restaurant grades) may work, hedged "There is little real evidence" (p. 743) (sister facet L, entry 1; OCR of an offprint, not yet checked against the journal). + +**Ben-Shahar and Schneider (2014), *More Than You Wanted to Know*, `benshahar2014more`.** *Record* (Princeton University Press; Crossref book DOI 10.2307/j.ctt5hhrqj per sister facet L, not re-resolved here). I found no open review. Cite the 2011 article for every quoted claim. + +**Loewenstein, Sunstein and Golman (2014), `loewenstein2014disclosure`.** *Full*, sister card (S2-verified). "In many cases, disclosure does not much affect the recipients of the information but does significantly affect the behavior of the providers" (abstract, p. 391); "A pervasive theme is that disclosure may have little effect on recipients but large effects on providers" (p. 398); fewer than 3% read website privacy disclosures (p. 399); "making disclosure simple and accessible may provide insufficient help to consumers who lack the tools to process that information" (p. 408); policy makers must guard against disclosure "implemented as a substitute for other, often more effective, regulatory interventions" (p. 412); and disclosure requirements "appear to have been less effective in changing recipient behavior than their proponents seem to assume" (p. 413). Against the thesis: simplified, standardised, comparative and personalised disclosure sometimes changed choices (Kling et al.'s Medicare letter, 28% against 17% switching, p. 407), and §4.6 proposes intermediaries who read for the consumer (p. 411). The authors call themselves proponents of disclosure (p. 395). + +**McDonald and Cranor (2008), `mcdonald2008cost`.** *Part* (author's pre-press, reread this session: abstract p. 1, conclusion p. 19; the journal is *I/S* 4(3), 543–568). A modelled estimate, not observed behaviour: "if website users were to read the privacy policy for each site they visit just once a year, what would their time be worth?" (p. 1). Reading "carries costs in time of approximately 201 hours a year, worth about $3,534 annually per American Internet user. Nationally ... about $781 billion annually" (pre-press p. 19). The point estimate of 244 hours a year equals "an average of 40 minutes a day" (pre-press p. 18; journal p. 563, per sister facet L). The authors set this against an online advertising market of about $21 billion in 2007: users "should give up an estimated $781 billion of their time to protect themselves from an industry worth substantially less" (pre-press p. 20). Solove (2013, p. 1889, n. 41) cites the $781 billion figure to journal p. 564. + +**Obar and Oeldorf-Hirsch (2020), `obar2018biggest`.** *Record* for the article (Crossref DOI 10.1080/1369118X.2018.1486870; *ICS* 23(1), 128–147). Its results, **as restated by the same authors** in their 2022 article (p. 4784), which I read: an experimental survey of undergraduates (N = 543) joining a fictitious network, NameDrop; "Seventy-four percent" took the clickwrap and skipped the privacy policy; the policy "should have taken about 30 minutes to read" and those who opened it averaged 76 seconds (median 14 s); the terms of service needed 16 minutes and drew an average 51 seconds (median 14 s); "Ninety-seven percent of participants agreed to the PP and 93% to the TOS, with 98% missing the gotcha clauses" (data sharing with the NSA; a first-born child as payment). The 2020 article's own phrase, quoted in 2022: participants aimed to "pursue the ends of digital production, without being inhibited by the means" (2020, p. 128, via 2022 p. 4784). Read the 2020 article before quoting its numbers from it directly. + +**Obar and Oeldorf-Hirsch (2022), `obar2022older`.** *Part* (open-access PDF; abstract, pp. 4780, 4784, 4795–4796). Replication with adults 50+ (N = 500): "77.6% chose the clickwrap"; average reading time about 70 seconds for the privacy policy and 81.4 seconds for the terms; "91.4% of participants accepted the NameDrop PP, which included data collection/sharing 'gotcha' clauses, while 83.4% accepted the TOS, including an extreme clause requiring users to provide a kidney or other 'redundant organ' in exchange for service" (abstract, p. 4779). The authors frame the result as burden: "a primary flaw of the data privacy transparency efforts of digital service providers is the impossible burden placed on the individual (McDonald & Cranor, 2008; Solove, 2012)", who stands "as both perpetual data subject and overwhelmed privacy savior" (p. 4780). Their remedy is placed on providers and policymakers, not readers: they "must do more to address the burden placed on individuals through deceptive interface designs like the clickwrap, and via TOS and privacy policies that are long and complicated" (p. 4796; see also p. 4795, "overburdened privacy savior"). + +**Solove (2013), `solove2013introduction`.** *Full* (HLR PDF, pp. 1880–1903). The most direct statement of the burden claim in the privacy literature. +- The regime: "rights to notice, access, and consent" that he names "privacy self-management", and which "is being tasked with doing work beyond its capabilities" (p. 1880). +- The cognitive chain: "(1) people do not read privacy policies; (2) if people read them, they do not understand them; (3) if people read and understand them, they often lack enough background knowledge to make an informed choice; and (4) if people read them, understand them, and can make an informed choice, their choice might be skewed by various decisionmaking difficulties" (p. 1888). +- The structural problem survives perfect readers: "Even assuming that people are fully informed and rational ... privacy self-management also faces serious structural problems" (p. 1888); "even if all companies provided notice and adequate choices, this data management problem would persist; the average person just does not have enough time or resources to manage all the entities that hold her data" (p. 1889). The image is the student whose five or six professors each assign a reasonable night's reading (p. 1889). +- Simplification trades against meaning: "making it simple and easy to understand conflicts with fully informing people about the consequences of giving up data" (p. 1885); education efforts "fail to address a deeper problem — privacy is quite complicated" (p. 1886). +- The burden in so many words: "Privacy self-management cedes substantial responsibility for preserving privacy to individuals" (p. 1893). +- Opt-in will fail: firms "will have the sophistication and motivation to find ways to generate high opt-in rates"; the iTunes agreement "often extends to more than fifty screens" (p. 1898); affirmative consent leads "to more buttons to click and more forms to sign, but not to more meaningful privacy protection" (pp. 1898–1899). +- **Against a strong reading of the thesis:** "For all its flaws, privacy self-management should not be abandoned. Providing people with notice, access, and the ability to control their data is key to facilitating some autonomy" (p. 1899). Its main effect may fall on firms: "perhaps the greatest practical impact of privacy self-management is not in informing individuals ... but in informing the companies that are collecting and using the data" (p. 1900). Self-management "is currently asked to bear far more weight than it can support" (p. 1900). Remedies: partial self-management, regulating downstream uses "at the time they are proposed", perhaps agency review (p. 1902), and "Moving Toward Substance over Neutrality", since "Consent often becomes a convenient way to reach outcomes without confronting the central values at stake" (pp. 1902–1903). + +**Nissenbaum (2011), `nissenbaum2011contextual`.** *Full*, algorithmacy-library card (`algorithmacy_design_ethics/literature/library/nissenbaum2011.md`, read twice against two copies). "There is considerable agreement that transparency-and-choice has failed" (p. 35); the transparency paradox, "transparency of textual meaning and transparency of practice conflict in all but rare instances" (p. 36); medical consent works only inside "a framework of supporting assurances" — "It is not the consent form itself that draws our signature ... but rather our faith in the system", whereas "In the online environment, by contrast, individual consent agreements must carry the entire weight of expectation" (p. 36); "burdening individuals with the full weight of protecting their privacy online through notice-and-choice is unlikely to yield success" (p. 43). Her remedy is substantive context norms embedded "in law and other specialized codes" (p. 45). + +## 2. The transparency critique in algorithmic governance + +**Ananny and Crawford (2018), `ananny2018seeing`.** *Abstract* (Crossref abstract verified on the algorithmacy card) plus two passages reached through citing works I read or that sit on verified cards. Abstract: "Being able to see a system is sometimes equated with being able to know how it works and govern it ... But can 'black boxes' ever be opened, and if so, would that ever be sufficient?"; the article presents "10 limitations" and "an alternative typology of algorithmic accountability". Via Selbst and Barocas (2018, p. 1094, n. 48, citing A&C at 974): "Transparency concerns are commonly driven by a certain chain of logic: observation produces insights which create the knowledge required to govern and hold systems accountable." Via Edwards and Veale (2017, p. 67, verified card): transparency places "a tremendous burden on individuals ... only then to find out they have little power to change things anyway, being 'disconnected from power.'" *Caution:* the "tremendous burden" wording is Edwards and Veale's paraphrase around a quoted phrase ("disconnected from power"); I could not check which words are Ananny and Crawford's. The ten limitations are unread. + +**Selbst and Barocas (2018), `selbst2018intuitive`.** *Part* (Fordham PDF, local; p. 1094 and its notes reread). **This is the cleanest located statement that transparency puts the work of understanding on the observer**: "The process of moving from observation to knowledge to accountability cannot be assumed; in many cases, the ability to leverage observations for accountability requires preexisting knowledge that allows observers to appreciate the significance of a disclosure" (p. 1094; n. 49 cites Burrell at 4). The same page: "Transparency of systems of decision-making is important, but incomplete", and, against making this an algorithm-only problem, "the problem of expertise—like the problem of secrecy—is not unique to algorithms" (p. 1094). Abstract (algorithmacy card): existing law addresses inscrutability, but this "may not help if the goal is to assess whether the basis for decision-making is normatively defensible"; one "must seek explanations of the process behind a model's development, not just explanations of the model itself" (p. 1086). + +**Burrell (2016), `burrell2016machine`.** *Full*, algorithmacy-library card (`burrell2016.md`, S2-checked against the Wayback copy of the version of record). Three opacities: secrecy (remediable by disclosure or audit), "technical illiteracy" (remediable by education), and the mismatch "between mathematical optimization in high-dimensionality ... and the demands of human-scale reasoning". End-user explanations listing top-weighted criteria "provide an understanding that is at best incomplete and at worst false reassurance". Her remedies include external audit and, "perhaps surprisingly", not using machine learning in some domains. The HTML has no page numbers; cite by section. + +**Edwards and Veale (2017), `edwards2017slave`.** *Full* (verified card). The search for a right to explanation "may be at best distracting, and at worst nurture a new kind of 'transparency fallacy'" (abstract); "Individuals are mostly too time-poor, resource-poor, and lacking in the necessary expertise to meaningfully make use of these individual rights" (p. 67); individual rights "rely too much on individual rights for what are too often group harms" (p. 81); they prefer "creating better systems, with less opacity, clearer audit trails, well and holistically trained designers, and input from concerned publics" (p. 82). + +**Kaminski (2019), `kaminski2019right`.** *Part* (preprint reread this session, pp. 18–19, 23). **The direct reply to Edwards and Veale, and it narrows the thesis.** "To summarize the right to explanation and accompanying transparency measures, as some have, as a 'transparency fallacy'—palliative measures requiring mere icons or simplistic explanations—is to both misrepresent their actual substance and mischaracterize the GDPR's overall transparency regime" (p. 19). The GDPR approximates Pasquale's "qualified transparency": "targeted revelations of different degrees of depth and scope, aimed at different recipients", including "internal company oversight, oversight by regulators, oversight by third parties, and communications to affected individuals" (p. 19). "Individuals might not have access to source code or data sets, but other parties do" (p. 23). Earlier rounds (N2) recorded her own caveat: reliance on "collaborative governance in the absence of significant public or third-party oversight is not a known recipe for success" (p. 25). + +**Kaminski and Urban (2021), `kaminski2021right`.** *Part* (verified card). A right to contest private-sector AI "is unlikely to include a neutral arbiter, at least at first bite. ... interested actors make the decisions" (p. 2039). Bearing: the reviewer of a disclosure-based safeguard is the deploying organisation, which fits the post's "us". + +## 3. Experiments on explanation and human oversight of AI + +All entries in this section rest on S2-verified cards in the sister arm's library (`/Users/ludwitt/iit-playground/wt-dxd-foundation/submissions/dxd_foundation/literature/library/.md`) or in the algorithmacy library, except where marked. **The direction of the evidence is mixed.** Explanations rarely improve accuracy over the AI's advice alone; they often raise acceptance of wrong advice; but some designs do lower overreliance, and every study is one person and one machine. + +**Poursabzi-Sangdeh et al. (2021), `poursabzi2021manipulating`.** *Full* (sister card; N = 3,800 over four pre-registered MTurk experiments). A clear two-feature model let people simulate its predictions better (Exp. 1: t(994) = −12.06), but people did not follow it more than a black box (t(994) = 0.67, p = 0.5), and those shown the clear model were *less* able to detect its large mistake on an unusual apartment (Exp. 1: F(1, 994) = 8.81, p = 0.003; replicated for one of two apartments in Exp. 2; absent in Exp. 3, where people guessed first; removed by an "outlier focus" message in Exp. 4). The authors' rule: "testing, not intuition" (p. 26). *Bearing:* seeing inside raised prediction of the model, not detection of its failures; the fix that worked was a flag the system raised, not more to read. + +**Bansal et al. (2021), `bansal2021whole`.** *Full* (sister card; 1,626 MTurk users, AI accuracy matched to human). Confidence alone produced complementary performance (Beer: team 0.89 against AI 0.84 and humans 0.82); no explanation condition beat it (0.89 against 0.88, z = −1.18, p = .24); explanations "increased the chance that humans will accept the AI's recommendation, regardless of its correctness" (abstract), better when the AI was right and worse when it was wrong (§5.1, Fig. 4B). *Narrows:* with immediate feedback after each item, some workers learned when to rely on the AI (§5.3), which is learning by acting, not by reading. + +**Buçinca, Malaya and Gajos (2021), `bucinca2021trust`.** *Full* (sister card; arXiv v1, N = 199). Cognitive forcing functions (AI advice on request, after one's own decision, or after a wait) beat simple explanations only on the six wrong-AI questions (correct source picked: no AI .49, explanation .08, forcing .27) and never reached the no-AI level; the authors found no support that forcing functions improve team performance overall (p. 188:16). People "performed best in conditions that they preferred and trusted the least" (p. 188:16), and benefits went mainly to high need-for-cognition participants, which the authors name "intervention-generated inequalities" (p. 188:12). *Bearing:* the design that helped adds friction and reorders the task; it does not add text. Its gains fall unevenly. + +**Vasconcelos et al. (2023), `vasconcelos2023explanations`.** *Full* (sister card; five studies, N = 731, maze task). **Against a blanket "explanations fail":** a highlighted-path explanation lowered overreliance in the hard task (50×50 maze; estimate 1.74, CI 0.89 to 2.76) though not in easy or medium mazes; a written step-list carrying the same content did no better than no explanation (Study 2); a salient explanation that coloured the path where it crossed a wall averaged 0% overreliance (Study 3, exploratory, p. 129:21); higher pay lowered overreliance (Study 4). Conclusion: "overreliance is not merely an immutable inevitability of cognition but at least, in part, a strategic choice" (p. 129:32). *Bearing:* explanations help when they make verification cheaper than doing the task, which is a change of form, not more reading; a text with the same content did nothing. + +**Schemmer et al. (2022), `schemmer2022meta`.** *Full* (sister card). Explanation over AI advice alone: SMD 0.07, 95% CI −0.15 to 0.30, 11 studies (p. 620); explanation-plus-AI over humans alone: 0.59 [0.39, 0.79], 33 studies (p. 620). "Just one article reports some form of appropriate reliance" (p. 624); pure acceptance increase "can be understood as a form of manipulating people to blindly accept AI advice" (p. 625). Quote both numbers together. + +**Vaccaro, Almaatouq and Malone (2024), `vaccaro2024when`.** *Full* (sister card; 106 experiments, 370 effect sizes). Human–AI combinations did worse than the better of human or AI alone (g = −0.23), worst in decision tasks (g = −0.27); explanation and confidence were not significant moderators (p. 2295); researchers "may wish to de-emphasize this line of inquiry" (p. 2297). A between-study null, not proof of no effect. + +**Lee et al. (2025), `lee2025impact`.** *Full* (sister card; survey of 319 knowledge workers, 936 tasks; self-report, cross-sectional). Confidence in the AI was associated with less self-reported critical thinking (β = −0.69), self-confidence with more (β = 0.26); critical thinking shifts "from information gathering to information verification" and toward "task stewardship" (§5.2); GenAI "may also reduce critical engagement, particularly in routine or lower-stakes tasks in which users simply rely on AI" (§1, p. 2). *Against the thesis on remedy:* the authors' own design advice is partly explanatory ("providing explanations of AI reasoning, suggesting areas for user refinement, or offering guided critiques", §6.1.2). "Our analysis does not establish causation" (§6.1.1). + +**Simkute et al. (2024), `simkute2024ironies`.** *Full* (sister card; preprint, narrative synthesis). With GenAI, "users' roles have shifted from producing output to evaluating it, often with little contextual information and situational awareness" (p. 2). Its first remedy is feedback "tied to carefully designed explainability features" (p. 13), so it documents the reading burden and then prescribes more legible output. + +**Bainbridge (1983), `bainbridge1983ironies`.** *Full* (Pergamon PDF reread this session, pp. 775–777; also the S2-verified algorithmacy-library card `algorithmacy_design_ethics/literature/library/bainbridge1983.md`). The designer "still leaves the operator to do the tasks which the designer cannot think how to automate" (p. 775); vigilance cannot be sustained "for more than about half an hour" (p. 776); where the computer uses finer criteria than a human can follow, "there is therefore no way in which the human operator can check in real-time that the computer is following its rules correctly ... The human monitor has been given an impossible task" (p. 776); the job is "very boring but very responsible, yet there is no opportunity to aquire [sic] or maintain the qualities required to handle the responsibility" (p. 776). Remedies: the computer should decide "using methods and criteria, and at a rate, which the operator can follow" (p. 777); "hands-on control for a short period in each shift" (p. 777); systems "should fail obviously" (p. 777); and "it is the most successful automated systems, with rare need for manual intervention, which may need the greatest investment in human operator training" (p. 777). *Bearing on wrong (c):* Bainbridge's remedy is scaffolded practice, not readable records. + +**Green (2022), `green2022flaws`.** *Full* (sister card; 41 policy documents). Oversight rules "rest on an uninterrogated assumption: that people are able to effectively oversee algorithmic decision-making" (abstract, p. 1); "Algorithmic transparency similarly reduces people's ability to detect and correct model errors" (p. 8, citing Poursabzi-Sangdeh et al.); the overseer becomes the "scapegoat" (p. 10); "The default assumption should be that human oversight is likely to be ineffective, unless proven otherwise" (p. 15). Green leaves better design "an open empirical question" (p. 11) and moves oversight to institutions (pp. 12–14). + +**Elish (2019), `elish2019moral`.** *Full* (sister card). The moral crumple zone: responsibility "may be misattributed to a human actor who had limited control over the behavior of an automated or autonomous system", protecting "the integrity of the technological system, at the expense of the nearest human operator" (abstract, p. 40); "While control has been effectively distributed, responsibility has not scaled accordingly" (p. 49). At Three Mile Island "the operators made incorrect decisions because they had incorrect information" (p. 44). + +**Crootof, Kaminski and Price (2023), `crootof2023humans`.** *Full* (sister card). The "MABA-MABA trap", "an assumption that is intuitive, wrong, and dangerous: that human-machine systems represent the best of both worlds" (p. 438); the "tautological challenge of relying on humans to monitor the performance of systems designed to improve on human performance" (p. 469); a human may be held responsible legitimately only "if the human in the loop has the power, information, judgment, and time to make the final decision" (pp. 482–83). + +*Bearing on the post.* The post's P15 asks that "Every meaningful model action must leave tamper-proof human readable evidence" and its P9 calls chain-of-thought transparency "non-negotiable" while saying it is not sufficient alone. This section's studies concern a reader of AI output or explanation at decision time; none tests after-the-fact audit of agent logs by trained staff (the sister card on Green says the same). The transfer is by analogy. + +## 4. Responsibilization and notice-and-choice as burden-shifting + +**Giesler and Veresiu (2014), `giesler2014creating`.** *Full* (card, verified: corrected). Responsibilization proceeds through "four distinct but interrelated processes through which responsibility is shifted away from the state and corporations and the responsible consumer subject is shaped" (p. 841); personalization moves a problem "from the systemic to the level of individual decision making" (p. 843); the fourth case is "the financially literate consumer" (p. 851), for whom the WEF "shifted the problem of financial debt away from the level of better financial regulation to a need for more conscious individual financial decision makers" (p. 851). + +**Shamir (2008), `shamir2008age`.** *Record* (Economy and Society 37(1), 1–19; no open copy). Known through Giesler and Veresiu: responsibilization is a "technique of the self that sets into action a reflexive subjectivity deemed suitable to partake in the deployment of horizontal authority and one which willingly bears the consequences of its actions" (Shamir 2008, p. 4, via Giesler and Veresiu p. 841; facet N4); and Shamir "describes corporations as one target of responsibilization" (p. 841). Do not quote Shamir directly until read. + +**Notice-and-choice named as burden-shifting.** The privacy literature says it in plain terms, which spares the paper an inference: Solove, privacy self-management "cedes substantial responsibility for preserving privacy to individuals" (2013, p. 1893); Nissenbaum, "burdening individuals with the full weight of protecting their privacy online through notice-and-choice is unlikely to yield success" (2011, p. 43); Obar and Oeldorf-Hirsch, "the impossible burden placed on the individual", the user as "overwhelmed privacy savior" (2022, p. 4780); Edwards and Veale, individuals "too time-poor, resource-poor, and lacking in the necessary expertise" (2017, p. 67). Ben-Shahar and Schneider pair disclosure with literacy education and report the education "has largely failed" (2011, p. 667, via sister OCR read). Giesler and Veresiu supply the sociological name, responsibilization, and the literacy case (financial literacy). **I found no work that applies the responsibilization frame to AI or algorithmic *disclosure* specifically** in the sources this facet and facets M4 and N4 searched (Crossref and web searches recorded there; M4 found only an unopened arXiv title on "AI literacy" and power-knowledge). + +## 5. The counter-literature: disclosure and transparency defended + +**Calo (2012), `calo2012against`.** *Part* (Notre Dame production file, journal pagination; pp. 1027–1035, 1047–1050, 1069–1072). The strongest defence of notice I read, and it cuts two ways. Its thesis: "experience as a form of privacy disclosure is worthy of further study before we give in to calls to abandon notice as a regulatory strategy" (p. 1027). Notice has a legitimating role in law (Fuller's "internal morality", Hayek on the rule of law, p. 1028); "Notice purports to respect the basic autonomy of the consumer or citizen by arming her with information and placing the ultimate decision in her hands" (p. 1049); and "notice is the worst regulatory mechanism, except for all of the alternatives" (p. 1047), since substitutes are "perceived to be invasive, ineffective, expensive, or arbitrary" (p. 1047). He concedes the textual case: "Studies show only marginal improvement in consumer understanding where privacy policies get expressed as tables, icons, or labels, assuming the consumer even reads them" (p. 1033). **His remedy leaves text behind:** "visceral" notice "does not rely exclusively on language or its symbolic equivalent. Rather, it is 'visceral,' in the sense of changing the consumers understanding by leveraging the very experience of a product or service" (p. 1033); a camera that must make a shutter sound is his first example (p. 1027). He keeps it a notice strategy, whose purpose is "to ensure consumers have the correct mental model of the product or service and can make decisions accordingly" (p. 1070). Conclusion: "Emerging strategies that do not necessarily rely on words or symbols to convey salient information to consumers may not be susceptible to the same withering critiques as more traditional notice" (pp. 1071–1072). *Bearing:* against the claim that disclosure is ethically empty (notice respects autonomy, and its alternatives have costs), and **for** wrong (b): the most cited defender of notice saves it only by moving it out of the text and into the experience of using the product. + +**Solove (2013).** *Full* (see §1). Defends keeping self-management for autonomy (p. 1899) and warns that the alternative, paternalism, "even more directly denies people the freedom to make consensual choices about their data" (pp. 1881–1882). His "consent dilemma" is the objection the paper must meet: if disclosure is not sufficient, the obvious replacement removes choice. + +**Kim and Routledge (2022), `kim2022why`.** *Part* (card, verified). Informed consent reconceived "as an assurance of trust for incomplete transactions" grounds "a right to ex post explanation" (p. 75); data subjects can trust firms "only when the relationship is predicated on the companies' readiness/commitment to be accountable" (p. 88). Their own limit: "We do not believe that a right to ex post explanation can fundamentally solve such problems" (p. 101). *Bearing:* a business-ethics argument that disclosure (here, ex post explanation) is owed to the person as a matter of trust, independent of whether it improves her decisions. The post's P20 ("timely disclosure to those affected") fits their ex post model. + +**Kaminski (2019).** *Part* (see §2). Calling the GDPR's explanation right a "transparency fallacy" misrepresents a layered regime of "targeted revelations ... aimed at different recipients" (p. 19). *Bearing:* disclosure regimes need not put the whole load on the individual reader; the reading can be distributed to regulators, auditors and internal reviewers. + +**Fung, Graham and Weil (2007), `fung2007full`.** *Part*, via the sister arm's facet L (entry 3; read in part from the Ash Center PDF, printed pagination). "Transparency systems alter decisions only when they provide pertinent information that enables people to substantially improve their choices without imposing significant additional costs" (p. 55); effective disclosure is "doubly embedded" in the user's and the discloser's routines (p. 65); grade cards "convert data and expert interpretation into simple normative signals such as stars or letter grades" (p. 57). *Bearing:* the targeted-transparency school places the success condition in the design of the disclosure, not in the reader's skill, which narrows the thesis from "disclosure fails" to "disclosure designed as text for a reader fails". + +**Loewenstein, Sunstein and Golman (2014)** and **Solove (2013, p. 1900)** both record that disclosure duties act on the discloser (the "telltale heart" effect; notices that force firms "to inspect their own practices"). *Bearing:* a disclosure regime can be ethically useful even when no one reads it, because it disciplines the organisation that writes it. That fits the post's P15 reading of evidence as an internal control. + +**Disclosure that changes behaviour.** Sister facet L (entries 24, 26, 27) records studies in which a single statement of an algorithm's rule changed how people used it: Cen et al. (2024; read in part there), likes used 1.9 times more when users were told the algorithm attends to likes; Jhaver et al. (2019; abstract only there), removal explanations reduced later removals; Tyler et al. (abstract only there), a procedural-justice message reduced violations. Facet L's summary: "disclosure changes behaviour, and often in the wrong direction or on the wrong party" (§(c) Q5). I did not reread these; cite facet L. + +**Informed-consent ethics.** O'Neill (2003), "Some limits of informed consent", *J Med Ethics* 29(1), 4–7 (`oneill2003some`): *Record* (Crossref). The PMC copy returned a bot challenge and I did not work around it. It is the standard medical-ethics source for the view that consent is a limited waiver rather than full autonomous choice and belongs in the objections section once read. Nissenbaum's medical-consent passage (2011, p. 36, §1) already supplies the point the paper needs: consent forms work in medicine because a "framework of supporting assurances" carries most of the weight. + +## Sorted: for, narrows, against + +### For the thesis +1. **(a) Burden on the reader.** The privacy and disclosure literatures state it in so many words: Solove (privacy self-management "cedes substantial responsibility for preserving privacy to individuals", p. 1893; it "is being tasked with doing work beyond its capabilities", p. 1880); Nissenbaum ("burdening individuals with the full weight", p. 43); Obar and Oeldorf-Hirsch ("the impossible burden placed on the individual", 2022, p. 4780); Edwards and Veale ("too time-poor, resource-poor, and lacking in the necessary expertise", p. 67); McDonald and Cranor (a modelled 201–244 hours a year; $781 billion nationally, pre-press pp. 18–19). +2. **(a) The burden survives a perfect reader.** Solove's structural argument: even "fully informed and rational" people cannot manage scale, aggregation and future harm (pp. 1888–1891). +3. **(a) Transparency assumes a knowing observer.** Selbst and Barocas: moving "from observation to knowledge to accountability cannot be assumed", and needs "preexisting knowledge that allows observers to appreciate the significance of a disclosure" (p. 1094), restating Ananny and Crawford's "chain of logic" (via p. 1094, A&C at 974). +4. **(a) People do not read.** 74% skipped the policy and 98% missed the gotcha clauses (Obar and Oeldorf-Hirsch 2020, as restated in 2022, p. 4784); 77.6% and 91.4%/83.4% among adults 50+ (2022, p. 4779); about 1 in 1,000 open boilerplate (Ben-Shahar and Schneider, p. 671). +5. **(a) The burden moves blame to the person.** Elish's moral crumple zone (p. 40), Green's "scapegoat" (p. 10), Crootof et al.'s "liability sponge" test (pp. 482–83). +6. **(b) Text is the weak form.** Solove: simplification "conflicts with fully informing people" (p. 1885); Ben-Shahar and Schneider: "complexity cannot be explained simply" (p. 713, sister OCR); Calo: tables, icons and labels give "only marginal improvement" (p. 1033), and his rescue of notice leaves "language or its symbolic equivalent" behind (p. 1033); Vasconcelos et al.: a written step-list did nothing where a visual highlight helped (Study 2). +7. **(b) Seeing inside does not help one catch failures.** Poursabzi-Sangdeh et al. (clear model, worse error detection); Bansal et al. (explanations raise acceptance regardless of correctness); Schemmer et al. (0.07 over AI advice alone); Vaccaro et al. (explanation not a moderator); Burrell ("at best incomplete and at worst false reassurance"). +8. **(c) No scaffolding, and literacy education as the failed complement.** Ben-Shahar and Schneider: financial-literacy education "has largely failed" (p. 667, sister OCR); Giesler and Veresiu: financial literacy as responsibilization that moves debt "away from the level of better financial regulation" (p. 851). Bainbridge: the monitor gets "no opportunity to aquire [sic] or maintain the qualities required" (p. 776); "the most successful automated systems ... may need the greatest investment in human operator training" (p. 777). + +### Narrows the thesis +1. **The post's charge must be "necessary but treated as enough", not "disclosure is worthless".** Solove keeps self-management for autonomy (p. 1899); Calo calls notice the worst mechanism "except for all of the alternatives" (p. 1047); Kim and Routledge ground ex post explanation in trust (pp. 75, 88). +2. **Disclosure works on the discloser.** Loewenstein et al. (pp. 391, 398, 403) and Solove (p. 1900): the main effect may be on the firm that must write the disclosure. P15's readable evidence may be valuable as internal discipline even if the affected people never read it. +3. **Disclosure can be layered across readers.** Kaminski (2019, pp. 19, 23): regulators, auditors and internal reviewers read deeper than the individual. Nadella's P15 audience is the enterprise, not the end user (SOURCE_v4 note), so "burden on the user" applies most to the organisation's staff as overseers (Bainbridge, Green, Elish) and least to the people the systems act on, who appear only at P20. +4. **Form, not disclosure as such, decides.** Fung et al. (simple signals embedded in routines, pp. 55–65), Vasconcelos et al. (cheap-to-verify explanations reduce overreliance), Calo (experiential notice), Poursabzi-Sangdeh et al. (an outlier flag fixed the detection loss). These support wrong (b) more than wrong (a): the problem is the reading form, and some non-textual disclosures help. +5. **The scaffolding remedy has its own costs.** Buçinca et al.: forcing functions helped on wrong-AI items only, were disliked, and benefited high-NFC people most ("intervention-generated inequalities", p. 188:12). Fernandes et al. (via sister facet L): financial-literacy interventions explain 0.1% of variance in behaviour and decay. A paper that demands scaffolding to algorithmacy must say why it will not repeat literacy education's weak record. +6. **The AI evidence is dyadic and at decision time.** None of the studies tests after-the-fact audit of agent logs, or a mediated triad; transfer to P15 is analogy. + +### Against the thesis +1. **Some explanations help.** Vasconcelos et al. (hard task; salient explanation at 0% overreliance); Bansal et al. (feedback let some workers learn when to rely); explanation-plus-AI beats humans alone (Schemmer et al., 0.59). Lee et al. recommend explanations and guided critiques as remedies (§6.1.2). +2. **Disclosure changes behaviour.** Cen et al. (1.9×), Jhaver et al., Tyler et al. (sister facet L). "Disclosure does nothing" is false; "disclosure does not reliably help the person" is the defensible claim. +3. **Abandoning disclosure has a cost to autonomy.** Solove's consent dilemma (pp. 1881–1882, 1894) and Calo (p. 1049): the alternatives to notice decide for people. The paper must say what replaces the reader's choice without removing it. +4. **The post itself says CoT transparency is not sufficient** (P9) and puts most weight on controls (SOURCE_v4 note). The literature's burden critique targets regimes that rely on the individual reader; the post's reader is mainly the enterprise's own staff. The charge that the post treats disclosure as *sufficient* needs that qualification. + +## What the paper can say + +(Claims the record supports, with sources. Notes for the drafter, not prose.) + +1. Mandated disclosure has been called "the single most common technique for protecting personal autonomy in modern society", and its critics find that it "chronically fails" (Ben-Shahar and Schneider 2011, pp. 647, 651). +2. Privacy law's notice-and-consent regime "cedes substantial responsibility for preserving privacy to individuals" and "is being tasked with doing work beyond its capabilities" (Solove 2013, pp. 1893, 1880); the same critique holds even for "fully informed and rational" readers, because of scale and aggregation (pp. 1888–1891). +3. Reading every privacy policy once a year would cost an estimated 201–244 hours per person and about $781 billion nationally, a modelled figure (McDonald and Cranor 2008, pre-press pp. 18–19). +4. In a controlled sign-up, 74% skipped the privacy policy and 98% missed clauses giving data to the NSA and a first-born child as payment (Obar and Oeldorf-Hirsch 2020, as restated in 2022, p. 4784); among adults over 50, 83.4% accepted terms requiring a kidney (2022, p. 4779). +5. Transparency presupposes an observer who already knows enough to see what a disclosure means (Selbst and Barocas 2018, p. 1094), which is the literacy precondition the oral-to-written analogy needs. +6. Human-factors research since Bainbridge (1983, pp. 776–777) holds that a person asked to monitor a system faster and finer than she is has "an impossible task", and that her competence has to be built by practice, not by records; this is the oldest statement of wrong (c). +7. When oversight fails, blame settles on the nearest human (Elish 2019, p. 40; Green 2022, p. 10), so a regime of readable evidence can shift responsibility as well as work. +8. In AI-assisted decisions, adding an explanation to the AI's advice did not improve accuracy on average (Schemmer et al. 2022, SMD 0.07; Vaccaro et al. 2024) and often raised acceptance of wrong advice (Bansal et al. 2021); explanations helped where they made checking cheaper than the task (Vasconcelos et al. 2023). +9. Literacy-style remedies have a weak record: financial-literacy education "has largely failed" (Ben-Shahar and Schneider 2011, p. 667) and is a vehicle of responsibilization (Giesler and Veresiu 2014, pp. 841, 851). +10. Defenders of notice concede the textual case and save notice by moving it into the experience of the product (Calo 2012, pp. 1033, 1071–1072), which is the move from a literacy-based experience the thesis asks for. +11. The paper should state the objection fairly: disclosure respects autonomy (Calo 2012, p. 1049), disciplines the discloser (Loewenstein et al. 2014, p. 398; Solove 2013, p. 1900), can be layered across expert readers (Kaminski 2019, pp. 19, 23), and the alternative to choice is paternalism (Solove's consent dilemma, p. 1894). + +## Gaps + +- I found no work that applies the responsibilization frame (Shamir; Giesler and Veresiu) to AI or algorithmic disclosure specifically, in the sources this facet and facets M4 and N4 searched. +- I found no experiment that tests after-the-fact audit of an AI agent's readable logs by organisational staff, the post's P15 setting, in the sister arm's library, facet L, or this facet's reading; every explanation study read here concerns one person and one AI at decision time. +- I found no study that compares a reading-based disclosure with a scaffolded-practice intervention (Bainbridge's "hands-on control") for oversight of AI systems, in the same sources. +- Unread and needed before quotation: Ananny and Crawford (2018) body (the ten limitations); Obar and Oeldorf-Hirsch (2020) itself; Shamir (2008); O'Neill (2003); Ben-Shahar and Schneider (2014). The author's library access is the route. + +## Cards written (all `verified: pending`) + +New cards in `literature/cards/`: `solove2013introduction` (full-text), `calo2012against` (partial-text), `selbst2018intuitive` (partial-text), `obar2022older` (partial-text), `mcdonald2008cost` (partial-text), `kaminski2019right` (partial-text), `bainbridge1983ironies` (full-text), `vasconcelos2023explanations` (full-text). Existing cards cited and left untouched: `benshahar2011failure`, `edwards2017slave`, `kaminski2021right`, `kim2022why`, `giesler2014creating`. Recommended for carding once read: `ananny2018seeing` (body), `obar2018biggest` (the 2020 article itself), `shamir2008age`, `oneill2003some`; and, from the sister library, `poursabzi2021manipulating`, `bansal2021whole`, `elish2019moral`, `green2022flaws` (their sister cards are S2-verified and can be cited by path meanwhile). diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V5_scaffolding.md b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V5_scaffolding.md new file mode 100644 index 00000000..f3bd9b4e --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V5_scaffolding.md @@ -0,0 +1,561 @@ +# V5 — Scaffolding: what it would mean to build a person's growth into a new competence, and who did it + +*Facet V5, research round 7, 2026-10-14. Agent-drafted (Claude). None of the text is the author's. Notes for the +drafter, not prose for the paper.* + +**What this facet covers.** The evidence for and against wrong (c) of the v4 thesis: that Nadella's post (and the +oral-to-literate transition before it) fails by "not scaffolding the person's growth" to the new competence +(algorithmacy). Five questions: (1) what "scaffolding" means in learning science (Wood, Bruner and Ross 1976; +Vygotsky's zone of proximal development; Collins, Brown and Newman's cognitive apprenticeship); (2) what it means +in design for algorithmic systems (seamful design, folk theories, explanatory debugging, boosts versus nudges); +(3) who historically built people's capacity to read, who paid, and with what result; (4) the capability +approach on literacy and the duties it implies; (5) the case against: paternalism, the weak record of +literacy-education programmes, and responsibilization. + +**Read-depth marks.** *Full* = whole text read this session or on a verified card; *Part* = the named pages or +sections read; *Abstract* = abstract or database summary only; *Record* = bibliographic record only. + +## Method notes + +Search log (appended as the run proceeds). + +| # | Tool | Query or target | Result | +| --- | --- | --- | --- | +| 1 | repo read | `v4/RESEARCH_BRIEF.md`, `v4/SOURCE_v4.md`; this arm's facets C, F, M2, M4, M5, N3, Z (grep for Freire, Cuba, scaffold, capability, Sunday school, Self-Taught); card list | Freire ch. 2, Cuba (M5), Sunday schools (C, not opened), Self-Taught (Margo review, carded), capability SEP entry (M4) already held. Cards `johansson1985popular`, `washington1901up`, `brandt1998sponsors`, `kamperman2026ai`, `margo2005williams`, `graff1991literacy`, `street2003whats` exist. | +| 2 | repo read | sister arm `dxd_foundation/literature/`: `library/hertwig2017nudging.md`, `library/ehsan2024seamful.md`, facets L, I, M, G | Hertwig and Grüne-Yanoff card (verified, corrected); Ehsan seamful-XAI card; Fernandes et al. read in part (WP); Eslami 2016 and DeVito 2021 abstract-level in facet I; Bucher 2017 bibliographic only (facet M, no authorised copy). Facet G records Nadella's Dec 2025 post calling AI "a scaffolding for human potential". | +| 3 | WebSearch | Wood, Bruner & Ross 1976 pdf; six functions | No authorised full text; Wiley PDF returns 403. Secondary only. | +| 4 | archive.org FTS (`be-api.us.archive.org/fts/v1/search`) | "beyond his unassisted efforts"; six function names | Dozens of secondary quotations, all attributing the definition to p. 90. The original not reachable. | +| 5 | archive.org (OAPEN mirror) | Belland 2017, *Instructional Scaffolding in STEM Education* (open access, CC BY-NC) | Downloaded `V5/belland2017.pdf`; read ch. 1 pp. 1–6 and ch. 2 pp. 17–28. | +| 6 | archive.org FTS over the lending scan `mindinsocietydev0000vygo` (not borrowed) | "zone of proximal development", "level of potential development", "more capable peers", "functions that have not yet matured", "good learning" | Snippets only; no printed page from the API. Contents snippet places ch. 6 at p. 79. | +| 7 | ERIC | Collins, Brown & Newman 1987, Technical Report 403 (ED284181) | Downloaded `V5/collins1987eric.pdf`; read report pp. 1–8. | + +## Findings + +### Q1. Scaffolding in learning science + +**Wood, Bruner and Ross (1976), `wood1976role`. Secondary (no original read).** *J. Child Psychology and +Psychiatry* 17(2), 89–100, DOI 10.1111/j.1469-7610.1976.tb00381.x (PubMed 932126; Wiley 403). The definition every +secondary source quotes, attributed to p. 90: scaffolding is a process "that enables a child or novice to solve a +problem, carry out a task, or achieve a goal which would be beyond his unassisted efforts" (wording as quoted in +ERIC ED446926 and ED442640, both giving p. 90; seen as archive.org full-text snippets, not in the original). The six +tutor functions — recruitment, reduction in degrees of freedom, direction maintenance, marking critical features, +frustration control, demonstration — are listed in that order across secondary sources (PMC4021139; Belland 2017, p. 22, groups recruitment, frustration control and direction maintenance as support for motivation, and +the other three as support for cognition). *Do not quote the original until the author supplies the PDF.* **Flag for the brief:** the +word "fading" is not Wood, Bruner and Ross's. Belland (2017, p. 21) says the removal of support "was later termed +'fading' by Collins et al. (1989)". + +**Belland (2017), `belland2017instructional`. Part** (ch. 1 pp. 1–6; ch. 2 pp. 17–28). *Instructional Scaffolding +in STEM Education*, Springer, open access, DOI 10.1007/978-3-319-02565-0. The most useful single source for what +"scaffolding" requires, because it draws the boundary the thesis needs. +- Definition: "interactive support that leverages what students already know to help them meaningfully participate + in and gain skill at tasks that are beyond their unassisted abilities" (p. 5, citing Wood et al. 1976 and others). +- What does **not** count: "a calculator does not qualify as a scaffold because while it supports current + performance, it cannot be reasonably expected to help users calculate independently ... in the future" (p. 5); + "simply telling students what to do or how to do it does not qualify as scaffolding" (p. 6); "Support that + continues indefinitely does not meet the scaffolding definition either" (p. 18); "A job aid does not meet the + definition" (p. 19); "If the tool does all or most of the work ... the possibility of skill gain is compromised" + (p. 19). +- Contingency: scaffolding pairs "dynamic assessment of the child's current performance characteristics and + provision of just the right support" (p. 18); it requires "intersubjectivity", a shared understanding of what + success looks like (p. 18). +- Fading "is designed to gradually transfer the responsibility for the performance of the target skill from the + scaffold provider to the scaffold receiver" (p. 21). +- The term is imprecise: "the lack of precision that has emerged in the term scaffolding ... an often-lamented + phenomenon (Pea, 2004; Puntambekar & Hübscher, 2005; Stone, 1998)" (p. 18). +- Effect: in Belland's meta-analysis "the average effect size of computer-based scaffolding was 0.46", above the + 0.13 and 0.16 averages for educational technology in mathematics and reading (p. 27). Population: STEM students, + K-12 to adult, problem-centred instruction. +- *Bearing:* **for** wrong (c) and **narrows** it. For: by this definition a disclosure is "telling", and a readable + record is a job aid; neither scaffolds. Narrows: an AI tool that does the work, and any permanent design aid — + including the sister arm's DXD conventions, which change what the mechanism asks of a person — is a calculator in + Belland's sense, not a scaffold. If the paper says the post "fails to scaffold", it must say it means support + aimed at the person's later independent competence, and it must accept that permanent re-representation is a + different remedy. + +**Vygotsky (1978), *Mind in Society*, ch. 6, `vygotsky1978mind`. Snippet** (archive.org full-text search over the +Harvard UP lending scan `mindinsocietydev0000vygo`; not borrowed; printed page not given by the API; the contents +snippet puts ch. 6, "Interaction between Learning and Development", at p. 79). Three overlapping snippets, joined by +me: the zone of proximal development "is the distance between the actual developmental level as determined by +independent problem solving and the level of potential development as determined through problem solving under +adult guidance or in collaboration with more capable peers." Other ch. 6 snippets: the zone "defines those +functions that have not yet matured but are in the process of maturation"; such functions are "the 'buds' or +'flowers' of development rather than the 'fruits'"; "the only 'good learning' is that which is in advance of +development"; learning "awakens a variety of internal developmental processes". *The common citation to p. 86 is +unverified here; check in the book.* +- *Bearing:* **for** and **narrows**. The zone is defined by the presence of someone more capable. For a reader of + print, the more capable peer was any literate neighbour. For an adaptive, proprietary system, the only party that + knows the system's workings is its operator. That supports placing the duty to scaffold on the holder of the + means (Q4), and it narrows the claim: ordinary users cannot scaffold one another in the way literate peers could. + +**Collins, Brown and Newman (1987/1989), `collins1987cognitive`. Part** (ERIC ED284181, BBN/CSR Technical Report +403, January 1987, report pp. 1–8; the 1989 chapter in Resnick (ed.), *Knowing, Learning, and Instruction*, is the +usual citation and was not read). Cognitive apprenticeship was built **for teaching reading and writing** — the +title is "Teaching the Craft of Reading, Writing, and Mathematics". +- Standard schooling "render[s] key aspects of expertise invisible to students" (p. 2). +- Apprentices learn by "observation, coaching, and practice, or what we, from the teacher's point of view, call + modelling, coaching, and fading"; "A key aspect of coaching is the provision of scaffolding, which is the support, + in the form of reminders and help, that the apprentice requires"; then "the master reduces his participation + (fades)" (pp. 2–3). +- Craft apprenticeship works where the process "is external and thus readily available to both student and teacher + for observation"; for cognitive skills it "requires the externalization of processes that are usually carried out + internally" (p. 4). +- In reciprocal teaching of reading, "students get as much support as they need to carry out the task, but no more. + Hints and modelling are then gradually faded out" (p. 7). +- *Bearing:* **for** wrong (c). Its premise is that reading itself had to be scaffolded, because the expert's + process is invisible; the parallel for algorithmacy is that the system's process is invisible to the person + and visible only to its operator. Externalising the process is the operator's move to make. The text is a + technical report on schoolchildren; it says nothing about firms or platforms. + +### Q2. Scaffolding in design for algorithmic systems + +Search additions: OpenAlex records and abstracts (DOI lookups and title searches) for Eslami et al. 2015 and 2016, +DeVito et al. 2017, Bucher 2017, Kulesza et al. 2015, Chalmers 2003; ACM full texts 403; IDEALS (UIUC) 403; +Glasgow Enlighten has no full text for Kulesza; City, University of London repository holds the accepted version +(`V5/kulesza2015.pdf`). + +**Kulesza, Burnett, Wong and Stumpf (2015), `kulesza2015principles`. Full** (accepted manuscript, City Research +Online eprint 13819, 11 pp.; pagination is the manuscript's, the IUI proceedings run pp. 126–137, DOI +10.1145/2678025.2701399). *Explanatory Debugging*: "the system explains the reasons for its predictions to its end +user, who in turn explains corrections back" (p. 2). Principles: be sound, be complete, but don't overwhelm (pp. +3–4); be actionable, be reversible, always honour user feedback, incremental changes matter (p. 4). Their earlier +study "relied on human instructors to help participants build useful mental models" (p. 3) — this paper moves +that tutoring into the system. Experiment: 77 participants from the community and university with no more than +introductory programming, 37 control (black-box labelling) and 40 treatment, 30 minutes, a hockey/baseball text +classifier (p. 8). Treatment participants "scored 52% higher on the mental model assessment" (W = 259, p < .001) +and each of their feedback actions was worth roughly two of the control group's (pp. 10–11); mental-model score +correlated with classifier F1, ρ = .282 (p. 11); NASA-TLX showed no extra workload (p. 10). +- *Bearing:* **for** wrong (c), with a narrow scope. It is the cleanest measured case of a system that scaffolds a + person's understanding through a two-way loop rather than a text to read, and it replaces the human instructor. + **Narrows:** the person owns the classifier and its goal; the system has no interest against the user; the task + is 30 minutes in a lab. It says nothing about systems run by a party whose interests diverge from the user's, + which is the post's case. + +**Eslami et al. (2016), `eslami2016first`. Abstract** (OpenAlex; ACM closed; DOI 10.1145/2858036.2858494, +CHI '16, pp. 2371–2382). Qualitative lab study, 40 Facebook users, 10 folk theories of curation. "Users who were +given a probe into the algorithm's operation via an interface that incorporated 'seams,' visible hints disclosing +aspects of automation operations, could quickly develop theories. Users made plans that depended on their +theories." The authors conclude that foregrounding automated processes "may increase interface design complexity, +but it may also add usability benefits." Also held in the sister arm (facet I, abstract level). +- *Bearing:* **for**, as design practice: seams are disclosures, but disclosures shaped so that a person forms and + tests a theory, which is closer to scaffolding than to a readable record. **Narrows:** this is the researchers' + probe (FeedVis), not a platform's design; no outcome beyond theory formation and stated plans is measured. + +**Eslami et al. (2015), `eslami2015always`. Abstract** (OpenAlex/IDEALS record; ACM DOI 10.1145/2702123.2702556, +CHI '15). 40 Facebook users; "more than half of the participants (62.5%) were not aware of the News Feed algorithm +at all"; after FeedVis, satisfaction two to six months later was similar, but "algorithmic awareness led users to +more actively engage with Facebook and bolstered their overall feelings of control on the site." +- *Bearing:* **narrows**. Awareness raised feelings of control and engagement; it did not show greater competence, + and "feelings of control" is the outcome Vaccaro et al. (2018, sister arm facet H) show can be a placebo. + +**Chalmers (2003), `chalmers2003seamful`. Abstract** (IEE Eurowearable '03, pp. 11–16, DOI 10.1049/ic:20030140). +Seamful design questions "the assumption that seamless integration ... is necessarily a design requirement"; when +seams show, "users perceive and appropriate them for their own uses"; he proposes "making it a deliberate policy +to reveal and use seams." Chalmers and MacColl's 2003 UbiComp workshop paper, the usual citation, was not located. +Ehsan et al. (2024) carry the idea into explainable AI; the sister arm's card (`dxd_foundation/literature/library/ +ehsan2024seamful.md`, partial, S2-corrected) finds the evidence formative (43 practitioners, no end-user test) and +notes that "strategic revelation and concealment" leaves the choice of seams with the builder. +- *Bearing:* **narrows**. Seamful design is the design tradition nearest to scaffolding, but its record is + design-method reports, and the builder decides which seams the person may see. + +**DeVito, Gergle and Birnholtz (2017), `devito2017algorithms`. Abstract** (OpenAlex; ACM 403; DOI +10.1145/3025453.3025659, CHI '17, pp. 3163–3174). Content analysis of 102,827 #RIPTwitter tweets: resistance to +algorithmic change "largely revolves around expectation violation, with folk theories acting as frames"; more +detailed folk theories go with more specific reactions. DeVito (2021) (`devito2021adaptive`, this arm N3 and sister +facet I, abstract): because platforms change, users "must constantly re-evaluate their folk theories." +- *Bearing:* **for**, indirectly. People build the competence themselves, unaided, from outcomes and rumour — + self-taught, like the freedpeople (Q3) — and the platform's changes undo what they built. That is the + condition under which a person cannot reach competence without the operator's help. **Narrows:** these are + descriptive studies; none tests an intervention. + +**Bucher (2017), `bucher2017algorithmic`. Abstract** (DOI 10.1080/1369118x.2016.1154086; ICS 20(1), 30–44; sister +facet M found no authorised copy). Tweets and interviews with 25 ordinary users; the "algorithmic imaginary — +ways of thinking about what algorithms are, what they should be and how they function" plays "a generative role in +moulding the Facebook algorithm itself." +- *Bearing:* **narrows**. Users are not empty vessels to be scaffolded; they already act on the system through + their imaginaries, and the system learns from them. Wrong (c) must not describe the person as competence-less. + +**Hertwig and Grüne-Yanoff (2017), `hertwig2017nudging`. Full (reused; sister arm card verified and corrected, +`dxd_foundation/literature/library/hertwig2017nudging.md`).** Boosts "target competences rather than immediate +behavior" (p. 977); reversibility test: if behaviour persists after the intervention is removed, it was more likely +a boost (p. 981); boosts must be "explicit, visible, and transparent" and so respect autonomy (p. 982); "The mere +provision of information is often not enough" (p. 977); boosts can equip people for "commercially constructed +choice architectures and industry nudges" (p. 981); whether boosts beat nudges "is an empirical question" (p. 978). +- *Bearing:* **for** wrong (c) as a frame: a disclosure regime is an "educative nudge" at most (p. 977), and the + boost is the named alternative that builds competence. **Narrows:** the authors claim no superiority for boosts, + and a short-term boost includes changing the representation (natural frequencies, p. 977), which covers what the + sister arm calls re-representation. The scaffold/calculator line in Belland (Q1) and the long-term/short-term + boost line here are the same line drawn twice. + +### Q3. Who historically scaffolded the move to reading, who paid, and what came of it + +Search additions: archive.org full-text search for the UNESCO Cuba report (Lorenzetto and Neys 1965: cited +everywhere, not digitised that I could find; UNESDOC returns 403 to scripts and to WebFetch); Bhola (1984) found as +an open microfiche scan (`micro_IA41948409_0395`, downloaded to `V5/bhola1984.pdf`, OCR text extracted; printed page += PDF page + 3); Boughton and Durnan (2014), open at the University of Malta repository (`V5/boughton_cuba.pdf`); +Williams, *Self-Taught*, and Freire and Macedo (1987) as full-text-search snippets over lending scans (not +borrowed); Alvord's 1866 Bureau report as reprinted in Fleming (1905), open scan `reconstructionof01flem`. + +**Bhola (1984), `bhola1984campaigning`. Part** (Cuba chapter pp. 88–105; ch. "Why literacy?" pp. 30–31; Brazil +chapter pp. 126–136). UNESCO, Paris, *Campaigning for Literacy: Eight National Experiences of the Twentieth +Century*. The book "advocates the literacy campaign" (Preface) — a sponsor's own assessment; Cuban figures are +official, via Fagen and Kozol. +- *Cuba 1961 — who scaffolded.* Schools closed for eight months; "everyone of 13 years and older, including + teachers" went to teach peasants, lodging "in the homes of illiterates, teaching small groups of two to three + adults" (p. 92). About 100,000 instructors were trained at Varadero; the final force was 268,420 (popular + instructors 120,632; workers' brigades 13,016; Conrado Benítez student brigades 100,000; teachers 34,772) (p. 97). + No classes: tutors met learners at home, a ratio "of 1:3 and sometimes 1:2", which "gave adult illiterates greater + confidence" (p. 98). Three tests: initial, intermediate (monitoring), final (p. 102). Itinerant "Study Coaches" + went to the most isolated learners (p. 100). +- *Who paid.* The state's National Literacy Commission had a finance section (p. 95); each brigadista received a kit + and "ten pesos a month for incidental expenses" (p. 98); the labour was mobilised, not hired. +- *Outcomes.* Illiteracy of 23.6 per cent (age 10+) among 979,207 people; "made 707,212 literate during 1961, + resulting in an illiteracy rate of only 3.9 per cent" (p. 100), at "approximately" the end-of-first-grade level + (p. 100). Follow-up: *El Placer de Leer* distributed free, rural libraries, the *Seguimiento* course from February + 1962; "some 586,000 people returned to school" for sixth grade, about 65,500 graduating by 1966 (pp. 102–103). + The municipal files "merely provided numerical data on the number of persons declared literate" (p. 104). +- *Whom it served.* The primer *Venceremos*'s lessons included "The Revolution; Fidel is Our Leader" (p. 100); the + follow-up meant "recruitment through education of the erstwhile underprivileged into the service of the + revolution" (p. 102); adult-education objectives include "the Marxist-Leninist conception of the world" (p. 103); + Bhola's lesson 2: "Literacy campaigns cannot be justified in purely economic terms. Ideological justifications are + necessary" (p. 104). +- *Brazil's MOBRAL (1967–80).* Funded by government, 6.75 per cent of the Sports Lottery and a 2 per cent + corporate tax deduction; US$66 million in 1979 (p. 129). Its reading method has "the same" technical structure as + Freire's, but the generative themes "set the two methods apart" (p. 130) — the military regime kept Freire's + technique and dropped his content. External evaluations found instructor training "lasting only a few days" and + "failure, in many cases, in follow-up" (p. 134); "critics have not always been convinced of the validity of + official statistics" (p. 136). +- *General.* "literacy has not always been used exclusively for conscientization but also for mystification and for + creating and reinforcing dependencies" (pp. 30–31). +- *Bearing:* **for** wrong (c) — the record contains large, measured, sponsor-funded scaffolding of adults (one tutor + to two or three learners, in the learner's home, with tests and follow-up), so "not scaffolding" is a choice, not + a historical impossibility. **Narrows** — every well-funded scaffold in this record served its sponsor's + project (revolution in Cuba; a military regime's development plan in Brazil), which is Brandt's point; and the + measured outcome was a first-grade reading level. + +**Boughton and Durnan (2014), `boughton2014cubas`. Part** (pp. 325–330, 345–346). *Postcolonial Directions in +Education* 3(2), 325–359. Cuba's *Yo, Sí Puedo* model descends from 1961 and has "mobilised over six million +people in twenty-nine countries" (abstract, authors' report); "the enthusiasm for mass campaigns ... came up +against opposition from the World Bank, which used its increasing control over international education aid funding +in the 1980s to call a halt" (p. 327, citing Jones 1990). Cuba's 1961 campaign "pre-dated Freire's initial work in +Brazil, and was one of its inspirations" (p. 345). Authors are participant-researchers in Timor-Leste's campaign. +- *Bearing:* **for** (scaffolding at scale is a funding decision that funders have refused) and **narrows** + (written by advocates). + +**Alvord (1866), report of the inspector of Freedmen's Bureau schools, `alvord1866report`. Part** (as reprinted in +Fleming 1905, *The Reconstruction of the Seceded States, 1865–76*, NY State Education Dept syllabus, pp. 144–146; +original: report of 1 January 1866, Senate Ex. Doc. 39th Cong., 1st Sess.). "in very many places I have found +what I will call 'native schools,' often rude and very imperfect, but there they are, a group, perhaps, of all ages, +trying to learn. Some young man, some woman, or old preacher, in cellar, or shed, or corner of a negro meetinghouse, +with the alphabet in hand, or a torn spelling book, is their teacher" (Fleming p. 145). Adults "gladly avail +themselves of the evening and Sabbath schools"; in colored regiments "nearly every man can now read" where few +knew their letters, with officers and chaplains as teachers (p. 144). +- *Bearing:* **for** — the freedpeople scaffolded one another before any sponsor arrived; the state's own + inspector recorded it. This matches Washington's "a whole race trying to go to school" (`washington1901up`, p. 30) + and the community-paid teacher (p. 29). + +**Williams (2005), *Self-Taught*, `williams2005self`. Snippet** (archive.org full-text search over lending scan +`selftaughtafrica0000will`; not borrowed; no printed pages). Freedpeople "constructed school buildings, provided +churches for use as schools, paid tuition when they could afford to, and frequently operated completely +self-sustaining schools"; one school "With 150 students ... was fully self-sustaining, receiving no assistance from +the Freedmen's Bureau"; Williams quotes Morris that "Black education was a cooperative venture involving the +Freedmen's Bureau, benevolent societies, and a corps of teachers that by July 1870 numbered 3,500"; she records +"wrangling with northern white missionaries over control of schools". The arm's card `margo2005williams` (verified) +adds the 1868 federal finding that Black teachers "accounted for slightly more than half of those teaching +ex-slaves" (p. 590) and the census decline in illiteracy from 84 per cent (1870) to 26 per cent (1920) (p. 590). +- *Bearing:* **for** and **narrows**. For: the people denied literacy built their own scaffolds and paid for them, + and the sponsors (Bureau, missionaries) arrived second and fought for control. Narrows: the strongest historical + scaffold was self-organised, not supplied by the holder of the means — which cuts against making the operator + the scaffolder and toward the person's own collective action. + +**Freire and Macedo (1987), *Literacy: Reading the Word and the World*, `freire1987literacy`. Snippet** (lending scan +`literacyreadingw00frei`; not borrowed; no printed pages, except a running head placing "Literacy in Guinea-Bissau +Revisited" at about pp. 95–99). "Reading the world always precedes reading the word, and reading the word implies +continually reading the world" (two overlapping snippets, joined). On Guinea-Bissau, the +dialogue discusses the decision to "adopt the Portuguese language as the main vehicle in the literacy campaign" +(snippet; who decided is not shown); Freire: "My method did not +fail, as has been claimed. Nor did it fail in Cape Verde or Sao Tome"; "What I could not do in Guinea-Bissau is +overstep the political limitations of the moment." Freire (1970, ch. 2) on "banking" education (p. 72) is held in +facet F (read in part). +- *Bearing:* **for** wrong (b) and (c) together: a campaign conducted in the colonial written language — not the + learners' spoken Creole — is the clearest historical case of teaching people in a medium that is not theirs, and + even Freire's defenders concede the outcome there was contested. **Narrows:** the snippets give Freire's defence, + not the critics' evidence; do not state outcomes for Guinea-Bissau from this source. + +**Sunday schools and the sponsor's limit (held in facet C).** Hannah More, 1801: "My plan of instruction is +extremely simple and limited ... I allow of no writing for the poor. My object is not to make fanatics, but to +train up the lower classes in habits of industry and piety" (Roberts 1835, vol. II, p. 72; `roberts1835memoirs`, +facet C entry 14, read in part). Laqueur (1976) not opened. +- *Bearing:* **narrows** — the sponsor scaffolded reading and withheld writing: the person could receive texts and + not answer them. This is the closest historical analogue to a regime that builds the capacity to read the + operator's records while withholding the means to act on the system. + +**Johansson (1985), `johansson1985popular` (card, full, corrected).** The Swedish church taught reading through the +home, tested at confirmation, communion, marriage and migration, and held people "to the book" by "worldly compulsion +and church discipline" (p. 64); peasants resisted "costly school systems" because reading "had no obvious practical +use" (p. 63). **Narrows:** the most successful pre-industrial reading campaign was coercive scaffolding by the +sponsor, with access to marriage as the sanction. + +**Brandt (1998), `brandt1998sponsors` (card, verified).** Sponsors "enable, support, teach, model, as well as +recruit, regulate, suppress, or withhold literacy—and gain advantage by it in some way" (p. 166); their interests +and the learner's "do not have to converge" (p. 166). **Narrows** every "for" item above: a firm that scaffolds its +users' algorithmacy is a sponsor in Brandt's sense and will set the terms. + +### Q4. The capability approach: literacy as a capability, and what follows for those who hold the means + +Search additions: archive.org full-text search over the Harvard UP lending scan of Nussbaum (2011) +`creatingcapabili0000nuss` (not borrowed; no printed pages); the SEP entry (Robeyns and Byskov) held locally from +facet M4 (`berthon_ethics_local/pdfs/M4/sep_capability-approach.txt`); Sen (1999) — the only archive.org copy is a +community upload I did not use; record only. + +**Nussbaum (2011), *Creating Capabilities*, `nussbaum2011creating`. Snippet.** The list's fourth capability, +"Senses, imagination, and thought", is to be exercised in a way "informed and cultivated by an adequate education, +including, but by no means limited to, literacy and basic mathematical and scientific training" (two snippets, +joined). The distinction that matters: *internal capabilities* (trained, "fluid and dynamic" traits of the person) +versus *combined capabilities* ("internal capabilities plus the social/political/economic conditions"). "A society +might do quite well at producing internal capabilities but might cut off the avenues through which people actually +have the opportunity" to function. A just order "must secure to all citizens at least a threshold level of these ten +Central Capabilities." Following Wolff and De-Shalit, a "fertile functioning ... tends to promote other related +capabilities"; "corrosive disadvantage" is its flip side. +- *Bearing:* **for** and **narrows**. For: literacy is on the list, and the duty to secure the threshold falls on + whoever orders social conditions. Narrows sharply: scaffolding builds *internal* capability; a person who has + learned the system's ways but has no avenue to act on it (no contest, no exit, no voice) lacks the *combined* + capability. So wrong (c) cannot stand alone — it needs wrong (a)/(b)'s point about the avenues, and the paper + should say scaffolding without an avenue to act is Nussbaum's case of a society that trains people and then + cuts them off. + +**Robeyns and Byskov, "The Capability Approach" (SEP), `robeyns2025capability`. Part** (§2.3, held from facet M4). +A *conversion factor* is "the degree in which a person can transform a resource into a functioning"; personal +conversion factors include "reading skills"; social and environmental factors include "public policies" and "the +means of transportation and communication"; Sen's "capability" refers "to an opportunity made feasible, and +constrained by, both internal (personal) and external (social and environmental) conversion factors" (§2.3). +- *Bearing:* **for**, as the paper's cleanest analytic frame. A disclosure is a resource. Its value to a person + depends on a personal conversion factor (the competence to use it — what scaffolding builds) and on environmental + factors (the design of the record and the system — what the deploying firm controls). A regime of disclosure that + attends to neither has not given anyone a capability. This frame joins wrongs (a), (b) and (c) without the + paper's needing to claim that disclosure is worthless. + +**Sen (1999), *Development as Freedom*, `sen1999development`. Record** (Knopf; not opened). Cite through Robeyns +and Byskov until a copy is read. + +### Q5. The case against wrong (c) + +Search additions: Consensus (two queries: AI/algorithmic-literacy interventions; platform duty to scaffold); +OpenAlex for Fernandes et al. (2014), Kaiser et al., Willis (2008); NBER w27057 PDF (`V5/kaiser2020nber.pdf`, read +printed pp. 1–4, 15–17, 21–22; printed page = PDF page − 1); Fernandes et al. June 2013 working paper (sister arm's local copy, `dxd_foundation_local/ +pdfs/L/fernandes2013wp.pdf`, re-paginated this session); Puppart et al. (arXiv 2503.10556, `V5/puppart2025.pdf`, +read pp. 1–9). Penn Law's repository (Willis) returned 403; ScienceDirect (Leichtmann) returned 403; Scholar +Gateway returned "Access denied". + +**Fernandes, Lynch and Netemeyer (2014), `fernandes2014financial`. Part** (working paper of 2 June 2013, pp. 2, 14, +22; published abstract via OpenAlex). The published meta-analysis covers "168 papers covering 201 prior studies" +(the 2013 working paper: 155 papers, 188 studies, p. 2). "interventions to improve financial literacy explain only +0.1% of the variance in financial behaviors studied, with weaker effects in low-income samples" (published +abstract); "at delays of 20 months or greater, there is no significant effect of even 24 hours of instruction" +(WP p. 14). Their alternative is not "design instead": defaults "work best when almost all consumers have similar +needs. When consumers' needs are heterogeneous, one needs to know something to decide for oneself. Here, +'just-in-time' financial education may have promise, alone and embedded in decision support systems" (WP p. 22). +- *Bearing:* **against** wrong (c) as decontextualised training ("teach people algorithmacy" in a course), and + **for** a narrower wrong (c): support delivered at the moment of decision, inside the system — which is what + scaffolding in Belland's sense is (support given "while students engage with problems", p. 18). + +**Kaiser, Lusardi, Menkhoff and Urban (2020/2022), `kaiser2022financial`. Part** (NBER w27057, 2020; published +*J. Financial Economics* 145(2), 2022, 255–272, DOI 10.1016/j.jfineco.2021.09.022, not read). 76 randomised +experiments, more than 160,000 people. The weighted effect is "at least three times as large as the weighted +average effect documented in Fernandes et al. (2014), which examined 13 Randomized Controlled Trials" (p. 3); with +heterogeneous true effects, about 0.10 SD, "approximately 5.5 times larger" (pp. 16–17); "one of the main findings +of Fernandes et al. (2014) is not confirmed in this larger sample of RCTs" (p. 16); effects similar to +behaviour-change interventions in health and energy (p. 17). They "do not find strong evidence to support a rapid +decay ... though we do not find support for the sustainability of long-run effects either" (p. 3); no difference +for low-income people (p. 3). +- *Bearing:* **narrows the strongest objection.** The 0.1% figure, which the brief calls "very relevant", has been + re-estimated on a larger RCT base: education has a real, small effect (about a tenth of a standard deviation), + comparable to other behaviour-change interventions. The paper should not cite Fernandes's 0.1% without Kaiser + et al.'s correction. Neither study concerns algorithmic systems. + +**Willis (2008), "Against Financial-Literacy Education", `willis2008against`. Abstract** (Penn Law repository record +via OpenAlex; full text 403; *Iowa Law Review* 94, 197). Financial-literacy education is "a necessary corollary to +the disclosure model of regulation"; the belief in it "lacks empirical support" and is "implausible, given the +velocity of change in the financial marketplace ... and the disparity between educators and financial services +firms in resources"; "For some consumers, financial education appears to increase confidence without improving +ability"; "the regulation through education model blames them for their plight, shaming them and deflecting calls +for effective market regulation"; consumers "generally do not serve as their own doctors and lawyers." +- *Bearing:* **against** — the strongest form of the objection. "Scaffold the user" is the education half of the + disclosure model, not its alternative, and it carries the same responsibilization as wrong (a). Willis's + conditions — a fast-changing product, a resource gap between educator and firm — fit adaptive AI systems better + than they fit mortgages. Her remedy is regulation, not competence. + +**Puppart et al. (2025), `puppart2025brief`. Part** (arXiv 2503.10556, preprint, pp. 1–9). Randomised; high-school +students; ChatGPT advice wrong on half of eight puzzles. Over-reliance in "52.1% (SD = 28.90) of trials with +incorrect recommendations" (p. 8); the text-based literacy intervention "did not significantly reduce over-reliance" +and raised under-reliance (abstract; p. 8). Sample n = 36 (Figure 2, p. 9) — very small; preprint. +- *Bearing:* **against** brief, text-based AI literacy (which is a disclosure about the system, not a scaffold). + +**Leichtmann et al. (2023), `leichtmann2023effects`. Abstract** (CHB 139, 107539; ScienceDirect 403). N = 410 +online, 2×2: in-task XAI explanations improved performance and trust calibration; "the educational intervention, +domain-specific (i.e., mushroom) knowledge, and AI knowledge had no effect on performance." +- *Bearing:* **against** up-front training, **for** support inside the task — the same split as Fernandes's + just-in-time point and Belland's definition. + +**Chiang and Yin (2022), `chiang2022exploring`. Abstract** (IUI '22, pp. 148–161). Randomised; an interactive, +model-specific tutorial reduced over-reliance only for people already able to do the task; "low-performing +individuals' reliance on the ML model is not affected by the presence or the type of user tutorial." +- *Bearing:* **narrows** — scaffolding as delivered helped those who least needed it; a Matthew effect the paper + must name. + +**Lin (2025), `lin2025vernacular`. Abstract** (Asian J. Communication 36(3), 255–274). 316 official Douyin +"algorithm explanation" videos and articles plus 22 diary-interviews: the platform's transparency discourse +"selectively presents algorithmic information and shifts responsibility to automated systems, content creators, and +state regulation" and "legitimises the platform control through a carefully crafted illusion of transparency." +- *Bearing:* **against** trusting the operator to scaffold — the operator's own "literacy" materials served the + operator (Brandt's sponsor, in the present). + +**Ramesh, Kameswaran, Wang and Sambasivan (2022), `ramesh2022platform`. Abstract** (FAccT '22, pp. 1917–1928). +29 financially stressed users of Indian instant-loan apps "assumed responsibility for their negative experiences, +thus releasing the high-powered loan platforms from accountability obligations"; the authors call for interventions +"that enhance agency of users" and caution against "a purely technical approach." +- *Bearing:* **for** wrong (a) and a **warning** for (c): responsibility migrates to the user even without anyone + asking it to; a scaffolding programme would have to work against that drift, not with it. + +**Paternalism.** I found no source that argues capacity-building is paternalistic in the way nudging is; Hertwig and +Grüne-Yanoff hold the reverse — boosts need the person's cooperation and are "more respectful of autonomy" (p. 982, +sister card). The paternalism risk in the record is different: the sponsor chooses the content (Cuba's primer; +MOBRAL's themes; More's no-writing rule) and treats the learner as empty (Freire's "banking", 1970, p. 72; Bucher's +users already have an imaginary). Kamperman and Graff (2026, card) add that the literacy myth makes failure "the +fault of the individual" (p. 36). Note that Kamperman and Graff is a critique, not evidence that programmes fail. + +**Prior work that already frames explanation as scaffolding (novelty check).** Lund et al. (2026), "Explainable +artificial intelligence as AI literacy scaffolding", *J. Documentation*, pp. 1–14, DOI 10.1108/jd-02-2026-0088, +`lund2026explainable`, **Abstract**: proposes "contingent support calibrated to the learner's Zone of Proximal +Development, progressive fading of explanatory support as competence grows, and the reframing of explanations as +entry points for inquiry rather than disclosure endpoints." Rohlfing et al. (2021), *IEEE TCDS* 13(3), 717–728, +`rohlfing2021explanation`, **Abstract**: explanation as a co-constructed social practice whose framework "relies on +the key concepts of monitoring and scaffolding". Oeldorf-Hirsch and Neubaum (held in this arm, M2/N3) set +"balancing users' expectations of algorithmic literacy with developers' responsibility" as a research direction. +- *Bearing:* **narrows** the novelty of wrong (c): the move "explanations should scaffold, not just disclose" is + in print (Lund et al. 2026, two months before the post). What I did not find is the historical argument — that + the oral-to-literate transition shows who scaffolded and whom it served. + +### A note on Nadella's own word + +The sister arm's facet G (`dxd_foundation/literature/facets/G_nadella_lineage.md`, read in full there) records that +Nadella's December 2025 post wants AI to be "a scaffolding for human potential vs a substitute", calls it "a product +design question we need to debate", and uses "scaffolds" for orchestration of models and agents. So the word is his. +The paper must say "the insider post (P1–P21) offers no scaffold", not "Nadella ignores scaffolding"; and it can +note that his "scaffolding" there means a system that supports people, which by Belland's test (support that +"continues indefinitely" is not scaffolding, p. 18) is a calculator, not a scaffold. + +## Sorted for, narrows, against (wrong (c): "not scaffolding the person's growth to algorithmacy") + +**For.** +1. Learning science draws the line the thesis needs: telling, a job aid and permanent support are not scaffolding; + scaffolding is contingent support toward later independent performance (Belland 2017, pp. 5–6, 18–21; Collins, + Brown and Newman 1987, pp. 2–3, 7). A disclosure or readable record is "telling". +2. Cognitive apprenticeship was invented for teaching reading, because the expert's process is invisible and must + be externalised (Collins et al. 1987, pp. 2, 4). The operator is the only party who can externalise an + adaptive system's process (Vygotsky's "more capable peer", ch. 6). +3. Measured systems that scaffold through a two-way loop work: Explanatory Debugging raised mental-model scores 52% + (Kulesza et al. 2015, pp. 10–11); seams let users form and plan from theories (Eslami et al. 2016, abstract); + computer-based scaffolding averages g = 0.46 (Belland 2017, p. 27). +4. History shows scaffolding of adults at scale was possible and was a sponsor's choice: Cuba 1961, 268,420 + tutors, 1:2–3 in learners' homes, 707,212 declared literate (Bhola 1984, pp. 92, 97–98, 100). +5. The capability approach makes a disclosure a resource whose value depends on personal and environmental + conversion factors (SEP §2.3); literacy is on Nussbaum's list (2011). +6. In one comparison, in-task explanations improved performance where an up-front educational intervention did not + (Leichtmann et al. 2023, abstract); Fernandes et al. recommend just-in-time support where needs differ (WP p. 22). + +**Narrows.** +1. "Fading" is Collins et al.'s word (1989), not Wood, Bruner and Ross's (Belland 2017, p. 21); the 1976 original + was not read. +2. By the scaffolding definition, permanent design aids (including the sister arm's re-representation) are + calculators, not scaffolds; the paper must choose which remedy it means or name both. +3. Every funded scaffold in the record served its sponsor: Cuba's primer ("Fidel is Our Leader", Bhola p. 100), + MOBRAL's Freire-without-Freire (p. 130), More's no-writing rule (Roberts 1835, p. 72), Sweden's compulsion + (Johansson 1985, p. 64). Brandt (1998, p. 166) names the pattern. An operator that scaffolds its users is a + sponsor. +4. The strongest historical scaffolds were self-organised (Alvord 1866 in Fleming 1905, p. 145; Williams 2005; + Washington 1901, pp. 29–30) — which points to the person's own collective action as well as to the holder's + duty. +5. Internal capability without an avenue to act is not a capability (Nussbaum 2011): (c) needs (a) and (b). +6. Users already act through folk theories and imaginaries (Bucher 2017; DeVito et al. 2017); they are not blank. +7. Interactive tutorials helped only people already competent (Chiang and Yin 2022). +8. "Explanations as scaffolds, not disclosure endpoints" is already in print (Lund et al. 2026; Rohlfing et al. + 2021). +9. Nadella has used "scaffolding" himself (sister facet G). + +**Against.** +1. Financial-literacy education explained 0.1% of variance in behaviour and decayed to nothing by 20 months + (Fernandes et al. 2014) — but see Kaiser et al. (2020/2022): on 76 RCTs, a real effect of about 0.10 SD, the 0.1% + "not confirmed". +2. Willis (2008): education is the disclosure model's corollary, raises confidence without ability, and blames the + person; with fast-changing products and a resource gap, competence-building is implausible and regulation is the + remedy. This is the strongest objection to wrong (c), and it applies with more force to adaptive AI. +3. Brief AI-literacy interventions did not reduce over-reliance (Puppart et al. 2025, n = 36) and education did not + improve performance (Leichtmann et al. 2023). +4. Operator-led "algorithm literacy" served the operator (Lin 2025, Douyin). +5. Responsibility drifts to users without anyone pushing it (Ramesh et al. 2022); a scaffolding duty could + accelerate that drift. + +## What the paper can say (notes for the drafter, with sources) + +1. In learning science, scaffolding is support that lets a person do now what they cannot yet do alone and that is + withdrawn as they become able to do it; telling a person what to do, a job aid, and support that never ends do + not count (Belland 2017, pp. 5–6, 18–21). +2. By that definition, a disclosure, a readable record and an emergency brake held by the deployer are not + scaffolds, and the insider post offers nothing that would count as one (post P9, P15, P19–P20; Belland 2017). +3. Collins, Brown and Newman designed cognitive apprenticeship to teach reading and writing because expert reading + is an invisible process that a teacher must externalise (1987, pp. 2–4); an adaptive system's process is + invisible to everyone but its operator. +4. Where systems have been built to explain themselves and take correction back, novices' understanding rose + measurably — by 52% on a mental-model test in one lab study (Kulesza et al. 2015, pp. 10–11) — though in a setting + where the system served the user alone. +5. Cuba's 1961 campaign shows that a sponsor can scaffold adults at scale — 268,420 tutors working with two or + three learners each in their homes, 707,212 people declared literate at a first-grade level — and that the + sponsor wrote its own politics into the primer (Bhola 1984, pp. 92, 97–98, 100). +6. The most striking scaffolds of the American transition were the freedpeople's own "native schools", which a + Freedmen's Bureau inspector found already running in 1866 and which communities paid for themselves (Alvord in + Fleming 1905, p. 145; Washington 1901, pp. 29–30; Margo 2005, p. 590). +7. Literacy's sponsors have scaffolded in their own interest: Hannah More taught reading and "allow[ed] of no + writing for the poor" (Roberts 1835, p. 72), and Brandt names sponsors as agents who teach or withhold literacy + "and gain advantage by it" (1998, p. 166). +8. On the capability approach, a disclosure is a resource whose worth depends on the person's competence and on + the design of the environment (Robeyns and Byskov, SEP §2.3), and training that leaves people no avenue to act + does not give them a capability (Nussbaum 2011). +9. The evidence on teaching people to cope with complex products is contested: one meta-analysis found financial + education explained 0.1% of variance in behaviour (Fernandes et al. 2014), and a later one on 76 randomised trials + found a small but real effect (Kaiser et al. 2022). +10. Willis (2008) argues that education is the disclosure model's companion and shifts blame to the consumer; the + paper must answer her by specifying in-task, operator-provided support that fades, not courses — and by + pairing it with the avenues to act (Fernandes et al. 2014 WP p. 22; Leichtmann et al. 2023). + +## Gap statement + +I found no work that uses the history of the oral-to-literate transition (Cuba 1961, freedpeople's schools, +Sunday schools, the Swedish church) to argue that firms deploying adaptive AI owe users scaffolding toward +competence, in Consensus (two queries), OpenAlex title searches, the arm's facets (C, F, M2, M4, M5, N3, Z) and the +sister arm's facets (G, I, L, M). The closest works argue the design move without the history (Lund et al. 2026; +Rohlfing et al. 2021) or the history without the design move (Kamperman and Graff 2026; Brandt 1998). This is a +search of those sources only; Scholar Gateway was unavailable. + +## Read-depth tally + +| Depth | Works | +| --- | --- | +| Full | Kulesza et al. 2015 (accepted MS); Hertwig and Grüne-Yanoff 2017 (sister card); Johansson 1985, Brandt 1998, Margo 2005, Kamperman and Graff 2026, Washington 1901 (arm cards, as carded) | +| Part | Belland 2017; Collins, Brown and Newman 1987; Bhola 1984; Boughton and Durnan 2014; Alvord 1866 (Fleming reprint); Fernandes et al. 2014 (WP); Kaiser et al. 2020; Puppart et al. 2025; Robeyns and Byskov (SEP); Freire 1970 (facet F) | +| Snippet | Vygotsky 1978; Williams 2005; Freire and Macedo 1987; Nussbaum 2011 | +| Abstract | Eslami et al. 2015, 2016; DeVito et al. 2017; Bucher 2017; Chalmers 2003; Willis 2008; Leichtmann et al. 2023; Chiang and Yin 2022; Lin 2025; Ramesh et al. 2022; Lund et al. 2026; Rohlfing et al. 2021 | +| Secondary | Wood, Bruner and Ross 1976 (definition via ERIC documents and Belland) | +| Record | Sen 1999; Collins, Brown and Newman 1989 chapter; Lorenzetto and Neys 1965; Laqueur 1976 | + +**For the author to supply:** Wood, Bruner and Ross (1976) PDF (p. 90 definition; the page listing the six functions); Vygotsky +(1978) ch. 6 printed page for the ZPD definition (commonly p. 86); Sen (1999) on literacy; Willis (2008) full text; +Lorenzetto and Neys (1965) UNESCO report. + +## Cards written (all `verified: pending`) + +`belland2017instructional` (partial-text), `collins1987cognitive` (partial-text), `vygotsky1978mind` (partial-text, +snippets), `kulesza2015principles` (full-text, accepted MS), `bhola1984campaigning` (partial-text), +`alvord1866report` (partial-text, Fleming reprint), `fernandes2014financial` (partial-text, WP), `kaiser2022financial` +(partial-text, NBER WP), `willis2008against` (abstract), `nussbaum2011creating` (partial-text, snippets), +`lund2026explainable` (abstract). Existing cards cited, not rewritten: `johansson1985popular`, `brandt1998sponsors`, +`washington1901up`, `margo2005williams`, `kamperman2026ai`; sister-arm card `hertwig2017nudging`. diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V6_prior_work_and_objections.md b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V6_prior_work_and_objections.md new file mode 100644 index 00000000..cd958637 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_facet_V6_prior_work_and_objections.md @@ -0,0 +1,227 @@ +# Facet V6: prior work and objections (research round 7, 2026-10-14) + +Facet V6, 2026-10-14. Agent-drafted. The author has not reviewed it, and none of the text is the author's. + +## What this facet covers + +The brief (`../../v4/RESEARCH_BRIEF.md`) and the author's v4 instruction (`../../v4/SOURCE_v4.md`) set the thesis: Nadella's post of 10 October 2026 treats disclosure (specific disclosures, readable records, design for legibility) as sufficient ethical design, and so repeats three mistakes of the oral-to-written transition: (a) placing the burden on the person who must read; (b) keeping that person in a literacy-based experience when the medium has changed; (c) not scaffolding the person's growth to the new competence (algorithmacy). This facet asks five things. + +1. **Gap search.** Has anyone argued (a) that disclosure or transparency is a literacy-based intervention unsuited to algorithmic mediation, (b) that AI ethics repeats the ethical failures of the oral-to-written transition, or (c) that firms owe users scaffolding toward a new competence? +2. **"AI as a return to orality."** Ong's secondary orality applied to chatbots and LLMs, and 2023–2026 claims that conversational AI is oral, not literate. This is the counterargument the paper must meet. +3. **The BEQ 2023 special issue on organized immaturity**, Martin (2019) in JBE, and Berthon's own informing/guiding/restricting. +4. **Algorithmic literacy, algorithmic competence, algorithmacy**: definitions and where algorithmacy sits. +5. **Objections**: paternalism; disclosure respects autonomy; scaffolding is itself responsibilization; the post addresses enterprises, not consumers; the history is analogy only. + +Earlier facets hold much of the ground and I do not redo it. N5 (`N5_prior_work_gap.md`) ran the business-ethics journal search and read Scherer et al. (2023), Martin (2019, abstract), Lindblom, Qadri et al.; Z (`Z_gap_search.md`) ran the first gap search and the AI-literacy rules (AI Act Art. 4 and the 2026 Omnibus); M4 and N4 read Giesler and Veresiu on responsibilization; L1 read Berthon in full. The sister arm's `Z_gap_search.md` resolved the 1995 use of "algorithmacy" (Oberschelp). Where I reuse a finding I name the file. New reading in this facet: Ong's secondary-orality passages (pp. 2, 11, 132–134), Heyd (2021), Arató and Balázs (2024), Marz (2026), the full Scherer et al. introduction on paternalism and on business's role, Martin (2019) in full, and the items listed under each question below. + +## Method notes + +- **Databases.** Crossref REST (`query.bibliographic`, title and metadata only); OpenAlex `search` (title, abstract, indexed full text); Consensus (semantic search over Semantic Scholar/OpenAlex records); web search (standard and extended). Scholar Gateway not used (it returned `ACCESS_DENIED` in Z and N5). Google Scholar, Web of Science, Scopus, PhilPapers, JSTOR were not searched; every gap statement is bounded by that. +- **Downloads** are in `/Users/ludwitt/iit-playground/berthon_ethics_local/pdfs/V6/`, each with a `pdftotext` file beside it. Scripts in my own scratch subfolder. Nothing from a download was executed. +- **Locally reused full texts**: `pdfs/N5_prior_work_gap/scherer2023guest.txt` (BEQ 33:3, printed pages); `pdfs/N4_responsibility_intermediaries/martin2019_nsfpar.pdf` (16 pp., the published JBE layout; PDF p. n = printed p. 834+n). +- **Read-depth key.** *Full*: whole text. *Part*: opened, sections named. *Abstract*: database abstract only. *Record*: database record verified, no text. +- **Locators.** Ong is the archive.org djvu text of a Routledge printing with New Accents pagination (card `algorithmacy_design_ethics/literature/library/ong1982.md`; S2-verified there). Page numbers come from the running heads in that text. + +## Q2. "AI as a return to orality": the counterargument + +### Search log (Q2) + +| # | Source | Query | Hits | Notes | +| --- | --- | --- | --- | --- | +| 1 | Consensus | secondary orality chatbots large language models conversational AI Ong | 20 | Chatbot and language-learning studies only; none uses Ong. | +| 2 | Web (extended) | `"secondary orality" ChatGPT OR "large language models" OR chatbots Ong article` | 9 | Clinical and EFL chatbot papers; nothing applying Ong to LLMs. | +| 3 | Crossref | `secondary orality artificial intelligence`; `secondary orality chatbot`; `orality large language models`; `return to orality generative AI`; `Ong orality literacy ChatGPT` | first 10 each | Colon 2026 (SSRN, "From Orality to Artificial Intelligence: How Communication Modalities Redistribute Cognitive Burden"; title only); Leslie 2025 (HDSR); Cánovas 2025 ("Oral Poetics and Multimodal Language Models", in *Rethinking Orality III*); Lee and Lui 2024 (podcasting); Holly 1995; Nyíri 1996. | +| 4 | OpenAlex | `"secondary orality" artificial intelligence` (231); `"secondary orality" chatbot` (19); `"orality" "large language model"`; `"orality" ChatGPT` | as shown | Heyd 2021 ("Tertiary Orality?"), Arató and Balázs 2024, Logan 2024 ("Eleven Kinds of Orality"), Martínez-Arboleda 2025. | +| 5 | Web (extended) | `"orality" generative AI chatbots Walter Ong "conversational" ... "new orality" OR "digital orality" OR "tertiary orality"` | 10 | Ryu 2023 (HCII, "tertiary digital orality" in VR); blog and Substack essays (Clark 2024; Bashir 2026, reporting Jasmine Sun's "The post-literate society"). | +| 6 | Web | `James Marriott "post-literate society" ... Ong` | 10 | Marz 2026 (*The Drift*), which names the "return to orality" writers and answers them from Ong. | +| 7 | OpenAlex | `"secondary literacy" Ong` | 44 | Pugen 2020 (*New Explorations*); not opened. | +| 8 | OpenAlex | `Ong "Hermeneutic Forever"` | 17 | Ong 1995, *Oral Tradition* 10(1): 3–26. The journal PDF did not download (HTTP 000). Not read. | + +### Findings (Q2) + +**Ong (1982), `ong1982orality` (existing key; repo card in `submissions/algorithmacy_design_ethics/literature/library/ong1982.md`). Part** (the secondary-orality passages; archive.org djvu text of a Routledge printing with New Accents pagination, saved at `pdfs/V6/ong_djvu.txt`). Ong defines secondary orality three times, and each time ties it to writing: +- p. 2 (this copy; Heyd cites p. 3): "The electronic age is also an age of 'secondary orality', the orality of telephones, radio, and television, which depends on writing and print for its existence." +- pp. 10–11: primary orality is "'primary' by contrast with the 'secondary orality' of present-day high-technology culture, in which a new orality is sustained by telephone, radio, television, and other electronic devices that depend for their existence and functioning on writing and print." +- p. 133: "This new orality has striking resemblances to the old in its participatory mystique, its fostering of a communal sense, its concentration on the present moment, and even its use of formulas ... But it is essentially a more deliberate and self-conscious orality, based permanently on the use of writing and print, which are essential for the manufacture and operation of the equipment and for its use as well." +- p. 133, on the computer: "the sequential processing and spatializing of the word, initiated by writing and raised to a new order of intensity by print, is further intensified by the computer, which maximizes commitment of the word to space and to (electronic) local motion and optimizes analytic sequentiality by making it virtually instantaneous." +- pp. 133–134: "Secondary orality is both remarkably like and remarkably unlike primary orality"; "We plan our happenings carefully to be sure that they are thoroughly spontaneous." +*Bearing:* **narrows the counterargument, and so supports the thesis.** The counterargument says chat is oral, so a literacy frame is the wrong frame for the critique. Ong's own concept answers it: the orality of electronic media rides on writing, and the computer intensifies the commitment of the word to space. A chat interface can be conversational on the surface while its operation (logs, prompts, system instructions, evidence trails) stays textual, which is the v4 line. The paper must not borrow Ong's cognitive "great divide" (contested; see the repo card's Caution: Street, Finnegan, Halverson, Scribner and Cole). + +**Heyd (2021), `heyd2021tertiary`. Full** (*Anglistik* 32(2): 131–147; Winter OA PDF, `pdfs/V6/heyd2021tertiary.pdf`). The nearest peer-reviewed work that extends Ong to talking machines. "Through technologies such as speech-to-text and text-to-speech recognition, and through the rise of digital assistants and artefacts, we are increasingly talking not just through, but with machines" (p. 131). She proposes "tertiary orality" for spoken CMC in which "boundaries between orality/scripturality and human/nonhuman agency become contested or irrelevant" (p. 139), and treats digital assistants such as Alexa and Siri as the clearest case: "users speak to them, and the devices speak back" (p. 138). On algorithms she adds that YouTube routines are "oral discourse created by humans, but optimized and structured toward being read and recommended by a machine" (p. 142). Footnote 5: "The term tertiary orality has been suggested before, but never fully theorized" (p. 136). LLM chat post-dates the article (2021); it does not discuss ethics, governance or disclosure. +She also records the field consensus that typed CMC is "conceptual orality in a medium-based scripturality setting" in Koch and Oesterreicher's terms: studies "all echo Murray's initial observation of written text displaying certain spoken-like features" (p. 133), and she argues that genuinely spoken (medium-based) digital orality is the under-theorized case (pp. 133–134). +*Bearing:* **against, in its strongest scholarly form.** It is the best citation for the claim that human-machine interaction is becoming oral. It also hands the paper its distinction: typed chat with an LLM is conceptually oral but medially written (p. 133), and only voice agents are oral in the medium. It also narrows the counterargument from inside: even Heyd's oral case is "structured toward being read ... by a machine" (p. 142), so the machine side stays a reader of text. + +**Arató and Balázs (2024), `arato2024communicative`. Full** (*Magyar Nyelvőr* 148(5): 601–617, English; OA PDF). They classify generated text as writing, not speech: "Computer-generated text is a new form of literacy, tertiary literacy. In 2023, chatGPT ... bursts into the public consciousness" (p. 607), and set out a stratification in which "in the 21st century, today, artificial intelligence is leading to the emergence of tertiary orality and tertiary literacy" (p. 609). They discuss AI in the Hungarian judiciary (section 6) but make no argument about disclosure. +*Bearing:* **narrows the counterargument.** A peer-reviewed linguistics paper places ChatGPT output under literacy, with voice interfaces as a separate "tertiary verbality". + +**Marz (2026), `marz2026return`. Full** (*The Drift*, Issue 17, 20 April 2026; essay, not refereed; HTML saved at `pdfs/V6/marz2026return.html`). The clearest published statement of the "return to orality" claim and a reply to it. Marz reports Derek Thompson's "handoff from a culture of literacy to a culture of orality", Joe Weisenthal's warning that "the return to orality" will "fundamentally rewire the logic engine of the human brain", and James Marriott's newsletter. Her reply: "to use Ong to claim that we risk a 'return' to 'pre-literate' orality is to ignore the main thrust of his work. Ong took pains to emphasize that literacy and orality have been inextricably linked since writing began." She reports Ong's 1996 interview: online writing "comes not in oral-aural format but through the eye, and thus is not directly oral at all", and his proposed term "secondary literacy". On LLMs: they "not only commodify text but entail a new type of writing that comes in the form of the prompt. You could call LLMs oral, because the chatbots simulate conversations with us. But such a claim, like all insistence that digital text is actually something else, would be, to borrow Brandt's phrase, 'a holdover from the social arrangements of a disintegrating print culture.'" Her close: literacy "hasn't been killed but taken hostage." The 1996 Ong interview is quoted second-hand; I did not find it. +*Bearing:* **narrows the counterargument; supports (b).** Marz names the popular form of the objection (Thompson, Weisenthal, Marriott) and answers it with Ong and with Brandt, whom the paper already cites (`brandt1998sponsors`). Her "the prompt" as "a new type of writing" is the sharpest available sentence for the paper's reply. It is an essay; cite it as a statement of the debate, not as evidence. + +**Popular and grey items (not cited; logged).** Bashir (2026, Substack, 3 Feb) reports that Jasmine Sun's essay "The post-literate society" "uses Ong to argue that chat interfaces succeed because they leverage 'thousands of years of honed evolutionary instinct'" (Bashir's words; Sun's essay not opened). Clark (2024, blog) says AI combines text, speech and image and so blurs Ong's division. Ryu (2023, HCII) uses "tertiary digital orality" for VR avatars. None is refereed scholarship on LLMs. + +**What I did not find.** I found no peer-reviewed work, 2023–2026, that argues at length that LLM chat is oral rather than literate, in Crossref, OpenAlex, Consensus and web search (rows 1–7). The claim lives in essays and newsletters (Thompson, Weisenthal, Marriott, Sun, as reported by Marz and Bashir). The scholarly items either extend Ong to voice assistants (Heyd 2021) or class generated text as a new literacy (Arató and Balázs 2024). Colon (2026, SSRN, title only) and Cánovas (2025, chapter, title only) were not opened and could change this. + +**Answer to Q2, for the drafter.** The counterargument has a real basis: Heyd shows talking-with-machines is a recognised linguistic trend, and voice agents are oral in the medium sense. The paper can meet it in three moves, each sourced: (1) Ong's secondary orality is by definition "based permanently on the use of writing and print" (p. 133), so even an oral interface leaves the governance layer textual; (2) the post's safeguards are texts (P9 chain-of-thought, P15 "human readable evidence", P20 disclosure), so whatever the interface, the remedy asks people to read; (3) Marz (2026) and Arató and Balázs (2024) class the prompt and generated text as writing. The concession the paper owes: if agents increasingly speak, a literacy-based safeguard is doubly misfit, which strengthens wrong (b) rather than refuting it. + +## Q1. Gap search + +### Search log (Q1) + +Facets N5 (25 rows, 14 Oct) and Z (62 rows, 13 Oct) logged the print-and-literacy-transition comparisons, the business-ethics journals and the AI-literacy rules. This log adds the three specific claims of v4. + +| # | Source | Query | Hits | Notes | +| --- | --- | --- | --- | --- | +| 9 | Consensus | transparency disclosure assumes literacy; reading-based interventions unsuited to algorithmic systems | 20 | Ananny and Crawford 2018; Moon et al. 2025; Ostinelli et al. 2024; Stamboliev 2023; Springer and Whittaker 2019; Eslami 2021. None calls disclosure a literacy-based intervention. | +| 10 | Consensus | corporate duty to educate users; algorithmic literacy; responsibility of platforms | 20 | **Docherty and Valderrama Barragán 2025 and 2026** (Meta's corporate curricula as responsibilization); Oeldorf-Hirsch and Neubaum; Gagrčin et al. 2024; DeVito 2021; Nichols and LeBlanc 2021. | +| 11 | OpenAlex | `"transparency" literacy assumption algorithmic accountability reading burden`; `disclosure presupposes literacy algorithmic` | 2,955; 2,079 | Nothing on point in the first 8 of each. | +| 12 | OpenAlex | `oral to written transition ethics artificial intelligence` | 15,785 | Arató and Balázs 2025 (a follow-up, not opened); nothing else on point. | +| 13 | OpenAlex | `Clanchy "memory to written record" artificial intelligence`; `"from memory to written record" algorithm`; `Clanchy written record trust artificial intelligence` | 41; 45; 58 | Archival-science and records-management items; Ong 1995; Vee 2013. No AI-ethics use of Clanchy. | +| 14 | OpenAlex | `duty to educate users platform algorithmic literacy corporate responsibility`; `corporate digital responsibility Lobschat` | 1,763; 490 | Lobschat et al. 2021 (JBR; ScienceDirect PDF blocked, not read); Wannhoff and Westner 2026. | +| 15 | OpenAlex | `scaffolding users competence AI design obligation` | 1,688 | **Rohlfing et al. 2021** (explanation as co-constructed, scaffolded social practice); Laux 2023. | +| 16 | OpenAlex | `Phaedrus writing artificial intelligence ChatGPT Socrates`; `"Phaedrus" "large language models"` | 88; 205 | **Kitzinger 2024**; **Calderón and Herrera 2025**; Ess 2026; Richards 2023; Coeckelbergh and Gunkel 2023; Karpouzis 2024. The Phaedrus-and-AI genre. | +| 17 | OpenAlex | `transition from oral to written law lessons algorithmic governance` | — | HTTP 429 (rate limit). Not retried. | +| 18 | Consensus | defence of mandated disclosure and transparency as respecting autonomy | 20 | **Bar-Gill 2016**; Ben-Shahar and Schneider (card exists); Sunstein 2012 (Storrs Lectures); Steffel et al. 2016; Cuypers et al. 2026; Nguyen 2021. | +| 19 | Consensus | paternalism objection to digital/AI literacy mandates and platform design | 20 | **Oleart and Flores Moleón 2026** (EU literacy policy shifts responsibility to citizens); Passlack et al. 2025; Pangrazio and Sefton-Green 2021. No paper on the paternalism of scaffolding. | +| 20 | Consensus | right to explanation grounded in autonomy and informed self-advocacy | 20 | **Vredenburgh 2022**; **Grote and Paulo 2025**; Cordasco and Véliz 2026 (in N5); Kim and Routledge (in arm); Edwards and Veale 2018; Karlan and Kugelberg 2025. | + +Blocked and not worked around: Springer (Grote and Paulo; Coeckelbergh and Gunkel), Wiley (Ostinelli et al., Cloudflare), MDPI (Oleart and Flores Moleón, "Access Denied"), Nomos/DNB resolver (Bar-Gill, bot challenge), Brill (Ess), ScienceDirect (Lobschat et al.), *Oral Tradition* (Ong 1995, no response). + +### Findings (Q1) + +**(a) Disclosure as a literacy-based intervention unsuited to algorithmic mediation.** + +- **Rohlfing et al. (2021), `rohlfing2021explanation`. Part** (IEEE *TCDS* 13(3): 717–728, CC BY; pp. 717–718, 723, 726 read). The XAI literature they criticise assumes "that 2) the mere availability of information will suffice to 3) warrant an understanding that 4) enables the explainee to act further. Clearly, the explainee is 'conceptualized as an independent actor, who makes autonomous decisions on the basis of information made available to them through transparency'" (p. 718, quoting Felzmann et al. 2019). Their first named limitation is the "passive explainee" (p. 717). Their alternative: explanation as a "social practice in which explainer and explainee co-construct understanding", built on "monitoring and scaffolding" (p. 717), and "Explaining Is Scaffolding" (§IV-B, p. 723): "The goal in scaffolding is to allow a less competent partner to participate in an interaction and to contribute to a task ... an important feature of scaffolds is that they are temporary." They warn: "We must be careful, however, to not add the empowerment of persons to a number of already existing burdens [86] in legal or ethical requirements" (p. 726). + *Bearing:* **already says part of (a) and (c), and narrows the novelty claim.** The XAI literature has named the information-suffices assumption and proposed scaffolded, interactive explanation in its place. What Rohlfing et al. do not say: that the assumption is a *literacy* assumption, that it repeats the oral-to-written transition, or that the firm owes the scaffolding as a duty. Their p. 726 burden warning is also the responsibilization objection (Q5) stated from inside the scaffolding camp. +- **Ananny and Crawford (2018), `ananny2018seeing`** (existing key, Abstract): "Being able to see a system is sometimes equated with being able to know how it works and govern it"; they list ten limits of the transparency ideal (abstract). The canonical critique; it does not frame transparency as a literacy demand. +- **Moon et al. (2025), `moon2025effects`. Abstract.** In a 2×2×2 experiment (n = 240), "When neither feature was present, algorithmic literacy had no significant impact. However, when at least one feature was present, literacy significantly influenced the dependent variables", so that "literacy creates a new dimension of the digital divide, where transparency benefits are unequally experienced" (abstract). *Bearing:* **for (a)**, with evidence: the benefit of a disclosure depends on the reader's literacy. Single experiment, perceptions only (perceived transparency, legitimacy, satisfaction). +- **Ostinelli, Bonezzi and Lisjak (2024), `ostinelli2024unintended`. Abstract.** "Merely believing that one can access an explanation of how an algorithm works can foster an illusory sense of understanding the algorithm, even when people do not actually access and read the explanation", yielding "unfounded reliance on algorithmic determinations" (abstract). *Bearing:* **for (a)**: a disclosure that is not read still changes behaviour, in the discloser's favour. +- **Grote and Paulo (2025), `grote2025minimalist`. Abstract** (Springer blocked). Against Vredenburgh, explanations for informed self-advocacy "are likely to be intellectually over-demanding since their correct interpretation will require a combination of domain-knowledge and statistical proficiency that cannot be presumed for laypersons", and "shifting the burden to detect inadequacies to laypeople makes scrutinizing explanations very costly for them"; they would replace the goal with "enabling public trust" (abstract). *Bearing:* **for (a) — wrong (a) stated in moral philosophy, 2025**: the burden argument exists. They do not call it a literacy burden or invoke history. +- **Green (2022), `green2022flaws`** (existing key; M2 read in part; local PDF p. 1). Human-oversight policies "rest on an uninterrogated assumption: that people are able to effectively oversee algorithmic decision-making" and "provide a false sense of security" (abstract, p. 1). *Bearing:* **for (a) on the staff side.** The post's "us" is the enterprise; its "human readable evidence" (P15) asks staff to read and oversee. Green's survey of 41 policies is the evidence that such reading fails. Scope: government algorithms. +- **Dembeck (2023), `dembeck2023flow`** (dxd card, Full, German). Current fake-news and critical-media-literacy education take the book-bound "Beitrag" as its starting point, and "es könnte aber sein, dass das an der Sache völlig vorbeigeht" (p. 255). *Bearing:* **for (b)**: the nearest printed statement that a book-bound literacy frame may not fit the digital. Hedged and short; no algorithmic third party. + +**(b) AI ethics repeats the ethical failures of the oral-to-written transition.** + +- **The Phaedrus-and-AI genre.** Calderón and Herrera (2025, `calderon2025and`, Part: abstract and §§3–4 headings, pp. 1–5 skimmed) use "Plato's myth of Theuth and Thamus" to ask whether chatbots can improve science, and note that "literature has found analogical similarities between some issues involved in Chatbots and that of Plato criticisms of writing, contained in the Phaedrus. However, the research area has been neglected" (abstract, p. 1). Their remedies are authorship rules and "Advertising transparency" with "formal responsibility for all contributions made using chatbots" (PDF p. 8; the call for "advertising transparency, accountability or responsibility" is at p. 6), i.e. disclosure. Kitzinger (2024, `kitzinger2024openai`, Abstract) argues the opposite use: "Plato's dialogue Phaedrus is often used as a cautionary tale about the futility of distrusting new communicative technologies. But in the context of LLMs ... the Phaedrus analogy is substantially misplaced", and "risks justifying the swift adoption of a commercial tool" (abstract). Ess (2026) and Richards (2023) are further entries (Records). *Bearing:* **narrows (b).** An oral-to-written comparison applied to AI ethics exists, but it is Plato's epistemic critique of writing (memory, wisdom, authorship), used either to warn about or to excuse the new tool. None I found treats the transition's *ethical failures toward people* (who bore the burden of the new medium, who was left out of the record, who was taught) as the template. +- **The literacy-transition comparisons** already logged: Kamperman and Graff 2026 (overselling; L2), Posada et al. 2021 (print and power), Qadri et al. 2026 (analogy method), Hänska 2021 (unread). See N5 §2. +- **Gap statement (b).** I found no work that argues that AI-ethics remedies repeat the ethical failures of the transition from oral to written proof (Clanchy's shift from memory to written record, the colonial treaty record) in Crossref, OpenAlex (until rate-limited), Consensus and web search (rows 9–20 here; N5 rows 5–11, 22–24; Z rows 7–36). The Phaedrus-and-AI genre uses the same transition for epistemic, not distributive, claims. + +**(c) Firms owe users scaffolding toward a new competence.** + +- **Scherer, Neesham, Schoeneborn and Scholz (2023), `scherer2023guest`. Part** (pp. 409–433 of the local OA PDF; N5 read pp. 409–416, 424–427, 432; I add pp. 421–423 and 428–431). Under "The Role of Business Organizations": "business organizations also have a responsibility to contribute to organizing individual and collective maturity" (p. 428), including "2) providing social rights by informing and training citizens in the appropriate use of social media and the pitfalls of manipulative algorithms; and 3) operating as a channel for political rights, whereby individuals are provided with opportunities and training in argumentation processes and collective will formation" (p. 429). *Bearing:* **already says (c), in BEQ.** The claim that firms owe people training toward a new competence is in the guest editors' introduction of the field's main special issue on AI and immaturity. The v4 paper cannot claim (c) as new; it can claim the historical argument for it and the distinction between training as CSR and scaffolding as a design duty. +- **Vredenburgh (2022), `vredenburgh2022right`. Part** (LSE accepted manuscript; AAM pp. 5, 18–22). The right to explanation is grounded in "informed self-advocacy", "a cluster of abilities to represent one's interests and values to decision-makers and to further those interests and values within an institution" (AAM p. 5). Explanations alone are not enough: "It is neither realistic nor desirable to expect individuals to use explanations to engage in informed self-advocacy within all the institutions that affect them, absent support ... Individuals require epistemic support in order to engage in tolerably costly informed self-advocacy" (AAM p. 18), so "Rights holders should be provided with access to free experts and advocates" (AAM p. 19), and "free experts could be AI systems designed to generate explanations in response to user queries" (AAM p. 21). Conclusion: "states and companies may need to invest more resources in providing individuals with free experts – human or algorithmic" (AAM p. 22). *Bearing:* **narrows (c) and answers part of (a).** A leading defender of explanation already says explanation is insufficient without supported interpretation, and puts the cost on decision-makers. Her remedy is an intermediary expert, not the person's own competence; the paper's "scaffolding growth" differs there. +- **Helberger, Pierson and Poell (2018), `helberger2018governing`** (existing key; N4, Part). "Platforms have an obligation to create the conditions that allow individual users to comply with their responsibilities" (p. 3), with Fahlquist's ordering principle (p. 3). *Bearing:* **narrows (c)**; an earlier statement of a platform duty to enable users. +- **The EU AI Act, Art. 4** (Z, R1–R2). Providers and deployers must take measures for the AI literacy of "their staff and other persons dealing with the operation and use of AI systems on their behalf"; the 2026 Omnibus weakened this to "support the development of" literacy with no guaranteed level (Regulation 2026/1744, Art. 1(5)). *Bearing:* **narrows (c) and supports the enterprise reading.** A legal duty to scaffold literacy exists, but it runs to the firm's own staff, not to the people the systems act on, and it was softened in July 2026. +- **Oeldorf-Hirsch and Neubaum (2023/2025), `oeldorf2023what`** (preprint, Part, N3/M2). Their first research direction is "balancing users' expectations of algorithmic literacy with developers' responsibility for algorithmic transparency" (preprint p. 2), and "the current reality is that individuals must endure most of effort in improving literacy" (p. 23). *Bearing:* **for (a) and (c)**: the literacy field names the allocation as unresolved. +- **Gap statement (c).** I found no work that grounds a firm's duty to scaffold users' growth toward a new competence in the record of a past literacy transition, in the sources above. The duty itself is stated (Scherer et al. 2023, p. 429; Vredenburgh 2022, AAM p. 22; Helberger et al. 2018, p. 3; AI Act Art. 4 for staff). + +**Gap statement (a).** I found no work that names disclosure or transparency as a *literacy-based* intervention, i.e. one that presupposes a reader, and argues it is misfit to algorithmic mediation because the medium is not text, in Crossref, OpenAlex, Consensus and web search (rows 9–20). The pieces exist separately: the information-suffices assumption (Rohlfing et al. 2021, p. 718), the burden on laypeople (Grote and Paulo 2025), the literacy dependence of transparency's benefit (Moon et al. 2025), and the book-bound frame of literacy education (Dembeck 2023). + +## Q3. The BEQ special issue, Martin (2019) and Berthon + +**Scherer et al. (2023): what remedies, and where v4 stands.** Read across pp. 421–432 (see also N5). The guest editors define the harm as "the reduction of individual capacities for public use of reason constrained by sociotechnological systems" (p. 409, via N5). Their remedies come in two families. *Disorganizing immaturity* restores negative freedom: Lanier's disconnection, Harcourt's "combination of education, awareness building, self-help, and technical devices" for avoiding surveillance, and "digital disobedience" (pp. 425–426). *Organizing maturity* strengthens positive freedom through four actors: the state (education first, then regulation "using informed consent provisions, the right to be forgotten and effective appeal processes", data property rights, watchdog NGOs, public forums; p. 427), the media (p. 427–428), social movements that alert, inform and educate (p. 428), and business, which should provide "social rights by informing and training citizens" (p. 429). On paternalism, they write that organized immaturity "can emerge simply as an effect of more or less subtle versions of paternalism", which "remains, in a political sense, a form of interference with one's liberty to make one's own decisions" (p. 422), and they name "technological paternalism specific to data capitalism" (p. 422). +*Where v4 stands.* (1) The special issue is Kantian: education toward maturity is the remedy, so it **supports (c)** and **narrows** its novelty. (2) It does not treat disclosure as the remedy: Schade's article, as the editors summarise it, shows GDPR transparency leading "paradoxically, to intransparencies and an infantilization of users (e.g., when one has to click through myriad detailed GDPR settings before being able to access a web page)" (p. 430). That **supports (a)** from inside BEQ. (3) The issue warns that paternalism produces immaturity (p. 422), which is the strongest in-venue form of the paternalism objection to scaffolding (Q5). (4) Ruehle (2023, BEQ 33(3): 502–531) argues digital workplace nudging, "in its paternalistic impetus, runs the risk of fostering tendencies of organized immaturity" (editors' summary, p. 430). The v4 paper should place itself as accepting the special issue's diagnosis and its call for business to train, while adding what the issue lacks: a historical argument that information-first remedies fail people who do not command the medium, and a distinction between training as CSR (Docherty, Q5) and scaffolding built into the system. + +**Martin (2019), `martin2019ethical`. Part** (local published layout, `martin2019_nsfpar.pdf`; printed pp. 835–836 and 844–846 read; N4 read the same sections). The argument: algorithms are value-laden and "influence the delegation of roles and responsibilities within these decisions", so "firms developing algorithms are accountable for designing how large a role individual will be permitted to take in the subsequent algorithmic decision" (abstract, p. 835). She rejects the excuse that opacity dissolves responsibility: "'It's complicated' or 'I do not know how it works' turns out to be an unsatisfying response" (p. 836), and "Creating inscrutable algorithms precludes users from taking responsibility for the ethical implications ... The design of the algorithm not only scripts what users can do but also the reasonable expectations of users to take responsibility" (p. 844). The ground: the firm "is knowledgeable as to the design decisions and is in a unique position to inscribe the algorithm with the value-laden biases as well as roles and responsibilities" (p. 844). Two passages cut both ways for v4. First, "if the role of the algorithm in a decision is minimized, by providing tools to allow users to revisit how the algorithm works, the firm would have more of a traditional handoff of a product with associated (minimal) responsibility" (p. 845): in Martin's model, giving users tools and understanding *moves accountability to the users*. Second, on transparency: "full transparency may be neither feasible nor desirable", it lets people "'game' the system", and "calls for transparency in algorithmic decision making may need to be targeted for a specific purpose or type of decision" (p. 846). Her "users" are the deploying decision-makers (judges using COMPAS, p. 844), which matches the post's enterprise reader. +*Bearing:* **supports** the design-duty premise (the designer allocates roles; the firm cannot hide behind opacity), **narrows** (c) (scaffolding, on Martin's own model, transfers accountability to the scaffolded user, which is the responsibilization objection with a JBE pedigree), and **supports** the enterprise reading of "the user". + +**Berthon's own frame** (`berthon2019addictive`, card exists; L1 full). Informing means "educate and disclose" (pp. 452, 457): Berthon, Pitt and Campbell put education and disclosure in one bin. On dashboards and labels, "empirical research suggests that information is necessary, but not sufficient to change behavior (e.g., Worsley 2002)" (p. 457). They treat informing as the first and least controversial rung (p. 457), restricting as "The most extreme public policy intervention" (p. 459), and note that "Consumers may prefer more user control rather than blanket mandates" (p. 462, via L1). *Bearing:* **for (a)** in the reader's own words, and a caution: Berthon's ladder places education *with* disclosure, so the v4 claim that scaffolding is categorically different from disclosure departs from his framework and should say so. His "more user control rather than blanket mandates" is the paternalism objection in his own paper. + +## Q4. Algorithmic literacy, algorithmic competence, algorithmacy + +- **Algorithmic literacy, two working definitions** (Oeldorf-Hirsch and Neubaum, preprint p. 6): DeVito (2021, p. 3): "the capacity and opportunity to be aware of both the presence and impact of algorithmically-driven systems on self- or collaboratively-identified goals, and the capacity and opportunity to crystalize this understanding into a strategic use of these systems to accomplish said goals"; Dogruel, Masur and Joeckel (2022 in print; online 2021, p. 4): "being aware of the use of algorithms in online applications, platforms, and services, knowing how algorithms work, being able to critically evaluate algorithmic decision-making as well as having the skills to cope with or even influence algorithmic operations". (Both quoted via the review's preprint; I did not open DeVito or Dogruel. Dogruel et al. is *Communication Methods and Measures* 16(2): 115–133, issue 2022, which is why the brief says "2022".) Ridley and Pawlick-Potts (2021) add "Position the individual as a co-constituent in algorithmic decision-making" (Z, B9a). Gagrčin et al. (2024) review 169 papers and frame acquisition as "experiential learning cycles" (abstract). +- **"Algorithmic competence"** occurs in Crossref titles for pupils' or teachers' skill with algorithms as procedures (Z row 39, titles only). It is not an established construct for living with algorithmic systems. +- **"Algorithmacy."** Oberschelp (1995, p. 121; dxd card, Full): "we have to preserve logicacy and algorithmacy by continuously training simple mental activities without artificial tools." One use, undefined, meaning a skill of carrying out algorithmic steps unaided. No other independent use (dxd Z §18; this arm's Z rows 37–43). Cousin to distinguish: Aneesh's "algocracy" (2009; N3), rule *by* algorithm. +- **Positioning.** Every definition above makes the competence the individual's, measured by awareness, knowledge, critique and coping; none contains a second party or a binding decision (N3 on Dogruel; Z on Ridley). DeVito's "capacity *and opportunity*" is the one that puts the environment in the definition, which is where v4's scaffolding claim attaches: opportunity is something a firm can provide or withhold. The paper can position algorithmacy as (i) a successor term on the literacy–numeracy pattern that Oberschelp also used, (ii) defined by relation (coordinating with others through an algorithmic intermediary) rather than by knowledge about algorithms, and (iii) not acquired by reading, which is why disclosure cannot build it (the v4 link between wrongs (b) and (c)). It must cite Oberschelp as a prior use in a different sense and DeVito/Dogruel as the definitions it departs from. + +## Q5. Objections the paper must meet, in their strongest published form + +**O1. Scaffolding is paternalism.** +- *Strongest version in venue:* Scherer et al. (2023, p. 422): paternalism, "no matter how much more effective it is in deploying superior expertise to protect an individual's interests ... remains, in a political sense, a form of interference with one's liberty to make one's own decisions", and organized immaturity "can emerge simply as an effect of more or less subtle versions of paternalism". Ruehle (2023) applies this to workplace nudges (via the editors, p. 430). +- *Welfarist version:* Sunstein (2012, Storrs Lectures, `sunstein2012storrs`, Abstract): "Official action may fail to respect heterogeneity, may diminish learning and self-help, may be subject to pressures from self-interested private groups ... and may reflect the same errors that ordinary people make" (abstract). He also holds that "choice architecture, and sometimes a form of paternalism, are inevitable" (abstract), which cuts the other way. +- *In the reader's own paper:* Berthon, Pitt and Campbell (2019, p. 462, via L1): "Consumers may prefer more user control rather than blanket mandates." +- *Reply the record supports:* scaffolding in Rohlfing et al.'s sense is contingent, learner-led and "temporary" (p. 723), which is the opposite of a standing substitution of judgment; and Scherer et al. themselves list business "training" as a remedy for immaturity (p. 429). The paper should define scaffolding as building the person's capacity to act, withdrawn as capacity grows, and distinguish it from nudging. + +**O2. Disclosure respects autonomy; it is the least paternalistic tool.** +- *Strongest statement found:* Bar-Gill (2016, `bargill2016information`, Abstract): "Disclosure mandates are often considered to be the least paternalistic of all regulatory techniques. Indeed, information provision is believed to enhance both autonomy and efficiency by facilitating more informed decisionmaking ... disclosure regulation ... is beyond reproach" (abstract; he states the received view in order to question it, and the chapter is unread). +- *Philosophical version:* Vredenburgh (2022): the right to explanation protects "informed self-advocacy" (AAM p. 5), and individuals "can reasonably reject institutional set-ups in which they cannot engage in informed self-advocacy" (AAM p. 5). Kim and Routledge (2022, in arm) ground an ex post right to explanation in informed consent. +- *Reply the record supports:* Vredenburgh herself concedes explanation is not usable "absent support" (AAM p. 18); Ben-Shahar and Schneider (card) document disclosure's failure; Steffel et al. (2016, abstract) find "disclosure does not make defaults less influential"; Ostinelli et al. (2024) find the mere prospect of an explanation breeds illusory understanding. The paper should concede that disclosure is necessary (Berthon's own "necessary, but not sufficient", p. 457) and argue only against sufficiency, which is the v4 claim. + +**O3. Scaffolding is itself responsibilization.** +- *Strongest version:* Docherty and Valderrama Barragán (2026, `docherty2026civilizing`, Part: pp. 1–3, 11, 16–18 of the online-first pagination, White Rose published version). Meta's digital-literacy curricula show "the growing appropriation of digital literacy discourses by technology companies like Meta to respond to and counter the criticisms made against their services" (p. 17); "by placing on individuals the burden of living well on platforms, Meta's social media corporate curriculum arranges users and technology in relationships of responsibilized control" (p. 17); "this burdening of the user to live well on social media perpetuates existing neoliberal logics of responsibilization" (p. 11); these curricula are "a performative practice of corporate social responsibility" (p. 2). Even their alternative carries the risk: "a critical digital literacy focused on a renewed personal political awareness devoid of the collective could serve simply to redistribute, rather than challenge, the primary responsibility of the individual" (p. 18). +- *Also:* Oleart and Flores Moleón (2026, Abstract): EU use of media and AI literacy "shifts responsibility from platforms to individual citizens" and so "empowers" Big Tech. Giesler and Veresiu (2014, card): the "financially literate consumer" as responsibilization (p. 851). Martin (2019, p. 845): tools that let users revisit how the algorithm works move accountability to them. Rohlfing et al. (2021, p. 726): do "not add the empowerment of persons to a number of already existing burdens". +- *Reply the record supports:* Helberger et al. (2018, p. 3) with Fahlquist: users can be held responsible only to "the extent that institutional agents have taken their forward-looking responsibility". Scaffolding that is a firm-held duty, built into the system and paired with controls, is the institution taking its forward-looking responsibility first; a curriculum handed to users instead of design change is what Docherty describes. **This objection is the most dangerous to wrong (c)** because the v4 thesis asks firms to do what Meta already does in name. The paper must say what distinguishes scaffolding from a corporate curriculum (in the system, not beside it; owed, not voluntary; paired with the firm's own controls and an independent check). + +**O4. The post addresses enterprises, not consumers.** +- *Evidence for the objection:* the post's "us" is the enterprise (SOURCE_v4 flag; DXD ruling D5). The AI Act's literacy duty also runs to "staff and other persons dealing with the operation and use of AI systems on their behalf" (Art. 4; Z R1), and Martin's "users" are the deploying decision-makers (p. 844). +- *Reply the record supports:* the three wrongs apply to the enterprise reader as well. Green (2022, p. 1) shows that policies resting on human oversight assume "that people are able to effectively oversee algorithmic decision-making", an "uninterrogated assumption"; that is wrong (a) for staff. The second reader, the people the systems act on, appears in the post only as the recipient of incident disclosure (P20). The paper should run the argument twice, once per reader, rather than speak of "the user". + +**O5. The history is analogy only.** +- *Strongest versions:* Qadri et al. (2026, card): gestalt analogies "fail to isolate the precise technical and social properties that carry the comparison" (N5). Kitzinger (2024, abstract): the Phaedrus analogy for LLMs "is substantially misplaced" and "risks justifying the swift adoption of a commercial tool". The medium differs (Heyd 2021: we now talk "with machines", p. 131). +- *Reply the record supports:* name the property that carries the comparison (Qadri et al.), which for v4 is the allocation of the burden of a new medium's competence; Ong (p. 133) supplies the continuity claim that electronic orality rests on writing, so the governance layer is still textual; and use the history as a record of what institutions did to people (Clanchy, the treaty record), not as a claim that history repeats. Kitzinger's point also warns the paper away from the "people feared writing too" move. + +## Sorted: for, narrows, against + +**For the thesis** +1. Ong (1982, pp. 2, 10–11, 133): secondary orality is "based permanently on the use of writing and print" — answers the orality counterargument. +2. Marz (2026): the "return to orality" writers misread Ong; LLMs "entail a new type of writing that comes in the form of the prompt". +3. Arató and Balázs (2024, p. 607): ChatGPT output is "a new form of literacy, tertiary literacy". +4. Rohlfing et al. (2021, p. 718): XAI assumes "the mere availability of information will suffice" (a). +5. Grote and Paulo (2025, abstract): explanations "intellectually over-demanding" for laypeople; the burden shifts to them (a). +6. Moon et al. (2025, abstract): transparency's benefits depend on the reader's literacy (a), (b). +7. Ostinelli et al. (2024, abstract): the prospect of an explanation creates illusory understanding (a). +8. Green (2022, p. 1): human oversight rests on an "uninterrogated assumption" (a, staff reader). +9. Scherer et al. (2023, p. 430, on Schade): GDPR transparency leads to "infantilization of users" (a). +10. Berthon, Pitt and Campbell (2019, p. 457): "information is necessary, but not sufficient" (a), in the reader's own work. +11. Oeldorf-Hirsch and Neubaum (preprint p. 23): "individuals must endure most of effort in improving literacy" (a), (c). +12. Dembeck (2023, p. 255): book-bound literacy education may miss the point of the digital (b). + +**Narrows the thesis (or its novelty)** +1. Scherer et al. (2023, pp. 428–429): business should provide "social rights by informing and training citizens" — (c) is stated in BEQ. +2. Vredenburgh (2022, AAM pp. 18–22): explanation needs "free experts and advocates"; companies may need to provide them — (c) in political philosophy, with the cost on decision-makers. +3. Rohlfing et al. (2021, p. 723): "Explaining Is Scaffolding" — the scaffolding remedy is in XAI already. +4. Helberger et al. (2018, p. 3): platforms must "create the conditions that allow individual users to comply". +5. AI Act Art. 4 and Omnibus 2026: a firm duty to build literacy exists, for staff only, and was softened. +6. Calderón and Herrera (2025) and the Phaedrus genre: oral-to-written comparisons with AI ethics exist, on epistemic grounds. +7. Martin (2019, p. 844): the designer's duty to allocate roles; her "users" are deployers. +8. Berthon (2019, pp. 452, 457): informing = "educate and disclose", so education and disclosure share a rung in the reader's framework. +9. Oberschelp (1995, p. 121): "algorithmacy" is in print, in another sense. + +**Against the thesis** +1. Heyd (2021, pp. 131, 138–139): interaction with machines is becoming oral ("tertiary orality"); voice assistants "speak back". +2. Docherty and Valderrama Barragán (2026, pp. 11, 17–18): firm-provided literacy is responsibilization and CSR performance — the strongest objection to (c). +3. Oleart and Flores Moleón (2026, abstract): EU literacy policy shifts responsibility to citizens and empowers Big Tech. +4. Martin (2019, p. 845): giving users tools to understand the algorithm moves accountability to them. +5. Scherer et al. (2023, p. 422): paternalism produces immaturity — applies to scaffolding done badly. +6. Bar-Gill (2016, abstract), Vredenburgh (2022, AAM p. 5), Kim and Routledge (2022): disclosure and explanation serve autonomy. +7. Kitzinger (2024, abstract) and Qadri et al. (2026): oral-to-written analogies for AI are often misplaced. +8. SOURCE_v4 flag, confirmed by this facet's reading: the post's reader is the enterprise; the AI Act's literacy duty is to staff. + +## What the paper can say (notes for the drafter, not prose) + +1. Ong defines secondary orality as dependent on writing and print, and says the computer intensifies the commitment of the word to space (Ong 1982, pp. 2, 133), so a conversational interface does not make the governance of AI oral. +2. The claim that LLM chat returns us to orality lives in essays and newsletters (Thompson, Weisenthal, Marriott, as reported by Marz 2026); I found no refereed 2023–2026 article that argues it at length, and the refereed linguistics either extends Ong to voice assistants (Heyd 2021) or classes generated text as a new literacy (Arató and Balázs 2024, p. 607). +3. The assumption that "the mere availability of information will suffice" to enable action has been named and criticised in explainable-AI research (Rohlfing et al. 2021, p. 718), and philosophers have argued that explanations are over-demanding for laypeople and shift the burden of scrutiny onto them (Grote and Paulo 2025). +4. Experimental work finds that transparency features benefit those already algorithmically literate (Moon et al. 2025) and that the mere prospect of an explanation produces illusory understanding (Ostinelli et al. 2024). +5. Human-oversight requirements rest on the "uninterrogated assumption" that people can oversee algorithms (Green 2022, p. 1); this is the staff-side form of wrong (a), and it applies to the post's "human readable evidence" (P15). +6. Business ethics already holds that firms should train people to deal with manipulative algorithms (Scherer et al. 2023, p. 429), and political philosophy that decision-makers should fund free experts to make explanations usable (Vredenburgh 2022, AAM pp. 18–22); the v4 contribution is the historical case for that duty and its placement in the system's design, not the duty itself. +7. Firm-run literacy can be responsibilization: Meta's curricula place "on individuals the burden of living well on platforms" (Docherty and Valderrama Barragán 2026, p. 17); the paper must distinguish scaffolding built into a system and owed by its operator from a curriculum handed to users. +8. Martin (2019) places on the designing firm the allocation of "how large a role individual will be permitted to take" (p. 835) and rejects opacity as an excuse (p. 836), but on her model tools that help users understand the algorithm shift accountability toward them (p. 845); v4 must pair scaffolding with the firm's retained accountability. +9. Oral-to-written comparisons with AI exist, through Plato's *Phaedrus* (Calderón and Herrera 2025; Kitzinger 2024), but they concern memory, wisdom and authorship; I found none that treats the transition's distribution of burden among people as the template. +10. Existing definitions of algorithmic literacy are individual competences (DeVito 2021 and Dogruel et al. 2022, via Oeldorf-Hirsch and Neubaum, preprint p. 6); DeVito's "capacity and opportunity" is the hook for a firm's duty to provide opportunity. "Algorithmacy" appears once in Oberschelp (1995, p. 121) in a different sense. +11. Berthon's own framework puts education and disclosure together under "informing" and judges information "necessary, but not sufficient" (Berthon, Pitt and Campbell 2019, pp. 452, 457); v4 extends his "not sufficient" and departs from his pairing of education with disclosure. + +## Counts by read depth + +| Depth | Count | Works | +| --- | --- | --- | +| Full | 3 | Heyd 2021; Arató and Balázs 2024; Marz 2026 | +| Part | 8 | Martin 2019 (pp. 835–836, 844–846); Ong 1982 (secondary-orality passages); Rohlfing et al. 2021; Scherer et al. 2023 (pp. 421–433 added); Vredenburgh 2022 (AAM); Docherty and Valderrama Barragán 2026; Calderón and Herrera 2025; Oeldorf-Hirsch and Neubaum (preprint, reused from N3) | +| Abstract | 12 | Moon et al. 2025; Ostinelli et al. 2024; Grote and Paulo 2025; Kitzinger 2024; Bar-Gill 2016; Sunstein 2012; Oleart and Flores Moleón 2026; Docherty and Valderrama Barragán 2025 (AoIR); Steffel et al. 2016; Passlack et al. 2025; Ananny and Crawford 2018 (reused); Gagrčin et al. 2024 | +| Record | 6 | Ruehle 2023; Holt and Wiedner 2023; Ess 2026; Richards 2023; Colon 2026; Cánovas 2025 | +| Reused from other facets or cards | — | Green 2022 (M2); Helberger et al. 2018 (N4); Berthon et al. 2019 (card); Giesler and Veresiu 2014 (card); Qadri et al. 2026 (card); Dembeck 2023 and Oberschelp 1995 (dxd cards); AI Act Art. 4 and Omnibus (Z) | + +Not opened (blocked or absent): Ong 1995 "Hermeneutic Forever" (journal PDF did not respond); Schade 2023 (still abstract, N5); Lobschat et al. 2021; Jasmine Sun's essay; Ong's 1996 interview (known only through Marz). + +## Cards written (all `verified: pending`) + +`heyd2021tertiary`, `arato2024communicative`, `marz2026return`, `ong1982orality` (secondary-orality pages only; the sister arm holds a fuller card), `rohlfing2021explanation`, `scherer2023guest`, `martin2019ethical`, `vredenburgh2022right`, `docherty2026civilizing`, `kitzinger2024openai`, `oeldorf2023what`. Cited from existing cards without adding: `berthon2019addictive`, `giesler2014creating`, `qadri2026ai`, `benshahar2011failure`; from the dxd arm's library: `oberschelp1995developing`, `dembeck2023flow`. diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_v4_RESEARCH_BRIEF.md b/submissions/berthon_ethics_paper/nadella/sources/bec_v4_RESEARCH_BRIEF.md new file mode 100644 index 00000000..457c3ad0 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_v4_RESEARCH_BRIEF.md @@ -0,0 +1,58 @@ +# Brief: research round 7 for version 4 (facets V1–V6, 2026-10-14) + +Arm: `/Users/ludwitt/iit-playground/wt-berthon-ethics/submissions/berthon_ethics_course/`. +Read `v4/SOURCE_v4.md` first: the author's instruction, verbatim, and the argument at its core. You are building +the evidence for a new version of the paper. It opens on Satya Nadella's X post of 10 October 2026, goes back to +the transition from **oracy to literacy**, and reads the post for echoes of that transition's ethical failures. +The author's claim is that the post treats disclosure (specific disclosures, readable records, design for +legibility) as sufficient ethical design, and that this repeats three mistakes of the oral-to-written +transition: (a) **placing the burden on the person** who must read; (b) **keeping the person in a +literacy-based experience** (texts to read) when the medium has changed; (c) **not scaffolding the person's +growth** to the new competence (algorithmacy). + +Your job is to find what the record supports, what narrows the claim, and what cuts against it. Report all three. +A finding against the thesis is as valuable as one for it. + +## What already exists (read before searching; do not redo it) +- `MANUSCRIPT_v3.md`: the verified v3 (five practices; Douglass; Waitangi; *Marshall*; Pragmática of 1503; + *Foster v Mackinnon*; *Hill v Gateway*; Edwards & Veale; Ben-Shahar & Schneider; Kamperman & Graff; Pettit, + Lovett, Young). `literature/cards/REFERENCES.md` indexes its 118 verified cards. +- `literature/facets/*.md` (rounds 1–6), especially A (documents and standing), G (third parties), M1–M5, N1–N5, + T_adjudication.md. `literature/references.bib` (617 entries). +- The sister arm on the same post: `/Users/ludwitt/iit-playground/wt-dxd-foundation/submissions/dxd_foundation/` + — facets `literature/facets/G_nadella_lineage.md`, `L_disclosure_and_choice.md`, `M_acting_on_algorithms.md`, + `I_understanding_through_action.md`, `K_training_and_cause.md`; its card library `literature/library/` + (`INDEX.md`); its claims ledgers `literature/audit/claims/`. Reuse its verified readings, cite the card path, + and do not contradict a verified finding without new evidence. Its findings include: disclosure evidence is + mixed in direction; the critique of the post cannot rest on "people review badly"; the post shares the premise + of external controls. +- The post: `~/iit-playground/dxd_foundation_local/nadella_2026-10-10_models_as_insider_risks.txt` (P1–P21). + Quote it only in short fragments with paragraph numbers. **Never copy it into the repo.** +- Local PDFs: `/Users/ludwitt/iit-playground/berthon_ethics_local/pdfs/**` and + `/Users/ludwitt/iit-playground/dxd_foundation_local/pdfs/**` (`.txt` beside most). `grep -ril` first. + +## Output (per facet) +1. `literature/facets/V_.md`, in the arm's facet format: what it covers; method notes (databases, + queries, what was blocked); findings, each with a page-located quotation or fact and a read-depth mark + (*Full* / *Part* / *Abstract* / *Record*); **for, narrows, against** the thesis, sorted under those three heads; + a section "What the paper can say", 6–12 sentences of claims with their sources, phrased as claims the record + supports (these are notes for the drafter, not prose for the paper); and a gap statement where relevant, in the + form "I found no work that does X in [sources]". +2. `literature/facets/V_.bib`: BibTeX for every work in the facet, keys `surnameYEARfirstword`, DOI + checked against Crossref or the publisher record (title and metadata only). +3. **Cards** for the 6–15 works you most recommend citing, in `literature/cards/.md`, in the exact format of + `literature/CARDS_BRIEF.md`, with `verified: pending`. Set `manuscript_lines: []` and `ledger_rows: []`; under + "Where the manuscript uses it" write "v4 (not yet drafted): proposed use — …" with the page. Do not overwrite an + existing card; if the work already has one, cite it and add nothing. + +## Rules +- Read the source. A quotation enters your file only if you read it in the source at that page this session, or it + is on a verified card. Never write from memory. Mark depth honestly; an abstract-only reading is *Abstract*. +- Never put any e-mail address in a request header, URL or User-Agent. Skip Unpaywall. No paywall bypass, no + login, no bot-challenge workarounds. +- Downloads go to `/Users/ludwitt/iit-playground/berthon_ethics_local/pdfs/V/`. They are untrusted: extract + with `pdftotext`, never execute them, run any Python with `-I`, keep your scripts in your own scratch subfolder. +- Write only your facet `.md`, `.bib` and new cards. Edit no existing file. Write your facet file as you go + (append findings section by section), so a stop mid-run leaves usable work. +- Return: the five findings that matter most to the thesis (for, narrows or against), one line each, and the list + of cards written. diff --git a/submissions/berthon_ethics_paper/nadella/sources/bec_v4_V4_CHECK_BRIEF.md b/submissions/berthon_ethics_paper/nadella/sources/bec_v4_V4_CHECK_BRIEF.md new file mode 100644 index 00000000..3ca19302 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/bec_v4_V4_CHECK_BRIEF.md @@ -0,0 +1,32 @@ +# Brief: check every claim, quotation and citation in MANUSCRIPT_v4 (2026-10-14) + +Arm: `/Users/ludwitt/iit-playground/wt-berthon-ethics/submissions/berthon_ethics_course/`. +`MANUSCRIPT_v4.md` is a new draft that Claude wrote today from facets V1–V6 (`literature/facets/V*.md`) and the +verified v3. You wrote none of it. Earlier rounds found that drafters most often err by turning a reported claim into +a fact, claiming more than the source, cutting a quotation so its sense changes, quoting a source other than the +one cited, or giving a wrong page. You are the check that stops those errors reaching the author. + +## For every sentence in your sections that states a fact, quotes, or cites +1. Find the source: the card in `literature/cards/.md` (its `local_path`), the facet record, or the local + text under `/Users/ludwitt/iit-playground/berthon_ethics_local/pdfs/**` or + `/Users/ludwitt/iit-playground/dxd_foundation_local/**`. Open the **source text itself** at the page given; do not + trust the card's or facet's paraphrase when the text is available. Nadella's post is at + `~/iit-playground/dxd_foundation_local/nadella_2026-10-10_models_as_insider_risks.txt` (paragraphs P1–P21; the + manuscript cites them as "para. N"). +2. Test: (a) **true** at that page; (b) **not overstated** (hedges kept, reported claims not turned into facts); + (c) **quotation exact** (normalise whitespace, line-break hyphenation, curly quotes only; ellipses marked; words + attributed to the source actually read); (d) **page or paragraph right**; (e) **the cited work has a reference + entry** and the entry matches the card (authors, year, title, venue, DOI). +3. Verdict per sentence: `pass` | `fail-true` | `fail-overstated` | `fail-quote` | `fail-locator` | `fail-ref` | + `cannot-check (why)`. For each fail give the **exact corrected sentence**, as small a change as possible. Do not + restyle prose; this is a fact check, not an edit. + +## Output +`v4/V4_CHECK_.md`: a table `| § | sentence (first 8 words) | verdict | evidence (source + page) | corrected sentence |`, +then `## Rates`. Write the file as you go so a stop mid-run leaves usable work. + +## Rules +- Read only. Do not edit the manuscript or any other file; write only your check file. +- Never put any e-mail address in a header, URL or User-Agent. No paywall bypass; skip Unpaywall. Downloads go to + `/Users/ludwitt/iit-playground/berthon_ethics_local/pdfs/v4check/`, untrusted (`pdftotext`; Python with `-I`). +- Return the rates and every fail in one line each. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_ABSTRACT_v4.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_ABSTRACT_v4.md new file mode 100644 index 00000000..5de8720e --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_ABSTRACT_v4.md @@ -0,0 +1 @@ +Human–machine interaction has moved from signals and screens to extended reality (Mourtzis & Tsakalos, 2026) and human–machine integration (Grahn, Wang, Wang, & Wang, 2026). Yet, whether on a factory headset or in an enterprise audit log (Nadella, 2026), the system still sends a person something to read so they can check the machine. That approach assumes a dyad: an operator paired with a tool (Parasuraman, Sheridan, & Wickens, 2000). On digital platforms, though, counterparts—a driver and a rider, or a buyer and a seller—coordinate through an algorithmic mediator that binds and optimizes, forming an irreducible algorithmic triad. Mandated disclosure (Ben-Shahar & Schneider, 2011) does not spare them from having to game the mediator to coordinate with the other person. What these settings beg for instead is the kind of cognitive offloading (Risko & Gilbert, 2016) seventh-century Irish scribes achieved when they broke up scriptio continua by adding spaces between words (Tunbridge, 1992; Saenger, 1997). Unspaced text forced a reader to sound out the line just to hear where each word ended, scaffolding literacy with oracy; word separation freed literacy by moving boundary detection from the voice to the page, exploiting a representational effect (Zhang & Norman, 1994) that enabled readers to internalize text. I argue we should look for a similar "smoking gun" design change that would free algorithmacy from literacy—enabling people to decide within irreducible triads through epistemic action (Kirsh & Maglio, 1994), strategic offloading (Li, 2017), and ecological feedback immediacy (Shneiderman, 1983; Kahneman & Klein, 2009). diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_ARGUMENT_MAP.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_ARGUMENT_MAP.md new file mode 100644 index 00000000..e211e019 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_ARGUMENT_MAP.md @@ -0,0 +1,70 @@ +# Argument map + +*Written by Claude on 2026-10-12 and revised on 2026-10-13 after the author's ten rulings (`SOURCE.md`), +facets I–K and the card library (`literature/facets/T_adjudication.md` §9). It maps the strongest +defensible form of the argument, which differs from the form the author first gave. None of it is the +author's text. The first-round map is in the git history of this file.* + +**Depth marks.** `[card]` the source has a verified card in `literature/library/`, which carries the +locators. `[ok]` an auditor confirmed the passage in the source. `[abs]` abstract only. `[lib]` needs a +library copy before it is quoted. `[inf]` an inference by the lead or a facet agent with no source behind +it. `[author]` the author's ruling or proposal. Source keys are in `literature/references.bib`. + +## The argument + +| # | Premise | Rests on | Mark | Weakest point | +| --- | --- | --- | --- | --- | +| P1 | Writing in the Roman and early medieval West ran without word boundaries, and readers supplied them | Saenger 1997, pp. 8, 85 | `[ok]` | Whether unspaced script burdened trained readers at all is disputed (Vatri, via Heilmann 2020, p. 181). Knox (1968) supports "normally read aloud" and says nothing of boundaries | +| P2 | Insular scribes of the late seventh century began separating words; the Continent adopted the practice from the mid-tenth century and it was standard by the mid-eleventh | Saenger 1997, pp. 83, 100, 176–177 | `[ok]` | The chronology is Saenger's. Tunbridge (1992), supervised by Parkes, is no independent test. That the first readers served were learners of Latin as a foreign language is Parkes's and Saenger's reading `[lib]` | +| P3 | Readers used to spaced script lose a third to a half of their speed when the spaces are removed | Winskel et al. 2009; Rayner et al. 1998; Mirault et al. 2019 (citing Rayner) | `[card]`; `[abs]`; `[card]` | Modern readers habituated to spaces; no test of ancient readers | +| P4 | For readers of unspaced scripts the effect of added spaces is mixed: shorter fixations on words and longer sentences in Thai, a gain in Tibetan, no gain for skilled Chinese adults, a loss for child learners of Chinese | Chantavarin 2025; Kohsom and Gobet 1997; Wang 2024; Bai et al. 2008; Bassetti and Masterson 2012 | `[card]` | Small samples (8 to 24 readers). The second-language learner studies that show gains are `[abs]` | +| P5 | Word separation re-represented the page: the same letters in a form that turns an inferential task into a perceptual one. It disclosed nothing and trained nobody | Zhang and Norman 1994, p. 88; Larkin and Simon 1987; Hutchins 1995 (article), pp. 282–283; Risko and Gilbert 2016, Box 3 | `[card]` | Applying these to a page is the paper's move; no human-factors source found treats the history of writing this way (facet Z) | +| P6 | Saenger claims that separation and three other devices gave Insular readers "shortcuts" to skills an ancient elite reached by "a prolonged and arduous grammatical apprenticeship". No source found claims that separation raised the number of readers, and it came as one of a complex of conventions | Saenger 1997, pp. x, 85; facets C, F, K | `[ok]` | One sentence, no measure. The histories of schooling show the alphabet, syllables and psalter continuing after separation (snippets) | +| P7 | The history is a guide to the move from one -acy to another and no proof of how the next will go | The author's ruling D10; Part 2 ruling | `[author]` | A guide licenses a question, not a design | +| P8 | Disclosure of an algorithm is an intervention in literacy: it hands the person a text. It has not been shown to change how well a person navigates a mediated coordination, and in several studies it changed nothing or did harm | The author's ruling D3; Jhaver 2019; Seira 2017; Poursabzi-Sangdeh 2021; Cabrera 2023; Korbak 2025, p. 7; P2B Art. 5(6) | `[author]`; `[card]` | A cue with its reason cut simulator collisions from 7 of 15 to 1 of 15 (Matsuo 2026), and explanations helped in a hard task (Vasconcelos 2023). Neither is a triad. No disclosure study in a triad with a behavioural outcome was found | +| P9 | People come to understand a mediator by deciding and taking in the results, provided the environment is valid and feedback is "both rapid and unequivocal" | The author's ruling D7; Weick 1988; Kahneman and Klein 2009, p. 524; Cook et al. (drivers, 14%); Heiland 2023; Dismukes and Berman 2010 | `[author]`; `[card]` | "Pilots don't spend time reading the situation" is contradicted (Hutchins 1995, pp. 283, 285). Part 1, line 66, says the mediator disrupts both conditions. No source gives "from reading and writing to deciding" as a historical shift | +| P10 | Two conventions are candidates for the most critical one. A rule in the mediator can remove a strategic task and show that it is gone. A convention can return the result of a decision quickly and plainly enough to learn from | Li 2017, §5.3 (36% against 7%); Rees-Jones and Skowronek 2018 (23% misreport); Rahmandad and Gary 2023 | `[card]` for the first; `[inf]` for the second | Matching clearinghouses have no objective of their own. The first candidate works because the person does not act on the mediator, which sits badly with P9. The second has no direct test | +| P11 | The triad is irreducible, and Φ gauges the irreducibility of any given triad | The author's ruling D9; the lab's results | `[author]` | Φ classifies a triad. It does not show that a convention helps the people in it. No convention has been tested in a triad | +| P12 | The owner of a platform's interface is an interested party and can change a convention alone. In courts, dealer foreign exchange and platforms the convention moved toward the weaker party only under outside force | Jensen and Meckling 1976; Suzor 2018; P2B Art. 3; FX Global Code, Principles 7, 9, 36, 39, 40; DSA Arts. 17, 20, 53–54; *Tumey v. Ohio*; Nagaraj Rao 2026 | `[card]` | Tyler: "no universally fair procedures" (1988, p. 131), so the settings cannot be ranked on one scale. "Draconian" is the author's word; Klonick calls platform moderation "impressively intricate and responsive" (p. 1664). The Jensen card rests on a scan of unclear provenance | +| P13 | Nadella's insider post places authority over a model with the deploying organisation. What it gives a person is a record to read and a brake; its other controls are the organisation's. It is silent on the people the model acts upon | Post, P5, P9, P14, P15, P17, P19, P20; the author's rulings D4 and D5 | `[ok]` against the capture | "His only move is disclosure" does not match P14, P17, P19. The practice he draws on supports review of records by people (CERT 2022, practice 12), and one study puts human readers of agent transcripts at AUC 0.65 (Kutasov 2025), so the critique cannot rest on "people review badly" | + +**Conclusion.** Word separation is a historical case of an intervention that is neither disclosure nor +training: it re-represented the page (P1–P6), and it guides the question without settling the answer +(P7). Disclosure, the usual remedy for an opaque algorithm, is a literacy-based intervention and has not +been shown to help a person navigate a mediated coordination (P8). Understanding there comes through +deciding, where feedback allows it (P9), which points to two candidate conventions (P10) in a triad +whose irreducibility Φ can gauge (P11). Whoever holds the convention, the record of courts, dealer +markets and platforms says the third party yields it only under outside force (P12). Nadella's insider +post addresses how an organisation contains a model it cannot read; what it offers a person is something +to read, and it does not address the people the model acts upon (P13). Decision Experience Design is the +discipline that looks for the missing conventions. + +## What the argument no longer claims + +| Dropped | Why | +| --- | --- | +| Spacing "ignited literacy" | No historian of literacy credits it; Saenger's book "is not a general history of literacy" (p. x); unspaced scripts support fluent reading (P4) | +| Spacing created silent reading | Knox (1968, pp. 432–435); Saenger himself has Ambrose reading silently (p. 8) | +| "The" move | Every parallel case is a bundle (facet F); the candidates in facet H are several | +| Nadella proposes "the exact opposite" | Shared premises and mechanisms (P9, P11 of the post); the difference is authority and seat | +| "Moving work to the artefact" marks the paper off from Nadella | A reference monitor moves work to the artefact too | +| Reading "no longer required the training" | The task was transformed and the scribe took on work (Hutchins 1995, pp. 280–283); schooling in letters and syllables continued (facet K). Saenger's p. 85 sentence is the nearest support and should carry the claim under his name | +| Spacing helped "most for learners" | Child learners of Chinese did worse with spaces (Bassetti and Masterson 2012); the modern learner evidence is mixed | +| People perform after-the-fact review badly, as the ground of the critique of the post | CERT practice endorses it; the direct evidence is one study (Kutasov 2025) | +| "Disclosure doesn't affect our ability to navigate it", stated flatly | Mixed in direction: some null, some harm, two gains outside the triad | +| "Pilots don't spend time reading the situation" | Both pilots monitor (Hutchins 1995); monitoring is the least-caught part of crew performance (Dismukes and Berman 2010) | + +## Where the argument is exposed + +1. **No convention has been tested in a triad with a behavioural outcome.** Every measured effect in + P8–P10 comes from a person and a system, or from a clearinghouse with no objective of its own. +2. **P9 and P10 pull apart.** The best-evidenced convention removes a task from the person; the author's + ruling has the person understand by acting. The paper has to say whether the two serve different + decisions or compete. +3. **P9 depends on a condition the author's own Part 1 says the mediator breaks.** The feedback + convention in P10 is the answer to that, and it is an inference `[inf]`. +4. **The precedent is a literacy-side intervention.** Word separation helped people read. P7 covers this + only if the paper keeps the history as a guide. +5. **The contrast with the post rests on whom the controls serve.** It no longer rests on a showing that + his human tasks fail. +6. **"Algorithmacy" has a prior printed use** in another sense (Oberschelp 1995, p. 121). diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_CLAIM.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_CLAIM.md new file mode 100644 index 00000000..3975ae61 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_CLAIM.md @@ -0,0 +1,34 @@ +# Claim — provisional + +*Stated by Claude on 2026-10-12 from the author's instruction. It is a claim under test. The author +rules on its final form after `literature/facets/T_adjudication.md`.* + +## The thesis as the author gave it + +The move from *scriptio continua* to spaces between words was the critical user-experience move that +ignited literacy. Algorithmacy needs a move of the same kind. Satya Nadella's "Models as Insider Risks +in the Super Intelligence Era" proposes the opposite. + +## The thesis in the form the research tests + +1. **Precedent.** Scribes changed the page so that the reader no longer had to find the words. The work + moved from the reader to the artefact. +2. **Consequence.** One of two readings holds, and the research decides which: + - *H-cost.* The move lowered the cost of reading and widened the group who could read without long + training. + - *H-cause.* The move caused the spread of silent reading and, through it, of wider literacy. +3. **Transfer.** Coordination through an algorithmic third party is today in the condition of unspaced + script: only the trained can do it. Decision Experience Design is the discipline that looks for the + move that takes the work off the person. +4. **Foil.** Nadella's answer to opaque models is observation, logging and containment by the + organisation. On the author's reading it asks people to read more, keeps a two-party picture of + principal and contained tool, and separates authority from engagement. + +## Do not claim + +- That word separation created silent reading, unless the research supports it. +- That a design change restructures cognition (standing ruling; see `RESEARCH_PLAN.md` §3). +- That the remedy for algorithmacy is helping users read the algorithm (standing ruling). +- That Nadella argues for transparency alone. He says it is "not sufficient or dependable". +- Any figure or source from the Gemini essays that a facet has not confirmed. +- Novelty for the analogy before facet Z reports. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT.md new file mode 100644 index 00000000..82552366 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT.md @@ -0,0 +1,178 @@ +# Draft — abstract, introduction and section outline + +> **Snapshot of the author's Google Doc "Human Factors Paper", taken 2026-10-13.** The Google Doc is the +> governing text; this file is a copy for the repo and may fall behind it. Gemini wrote the diagnostic, +> abstract, introduction and outline as a refactoring of Claude's earlier draft (in this file's git +> history, commit `5b4a7dc`), and the author adopted it. Claude then made 37 word-level edits in the doc +> at the author's instruction ("your judgement is fine"; "if you think the changes are necessary make +> them"): 22 factual corrections and 15 changes following the third research round. Bullets that begin +> "Key citations (added by Claude…)" or "Literature to add (Claude…)" are Claude's. Everything else is +> the adopted text. The evidence behind the edits is in `literature/facets/T_adjudication.md` §9–§10 and +> `literature/audit/AUDIT_07`–`11.md`. + +# The Space Between Decisions: Word Separation as a Guide to Designing for Algorithmacy + +# Part I: Critical Diagnostic Review of Prior Draft + +## 1. Core Structural and Stylistic Deficiencies + +- **Defensive Hedging vs. Theoretical Authority:** The previous draft operated defensively, repeatedly undercutting assertions with self-invalidating meta-disclaimers ("A guide is all the history can be," "The paper does not say in its own voice"). Academic theory papers in venues like *Organization Theory* require assertive, conceptually grounded arguments rather than anxious literature ledgering. +- **Mechanical "5C" Execution:** Lange and Pfarrer's (2017) 5C structure was treated as isolated bureaucratic boxes rather than a cumulative cognitive escalation. The narrative leap from medieval scribes to algorithmic mediation lacked the connective conceptual bridge: the distribution and locus of cognitive burden. +- **The Disclosure Category Error:** Prior drafts treated disclosure's inconsistent empirical outcomes as a laundry list of conflicting studies. In truth, disclosure fails because it commits a category error: deploying a literacy-based intervention (demanding passive textual reading) into an interactive, triadic coordination problem. +- **Subjugation of Core Theory:** Algorithmacy was introduced as an aside rather than developed as a primary theoretical construct. The foil—enterprise compliance models treating algorithms as insider risks—was summarized passively rather than dismantled as the institutional zenith of literacy dogma. + +# Part II: Refactored Abstract and Introduction + +## Abstract + +Human factors engineering traditionally resolves cognitively demanding tasks not by re-educating the operator, but by re-representing the task environment. The foundational historical precedent is medieval word separation: inserting spaces between words introduced no semantic data to the page, yet it offloaded lexical boundary-detection costs from inference to perception. Today, human coordination increasingly occurs through algorithmic third parties that mediate exchanges between independent actors while optimizing for distinct, proprietary objectives. Regulators and enterprise architects routinely address algorithmic opacity through disclosure: explainability dashboards, published rules, and audit logs. + +This paper argues that disclosure represents a fundamental category error: it applies a literacy-based intervention (demanding textual decoding) to an interactive coordination problem. Navigating irreducible algorithmic triads demands not literacy, but *algorithmacy*—the behavioral capacity to infer, test, and anticipate mediator dynamics through continuous, direct action. Drawing on distributed cognition, epistemic action, and mechanism design, I formulate Decision Experience Design (DXD) and articulate two candidate structural conventions: strategic offloading (eliminating tactical posturing) and radical feedback immediacy (returning valid, plain results of an action with minimal latency). Examining the institutional architectures of courts, dealer foreign-exchange markets, and digital platforms, I analyze who holds the power to enforce these conventions and specify the behavioral conditions required to test them experimentally. + +## Introduction: The 5Cs Structural Arc + +### Common Ground: The Representational Effect + +Human factors engineering has long recognized that when a task exceeds human cognitive capacity, the most robust solution is not to retrain the person, but to transform the representation of the task environment. Zhang and Norman (1994) formalized this principle as the representational effect: isomorphic representations of an identical formal structure evoke radically divergent cognitive behaviors. Hutchins (1995) demonstrated this dynamic within cockpit cognitive architectures, where physical speed bugs placed along the airspeed indicator allow pilots to substitute complex numerical comparisons with immediate judgments of spatial proximity. + +The most profound historical instantiation of this principle is the written page. Antiquity and early medieval manuscripts recorded language in continuous script (scriptio continua). Irish scribes initiated word separation by the late seventh century, Continental scribes adopted the practice in the tenth, and by the eleventh century the convention was universal (Saenger, 1997). The space introduced zero semantic information; every grapheme was already present. Yet by introducing visual white space, scribes transformed lexical boundary identification from an arduous phonological parsing task into automated visual pattern recognition, cutting cognitive overhead and easing rapid, silent reading (Saenger, 1997; cf. Rayner, Fischer, & Pollatsek, 1998). + +### Complication: The Irreducible Algorithmic Triad + +This foundational human-factors insight has failed to reach the systems through which contemporary economic and social life is coordinated. Riders and drivers, buyers and sellers, patients and providers, job applicants and employers increasingly transact across an algorithmic third party that evaluates both sides and enforces outcomes that neither participant directly controls. + +Confronted with opaque intermediaries, designers, corporate leaders, and regulatory bodies reflexively default to disclosure: publish model architectures, mandate explainability statements, and maintain immutable audit trails. Yet empirical evaluations of disclosure reveal profound systemic dysfunction: + +- **Delinquency and Confusion:** Credit card disclosure messages detailing repayment horizons increased delinquency among some borrowers rather than improving financial health (Seira, Elizondo, & Laguna-Müggenburg, 2017). +- **Illusion of Comprehension:** High-interpretability models with reduced feature counts impair users' ability to catch algorithmic errors during anomalous edge cases (Poursabzi-Sangdeh et al., 2021). +- **Paradoxical Perceptions:** Reading the posted rules in content moderation environments was associated with lower perceived procedural fairness among affected participants (Jhaver et al., 2019). +- **Dyadic Misalignment:** Studies demonstrating positive disclosure outcomes—such as driver intervention alerts in automated driving simulators (Matsuo et al., 2026)—occur in dyadic human-machine settings lacking an active, strategic third-party mediator. + +### Concern: The Literacy Category Error + +The persistent failure of disclosure stems from a foundational category error. Disclosure presents the human operator with a text—a rulebook, an explanation, a post-hoc audit log—and therefore demands a reader. It is fundamentally a literacy-based intervention. + +Coordinating through an adaptive algorithmic intermediary, however, is not a reading task. It is an interactive, multi-agent navigation problem within an irreducible triad. Solving a navigation problem requires algorithmacy: the active behavioral sensibility to probe, anticipate, and respond to mediated algorithmic dynamics. + +A remedy tailored for readers merely increases the cognitive tax on the people it purports to protect, leaving their underlying strategic decision paralyzed. Furthermore, disclosure tacitly consolidates institutional power. Satya Nadella's (2026) proposal to treat frontier models as "insider risks" exemplifies this dogma. By demanding tamper-proof human-readable records, model halt switches, and organizational separation between intelligence supply and authority, the framework serves the enterprise principal seeking internal operational defense. To the human transacting through the system, it offers only an autopsy record after failure has occurred. While medieval scribes answered to a church and court that ordered books made correct for their readers, digital platforms possess independent extraction objectives, maintaining terms of service that preserve unfettered corporate discretion (Suzor, 2018). + +### Course of Action: Decision Experience Design and Algorithmacy + +To escape the literacy trap, human-system design must return to the scribal page and re-examine word separation not as a cultural artifact, but as a cognitive design guide. Understanding an algorithmic intermediary does not come from passive inspection of its documentation; it is enacted through direct, iterative decision-making where feedback is rapid, unequivocal, and causally coupled (Kahneman & Klein, 2009; Weick, 1988). + +I term this paradigm Decision Experience Design (DXD). Rather than outputting descriptive explanations, DXD re-engineers the decision interface itself. I identify two primary candidate conventions: + +1. **Strategic Offloading:** Designing mechanism representations that structurally eliminate the need for tactical manipulation and visually signal that strategic gaming is obsolete. +2. **Radical Feedback Immediacy:** Collapsing the temporal and cognitive latency between user actions and intermediary state adjustments to enable rapid naturalistic calibration, where the feedback is valid. + +### Contribution: Advancing Human Factors and Organizational Coordination + +This paper makes three primary contributions: + +1. **Establishes the Historical Mechanics of Re-Representation:** Demonstrates how word separation functioned as cognitive offloading without overclaiming historical literacy leaps. +2. **Identifies the Literacy vs. Algorithmacy Boundary:** Provides an analytical framework that sorts regulatory and design remedies by cognitive mode, explaining why disclosure mechanisms so often falter. +3. **Articulates Decision Experience Design (DXD):** Defines the operational conventions of DXD, analyzes the political economy of institutional enforcement across courts, financial markets, and platforms, and establishes the precise behavioral protocol required to test triadic systems experimentally. + +# Part III: Complete Reconstructed Manuscript Outline + +## Section 1: Introduction: The Space Between Decisions + +- **Conceptual Narrative:** The 5Cs structural arc establishing the progression from representational human factors to algorithmic mediation. +- **Core Argument:** The shift from dyadic tools to triadic systems necessitates replacing passive literacy interventions with active algorithmacy frameworks. +- Key citations (added by Claude, from the verified card library): Zhang & Norman (1994, p. 88); Hutchins (1995, pp. 282–283); Saenger (1997, pp. 83, 85, 100, 176–177); Rayner, Fischer, & Pollatsek (1998; abstract only); Kahneman & Klein (2009, p. 524); Nadella (2026); Suzor (2018); Hunt (2026, Part 1 of the DXD series). + +## Section 2: The Scribal Precedent: Word Separation as Cognitive Offloading + +- **The Historical Shift:** Detailed examination of *scriptio continua* in Greco-Roman antiquity, early Insular innovations (7th century), Continental adoption (10th century), and standardization (11th century) (Saenger, 1997; Parkes, 1992). +- **Information Neutrality:** Proof that word spacing introduced zero semantic information; letters remained identical while cognitive access mechanisms transformed. +- **Historiographical Boundaries:** Rejection of naive democratization narratives; spacing did not replace formal instruction, but provided cognitive shortcuts for lexical processing. +- **The Saccadic Shift:** Transition from acoustic, subvocalized decoding to silent, automated visual boundary detection (Saenger, 1997; disputed by Knox, 1968, and Vatri, 2012; modern eye-movement evidence in Rayner et al., 1998; Mirault et al., 2019). +- Key citations (added by Claude, from the verified card library): Saenger (1997, pp. x, 8, 11, 83, 85, 100, 176–177; wording confirmed, pages need a print check); Parkes (1992; library copy needed); Knox (1968, p. 435); Heilmann (2020, pp. 181–182); Vatri (2012; known through Heilmann); Svenbro (1999, in Cavallo & Chartier); O'Donnell (2000); Tunbridge (1992); Cribiore (1996, p. 48; library copy needed); Johnson (2000, 2010); Clanchy (1993) and Cressy (1980) on what drove literacy. +- Literature to add (Claude; third-round research, audited): who enforced the conventions of the page. Charlemagne, De litteris colendis (letter to Baugulf); Admonitio generalis (789), c. 72; Alcuin, carm. XCIV; Westwell (2018). These show ruler and church ordering correct books, and none of them mentions word separation. They would replace the unsourced claim that scribes and readers shared interests, here and in Section 7. + +## Section 3: The Human Factors of Re-Representation + +- **Distributed Cognition Foundations:** Zhang and Norman's (1994) representational effect; converting internal search and memory storage into external perceptual cues. +- **Perceptual Substitution:** How external spatial configurations substitute perceptual recognition for logical inference (Hutchins, 1995; Larkin & Simon, 1987). +- **Cognitive Offloading Mechanics:** Risko and Gilbert (2016) on environmental offloading; Kirsh (1995) on the intelligent use of space. +- **Cross-Orthographic Asymmetries:** The directional asymmetry of conventions—removing spaces cripples fluent readers, but adding spaces to non-spaced writing systems (e.g., Chinese, Thai) yields complex, non-linear effects (Bai et al., 2008; Winskel et al., 2009). +- **The Fundamental Divergence:** A manuscript page is passive, static, and cooperative; an algorithmic mediator is dynamic, adaptive, and possesses private extraction incentives. +- Key citations (added by Claude, from the verified card library): Zhang & Norman (1994, p. 88); Larkin & Simon (1987); Hutchins (1995, pp. 282–283); Kirsh & Maglio (1994, p. 513); Kirsh (1995); Risko & Gilbert (2016, Box 3); Scaife & Rogers (1996, p. 210); Vicente & Rasmussen (1992). Removing spaces: Rayner, Fischer, & Pollatsek (1998; abstract only); Mirault et al. (2019). Adding spaces: Bai et al. (2008); Winskel et al. (2009); Kohsom & Gobet (1997); Chantavarin & Leung (2025); Wang et al. (2024); Bassetti & Masterson (2012). +- Literature to add (Claude; third-round research, audited): display-design principles the outline has not used. The proximity compatibility principle (Wickens & Carswell, 1995); the split-attention effect from cognitive load theory (effect size not yet verified); Norman's gulfs of execution and evaluation (his texts not yet read), under which Convention A narrows the gulf of execution and Convention B the gulf of evaluation. Caution from the nudge literature: Mertens et al. (2022, as corrected) class re-representing information ("translation") with informational interventions, the weaker category. + +## Section 4: The Failure of Disclosure: A Literacy Intervention in an Algorithmic Arena + +- **The Literacy Fallacy:** Explainable AI (XAI) and disclosure mandates treat transacting agents as literary critics analyzing post-hoc prose rather than actors navigating live systems (Ben-Shahar & Schneider, 2011; Schemmer et al., 2022; Vaccaro, Almaatouq, & Malone, 2024). +- **Empirical Failure Synthesis:** + - Regulatory disclosure backfiring in finance (Seira et al., 2017). + - Interpretability paradoxes in algorithmic edge cases (Poursabzi-Sangdeh et al., 2021). + - Procedural fairness paradoxes in platform moderation (Jhaver et al., 2019). + - Dyadic vs. triadic boundary conditions (Matsuo et al., 2026). +- **Critique of the Enterprise Paradigm:** Close reading of Satya Nadella's (2026) "Models as Insider Risks": + - Recognition of "neuralese" and the insufficiency of chain-of-thought visibility. + - Diagnosis of the enterprise solution: audit logs, brake switches, and human verification (CERT, 2022). + - Theoretical flaw: Resolves principal-agent compliance for the firm while abandoning the external user to post-hoc autopsy notices. +- Key citations (added by Claude, from the verified card library): Seira, Elizondo, & Laguna-Müggenburg (2017, pp. 292, 295); Poursabzi-Sangdeh et al. (2021); Jhaver et al. (2019); Cabrera et al. (2023); Buçinca et al. (2021); Ananny & Crawford (2018). Where disclosure helped: Matsuo et al. (2026); Vasconcelos et al. (2023). Where the field stops short of disclosure: Korbak et al. (2025, p. 7); Regulation (EU) 2019/1150, Art. 5(6). On the insider post: Nadella (2026); CERT (2022, practice 12); Chan et al. (2025); Kutasov et al. (2025); Parasuraman & Manzey (2010); Green (2022); Rieder & Hofmann (2020). +- Literature to add (Claude; third-round research, audited): (1) The law-and-economics critique of mandated disclosure states the paper's mechanism: people do not read, understand or use disclosures (Ben-Shahar & Schneider, 2011, p. 665; Loewenstein, Sunstein, & Golman, 2014; Fung, Graham, & Weil, 2007). Disclosures that work are simple signals and often act on the discloser, not the reader (Ben-Shahar & Schneider, 2011, pp. 743–748). (2) Explanations do not improve decisions on average: Schemmer et al. (2022; SMD 0.07, 95% CI −0.15 to 0.30); Vaccaro, Almaatouq, & Malone (2024; 370 effect sizes). (3) Layout moves consent while the information stays fixed (Nouwens et al., 2020). (4) The nearest rival position is nudging against boosting (Hertwig & Grüne-Yanoff, 2017, pp. 977, 981); the measured case for structure over information is modest (Mertens et al., 2022, as corrected; Maier et al., 2022). (5) Disclosure often changes behaviour without improving outcomes, so the claim should be about performance: a one-line statement that a mechanism is strategy-proof raised sincere bidding from 20.6% to 46.9% and efficiency by about one point (Masuda et al., 2022). Split "disclosure" into presence of an algorithm, its objective, a property of its rule, and the rule's text. + +## Section 5: From Literacy to Algorithmacy: Epistemic Action in Triadic Systems + +- **Sensemaking Through Probing:** Karl Weick's (1988) enacting sensemaking—understanding opaque systems requires action before comprehension. +- **Epistemic Actions:** Kirsh and Maglio (1994) on actions executed to uncover latent system states rather than advance pragmatic task completion (by extension, gig workers probing surge boundaries; Lee et al., 2015). +- **Conditions for High-Validity Learning:** Kahneman and Klein (2009) and Rahmandad & Gary (2023) on ecological validity, unambiguous feedback, and temporal latency. +- **The Mediated Distortion:** How platforms withhold their rules, by their own account to prevent gaming and on Rahman's (2019) reading to keep control, while noise and delayed payoffs, whatever their cause, cripple human learning loops (Chen, Mislove, & Wilson, 2015; Lee et al., 2015). +- Key citations (added by Claude, from the verified card library): Weick (1988, pp. 305–307); Kirsh & Maglio (1994, pp. 513–515); Kahneman & Klein (2009, p. 524); Rahmandad & Gary (2023, pp. 13, 17); Brehmer & Nählinder (2004); Smith & Hancock (1995, p. 141). Learning on platforms: Cook et al. (2021; working-paper version); Heiland (2023); Rosenblat & Stark (2016). Monitoring against acting: Hutchins (1995, pp. 283, 285); Dismukes & Berman (2010). +- Literature to add (Claude; third-round research, audited): (1) Intent. Platforms withhold their rules; the stated reason is to prevent gaming, and Rahman (2019 dissertation, pp. 128–129; the 2021 ASQ article not yet read) also reads the secrecy as keeping control with the platform. No source shows intent behind noise or delay: the one traced case, the jitter in Uber's surge prices, was a bug by Uber's account (Chen, Mislove, & Wilson, 2015). The "Mediated Distortion" bullet should be split accordingly. (2) Strategic classification models people adapting to a decision rule (Hardt et al., 2016); where the rule is hidden, people infer it from other people's outcomes (Ghalme et al., 2021; Bechavod et al., 2022), so learning is social as well as enacted. (3) Recourse: Ustun, Spangher, & Liu (2019); Karimi et al. (2022). (4) What drivers could and could not learn by acting (Lee et al., 2015); collective sensemaking and resistance (Kellogg, Valentine, & Christin, 2020; Cameron & Rahman, 2022). (5) The counter-literature: people do learn markets by experience (List, 2003; Cook et al., 2021), subject to attrition (Seru, Shumway, & Stoffman, 2010). + +## Section 6: Decision Experience Design (DXD): Candidate Structural Conventions + +- **Convention A: Strategic Offloading:** + - *Mechanism:* Making strategic manipulation structurally unnecessary via interface-level obvious strategy-proofness (Li, 2017; Rees-Jones & Skowronek, 2018). + - *Implementation:* The interface visually proves dominant-strategy truthfulness, reducing game-theoretic second-guessing. + - *Limitations:* User inertia and residual suspicion in high-stakes environments; about four in ten participants still deviate under the best presentation tested (Gonczarowski, Heffetz, Ishai, & Thomas, 2024). +- **Convention B: Radical Feedback Immediacy:** + - *Mechanism:* Collapsing the temporal and causal gap between choice and mediator state response. + - *Implementation:* Real-time visual heatmaps, dynamic constraint curves, and immediate micro-signals replacing terminal logs. + - *Limitations:* Platform resistance to surrendering tactical information rents; frequent feedback on noisy outcomes lowers performance (Larson, List, & Metcalfe, 2016; Kluger & DeNisi, 1996), so validity comes before speed. +- **Functional Separation:** Convention A governs market clearing and allocation matching; Convention B governs real-time execution and dispatch environments. +- Key citations (added by Claude, from the verified card library): Convention A: Li (2017, §5.3; working-paper version); Rees-Jones & Skowronek (2018); Ashlagi & Gonczarowski (2018); Gonczarowski, Heffetz, & Thomas (2023; theory only in the version read); Arteaga et al. (2022) on where the removed task goes; Kling et al. (2012). Convention B has no tested intervention behind it yet: Kahneman & Klein (2009, p. 524) for the condition; Shneiderman (1983); Harper et al. (2015); Kuhl et al. (2023), which did not replicate its own gain. Against bounded control beating free control: Dietvorst et al. (2018). +- Literature to add (Claude; third-round research, audited): Convention A. Gonczarowski, Heffetz, Ishai, & Thomas (2024; N = 542) is the experiment on presentation: a menu description raised understanding (71% against 58% and 56%) more than it raised straightforward play (59%, against 56% for the traditional description and 48% for the control), so about four in ten still deviate under the best presentation tested. Masuda et al. (2022): stating that the mechanism is strategy-proof raised sincere bidding from 20.6% to 46.9%. Convention B. The evidence puts validity before speed. Real-time feedback helps on a steady quantity the person controls (Tiefenbeck et al., 2018). On noisy, revisable choices it harms: traders given frequent price feedback invested about 25% less in the risky asset and earned about 35% less (Larson, List, & Metcalfe, 2016), and more than a third of feedback interventions lower performance (Kluger & DeNisi, 1996, 1998). "Radical feedback immediacy" should be restated as valid and plain first, then quick. Platform-run tests exist: a driver leaderboard (Ye et al., 2022; +1.66% weekly revenue) and a seller dashboard (Bar-Gill, Brynjolfsson, & Hak, 2024; +3.6% revenue). No controlled test of upfront fares, destination displays or surge maps was found. + +## Section 7: Institutional Custody: Who Enforces the Convention? + +| Institutional Realm | Primary Enforcement Authority | Historical / Structural Mechanism | Core Structural Vulnerability | +| --- | --- | --- | --- | +| **Courts** | Statutory mandates & judicial oversight | Procedural rules; judge disqualification under conflicts (*Tumey v. Ohio*) | High procedural friction; systematic bias toward repeat players (Galanter, 1974) | +| **Dealer FX Markets** | Voluntary industry coalitions | Conduct codes (FX Global Code) | Absence of binding statutory force; vulnerable to benchmark manipulation (FCA, 2014) | +| **Digital Platforms** | Private corporate governance | Unilateral terms of service (ToS); proprietary algorithms | Direct profit extraction conflicts; broad proprietor discretion (Suzor, 2018) | + +- **Theoretical Conclusion on Enforcement: The conventions of the written page were enforced by ruler and church, institutions with standing over the scribe. Algorithmic platforms operate with extractive objectives. DXD conventions are unlikely to emerge through voluntary corporate self-regulation; they require statutory mandates (e.g., EU Digital Services Act, Platform-to-Business Regulation, Platform Work Directive 2024/2831) or collective bargaining pressure.** +- Key citations (added by Claude, from the verified card library): Courts: Tumey v. Ohio (1927); Mathews v. Eldridge (1976); 28 U.S.C. § 2073; Galanter (1974, pp. 97–103); Tyler (1988, p. 131); Colvin (2011). Dealer FX: Global Foreign Exchange Committee (2024), Principles 7, 9, 36, 39, 40; FCA (2014); FSB (2014); NYDFS (2015). Platforms: Suzor (2018); Klonick (2018, pp. 1664–1665); Regulation (EU) 2019/1150, Art. 3; Regulation (EU) 2022/2065, Arts. 17, 20, 53–54; Seattle Office of Labor Standards (2024); Oversight Board (2022, p. 9); Douek (2024); Nagaraj Rao et al. (2026). Interested owner: Jensen & Meckling (1976; source copy needs a check). +- Literature to add (Claude; third-round research, audited): (1) Statute that goes beyond disclosure. Directive (EU) 2024/2831 on platform work, in force since 1 December 2024 with transposition due by 2 December 2026: a human must take suspension and termination decisions (Art. 10(5)); a decision that infringes a worker's rights must be rectified within two weeks or compensated (Art. 11(3)); a discriminating system is to be modified or discontinued if appropriate (Art. 10(3)); some data processing is banned (Art. 7); worker representatives are consulted with an expert (Art. 13). It gives no veto. (2) Co-determination: BetrVG §87(1) no. 6 binds the employer through arbitration. (3) Standards bodies bind no implementer (IAB, RFC 8890). (4) Governance models and their limits: Gorwa (2019); Balkin (2016); Khan & Pozen (2019). (5) The Carolingian sources listed under Section 2 show the page's conventions were enforced by ruler and church, which supports this section and replaces the first sentence of the conclusion. (6) A test the section could adopt, which is Claude's inference and not a finding: can the owner change the convention alone? + +## Section 8: Testing the Irreducible Triad: Experimental Boundary Conditions + +- **Formalizing Irreducibility:** Measuring system coordination through Integrated Information (Phi) to gauge triadic non-decomposability into isolated dyadic pairs, with the measure, state coding, lag and partition fixed in advance and validated on the lab's Boolean models. +- **Multi-Agent Laboratory Architecture:** Two interacting human participants (Party A and Party B) coordinating exchange through an algorithmic mediator (M) optimizing a distinct objective function (cf. Casella, Friedman, & Perez Archila, 2024, whose mediator maximises agreement and shows its rule). +- **Experimental Treatments:** + - *Control:* Baseline black-box matching with terminal payout. + - *Treatment 1 (Literacy / Disclosure):* Complete textual disclosure of matching heuristics, rules, and historical audit logs. + - *Treatment 2 (DXD / Strategic Offloading):* Interface providing structural strategy-proof representations. + - *Treatment 3 (DXD / Radical Feedback Immediacy):* Continuous real-time visual constraint signaling, with a discrete-time arm carrying the same information. +- **Measured Dependent Variables:** Pareto coordination efficiency, error-correction latency, surplus distribution, and strategic manipulation rates. +- Key citations (added by Claude, from the verified card library): Nearest existing designs: Li (2017); Rees-Jones & Skowronek (2018); Hohenstein et al. (2023); Terrucha et al. (2025); Zhu et al. (2026); Liu (2026). Disclosure arm: Matsuo et al. (2026); Poursabzi-Sangdeh et al. (2021). Feedback arm: Rahmandad & Gary (2023). Irreducibility: the lab's own integrated-information results (dissertation Papers 2 and 3), which are computed on Boolean models and have not yet been run on human participants. No experiment in the library tests any convention in a triad with a behavioural outcome. +- Literature to add (Claude; third-round research, audited): (1) The closest precedent is Casella, Friedman, & Perez Archila (working paper, 2024): two people bargain through a computer mediator that maximises the chance of agreement, with its rule shown on screen every round and no condition that varies what is shown. No study found gives the mediator an objective of its own or varies what participants are told about it. (2) Related results: Normann & Sternberg on an algorithm among human firms; Ishowo-Oloko et al. (2019) on disclosing a bot's identity; Masuda et al. (2022). (3) Integrated information on human groups: Engel & Malone (2018; 68 groups, a proxy measure, r = .37 with collective intelligence, no mediator). Barrett et al. (2026, preprint) write that Phi "has not been computed on any real physical system", only proxies. The section should name the measure, the coding of behaviour into states, the lag, the partition, and the lab's Boolean models as the validation set. (4) Design additions: a discrete-time arm with the same information as the continuous-feedback arm; a decision on whether participants are told the mediator's payoffs; the deception norm that an undescribed rule is acceptable and a false statement is not (Charness, Samek, & van de Ven, 2022); power at the level of the triad; joint surplus preregistered as the primary outcome. + +## Section 9: Boundary Conditions, Limitations, and Research Agenda + +- **Historical Scope Limits:** Distinguishing static orthographic reform from dynamic algorithmic learning models. +- **Conceptual Disambiguation:** Differentiating sociological algorithmacy (triadic sociotechnical navigation) from an earlier, undefined use in informatics teaching (e.g., Oberschelp, 1995). +- **Future Empirical Horizons:** Extending DXD beyond commercial matching platforms to clinical decision triage, algorithmic bail systems, and autonomous multi-agent networks. +- Key citations (added by Claude, from the verified card library): Limits of the precedent: Vicente & Rasmussen (1992); Scaife & Rogers (1996); Street (2003); Scribner & Cole (1981; pages unconfirmed). Earlier printed use of the word: Oberschelp (1995, p. 121). Unread and needed from a library: Saenger (1997) and Parkes (1992) beyond the confirmed passages; Vatri (2012); Gavrilov (1997); Burnyeat (1997). +- Literature to add (Claude; third-round research, audited): limits the new research adds. Integrated information is untested on human triads. The nudge literature classes re-representation with informational interventions (Mertens et al., 2022, as corrected). The claim that working disclosures are re-representations is the paper's argument and no source states it. Two reviews of human-algorithm strategic interaction were not opened (March, 2021; Chugunova & Sele, 2020), so the statement that no triadic disclosure experiment exists is limited to the sources searched. + +# Part IV: Immediate Manuscript Execution Checklist + +- **Archival Verification:** Verify physical page references and exact effect sizes in Saenger (1997) and Rayner et al. (1998). +- **Theoretical Drafting:** Draft Section 4 focusing on the category error of disclosure and the critique of enterprise insider-risk compliance models. +- **Formal Parameterization:** Define the explicit payoff matrices and Integrated Information (Phi) boundary metrics for the triadic laboratory experiment in Section 8. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v2.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v2.md new file mode 100644 index 00000000..37d55f80 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v2.md @@ -0,0 +1,783 @@ +# The Space Between Decisions: Word Separation as a Guide to Designing for Algorithmacy + +> **Version 2, 2026-10-13.** This is the author's full draft of 2026-10-13, repaired by Claude at his +> instruction ("rehab it and create a new file for the updated version"). The author's wording is kept +> wherever the sources carry it. Claude changed sentences only where a source contradicted them, where a +> citation did not match the published record, or where the review (`REVIEW_v2_2026-10-13.md`) found an +> argument that would not survive a referee. The reference list is rebuilt from the project's verified +> bibliography and from Crossref. Two sections at the end are Claude's and are not part of the paper: +> "Changes from version 1" and "Still unverified". The author's read-aloud is the last gate. + +## Abstract + +Human factors engineering has a standing answer to a task that only the trained can perform: change the representation of the task environment and leave the person as they are. This paper takes the oldest case of that principle, the medieval scribes' move from continuous script (*scriptio continua*) to spaces between words, as a guide to designing for algorithmic mediation. People increasingly coordinate through algorithmic third parties—in ride-hailing, labor markets, commerce, and content governance—that read both sides of an exchange and commit outcomes neither side controls. The prevailing design and regulatory remedy is disclosure: explanations, published rules, and audit logs, exemplified by a recent proposal to govern frontier models as "insider risks." I argue that disclosing rules and rationales commits a category error by applying a literacy-based intervention—handing the person a text and asking for a reader—to an interactive coordination problem. Navigating irreducible triads draws instead on algorithmacy, a sensibility exercised in deciding and taking in what returns. Integrating distributed cognition, visual search, and mechanism design, I formulate Decision Experience Design (DXD), specify two structural conventions—strategic offloading and ecological feedback immediacy—examine what institutions can enforce them, and set out a four-claim research agenda. + +## 1. Introduction + +Human factors has a standing answer to a task that only the trained can do: change the representation of the task and leave the person as they are. Zhang and Norman (1994) gave the principle its name—the representational effect, in which "different isomorphic representations of a common formal structure can cause dramatically different cognitive behaviors" (p. 88)—and Hutchins (1995) showed it at work in a cockpit, where movable markers on the airspeed indicator let pilots reach "the same results" by "judgments of spatial proximity" (p. 282) that they would otherwise reach by reading, remembering, and comparing numbers. The oldest instantiation of this principle sits on the written page. Antiquity and early medieval manuscripts recorded text in continuous script (*scriptio continua*), which on Saenger's (1997) account left readers to find the boundaries between words by sounding the text out. Beginning with Irish scribes in the late seventh century and spreading across Continental scriptoria by the eleventh, copyists inserted blank spaces between words (Saenger, 1997). The space added no semantic information, since every letter was already on the page, yet it changed what a reader had to do with the letters; modern readers deprived of word spacing lose roughly half their reading rate (Rayner, Fischer, & Pollatsek, 1998). + +That representational answer has not reached the decisions people now make through algorithmic intermediaries. Riders and drivers, buyers and sellers, applicants and employers, and creators and audiences increasingly coordinate across a third-party system that evaluates both sides and commits outcomes that neither party directly controls. When these intermediaries are opaque, regulatory mandates and HCI design interventions alike default to disclosure: publishing rules, generating post-hoc explanations, and logging decisions for inspection. Yet across empirical evaluations, disclosure often fails to improve human decision-making and sometimes degrades it. A repayment disclosure on credit-card statements increased delinquency among some borrowers (Seira, Elizondo, & Laguna-Müggenburg, 2017); a clear model with few features left users less able to detect and correct the model's mistake on an unusual case (Poursabzi-Sangdeh et al., 2021); and having read a community's posted rules went with lower perceived fairness among users whose posts were removed (Jhaver, Appling, Gilbert, & Bruckman, 2019). Explanations do not improve decision performance on average (Schemmer et al., 2022), and across 370 effect sizes in human–AI studies the presence of an explanation made no significant difference (Vaccaro, Almaatouq, & Malone, 2024). Where explanations or alerts do help—as in automated driving takeovers (Matsuo, Liu, Hiraoka, & Wada, 2026) or cognitive forcing interfaces (Buçinca, Malaya, & Gajos, 2021)—they operate inside dyadic human–machine loops. In neither of those settings does the algorithm issue a determination binding two parties before either acts. + +The failure is not an empirical anomaly; it is a category error. A disclosure of rules and rationales hands the person a text—a rule, a rationale, a record—and so asks for a reader, making it an intervention in literacy. Deciding through an adaptive third party draws on something else: algorithmacy, which I define as the sensibility to navigate irreducible triads, as literacy is the sensibility that reading and writing require. A remedy built for readers adds to the reading of the people it means to help and leaves their decision where it was. It also settles, without argument, who holds the remedy. Satya Nadella's (2026) proposal to govern frontier models as "insider risks" illustrates the enterprise endpoint of this logic, asking that every meaningful model action leave "tamper-proof human readable evidence," that an authorized person be able to halt a model in mid-task, and that organizations "separate the supply of intelligence from the authority over it." Each provision serves the deploying organization's internal control, whereas the people on whom the model acts enter the framework only once, as recipients of a disclosure after something has failed. Unlike medieval scribes, who answered to a church and court that ordered books made correct for their readers, a digital platform that owns the interface between two parties optimizes for objectives of its own under terms of service that reserve "broad, unfettered discretion to platform owners" (Suzor, 2018). + +The focus and contribution of this paper is to shift the design of mediated decision environments from literacy-based disclosure to algorithmacy-based re-representation. Treating the history of word separation as an analytic guide rather than a general history of literacy (Saenger, 1997, p. x), I restate the scribes' move in the vocabulary of distributed cognition and visual search, distinguishing structural re-representation from both disclosure and training. I then argue that in triadic systems, people come to understand a mediator not by reading about it, but by acting through it and taking in what returns—a process of enacted sensemaking and epistemic action (Kirsh & Maglio, 1994; Weick, 1988) that succeeds only where a valid environment supplies "feedback that is both rapid and unequivocal" (Kahneman & Klein, 2009, p. 524). From that premise, I formulate Decision Experience Design (DXD) and specify two candidate conventions that do for mediated decisions what word spacing did for the line of text: strategic offloading, which removes a tactical calculation from the mechanism and makes its removal plain on the interface, and ecological feedback immediacy, which returns valid, low-latency state changes as the person acts. + +The remainder of the paper proceeds in eight sections. I first review HCI and CSCW scholarship on transparency, trust calibration, and "revealing design" (§2), restate the scribal precedent in human-factors terms (§3), and show why disclosures of rules and rationales fail when transplanted from dyadic tools to triadic mediators (§4). I then define algorithmacy as enacted sensemaking in irreducible triads (§5) and articulate the two DXD conventions, each paired with a design opening (§6). Finally, I introduce a unilateral-modification test to compare convention enforcement across courts, dealer foreign-exchange markets, and platforms (§7), set out a four-claim research agenda anchored by a multi-agent experimental protocol (§8), and state the framework's scope (§9). + +## 2. Background: Transparency, Explanation, and Legibility in HCI + +### 2.1 The Transparency and Trust-Calibration Turn in HCI and CSCW + +HCI has spent three decades operating under the assumption that opacity is an informational defect to be repaired by exposure. When early collaborative filtering algorithms began making recommendations, researchers observed that users resisted suggestions whose origin they could not discern. The initial remedy was straightforward: attach an explanation showing neighboring user ratings or item similarities (Herlocker, Konstan, & Riedl, 2000; Knijnenburg et al., 2012). As statistical learning models grew more complex, this logic expanded into explainable artificial intelligence (XAI), producing feature-attribution summaries, counterfactual explanations, descriptions of model behavior, and content-moderation notices (Wischnewski, Krämer, & Müller, 2023; Cabrera, Perer, & Hong, 2023; Jhaver et al., 2019; Ehsan et al., 2021). The implicit cognitive model was classical: if an interface can render the internal state, training weights, or decision boundaries of an automated system visible, the human operator will form an accurate mental model, calibrate their trust appropriately, and make better decisions (Lee & See, 2004; Körber, 2018). + +Empirical reality has repeatedly unsettled this optimism. Instead of calibration, transparency interventions often produce a paradox: exposing internal system mechanics increases what the user must process without improving decision quality, and can induce misplaced reliance (Parasuraman & Manzey, 2010). When people are shown explanations for automated recommendations, the explanations can raise acceptance of the recommendation whether or not it is correct (Bansal et al., 2021; Chen, Liao, Wortman Vaughan, & Bansal, 2023), and explanations reduce overreliance only under particular conditions, such as a hard task with an explanation that is cheap to check (Vasconcelos et al., 2023). Prolonged reliance on automated aids shifts the human operator from active control toward passive monitoring, with costs to vigilance and to manual skill (Metzger & Parasuraman, 2001; Casner, Geven, Recker, & Schooler, 2014). + +Recent investigations into human collaboration with large language models and generative agents report related problems. Text-generation tools impose an output-inspection cost (Mozannar, Bansal, Fourney, & Horvitz, 2024; Simkute et al., 2024). Operators must read through fluent, probabilistically generated text to catch subtle errors or misalignment, and knowledge workers report reduced critical engagement when they do (Lee et al., 2025). The interface presents an uninterrupted token stream, leaving the user to check for errors after the fact. + +### 2.2 Situated Rationale vs. Post-Hoc Audit Trails and "Revealing Design" + +Recognizing that post-hoc textual explanations fail to support real-time execution, contemporary interaction design has begun exploring in-flow friction and situated scaffolding. Buçinca, Malaya, and Gajos (2021) showed that cognitive forcing functions—such as requiring users to formulate an independent choice before viewing an AI recommendation—reduced overreliance compared with simple explanations on the questions where the AI was wrong. Similarly, Disley and Khan (2026) reframe system representations as an active "design surface," showing that selectively suppressing a concept in a model can introduce productive cognitive friction. In software development environments, annotation made at the point of use preserves context that is hard to reconstruct afterward (Horvath et al., 2022). Dialogue systems similarly depend on economy in interaction and discoverable error-correction rather than lengthy explanatory turns (Pucci, Piro, & Matera, 2027). + +Critical scholars in CSCW and HCI have taken this critique a step further, interrogating the structural relations embedded in commercial platforms. Özçetin and Wiltse (2023) critique how Terms of Service (ToS) rely on static, legalistic text to govern dynamic, asymmetric relations. They advocate for "revealing design," an approach aimed at making platform entanglements "legible" to participants. + +Yet revealing design, for all its critical insight, stops short of escaping the literacy paradigm. To make an entanglement legible is still to treat the user as a critical reader. It presumes that if we give a gig worker or digital consumer an expressive visual diagram or a richer conceptual breakdown of their position, we have equipped them to act. But understanding that one is entangled does not change the actions required to accept a dispatch, price a product, or contest a penalty. A person sitting at an interface does not need a more transparent picture of their subordination; they need an interaction surface whose mechanics alter what they have to calculate to coordinate their lives. + +### 2.3 From Human–Machine Dyads to Irreducible Algorithmic Triads + +The deeper reason why transparency paradigms fail is structural: research on transparency and explanation remains oriented around dyadic interaction. In the classic cognitive engineering literature, interaction is modeled as a two-node loop: a human operator paired with a machine aid (Sheridan & Verplank, 1978; Parasuraman, Sheridan, & Wickens, 2000). The pilot monitors the autopilot; the clinician reviews a diagnostic classifier; the driver takes over from an automated steering system (Matsuo et al., 2026). In a dyad, the machine's role is auxiliary. Its explicit functional purpose is to support, inform, or substitute for the human operator in executing an external task. + +``` +Classic HCI Dyad: +[Human Operator] <====== Interaction Loop ======> [Machine / Tool] +``` + +Coordination across contemporary digital platforms does not fit this schema. In ride-hailing, digital labor platforms, electronic financial markets, and content clearinghouses, the architecture is triadic. Two human counterparts—a rider and a driver, a freelance specialist and an employer, a buyer and a seller—coordinate through a third node: an active algorithmic mediator (M). HCI and CSCW have studied settings with more than one person and a system, including AI-mediated communication, where suggested replies change how two correspondents write and regard each other (Hohenstein et al., 2023). The transparency literature has not taken the same step. + +``` +Irreducible Algorithmic Triad: + ┌─────────────────────────────┐ + │ Algorithmic Mediator │ + │ (Optimizing Objective J_M) │ + └──────────────┬──────────────┘ + ▲ ▲ + Mediation Probe│ │Boundary Reaction + │ │ + ┌────────────────────┴─┴────────────────────┐ + │ │ + ┌─────────┴─────────┐ ┌─────────┴─────────┐ + │ Party A │◄──Coordination Move──►│ Party B │ + │ (Worker / Driver) │ │ (Client / Rider) │ + └───────────────────┘ └───────────────────┘ +``` + +This mediator is not an advisory instrument assisting one party. It reads inputs from both sides, filters information symmetrically or asymmetrically, applies an independent objective function (J_M)—such as platform profit maximization, throughput balancing, or dispute minimization—and commits a binding allocation that neither party directly controls (Lee, Kusbit, Metsky, & Dabbish, 2015). Party A's payoff depends directly on Party B's move, yet Party A can only perceive, probe, or influence Party B through the transformation imposed by M; in that sense the triad does not decompose into two independent dyads. Transplanting dyadic transparency tools into a triadic mediator yields incoherence: explaining to Party A why the algorithm made an allocation does not eliminate the tactical necessity for Party A to second-guess what Party B will do through that same mediator. + +## 3. Re-Representation: The Scribal Precedent in Human-Factors Terms + +### 3.1 The Page Before and After Word Separation + +To understand how an interaction surface can be transformed without changing the underlying information, we can look to the history of the written page. For more than a millennium across Greco-Roman antiquity and the early Middle Ages, texts were transcribed in *scriptio continua*. Words were strung together in unbroken lines without spaces between them. + +Because *scriptio continua* provided no spatial demarcations between lexical units, Saenger (1997) argues, reading was laborious: a reader identified words by pronouncing or subvocalizing the characters, relying on sound to reconstruct lexical boundaries. Classicists dispute how heavy that burden was for trained readers (Heilmann, 2020), and the argument here does not depend on settling it. + +Between the late seventh and eleventh centuries, an uncoordinated ergonomic change unfolded across monastic scriptoria. Scribes in Ireland and Anglo-Saxon Britain learned Latin as a foreign language, not as a spoken one, and Insular copyists began inserting gaps between words (Saenger, 1997; Parkes, 1992). Over the next three centuries, this spatial convention reached Continental scriptoria, becoming standard by the eleventh century alongside developments in punctuation and layout (Parkes, 1992; Saenger, 1997). + +Before drawing human-factors conclusions from this shift, we must set strict historiographical boundaries. Spacing did not single-handedly generate modern literacy. Mass literacy required centuries of institutional change, legal documentation, administrative centralization, and the invention of movable type (Clanchy, 1993; Cressy, 1980; Eisenstein, 1969). Nor was silent reading impossible in the ancient world: classical accounts show that skilled readers could and did read silently long before word spacing appeared (Knox, 1968; Gavrilov, 1997; Burnyeat, 1997; Heilmann, 2020). + +The scribal convention did something more specific: it altered the sensory mechanics of the task. The space between words added no semantic data. Every letter on the page remained unchanged. Yet by altering the visual layout of the page, scribes altered what a reader had to do to extract meaning. Saenger's own statement of the effect is that word separation and related devices in Insular books "provided shortcuts for achieving the reading skills that an elite among the ancients had mastered only through a prolonged and arduous grammatical apprenticeship" (1997, p. 85). + +### 3.2 Cognitive Offloading, Visual Search, and Perceptual Substitution + +In the language of cognitive engineering, the introduction of inter-word spacing was an act of environmental cognitive offloading (Risko & Gilbert, 2016; Kirsh, 1995) that exploited the representational effect (Zhang & Norman, 1994). Zhang and Norman demonstrated that across formally isomorphic tasks—such as variations of the Tower of Hanoi—the external representation dictates the cognitive rules of engagement. When constraints are embedded directly into physical space, they do not need to be held in working memory or checked through conscious mental inference. + +Hutchins (1995) documented the same principle on the flight deck: pilots set movable speed bugs on the rim of the airspeed indicator, and the comparison of speeds is externalized into the spatial proximity between a needle and a marker. + +``` +Representational Shift on the Page: + +Scriptio Continua: +[L][E][X][I][C][A][L][B][O][U][N][D][A][R][Y] +▲ ▲ +└──────── Working Memory / Phonological ────┘ + Reconstruction Overhead + +Word Spacing: +[L][E][X][I][C][A][L] [B][O][U][N][D][A][R][Y] + ▲ ▲ ▲ + └── Visual ┴── Eye ───┘ + Saccade Targeting +``` + +On this reading, *scriptio continua* left boundary detection to the reader, who had to hold ambiguous strings of sound while testing competing segmentations, and word spacing substituted visual segmentation for that inference. The interpretation is this paper's; no study has measured it for medieval readers. + +The visual system processes spatial structure across the field in parallel. Under Wolfe's (2021) Guided Search model, bottom-up feature salience guides the allocation of focal attention. By carving white channels between character clusters, spacing gives the oculomotor system targets for the next saccade, and removing the spaces changes where the eyes land as well as how fast words are identified (Rayner et al., 1998). Deprived of these spatial markers, modern readers lose roughly half their reading rate (Rayner et al., 1998). Spacing let the geometry of the page do the work of lexical segmentation. + +### 3.3 Cross-Orthographic Asymmetries and the Structural Disanalogy + +The interaction between orthography and human vision reveals that representational offloading is not a simple, universal trick. Cross-linguistic reading experiments uncover a directional asymmetry. When spaces are removed from an alphabetic script like English, reading performance drops sharply. Yet when spaces are inserted into scripts that are normally written without them—such as Chinese or Thai—the effect is mixed (Bai et al., 2008; Winskel, Radach, & Luksaneeyanawin, 2009; Chantavarin & Leung, 2025). + +Skilled readers of Chinese read word-spaced text no faster than normal unspaced text, 5,239 against 5,221 milliseconds a sentence (Bai et al., 2008). Thai readers fixate a spaced word for less time and take longer over the sentence (Winskel et al., 2009; Chantavarin & Leung, 2025). Learners do not reliably gain: child learners of Chinese read spaced text more slowly and less accurately (Bassetti & Masterson, 2012), and for adult second-language learners the difference between spaced and unspaced text was not significant (Cui, 2023). A convention pays once readers have come to depend on it. An ergonomic intervention is not an abstract virtue; it must match the visual search affordances of the medium to the cognitive tasks of the user. + +Here lies the decisive structural disanalogy between the medieval page and the modern platform. A manuscript page is passive, static, and cooperative. Once ink dries on parchment, the page has no agency. It does not reorder its words, change its layout when the reader blinks, or extract behavioral surplus from the reader's gaze. Furthermore, the scribes who made it worked under institutional discipline. + +Carolingian rulers and church authorities ordered scriptoria to produce correct books: Charlemagne's letter *De litteris colendis* measures monasteries' writing against a standard of correctness, the *Admonitio generalis* of 789 (c. 72) orders books corrected and copied with care, and Alcuin's *Carmen* XCIV instructs scribes to mark sense units for the lector. Westwell (2018) cautions that uniformity imposed from above can be overstated. None of these texts concerns word separation, which the Insular scribes had adopted earlier. What they show is who stood behind the page: an authority with standing over the scribe. + +An algorithmic intermediary is dynamic, adaptive, and governed by the commercial incentives of its owner. The platform controls the interface, observes user probes, and can rewrite its ranking rules, pricing multipliers, and allocation weights. Where the medieval page was a stable, cooperative surface held in check by institutional authority, the platform interface is an active, contested site. + +## 4. Disclosure as a Literacy-Based Intervention + +### 4.1 Taxonomizing Disclosure and the Law-and-Economics Critique + +When algorithmic mediation produces unfair or exploitative outcomes, public regulators, legal bodies, and design standards routinely fall back on disclosure. Mandating that platforms explain themselves has become the universal compromise. To understand why this fails, we must disaggregate "disclosure" into four distinct targets: + +``` + ┌────────────────────────────────────────────────────────┐ + │ Target 1: Presence Disclosure │ + │ ("An automated algorithm is in use") │ + └───────────────────────────┬────────────────────────────┘ + │ + ┌───────────────────────────▼────────────────────────────┐ + │ Target 2: Objective Function Disclosure │ + │ ("System maximizes driver utilization and margin") │ + └───────────────────────────┬────────────────────────────┘ + │ + ┌───────────────────────────▼────────────────────────────┐ + │ Target 3: Formal Rule Property Disclosure │ + │ ("Mechanism is obviously strategy-proof") │ + └───────────────────────────┬────────────────────────────┘ + │ + ┌───────────────────────────▼────────────────────────────┐ + │ Target 4: Full Textual Rule & Rationale │ + │ (Terms of Service, XAI feature cards, audit logs) │ + └────────────────────────────────────────────────────────┘ +``` + +1. Target 1 (Presence): Notifying users that an algorithm or bot is active (Ishowo-Oloko et al., 2019). +2. Target 2 (Objective Function): Stating the objective the mediator is optimizing (e.g., platform margin maximization vs. wait-time minimization). +3. Target 3 (Formal Rule Property): Informing users of an actionable structural property of the mechanism (e.g., that an auction is strategy-proof; Masuda et al., 2022). +4. Target 4 (Full Textual Rule and Rationale): Providing full terms of service, technical documentation, or post-hoc feature-attribution scores explaining an individual allocation. + +The category error named in this paper lies in Target 4. Target 3 behaves differently: a one-line statement of a property changes what people do (§6.2), and it does so as a signal that a calculation is unnecessary, with nothing to study. + +The legal and behavioral critique of mandated disclosure explains why Target 4 repeatedly fails. Ben-Shahar and Schneider (2011) reviewed disclosure regimes across finance, healthcare, and consumer contracts and concluded that mandated disclosure, among the most common regulatory techniques, regularly fails because the people it addresses do not read disclosures, do not understand them when they do, and do not use them when they understand (p. 665). Psychology explains much of the gap between what is disclosed and what is done with it (Loewenstein, Sunstein, & Golman, 2014), and disclosure works only where it is embedded in the routines of users and disclosers alike (Fung, Graham, & Weil, 2007). + +The disclosures that do alter outcomes tend to be simple, low-cost signals, such as letter grades for hygiene posted in restaurant windows, and they act in good part on the discloser, who changes its conduct to avoid publishing a poor grade (Ben-Shahar & Schneider, 2011, pp. 743–748; Loewenstein et al., 2014, p. 403). Ben-Shahar and Schneider hedge even this case. + +### 4.2 The Empirical Record Across Finance, XAI, and Platform Governance + +The failure of disclosure is documented across empirical human–AI interaction. In a meta-analysis of XAI interventions, explanations yielded no reliable improvement in human decision performance (Schemmer et al., 2022, reporting a pooled standardized mean difference of 0.07, 95% CI [−0.15, 0.30]), and across 370 effect sizes the presence of an explanation was not a significant moderator of human–AI performance (Vaccaro et al., 2024). + +Field experiments in consumer finance show that disclosure can backfire. When a Mexican bank added messages to credit-card statements, a message stating the interest rate changed nothing, and a message stating how many months the balance would take to repay raised delinquency among two groups of clients by 0.011 points, 7.4% of the mean (Seira et al., 2017, pp. 292, 295). + +In machine learning interfaces, Poursabzi-Sangdeh et al. (2021) showed that presenting users with a clear model with few features reduced their ability to detect and correct the model's mistake on an unusual case, in three of four experiments. The authors conjecture that the visible detail overloaded participants. + +In platform governance, Jhaver et al. (2019) surveyed 907 Reddit users whose posts had been removed. Users who had read the community's rules beforehand judged the removal as less fair (β = −.314), while users who received an explanation of the removal judged it fairer (β = .384). The rulebook and the situated reason pulled in opposite directions. + +Furthermore, Nouwens et al. (2020) demonstrated in cookie consent interfaces that layout alone moves behavior: removing the opt-out button from the first page raised consent by 22 to 23 percentage points. Users respond to the spatial architecture of the interface. + +### 4.3 Enterprise Compliance as the Zenith of Literacy Dogma + +The apex of literacy-based governance is on display in modern enterprise AI proposals. Satya Nadella's (2026) post, "Models as Insider Risks in the Super Intelligence Era," provides a case study. Confronted with the unpredictability of autonomous frontier agents, Nadella proposes treating AI systems like insiders who may be mistaken or compromised, and sets out seven principles. Three bear on what a person is given: + +1. Auditability: every meaningful model action must leave "tamper-proof human readable evidence." +2. Interruption: an authorized person must be able to pause or shut down a model in mid-task, which he likens to an emergency brake. +3. Separation: organizations must "separate the supply of intelligence from the authority over it." + +This framework addresses internal principal–agent control for the deploying organization, in the tradition of insider-threat practice (CERT National Insider Threat Center, 2022). Its other principles—model diversity, independent controls, independent auditability—are likewise held by the organization. + +``` +Enterprise Insider Risk Architecture: +┌────────────────────────────────────────────────────────┐ +│ Deploying Firm │ +│ ┌──────────────────┐ ┌───────────────────┐ │ +│ │ Intelligence │ │ Executive │ │ +│ │ Provider (Model) │ │ Authority / Halt │ │ +│ └────────┬─────────┘ └─────────▲─────────┘ │ +│ │ │ │ +│ └──────► Human-Readable ────────┘ │ +│ Audit Logs │ +└───────────────────────────┬────────────────────────────┘ + │ Binding Triadic Allocation + ▼ +┌────────────────────────────────────────────────────────┐ +│ External Transacting Parties │ +│ (Gig Worker / Consumer / Displaced User) │ +│ │ +│ Status: Receive disclosure of an incident after │ +│ an adverse determination has occurred │ +└────────────────────────────────────────────────────────┘ +``` + +For the external humans transacting across the model, Nadella's framework changes little. The gig worker who is deactivated, the small business whose account is suspended, or the consumer denied a transaction does not hold the brake. They cannot interrupt the model in mid-inference. The framework promises them "timely disclosure" when the system fails or is compromised. + +For the person on the receiving end of an algorithmic determination, a disclosure after the fact is an autopsy report. It asks them to read the record of what happened to them and leaves them unable to reshape the interaction. + +## 5. Algorithmacy: Epistemic Action in Irreducible Triads + +### 5.1 Defining Algorithmacy + +I define algorithmacy as the behavioral sensibility required to navigate irreducible algorithmic triads: the capacity to coordinate with a counterpart through an active, optimizing mediator whose internal rules shift dynamically. + +``` + ┌─────────────────────────────┐ + │ Historical Transition │ + └──────────────┬──────────────┘ + │ + ┌──────────────────────────────────────┴──────────────────────────────────────┐ + ▼ ▼ +┌─────────────────────────────────┐ ┌─────────────────────────────────┐ +│ Literacy │ │ Algorithmacy │ +├─────────────────────────────────┤ ├─────────────────────────────────┤ +│ Mode: Evaluative, textual, │ │ Mode: Enacted, exploratory, │ +│ reflective │ │ epistemic │ +│ Medium: Static pages, documents,│ │ Medium: Dynamic, optimizing, │ +│ post-hoc logs │ │ triadic interfaces │ +│ Cognitive Demands: Working │ │ Cognitive Demands: Dynamic │ +│ memory, phonological loop, │ │ calibration, boundary │ +│ semantic parsing │ │ probing, feedback integration │ +│ Stance: Reader │ │ Stance: Decision-maker │ +└─────────────────────────────────┘ └─────────────────────────────────┘ +``` + +Algorithmacy is distinct from dyadic "AI literacy." The latter typically refers to conceptual knowledge: understanding what neural networks are, how training data shapes bias, or how to compose structured text prompts in chat windows. + +Nor does it match the one earlier printed use of the word that I have found. Oberschelp (1995), writing on informatics teaching, urges that "we have to preserve logicacy and algorithmacy by continuously training simple mental activities without artificial tools" (p. 121). He does not define the term; in context it names the skill of carrying out algorithmic steps by hand. + +Algorithmacy as I use it is an active coordination competency. Just as literacy allows a reader to traverse an author's text without demanding that the reader build a printing press, algorithmacy allows an individual to formulate choices, infer hidden constraints, and protect their welfare across an algorithmic intermediary without requiring them to inspect its source code. + +### 5.2 Enacted Sensemaking and Epistemic Action + +In an irreducible triad, people do not come to understand a mediator by reading about it. They understand it by acting through it and observing what happens. This process is grounded in Weick's (1988) account of enacted sensemaking: in ambiguous, dynamic environments, "people think by acting" (p. 305). Actors take action, create consequences in the environment, and then look at the results to infer what kind of system they are dealing with. Weick adds a warning that belongs in any design that relies on this: action that clarifies a situation can also intensify it. + +In cognitive science, exploratory moves of this kind are known as epistemic actions (Kirsh & Maglio, 1994). Kirsh and Maglio showed that in Tetris, players do not rotate and translate falling pieces only to place them (pragmatic action). They also move them to learn something sooner than they could work it out in the head. Epistemic actions are "physical actions that make mental computation easier, faster, or more reliable" (p. 513). + +``` +Triadic Epistemic Probing: +[Worker / User] ──── (1) Epistemic Probe: Move / Micro-bid ────► [Algorithmic Mediator] + │ + (2) Evaluates Probe + Against Objective J_M + │ +[Worker / User] ◄─── (3) Boundary Reaction / State Update ──────────────┘ + │ +(4) Dynamic Recalibration of Strategy +``` + +Transacting workers and consumers engage in epistemic actions every day, and extending Kirsh and Maglio's concept to them is this paper's move. Ride-hail drivers do not learn dispatch and surge behavior from company documentation. They form working theories from what the system does in response to their own moves, and they trade those theories in online forums (Lee et al., 2015; Kellogg, Valentine, & Christin, 2020). Workers pool what they observe and push back together (Cameron & Rahman, 2022). Formal models of strategic classification under opacity describe the same situation: when a decision rule is hidden, people infer it from the observed outcomes of others (Ghalme et al., 2021). Learning a mediator is social as well as enacted. + +These probes are epistemic actions designed to extract information from an opaque mediator. Transacting parties navigate algorithmic environments through continuous behavioral experimentation. + +### 5.3 High-Validity Learning Environments vs. Mediated Distortion + +For enacted sensemaking to succeed, the environment must meet the two conditions Kahneman and Klein (2009) set for skilled intuition: an environment regular enough to be learned, and an opportunity to learn it through "feedback that is both rapid and unequivocal" (p. 524). + +When feedback is delayed or noisy, calibration suffers, and in different ways. Rahmandad and Gary (2023) found that delay pushes decision-makers toward options that pay off quickly and can drive convergence to inefficient strategies, while noise does not produce that bias and instead pulls choices toward the middle of the strategy space. + +Algorithmic intermediaries often fail these conditions. To analyze mediated distortion, we must separate two distinct drivers: + +1. Strategic Rule Secrecy: Platforms conceal evaluation criteria, ranking parameters, and matching formulas. Platform operators justify this secrecy as necessary to prevent "gaming," and Rahman (2019) reads it as also keeping control of a stratified workforce with the platform. +2. Technical Noise and Latency: Algorithmic mediators introduce jitter, batching delays, and stochastic allocations. Not all distortion is intended. The rapid fluctuation in Uber's surge multipliers measured by Chen, Mislove, and Wilson (2015) was, by Uber's own account, a bug. No source I have found shows that a platform introduces noise or delay on purpose. + +Regardless of whether distortion originates in corporate strategy or system architecture, its cognitive result is the same. When an intermediary combines hidden rules with noisy, delayed feedback, it removes the conditions for learning by acting. The user probes the system, receives an ambiguous response later, and cannot distinguish whether their action failed, the market shifted, or the algorithm penalized them. + +Under these conditions, algorithmacy cannot flourish. The user is left relying on folk theories and defensive gaming. + +## 6. Decision Experience Design (DXD): Two Candidate Conventions + +### 6.1 Architectural Framing via Norman's Gulfs of Action + +Decision Experience Design (DXD) shifts the design agenda from writing textual explanations to restructuring the decision interface. DXD grounds its interventions in Norman's (1988) framework of action: + +``` + ┌─────────────────────────────────────────┐ + │ The Transacting Human │ + └──────────────┬──────────────────▲───────┘ + │ │ + Formulate Intent │ │ Evaluate State + │ │ + ┌────────────────▼┐ ┌┴────────────────┐ + │ Gulf of │ │ Gulf of │ + │ Execution │ │ Evaluation │ + └────────────────┬┘ └▲────────────────┘ + │ │ + Bridged by: │ │ Bridged by: + Strategic │ │ Ecological + Offloading │ │ Feedback + (Convention A) │ │ Immediacy + │ │ (Convention B) + │ │ + ┌──────────────▼──────────────────┴───────┐ + │ Algorithmic Mediator │ + └─────────────────────────────────────────┘ +``` + +1. The Gulf of Execution: The cognitive distance between a user's goal and the actions permitted by the interface. Convention A (Strategic Offloading) narrows the Gulf of Execution by eliminating tactical gaming calculations from the transaction mechanism and showing on the display that they have been eliminated. +2. The Gulf of Evaluation: The cognitive distance between the system's state and the user's ability to assess that state. Convention B (Ecological Feedback Immediacy) narrows the Gulf of Evaluation by delivering valid, noise-filtered, low-latency feedback as the user formulates their choice. + +### 6.2 Convention A: Strategic Offloading (Removing a Tactical Task and Showing It Is Gone) + +In classical mechanism design, an allocation mechanism is strategy-proof if reporting one's true valuation or preference is a dominant strategy: no participant can gain by lying, shading their bid, or misrepresenting their constraints. Yet theoretical strategy-proofness rarely translates into sincere human behavior. In a laboratory Vickrey auction, only about a fifth of bids were sincere without advice (Masuda et al., 2022). + +To address this, Li (2017) formalized obviously strategy-proof (OSP) mechanisms. A mechanism is obviously strategy-proof if, for any deviating strategy, the best possible outcome under the deviation is no better than the worst possible outcome under the truth. OSP mechanisms can be recognized as safe by agents who lack contingent reasoning capabilities, and in Li's experiments far fewer games ended off the dominant-strategy outcome under the obviously strategy-proof version of the same rule. + +Convention A embeds this property in the visual interface. It removes the tactical calculation—eliminating the need to second-guess what other participants are doing—and uses the interface to make that removal obvious. + +Recent experimental economics calibrates the convention and its limit. Gonczarowski, Heffetz, Ishai, and Thomas (2024; N = 542) tested whether describing a matching mechanism as a personalized menu, in place of the traditional description of its algorithm, increased understanding and straightforward play. Understanding of strategy-proofness rose to 71%, against 58% and 56% under the other descriptions. Straightforward play reached 59% under the menu description, against 56% under the traditional description, a difference that is not significant, and 48% under a control that conveyed almost nothing. + +The boundary is plain: under the best description tested, about four in ten participants still did not play straightforwardly, and better understanding moved play only a little. + +This is why Masuda et al. (2022) matter: a plain statement that sincere bidding is a dominant strategy raised sincere bids from 20.6% to 46.9%, while efficiency rose by about one percentage point. Convention A must therefore pair the structural elimination of the tactical task with an interface that visibly confirms the task is gone, and its designers should expect behavior to change more than outcomes. + +``` +Convention A Interface Transformation: + +Old Gaming Interface: +┌────────────────────────────────────────────────────────┐ +│ Enter Your Hourly Rate Bid: [ $ 45.00 / hr ] │ +│ │ +│ [ Warning: Other workers may undercut your bid. ] │ +│ [ Consider platform commission (20%) and surge index. ]│ +└────────────────────────────────────────────────────────┘ +Result: Tactical anxiety; complex shade calculation. + +DXD Strategic Offloading Interface: +┌────────────────────────────────────────────────────────┐ +│ Guaranteed Clearing Menu: │ +│ │ +│ ● [ $ 52.00 / hr ] ── Guaranteed Dispatch in Tier 1 │ +│ ○ [ $ 48.00 / hr ] ── Guaranteed Dispatch in Tier 2 │ +│ │ +│ [ Strategy Notice: Ascending Clock Active. ]│ +│ [ Sincere preference revelation is dominant. No worker ]│ +│ [ can undercut your allocated tier. ]│ +└────────────────────────────────────────────────────────┘ +Result: Tactical calculations offloaded to mechanism display. +``` + +**Design Opening: What if...?** +What if freelance labor dispatch and procurement platforms abandoned open-ended bid-entry boxes—which force workers to calculate speculative price-shading—and replaced them with ascending-clock menus that display clearing tiers where compensation is guaranteed and tactical undercutting is made impossible by design? + +### 6.3 Convention B: Ecological Feedback Immediacy (Valid and Plain First, Rapid Second) + +Convention B addresses the Gulf of Evaluation by shortening the temporal and causal distance between an actor's exploratory probe and the mediator's response. In classical graphical user interfaces, direct manipulation relies on immediate visual feedback: when a user drags a file, the icon tracks the cursor in real time (Shneiderman, 1983). In algorithmic mediation, that immediate coupling is routinely broken. A user submits a bid, posts a job, or positions a car, and waits in an informational vacuum while backend algorithms batch, calculate, and commit. + +However, in designing Convention B, ecological validity strictly precedes speed. Exposing high-frequency feedback on noisy or volatile data can damage human decision-making. Larson, List, and Metcalfe (2016) found that professional traders shown prices second by second invested about 25% less in the risky asset and earned about 35% less than traders shown prices every four hours, a pattern they attribute to myopic loss aversion. + +Similarly, Kluger and DeNisi (1996, 1998) showed in a meta-analysis that more than one-third of feedback interventions reduce performance. Providing uncurated, rapid feedback on a volatile algorithm hands the user a stream of noise. + +Therefore, Convention B specifies feedback that is valid and noise-filtered first, plain second, and low-latency third. It provides pre-commit sandboxes and boundary contours that let participants test how the mediator will treat a move before committing to an irreversible choice. No experiment has yet tested this convention in a mediated triad; it follows from the conditions in §5.3 and is offered as a hypothesis. + +``` +Convention B: Pre-Commit Boundary Feedback + +Opaque Batched Flow: +[ Formulate Move ] ──► [ Commit Action ] ──► [ Silent Batching Delay ] ──► [ Irreversible Outcome ] + (Zero visibility) (Post-hoc shock) + +DXD Pre-Commit Sandbox Flow: +[ Formulate Move ] + │ + ▼ (Exploratory Probe) +┌────────────────────────────────────────────────────────┐ +│ Pre-Commit Counterfactual Contour Sandbox │ +│ │ +│ Candidate Parameter: [ X = 42 ] │ +│ Mediator Boundary Reaction: [ Cleared: P(Match) = 88%│ +│ Noise-Filtered Confidence Band: [ ± 3% ] │ +│ │ +│ [ Visual contour shows parameter boundary line ] │ +└────────────────────────────────────────────────────────┘ + │ + ▼ (Informed Commitment) +[ Commit Final Action ] ──► [ Deterministic Execution ] +``` + +**Design Opening: What if...?** +What if gig drivers and marketplace sellers were provided with noise-filtered contour maps that showed, as they dragged a price or radius slider, how the platform's clearing mechanism would allocate orders across candidate locations before they committed to a shift or contract? + +## 7. Institutional Custody: Who Can Hold a Convention? + +### 7.1 The Unilateral-Modification Custody Test + +A representational convention is not an abstract piece of interface code; it is a promise between the mediator and the transacting parties. For a convention to hold, participants must trust that the interface will not deceive them. This introduces the problem of custody: who ensures that the interface remains faithful? + +To evaluate institutional stability, I formulate the Unilateral-Modification Custody Test: + +*Can the party that operates the intermediary unilaterally alter, suspend, or distort the representational convention without the informed consent of the transacting parties?* + +This test is grounded in the economic theory of agency costs (Jensen & Meckling, 1976). When an intermediary has an independent financial interest in the transaction, its operator faces continuing incentives to alter the representation to capture surplus from transacting users. + +### 7.2 Comparative Analysis Across Three Institutional Realms + +To evaluate how custody functions in practice, I examine three institutional environments: + +| Institutional Realm | Primary Enforcement Authority | Structural & Procedural Mechanism | Unilateral-Modification Vulnerability | +| --- | --- | --- | --- | +| Courts of Law | Constitutional oversight & statutory civil procedure | Rules Enabling Act (28 U.S.C. § 2073); judicial disqualification (*Tumey v. Ohio*, 1927); procedural due process (*Mathews v. Eldridge*, 1976) | Low (rule-bound): a judge cannot alter evidentiary or procedural rules mid-trial to advantage one party | +| Dealer FX Markets | Voluntary industry code; regulators after the fact | FX Global Code (GFXC, 2024): Principle 17 on last look; Principles 9, 36, and 39 on what is disclosed to clients; Principle 40 on agreeing changes to terms | Moderate (soft law): the Code is voluntary; benchmark manipulation was punished by regulators (FCA, 2014; NYDFS, 2015) | +| Digital Platforms | Private corporate ownership | Unilateral Terms of Service (Suzor, 2018; Klonick, 2018); proprietary interface updates; advisory boards (Oversight Board, 2022) | High (unilateral): the owner sets and changes rankings, algorithms, and interface conventions, subject to statute where one applies | + +In judicial systems, procedural conventions are insulated from unilateral tampering. Under the Rules Enabling Act (28 U.S.C. § 2073), federal procedural rules are prescribed through committees whose meetings are noticed to the public. The Supreme Court ruled in *Tumey v. Ohio* (1927) that a judge who has a direct pecuniary interest in the outcome violates due process. Repeat litigants hold systemic advantages (Galanter, 1974), and employees fare worse in arbitration than in court (Colvin, 2011), but the judge cannot rewrite procedural conventions during closing arguments. + +In foreign-exchange dealer markets, banks act as both liquidity providers and counterparties on their own platforms. "Last look" lets a dealer hold a client's order briefly and decide whether to accept it; the recitals of a 2015 consent order state that one bank used the practice to reject trades that had moved against it and did not tell clients why (NYDFS, 2015). + +Regulators fined five banks £1.1 billion in 2014 for manipulating benchmark rates (FCA, 2014; FSB, 2014). The industry's code followed. The FX Global Code asks dealers to disclose their last-look practice and to agree substantial changes to terms before a transaction (GFXC, 2024, Principles 17 and 40), and adherence to it is voluntary. A soft-law convention holds as far as reputation and the threat of a regulator carry it. + +Digital platforms fare worst on the custody test. In all fourteen sets of platform terms that Suzor (2018) examined, the terms "provided broad, unfettered discretion to platform owners," and appeals were thin when Klonick (2018) studied them. Bodies such as the Meta Oversight Board (2022) have narrow remits that do not reach matching algorithms, dynamic pricing, or dispatch interfaces. + +Technical standards bodies cannot fill the gap: adoption of their standards is voluntary, and end users take little part in writing them (Nottingham, 2020). If a platform can rewrite its matching conventions with a software update, a designer's promise of "strategic offloading" can be withdrawn whenever margins compress. + +### 7.3 External Compulsion and Statutes Beyond Disclosure + +Because platforms fail the custody test, Decision Experience Design conventions cannot survive through voluntary designer ethics alone. They require external legal compulsion. + +First-generation platform regulations relied largely on disclosure and procedure. The EU Platform-to-Business Regulation (Regulation (EU) 2019/1150) requires fifteen days' notice of changes to terms (Art. 3) and a description of the main ranking parameters (Art. 5), and it states that providers are not required to disclose their algorithms (Art. 5(6)). The Digital Services Act (Regulation (EU) 2022/2065) requires a statement of reasons for restrictions (Art. 17) and an internal complaint-handling system (Art. 20), and gives users a complaint to a regulator and a claim for damages (Arts. 53–54). These measures give notice, reasons, and recourse. They leave the owner free to set the convention. + +Recent statutory frameworks are beginning to move beyond disclosure toward structural intervention: + +1. The EU Platform Work Directive (Directive (EU) 2024/2831; in force since 1 December 2024; transposition due by 2 December 2026): bans certain processing of personal data (Art. 7); requires a system found to discriminate to be modified or discontinued where appropriate (Art. 10(3)); requires that decisions to restrict, suspend, or terminate an account be taken by a human (Art. 10(5)); requires rectification within two weeks, or compensation, where a decision infringes a worker's rights (Art. 11(3)); and establishes consultation of worker representatives, assisted by an expert (Art. 13). It gives workers no veto. +2. German Works Constitution Act (BetrVG § 87(1) no. 6): grants works councils co-determination over technical systems designed to monitor employee behavior or performance. An employer cannot introduce or modify such a system without works-council agreement or a binding decision of the conciliation committee (§ 87(2)). + +These statutes provide the institutional framework needed to sustain DXD conventions. Just as Carolingian scriptoria were directed by royal and ecclesiastical authority to produce correct books, digital platforms will maintain honest representational conventions only when external law removes their power to alter them at will. + +## 8. A Research Agenda for Triadic Decision Design + +### 8.1 Four Falsifiable Claims + +In the manner of a framework and research agenda (Terveen & McDonald, 2005), I set out four claims, each stated so that data could count against it: + +- **Claim 1 (Irreducibility).** In a triad comprising Party A, Party B, and Mediator M, a proxy measure of integrated information computed on the joint behavior of all three will exceed the same measure computed on each person–mediator pair, and the excess will shrink when M's determinations are made independent of one party's moves. + - *Operationalization.* Integrated information (Φ) has so far been applied to human groups only by proxy: Engel and Malone (2018) computed a proxy on who was active when in 68 groups and found it correlated with collective intelligence (r = .37), in groups with no mediator, and Barrett et al. (2026) note that Φ itself has not been computed on a real physical system. I therefore fix the measure, the coding of behavior into discrete states, the time lag, and the partition in advance, validate them on small Boolean models where exact values can be computed, and treat the application to human triads as untested. +- **Claim 2 (Representational conventions outperform textual disclosure).** In a triadic allocation game where Mediator M optimizes an independent objective (J_M ≠ J_A, J_B), representing the system through Convention A (menu-based strategic offloading) produces higher joint surplus and less tactical distortion than full textual rule disclosure. + - *Operationalization.* The closest precedent is Casella, Friedman, and Perez Archila (2026), where two participants bargain through a computer mediator that maximizes the chance of agreement and shows its rule every round. I give M a margin of its own and compare joint surplus between rulebook arms and menu arms. +- **Claim 3 (Validity governs the sign of feedback immediacy).** The effect of feedback immediacy on error correction depends on mediator noise. Under low noise, continuous feedback speeds convergence; under high noise, continuous feedback degrades choices, and discrete filtered feedback yields higher joint surplus. + - *Operationalization.* I test the boundary conditions of Larson, List, and Metcalfe (2016) by comparing continuous boundary feedback against discrete-round filtered feedback across high-jitter and low-jitter mediator regimes. +- **Claim 4 (Custodial drift degrades sincere play).** Conventions subject to unannounced mediator parameter drift decay into defensive, manipulative play, whereas conventions the mediator has committed not to change within a session sustain sincere coordination over repeated rounds. + +### 8.2 Five-Arm Multi-Agent Experimental Protocol + +To test these claims, I specify a controlled laboratory experiment. The protocol follows experimental-economics norms: no false statements to participants (Charness, Samek, & van de Ven, 2022), power calculated at the triad level (the independent unit of observation is the three-node triad, not the individual participant), and joint surplus preregistered as the primary outcome variable. + +``` +Experimental Architecture (5-Arm Triadic Protocol): + + [ Participant A ] <═══════> [ Mediator M ] <═══════> [ Participant B ] + ║ + Varying Interface / Representation Condition: + ═════════════════════════════════╬═════════════════════════════════ + Arm 1: Control (Black-Box Interface; Terminal Payout Only) + Arm 2: Disclosure (Full Written Rulebook, Margin Formulas, Audit Logs) + Arm 3: DXD Convention A (Ascending Clock / Strategic Offloading Menus) + Arm 4a: DXD Convention B (Continuous Low-Latency Boundary Contours) + Arm 4b: Discrete Control (Identical Filtered Feedback Between Rounds) +``` + +1. Arm 1 (Control — Black-Box): Participants submit raw bids or demands through an unadorned interface, receiving only terminal round payoffs with no explanatory rules. +2. Arm 2 (Literacy — Full Disclosure): Participants receive complete documentation of M's objective function, matching formulas, and an itemized post-round audit log explaining how both inputs were scored. +3. Arm 3 (DXD Convention A — Strategic Offloading): Participants interact with personalized menus where dominant strategies are pre-cleared, visually confirming that strategic undercutting has been removed. +4. Arm 4a (DXD Convention B — Continuous Immediacy): Participants receive real-time, pre-commit visual boundary contours that update as they adjust their sliders, showing clearing odds before commitment. +5. Arm 4b (Discrete Control for Convention B): Participants receive identical noise-filtered boundary data, delivered only at discrete intervals between rounds, isolating the effect of temporal immediacy from informational content. + +## 9. Conclusion and Scope Boundaries + +### 9.1 Synthesis + +Human factors engineering established long ago that when an interface demands heroic human performance, the solution is not to train the person or hand them an instruction manual; the solution is to re-represent the task environment. The scribes of the early Middle Ages demonstrated this on the manuscript page. By replacing continuous script with spaces between words, they changed what a reader had to do with the same letters. They did not alter human biology; they altered the visual geometry of the page. + +Today, human–computer interaction and algorithmic regulation stand where reading stood before word separation. Faced with opaque, complex algorithmic intermediaries, our default response has been to demand more literacy: more documentation, more transparency dashboards, more explainability cards, and more post-hoc audit logs. + +This is a category error. Rulebooks and rationales treat human beings as readers of algorithms rather than deciders within mediated environments. Navigating irreducible triads requires algorithmacy—an enacted sensibility that operates through epistemic probing and the integration of valid feedback. + +Decision Experience Design (DXD) provides an alternative path. By embedding strategic offloading and ecological feedback immediacy directly into the interaction surface, DXD looks for the modern equivalent of the space between words: an interaction architecture that lets people coordinate across algorithms with clarity, autonomy, and agency. + +### 9.2 Explicit Scope Boundaries + +To ensure theoretical precision, this framework establishes five explicit scope boundaries: + +1. Analytical Guide, Not Historical Determinism: Word separation is used strictly as an ergonomic guide for distributed cognition, not as a monocausal explanation of historical literacy or intellectual development (Street, 2003; Scribner & Cole, 1981). +2. Asymmetric Empirical Maturity: Convention A rests on experiments in mechanism design that show a real but partial effect (Li, 2017; Gonczarowski et al., 2024). Convention B and the measurement of triadic integrated information remain hypotheses awaiting the five-arm protocol. +3. Choice Architecture Boundary: DXD is a neighbor of nudging and boosting and must be placed against them (Hertwig & Grüne-Yanoff, 2017). A nudge alters how choices are presented within an unchanged game; DXD changes what the mechanism requires a person to calculate and shows that it has. The nudge literature classes re-presentation of information with informational interventions, the weaker kind (Mertens, Herberz, Hahnel, & Brosch, 2022), and DXD's claim to differ rests on Claim 2. +4. Theoretical Status of Working Disclosures: The argument that working disclosures succeed insofar as they function as simple perceptual signals is this paper's theoretical claim, not an established consensus in prior literature. +5. Institutional Boundary: Interaction conventions cannot withstand commercial pressure without external institutional anchors. Where the Carolingian page was sustained by church and court, modern algorithmic conventions will require statutory co-determination and enforceable legal mandates. + +## References + +Bai, X., Yan, G., Liversedge, S. P., Zang, C., & Rayner, K. (2008). Reading spaced and unspaced Chinese text: Evidence from eye movements. *Journal of Experimental Psychology: Human Perception and Performance, 34*(5), 1277–1287. https://doi.org/10.1037/0096-1523.34.5.1277 + +Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. (2021). Does the whole exceed its parts? The effect of AI explanations on complementary team performance. In *Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems* (pp. 1–16). ACM. https://doi.org/10.1145/3411764.3445717 + +Barrett, A. B., Milinkovic, B., Mediano, P. A. M., Rosas, F. E., Bor, D., Barnett, L., & Seth, A. K. (2026). *Integrated information theory: The good, the bad and the misunderstood* [Preprint]. + +Bassetti, B., & Masterson, J. (2012). Effects of removing morphemic information and adding interword spacing on reading in Chinese experienced and inexperienced readers. *Reading and Writing, 25*(9), 2291–2314. https://doi.org/10.1007/s11145-012-9358-7 + +Ben-Shahar, O., & Schneider, C. E. (2011). The failure of mandated disclosure. *University of Pennsylvania Law Review, 159*, 647–749. + +Bundesministerium der Justiz. (n.d.). *Betriebsverfassungsgesetz (BetrVG)* [Works Constitution Act], § 87. + +Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. *Proceedings of the ACM on Human-Computer Interaction, 5*(CSCW1), Article 188. https://doi.org/10.1145/3449287 + +Burnyeat, M. F. (1997). Postscript on silent reading. *The Classical Quarterly, 47*(1), 74–76. https://doi.org/10.1093/cq/47.1.74 + +Cabrera, Á. A., Perer, A., & Hong, J. I. (2023). Improving human-AI collaboration with descriptions of AI behavior. *Proceedings of the ACM on Human-Computer Interaction, 7*(CSCW1), 1–21. https://doi.org/10.1145/3579612 + +Cameron, L. D., & Rahman, H. (2022). Expanding the locus of resistance: Understanding the co-constitution of control and resistance in the gig economy. *Organization Science, 33*(1), 38–58. https://doi.org/10.1287/orsc.2021.1557 + +Casella, A., Friedman, E., & Perez Archila, M. (2026). On the fragility of mediation: Theory and experimental evidence. *Journal of the European Economic Association, 24*(3), 801–850. https://doi.org/10.1093/jeea/jvaf049 + +Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W. (2014). The retention of manual flying skills in the automated cockpit. *Human Factors, 56*(8), 1506–1516. https://doi.org/10.1177/0018720814535628 + +CERT National Insider Threat Center. (2022). *Common sense guide to mitigating insider threats* (7th ed.). Software Engineering Institute, Carnegie Mellon University. + +Chantavarin, S., & Leung, T. T.-C. (2025). The effects of interword spacing and morphological complexity in reading Thai: An eye-tracking study. *Language and Cognition, 17*. https://doi.org/10.1017/langcog.2025.10028 + +Charness, G., Samek, A., & van de Ven, J. (2022). What is considered deception in experimental economics? *Experimental Economics, 25*(2), 385–412. https://doi.org/10.1007/s10683-021-09726-7 + +Chen, L., Mislove, A., & Wilson, C. (2015). Peeking beneath the hood of Uber. In *Proceedings of the 2015 Internet Measurement Conference* (pp. 495–508). ACM. https://doi.org/10.1145/2815675.2815681 + +Chen, V., Liao, Q. V., Wortman Vaughan, J., & Bansal, G. (2023). Understanding the role of human intuition on reliance in human-AI decision-making with explanations. *Proceedings of the ACM on Human-Computer Interaction, 7*(CSCW2), 1–32. https://doi.org/10.1145/3610219 + +Clanchy, M. T. (1993). *From memory to written record: England 1066–1307* (2nd ed.). Blackwell. + +Colvin, A. J. S. (2011). An empirical study of employment arbitration: Case outcomes and processes. *Journal of Empirical Legal Studies, 8*(1), 1–23. https://doi.org/10.1111/j.1740-1461.2010.01200.x + +Cressy, D. (1980). *Literacy and the social order: Reading and writing in Tudor and Stuart England*. Cambridge University Press. + +Cui, Y. (2023). Eye movements of second language learners when reading spaced and unspaced Chinese texts. *Frontiers in Psychology, 14*, Article 783960. https://doi.org/10.3389/fpsyg.2023.783960 + +Directive (EU) 2024/2831 of the European Parliament and of the Council of 23 October 2024 on improving working conditions in platform work. *Official Journal of the European Union*. + +Disley, S., & Khan, H. (2026). Designing cognitive friction with concept suppression in large language models. *Human–Computer Interaction*. Advance online publication. https://doi.org/10.1080/07370024.2026.2708617 + +Ehsan, U., Liao, Q. V., Muller, M., Riedl, M. O., & Weisz, J. D. (2021). Expanding explainability: Towards social transparency in AI systems. In *Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems* (pp. 1–19). ACM. https://doi.org/10.1145/3411764.3445188 + +Eisenstein, E. L. (1969). The advent of printing and the problem of the Renaissance. *Past & Present, 45*(1), 19–89. https://doi.org/10.1093/past/45.1.19 + +Engel, D., & Malone, T. W. (2018). Integrated information as a metric for group interaction. *PLOS ONE, 13*(10), Article e0205335. https://doi.org/10.1371/journal.pone.0205335 + +Financial Conduct Authority. (2014, November 12). *FCA fines five banks £1.1 billion for FX failings and announces industry-wide remediation programme* [Press release]. + +Financial Stability Board. (2014). *Foreign exchange benchmarks: Final report*. + +Fung, A., Graham, M., & Weil, D. (2007). *Full disclosure: The perils and promise of transparency*. Cambridge University Press. + +Galanter, M. (1974). Why the "haves" come out ahead: Speculations on the limits of legal change. *Law & Society Review, 9*(1), 95–160. https://doi.org/10.2307/3053023 + +Gavrilov, A. K. (1997). Techniques of reading in classical antiquity. *The Classical Quarterly, 47*(1), 56–73. https://doi.org/10.1093/cq/47.1.56 + +Ghalme, G., Nair, V., Eilat, I., Talgam-Cohen, I., & Rosenfeld, N. (2021). *Strategic classification in the dark* [Preprint]. arXiv. + +Global Foreign Exchange Committee. (2024). *FX Global Code: A set of global principles of good practice in the foreign exchange market* (December 2024 update). + +Gonczarowski, Y. A., Heffetz, O., Ishai, G., & Thomas, C. (2024). *Describing deferred acceptance and strategyproofness to participants: Experimental analysis* (Working Paper No. 33020). National Bureau of Economic Research. https://doi.org/10.3386/w33020 + +Heilmann, J. (2020). Reading early New Testament manuscripts: Scriptio continua, "reading aids", and other characteristic features. In *Material aspects of reading in ancient and medieval cultures* (pp. 177–196). De Gruyter. https://doi.org/10.1515/9783110639247-011 + +Herlocker, J. L., Konstan, J. A., & Riedl, J. (2000). Explaining collaborative filtering recommendations. In *Proceedings of the 2000 ACM Conference on Computer Supported Cooperative Work* (pp. 241–250). ACM. https://doi.org/10.1145/358916.358995 + +Hertwig, R., & Grüne-Yanoff, T. (2017). Nudging and boosting: Steering or empowering good decisions. *Perspectives on Psychological Science, 12*(6), 973–986. https://doi.org/10.1177/1745691617702496 + +Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J., & Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. *Scientific Reports, 13*, Article 5487. https://doi.org/10.1038/s41598-023-30938-9 + +Horvath, A., Liu, M. X., Hendriksen, R., Shannon, C., Paterson, E., Jawad, K., Macvean, A., & Myers, B. A. (2022). Understanding how programmers can use annotations on documentation. In *CHI Conference on Human Factors in Computing Systems* (pp. 1–16). ACM. https://doi.org/10.1145/3491102.3502095 + +Hutchins, E. (1995). How a cockpit remembers its speeds. *Cognitive Science, 19*(3), 265–288. https://doi.org/10.1207/s15516709cog1903_1 + +Ishowo-Oloko, F., Bonnefon, J.-F., Soroye, Z., Crandall, J., Rahwan, I., & Rahwan, T. (2019). Behavioural evidence for a transparency–efficiency tradeoff in human–machine cooperation. *Nature Machine Intelligence, 1*(11), 517–521. https://doi.org/10.1038/s42256-019-0113-5 + +Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs and ownership structure. *Journal of Financial Economics, 3*(4), 305–360. https://doi.org/10.1016/0304-405X(76)90026-X + +Jhaver, S., Appling, D. S., Gilbert, E., & Bruckman, A. (2019). "Did you suspect the post would be removed?": Understanding user reactions to content removals on Reddit. *Proceedings of the ACM on Human-Computer Interaction, 3*(CSCW), Article 192. https://doi.org/10.1145/3359294 + +Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. *American Psychologist, 64*(6), 515–526. https://doi.org/10.1037/a0016755 + +Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. *Academy of Management Annals, 14*(1), 366–410. https://doi.org/10.5465/annals.2018.0174 + +Kirsh, D. (1995). The intelligent use of space. *Artificial Intelligence, 73*(1–2), 31–68. https://doi.org/10.1016/0004-3702(94)00017-U + +Kirsh, D., & Maglio, P. (1994). On distinguishing epistemic from pragmatic action. *Cognitive Science, 18*(4), 513–549. https://doi.org/10.1207/s15516709cog1804_1 + +Klonick, K. (2018). The new governors: The people, rules, and processes governing online speech. *Harvard Law Review, 131*, 1598–1670. + +Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. *Psychological Bulletin, 119*(2), 254–284. https://doi.org/10.1037/0033-2909.119.2.254 + +Kluger, A. N., & DeNisi, A. (1998). Feedback interventions: Toward the understanding of a double-edged sword. *Current Directions in Psychological Science, 7*(3), 67–72. https://doi.org/10.1111/1467-8721.ep10772989 + +Knijnenburg, B. P., Willemsen, M. C., Gantner, Z., Soncu, H., & Newell, C. (2012). Explaining the user experience of recommender systems. *User Modeling and User-Adapted Interaction, 22*(4–5), 441–504. https://doi.org/10.1007/s11257-011-9118-4 + +Knox, B. M. W. (1968). Silent reading in antiquity. *Greek, Roman, and Byzantine Studies, 9*(4), 421–435. + +Körber, M. (2018). *Theoretical considerations and development of a questionnaire to measure trust in automation* [Preprint]. PsyArXiv. https://doi.org/10.31234/osf.io/nfc45 + +Larson, F., List, J. A., & Metcalfe, R. D. (2016). *Can myopic loss aversion explain the equity premium puzzle? Evidence from a natural field experiment with professional traders* (Working Paper No. 22605). National Bureau of Economic Research. https://doi.org/10.3386/w22605 + +Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In *Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems* (pp. 1–22). ACM. https://doi.org/10.1145/3706598.3713778 + +Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. *Human Factors, 46*(1), 50–80. https://doi.org/10.1518/hfes.46.1.50.30392 + +Lee, M. K., Kusbit, D., Metsky, E., & Dabbish, L. (2015). Working with machines: The impact of algorithmic and data-driven management on human workers. In *Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems* (pp. 1603–1612). ACM. https://doi.org/10.1145/2702123.2702548 + +Li, S. (2017). Obviously strategy-proof mechanisms. *American Economic Review, 107*(11), 3257–3287. https://doi.org/10.1257/aer.20160425 + +Loewenstein, G., Sunstein, C. R., & Golman, R. (2014). Disclosure: Psychology changes everything. *Annual Review of Economics, 6*, 391–419. https://doi.org/10.1146/annurev-economics-080213-041341 + +Masuda, T., Mikami, R., Sakai, T., Serizawa, S., & Wakayama, T. (2022). The net effect of advice on strategy-proof mechanisms: An experiment for the Vickrey auction. *Experimental Economics, 25*(3), 902–941. https://doi.org/10.1007/s10683-021-09736-5 + +*Mathews v. Eldridge*, 424 U.S. 319 (1976). + +Matsuo, R., Liu, H., Hiraoka, T., & Wada, T. (2026). An educational human machine interface providing request-to-intervene trigger and reason explanation for enhancing the driver's comprehension of ADS's system limitations. *IEEE Transactions on Human-Machine Systems, 56*(3), 650–661. https://doi.org/10.1109/thms.2026.3679748 + +Mertens, S., Herberz, M., Hahnel, U. J. J., & Brosch, T. (2022). The effectiveness of nudging: A meta-analysis of choice architecture interventions across behavioral domains. *Proceedings of the National Academy of Sciences, 119*(1), Article e2107346118. https://doi.org/10.1073/pnas.2107346118 (Correction published 2022.) + +Metzger, U., & Parasuraman, R. (2001). The role of the air traffic controller in future air traffic management: An empirical study of active control versus passive monitoring. *Human Factors, 43*(4), 519–528. https://doi.org/10.1518/001872001775870421 + +Mozannar, H., Bansal, G., Fourney, A., & Horvitz, E. (2024). Reading between the lines: Modeling user behavior and costs in AI-assisted programming. In *Proceedings of the CHI Conference on Human Factors in Computing Systems* (pp. 1–16). ACM. https://doi.org/10.1145/3613904.3641936 + +Nadella, S. (2026, October 10). *Models as insider risks in the super intelligence era* [Post]. X. https://x.com/satyanadella/status/2108931348857827686 + +New York State Department of Financial Services. (2015, November 17). *In the matter of Barclays Bank PLC, Barclays Bank PLC, New York Branch: Consent order under New York Banking Law § 44*. + +Norman, D. A. (1988). *The psychology of everyday things*. Basic Books. + +Nottingham, M. (2020). *The Internet is for end users* (RFC 8890). Internet Architecture Board. + +Nouwens, M., Liccardi, I., Veale, M., Karger, D., & Kagal, L. (2020). Dark patterns after the GDPR: Scraping consent pop-ups and demonstrating their influence. In *Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems* (pp. 1–13). ACM. https://doi.org/10.1145/3313831.3376321 + +Oberschelp, W. (1995). Developing analytical competency through informatics teaching. In D. Watson & D. Tinsley (Eds.), *Integrating information technology into education* (pp. 115–126). Springer. https://doi.org/10.1007/978-0-387-34842-1_12 + +Oversight Board. (2022). *Q1 2022 transparency report*. + +Özçetin, S., & Wiltse, H. (2023). Terms of entanglement: A posthumanist reading of Terms of Service. *Human–Computer Interaction, 40*(1–4), 171–194. https://doi.org/10.1080/07370024.2023.2281928 + +Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. *Human Factors, 52*(3), 381–410. https://doi.org/10.1177/0018720810376055 + +Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. *IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 30*(3), 286–297. https://doi.org/10.1109/3468.844354 + +Parkes, M. B. (1992). *Pause and effect: An introduction to the history of punctuation in the West*. Scolar Press. + +Poursabzi-Sangdeh, F., Goldstein, D. G., Hofman, J. M., Wortman Vaughan, J., & Wallach, H. (2021). Manipulating and measuring model interpretability. In *Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems* (Article 237). ACM. https://doi.org/10.1145/3411764.3445315 + +Pucci, E., Piro, L., & Matera, M. (2027). From tabbing to talking: LLM-driven conversational agents for web form interaction. *International Journal of Human-Computer Studies, 218*, Article 103925. https://doi.org/10.1016/j.ijhcs.2026.103925 + +Rahman, H. A. (2019). *Invisible cages: Algorithmic evaluations in online labor markets* [Doctoral dissertation, Stanford University]. + +Rahmandad, H., & Gary, M. S. (2023). Delays impair learning and can drive convergence to inefficient strategies. *Organization Science, 34*(6), 2392–2414. https://doi.org/10.1287/orsc.2020.1405 + +Rayner, K., Fischer, M. H., & Pollatsek, A. (1998). Unspaced text interferes with both word identification and eye movement control. *Vision Research, 38*(8), 1129–1144. https://doi.org/10.1016/S0042-6989(97)00274-5 + +Regulation (EU) 2019/1150 of the European Parliament and of the Council of 20 June 2019 on promoting fairness and transparency for business users of online intermediation services. *Official Journal of the European Union*. + +Regulation (EU) 2022/2065 of the European Parliament and of the Council of 19 October 2022 on a Single Market For Digital Services (Digital Services Act). *Official Journal of the European Union*. + +Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. *Trends in Cognitive Sciences, 20*(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002 + +Rules Enabling Act, 28 U.S.C. § 2073. + +Saenger, P. (1997). *Space between words: The origins of silent reading*. Stanford University Press. + +Schemmer, M., Hemmer, P., Nitsche, M., Kühl, N., & Vössing, M. (2022). A meta-analysis of the utility of explainable artificial intelligence in human-AI decision-making. In *Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society* (pp. 617–626). ACM. https://doi.org/10.1145/3514094.3534128 + +Scribner, S., & Cole, M. (1981). *The psychology of literacy*. Harvard University Press. + +Seira, E., Elizondo, A., & Laguna-Müggenburg, E. (2017). Are information disclosures effective? Evidence from the credit card market. *American Economic Journal: Economic Policy, 9*(1), 277–307. https://doi.org/10.1257/pol.20140404 + +Sheridan, T. B., & Verplank, W. L. (1978). *Human and computer control of undersea teleoperators* [Technical report]. Massachusetts Institute of Technology, Man-Machine Systems Laboratory. https://doi.org/10.21236/ada057655 + +Shneiderman, B. (1983). Direct manipulation: A step beyond programming languages. *Computer, 16*(8), 57–69. https://doi.org/10.1109/mc.1983.1654471 + +Simkute, A., Tankelevitch, L., Kewenig, V., Scott, A. E., Sellen, A., & Rintel, S. (2024). Ironies of generative AI: Understanding and mitigating productivity loss in human-AI interaction. *International Journal of Human–Computer Interaction*. Advance online publication. https://doi.org/10.1080/10447318.2024.2405782 + +Street, B. V. (2003). What's "new" in New Literacy Studies? Critical approaches to literacy in theory and practice. *Current Issues in Comparative Education, 5*(2), 77–91. + +Suzor, N. (2018). Digital constitutionalism: Using the rule of law to evaluate the legitimacy of governance by platforms. *Social Media + Society, 4*(3). https://doi.org/10.1177/2056305118787812 + +Terveen, L., & McDonald, D. W. (2005). Social matching: A framework and research agenda. *ACM Transactions on Computer-Human Interaction, 12*(3), 401–434. https://doi.org/10.1145/1096737.1096740 + +*Tumey v. Ohio*, 273 U.S. 510 (1927). + +Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. *Nature Human Behaviour, 8*(12), 2293–2303. https://doi.org/10.1038/s41562-024-02024-1 + +Vasconcelos, H., Jörke, M., Grunde-McLaughlin, M., Gerstenberg, T., Bernstein, M. S., & Krishna, R. (2023). Explanations can reduce overreliance on AI systems during decision-making. *Proceedings of the ACM on Human-Computer Interaction, 7*(CSCW1), Article 129. https://doi.org/10.1145/3579605 + +Weick, K. E. (1988). Enacted sensemaking in crisis situations. *Journal of Management Studies, 25*(4), 305–317. https://doi.org/10.1111/j.1467-6486.1988.tb00039.x + +Westwell, A. (2018). *The dissemination and reception of the Ordines Romani in the Carolingian church, c. 750–900* [Doctoral dissertation, University of Cambridge]. https://doi.org/10.17863/cam.23255 + +Winskel, H., Radach, R., & Luksaneeyanawin, S. (2009). Eye movements when reading spaced and unspaced Thai and English: A comparison of Thai–English bilinguals and English monolinguals. *Journal of Memory and Language, 61*(3), 339–351. https://doi.org/10.1016/j.jml.2009.07.002 + +Wischnewski, M., Krämer, N., & Müller, E. (2023). Measuring and understanding trust calibrations for automated systems: A survey of the state-of-the-art and future directions. In *Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems*. ACM. + +Wolfe, J. M. (2021). Guided Search 6.0: An updated model of visual search. *Psychonomic Bulletin & Review, 28*(4), 1060–1092. https://doi.org/10.3758/s13423-020-01859-9 + +Zhang, J., & Norman, D. A. (1994). Representations in distributed cognitive tasks. *Cognitive Science, 18*(1), 87–122. https://doi.org/10.1207/s15516709cog1801_3 + +--- + +## Changes from version 1 (Claude's notes; not part of the paper) + +**Reference list.** Rebuilt. Thirty-nine entries were replaced with the published record (the full table +is in `REVIEW_v2_2026-10-13.md` §2a). Entries the text never cited were removed (Arteaga et al.; Ashlagi +& Gonczarowski; Balkin; Bar-Gill et al.; Bechavod et al.; Brehmer & Nählinder; Crabtree; Dietvorst et +al.; Dix; Douek; Hardt et al.; Lee & Moray; Rees-Jones & Skowronek; Scaife & Rogers; Thorgeirsson et al.; +Vicente & Rasmussen; Wickens & Carswell). Entries were added for every work the text cites. Three +citations were removed because no record could be found or the record did not match the claim: +Kosmyna et al. (2025), Thorgeirsson et al. (2026), and Sáenz & De Russis (2027), whose only Crossref +match is a paper with a different title. + +**Statements corrected against their sources.** + +| Section | Version 1 | Version 2 | +| --- | --- | --- | +| 1, 4.2 | Repayment disclosures "increased delinquency among borrowers"; "Mexican credit card companies"; "made the debt feel insurmountable" | One bank's messages; delinquency rose for two client groups by 0.011 points (7.4% of the mean); the invented mechanism is removed | +| 1, 4.2 | "High-interpretability models impaired users' ability to correct anomalous model mistakes"; "an illusion of competence" | A clear model left users less able to correct the mistake on an unusual case, in three of four experiments; the authors' own conjecture (overload) is given | +| 4.2 | Explained removals judged less fair than unexplained ones | Rule-readers judged removal less fair (β = −.314); those given an explanation judged it fairer (β = .384) | +| 4.2 | Layout shifted consent "by orders of magnitude"; text "produced no detectable effect" | Removing the opt-out button raised consent by 22 to 23 points; the claim about text is removed | +| 1, 4.2 | "Explanations do not improve decision performance on average (Vaccaro…)" | Schemmer et al. for the average effect; Vaccaro et al. for explanation not moderating performance | +| 3.1 | Reading "confined to an elite, highly trained scribal caste"; Insular languages "shared no linguistic roots with Latin"; distinctiones and "codex page indexing" standardized by the eleventh century | Saenger's claims attributed to him and marked as disputed; Insular scribes learned Latin as a foreign language; the unsourced details are removed; his p. 85 sentence is quoted | +| 3.2 | Boundary detection in "the phonological loop"; saccade claims unsourced | Marked as this paper's interpretation; landing-position effect cited to Rayner et al. | +| 3.3 | Spaces help children and learners; spaces slow skilled Chinese readers (Bai) | Bai: no difference; Thai: faster words, slower sentences; child learners worse (Bassetti & Masterson); adult learners not significant (Cui) | +| 3.3, 7.3 | Carolingian edicts cited to Westwell; scribes "held accountable to readers" | The three primary texts named; Westwell's caution added; a sentence states that none concerns word separation | +| 4.1 | A quotation from Ben-Shahar & Schneider not found at p. 665; "a bright caloric label" | Their mechanism paraphrased from p. 665; their own example (restaurant grades) with its hedge | +| 4.1 | Only Target 4 said to fail, while the thesis condemned all disclosure | The abstract, introduction and 4.1 now locate the category error in Target 4 and say why Target 3 differs | +| 4.3 | "Three pillars"; "halt switches" in quotation marks; affected people "handed the… audit log" | Seven principles, three of which bear on a person; his emergency-brake image; affected people are promised "timely disclosure" | +| 5.1 | Oberschelp "treated algorithmacy as elementary procedural programming literacy" | His one sentence quoted (p. 121); undefined; a hand-computation skill | +| 5.2 | Drivers with a second phone driving "two blocks north"; Ghalme et al. cited for sharing screenshots | What Lee et al. and Kellogg et al. report; Ghalme et al. cited for inferring a hidden rule from others' outcomes | +| 5.3 | Kahneman & Klein quoted beyond the verified words; platforms "deliberately disrupt ecological validity"; "an engineering synchronization bug in their geospatial grid" | Only "feedback that is both rapid and unequivocal" is quoted; intent is claimed for secrecy and denied for noise; the bug is Uber's own account | +| 5.3 | Rahmandad & Gary: failure "even after hundreds of trials… superstitious heuristics" | Delay and noise have different effects, as the paper reports | +| 6.2 | "Comprehension from 56% to 71%. Sincere play increased from 48% to 59%" | Understanding 71% against 58% and 56%; play 59% against 56% (not significant) and 48% (control) | +| 6.2 | Masuda et al. "pairing a structural representation with a clear confirmation" | A statement that sincere bidding is dominant; 20.6% to 46.9%; efficiency up about one point | +| 6.3 | Larson et al. in the *Journal of Finance*; "financial traders… real-time price telemetry" | An NBER working paper on professional traders; second-by-second against four-hourly prices | +| 7.2 | Code followed by fines; Principles 7, 9, 36, 39, 40 "governing last-look and pre-hedging" | Fines in 2014–15, Code afterward; Principle 17 for last look and the others for what they cover | +| 7.2 | RFC 8890 "explicitly declare[s]" standards do not bind platforms | Standards are voluntary and end users take little part; cited to Nottingham (2020) | +| 7.3 | P2B and DSA "merely" require plain-language description, "changing nothing" | Their notice, reasons, complaint and damages provisions stated; P2B's no-algorithm-disclosure clause added | +| 7.3 | Directive article summaries without their qualifiers | "Where appropriate" (Art. 10(3)); rights infringed (Art. 11(3)); no veto | +| 8.1 | Claim 1: "strictly positive integrated information (Φ > 0) that cannot be factored" | A comparison the data could fail, with the proxy status of the measure stated | +| 8.1 | Claim 4: "verifiable cryptographic or co-governance lock-in" | A commitment not to change the convention within a session | +| 9.1 | "where reading stood before the eighth century" | "before word separation" | +| 9.2 | Convention A "rests on established empirical mechanism design"; one sentence on nudging | "A real but partial effect"; nudge and boost named as the neighbor, with Mertens et al.'s classification | + +**Structure and voice.** The roadmap is prose. Where "we" meant the author it is now "I"; the inclusive "we" that addresses the reader is kept. Convention B is "ecological +feedback immediacy" everywhere. The claim that HCI is dyadic is narrowed to transparency research, with +AI-mediated communication acknowledged. "Importantly," and "Notably," are removed. The diagrams are kept +as they were, inside code blocks; they still need redrawing as figures. + +**Not done.** The paper is not lengthened. The body runs to about 7,600 words excluding diagrams and references, against 9,000 to 17,000 in the four recent articles read from the journal. Sections +2, 5 and 7 are the ones to grow. + +## Still unverified (Claude's notes; not part of the paper) + +- **Records confirmed at Crossref, texts not read.** The sentences that cite these rest on the earlier + draft's description of them: Herlocker et al. (2000); Knijnenburg et al. (2012); Wischnewski et al. + (2023; Crossref shows the preprint, and the CHI venue is from the draft); Ehsan et al. (2021); Lee & + See (2004); Körber (2018); Bansal et al. (2021); Chen et al. (2023); Metzger & Parasuraman (2001); + Casner et al. (2014); Mozannar et al. (2024); Simkute et al. (2024); Lee et al. (2025); Horvath et al. + (2022); Pucci et al. (2027); Sheridan & Verplank (1978); Wolfe (2021); Hohenstein et al. (2023, carded + for a different point); Eisenstein (1969). +- **Terveen & McDonald (2005)** is cited as the form the agenda follows and has not been read. +- **Norman (1988)** is cited for the gulfs of execution and evaluation and has not been opened. +- **Rayner et al. (1998)** was read as an abstract only. +- **Saenger (1997) and Parkes (1992).** Saenger's quoted wording is confirmed; his pages need a print + check. Parkes is known from snippets. +- **Rahman (2019)** is the dissertation; the published article (2021) was not opened. +- **Gonczarowski et al. (2024), Larson et al. (2016)** were read as working papers. +- **Barrett et al. (2026)** is a preprint; the bibliography holds no DOI or archive number for it. +- **The interface mock-ups and their numbers** (rates, tiers, "P(Match) = 88%") are illustrations. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v3.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v3.md new file mode 100644 index 00000000..50589026 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v3.md @@ -0,0 +1,904 @@ +# The Space Between Decisions: Word Separation as a Guide to Designing for Algorithmacy + +> **Version 3, 2026-10-14.** The author's wording from version 2 is kept where a sourced fact still +> holds. The 14 October claim check corrected contradicted and overstated sentences. A later pass the +> same day retied the paper to one argument in three moves. A further pass the same day reads Grahn +> et al. (2026) and Mourtzis and Tsakalos (2026) before the argument against Nadella, and puts the +> oracy-to-literacy change after that argument. New joining sentences are marked in the +> end notes. Version 2 is unchanged, so the claim ledgers' line numbers still match it. The notes at +> the end are not part of the paper. The author's read-aloud is the last gate. + +## Abstract + +Two reviews in the first issue of *Human-Machine Interaction* state where the field stands in 2026. Grahn, Wang, Wang, and Wang (2026) move the subject from interaction to integration, and the remedies they keep are oversight, interpretable outputs, and an education in working with machines. Mourtzis and Tsakalos (2026) trace the interface from levers to screens to extended reality, and the human in their model still receives information in order to decide. Nadella (2026) offers the enterprise version of that remedy: every meaningful model action should leave "tamper-proof human readable evidence," an unending chain of content. This paper argues that the vision is literacy, carried into a triadic platform. What it asks for is the change that occurred when oracy gave way to literacy: one alteration of the representation that took the finding of a boundary off the ear and put it on the page. Decision Experience Design names two candidates for that change and sets out what would count as finding it. + +## 1. Introduction + +Grahn, Wang, Wang, and Wang (2026) and Mourtzis and Tsakalos (2026) review human–machine interaction as it stands. Grahn et al. treat the classical model as a separated operator who monitors discrete levels of automation through an external display, and they say that model strains once there is no display to monitor and no stable boundary between person and machine. They order the field by what it costs to leave the coupling. At an external interface the person steps away from a screen. At what they call Level 4b, institutionally mandated AI, the person cannot decline the system without declining a role, and their examples include work allocation on labor platforms. The principles they set down for that future are meaningful human control, a stated human benefit, inclusion of people who do not adopt the technology, foresight, and an education in collaborating with machines. One educational obligation they transfer to that societal level is civic AI literacy. Mourtzis and Tsakalos date the same arc to the industrial revolutions. The human went from operator, to overseer, to supervisor, and is now to become a colleague. The interfaces that carry the arc are levers, buttons, screens, dashboards, decision-support tools, and extended reality. In the collaborative model they draw, the extended-reality interface "provides the human with the necessary information to perform tasks and make decisions in real-time" (Section 5.4). Both reviews are current. Both still hand the person something to take in. §2 reads them against the transparency record. + +Nadella's (2026) post is the argument this paper makes against that vision, read in §2.5. He asks organizations to "separate the supply of intelligence from the authority over it," to leave "tamper-proof human readable evidence" of every meaningful model action, and to let an authorized person halt the model. Containment and the brake are real. They are the oversight Grahn et al. still ask for, and the information Mourtzis and Tsakalos still route to a human, written for the firm that deploys the model. The human job the post leaves is to read a chain. The people the model acts on enter once, as readers of a notice after a failure. We are already in a triadic platform, a rider and a driver, a buyer and a seller, a creator and an audience, and the vision on offer is still literacy. + +The paper is a search for the spaces-between-words moment in that situation. *Scriptio continua* only really makes sense read aloud, so the ear can find the boundary the eye cannot. The transition from oracy to literacy was the moment that work moved. The smoking gun, in this paper, is one change in the representation that takes a task off the person and puts it on the surface. On the written page, that change was the space between words. The space moved the work onto the page, so the line could be taken in by looking (Saenger, 1997; Busch, 2002). Modern readers of English deprived of the space lose about 30 to 50 percent of their reading rate (Rayner, Fischer, & Pollatsek, 1998; Oralova & Kuperman, 2021; McGowan, White, Jordan, & Paterson, 2014). The question is what the space would be. + +The space is not the invention of literacy, and it was not a blank between every word from the late seventh century. Latin divided words with interpuncts until about AD 100, and continuous script after that lasted about six centuries, with learner texts still word-divided (Dickey, 2017; Busch, 2002). Irish manuscripts already show some separation in the second quarter of the seventh century; Insular scribes grouped prepositions with their nouns, and fully consistent separation is not found before the twelfth (Tunbridge, 1992). Mass literacy took centuries of institutional change beyond the space (Clanchy, 1993; Cressy, 1980). Those limits do not touch the claim. The smoking gun is the change in what the reader had to do. §3 states the history at that strength. + +The paper makes that argument in three moves. First, the 2026 reviews, and the transparency record beside them, still hand people a text (§2). Second, Nadella's post is that remedy at enterprise scale: readers, a chain, and a brake held inside the firm (§2.5). Third, the change to look for is the one the transition from oracy to literacy already contains, and Decision Experience Design names two candidates and a test (§3, §5–§8). A candidate is the space only if it takes a task off the person in the triad and shows, on the surface, that the task is gone. Strategic offloading removes a tactical calculation and makes the removal plain. Ecological feedback immediacy returns a valid state as the person acts, rather than a log to be read afterward. Neither is announced here as the space. + +The sections follow those three moves. §2 reads the two reviews and the HCI record that still hands people a text, and §2.5 is the reading of the post. §3 is the oracy-to-literacy change. §4 is the regulatory record of the same text. §5 names the sensibility a space would call on. §6 holds the two candidates to the test taken from the history. §7 states the condition any candidate has to meet: a space the operator can withdraw is not a space. §8 is how we would know. §9 states the scope and returns to the question. + +## 2. Background: Transparency, Explanation, and Legibility in HCI + +### 2.1 Where the 2026 Reviews Leave the Field + +The two reviews are the context for the argument, and they do not yet contain it. Grahn et al. (2026) are explicit about the model they inherit. Parasuraman, Sheridan, and Wickens, and Endsley after them, describe "a human operator, clearly separated from the machine," who "monitors discrete levels of automation through external displays" (p. 2). Grahn et al. say that scaffolding should be extended, not presupposed. At a closed-loop neural interface "there is no discrete level of automation to select, no external display through which oversight is exercised, and no stable boundary at which the human ends and the machine begins" (p. 7). At infrastructural scale, "the classical assumption of a supervisory operator standing outside the system is least tenable" (p. 8). Between those two strains sits the case this paper is about. Their Level 4b is "institutionally mandated AI," and the examples include "work allocation on labor platforms" (Table 2). A person in that class cannot decline the system without declining the role. + +What they put in place of the strained display is still something a person takes in or authorizes. At the external interface, outputs should be "transparent and interpretable to human operators" (Table 1). Agency, as they use Santoni de Sio and van den Hoven, requires that behavior be traceable to a human's reasons, plus override, opt-out, and human authorization of critical decisions (Section 4). Of the educational obligations they map onto the societal level, one transferred cell is "civic AI literacy." In the discussion, explainable AI is "essential for maintaining the principle of human agency" (p. 16). They also record that most deployed collaboration is still asymmetric: "the AI proposes and the human merely approves" (p. 10). The review is a narrative synthesis, purposive rather than exhaustive, and it does not mention a two-sided allocation or Nadella. The inference is this paper's. A field that has seen the external display fail, and that answers with interpretable outputs, authorization, and literacy, has not yet changed what the person in the triad has to calculate. + +Mourtzis and Tsakalos (2026) supply the shop-floor history of the same habit. Across four industrial revolutions the human goes from operator, to overseer, to supervisor, to colleague (Section 2). The interfaces go with the role: levers and wheels, then buttons, then screens and graphical interfaces, then "visualization dashboards, Augmented Reality (AR), and decision-support tools" built "to assist humans in understanding machine behavior" (Section 2). Industry 4.0, on their account, makes the human's work "cognitive and supervisory," and names trust and situational awareness as what that work requires. Industry 5.0 is supposed to make the human a colleague, and extended reality is supposed to cut "manual effort and cognitive workload" (Section 2). The colleague in the model they then draw still decides from a display. The extended-reality interface "provides the human with the necessary information to perform tasks and make decisions in real-time" (Section 5.4). The conclusions give the three layers the same split: sensors let the robot perceive, "XR interfaces provide instructions and support decision-making for humans collaborating with robots," and the robot does the precision work. Their state of the art is 87 publications from 2021 to 2026, and the sensor papers in it mostly sit at technology readiness levels 3 to 4 (Section 5.1; Table 2). The review is about a worker and a machine, not a rider and a driver. What it shows is that the current interface program, even when it calls the human a colleague, still hands that human information and instructions. + +The transparency record below is the same remedy, measured. Disclosure often fails. Where a text helps, it still asks for a reader inside a dyad. + +### 2.2 The Transparency and Trust-Calibration Turn in HCI and CSCW + +HCI's transparency turn hands people a text. For about twenty-six years, since Herlocker, Konstan, and Riedl (2000), a line of that work has treated opacity as an informational defect to be repaired by exposure. Herlocker et al. called recommender systems "black boxes, computerized oracles that give advice, but cannot be questioned" (p. 1). The remedy they proposed was an explanation. A scoping review of trust-calibration studies reports mixed and null findings (Wischnewski, Krämer, & Müller, 2023). The hope attached to the text is that an operator who can see the system's state will form an accurate mental model, calibrate trust, and decide better. Körber (2018) treats trust as a mainly affective response and notes that operators may not form a complete mental model. + +Disclosure often fails, and where a text helps it still asks for a reader inside a dyad. Feature-based explanations can raise acceptance of a recommendation whether or not it is correct; example-based explanations in the same study reduced overreliance (Chen, Liao, Wortman Vaughan, & Bansal, 2023). Explanations reduce overreliance when the task is hard and the explanation is cheap to check (Vasconcelos et al., 2023). Confidence information shown with recommendations led to smaller performance decrements when the aid was wrong (Parasuraman & Manzey, 2010, reviewing McGuirl & Sarter, 2006). Verification of one code assistant's suggestions took 22.4% of session time, in a lab study of 21 programmers on short tasks (Mozannar, Bansal, Fourney, & Horvitz, 2024). Knowledge workers report reduced critical engagement with generative tools (Lee et al., 2025). In each of these settings the person is still reading, and the loop is still one person and one system. + +### 2.3 Situated Rationale vs. Post-Hoc Audit Trails and "Revealing Design" + +Friction at the moment of choice is still, in these studies, work a person does with a text inside a dyad. Buçinca, Malaya, and Gajos (2021) showed that cognitive forcing functions—such as requiring users to formulate an independent choice before viewing an AI recommendation—reduced overreliance compared with simple explanations on the questions where the AI was wrong. Disley and Khan (2026) reframe system representations as an active "design surface" and show that selectively suppressing a concept in a model can introduce friction. Horvath et al. (2022) studied annotation in browser documentation. The anchor carried context for later readers, and the annotator did not benefit (p = 0.78). Pucci, Piro, and Matera (2027) report tensions around verbosity, pacing, and control for language-model agents that fill web forms, in a study of 18 blind and visually impaired participants. The person is still working through a representation. + +Özçetin and Wiltse (2025) critique how Terms of Service rely on static, legalistic text to govern dynamic, asymmetric relations. They advocate for "revealing design," an approach aimed at making platform entanglements "legible" to participants. Their later moves design an interaction surface for contestation, including what they call appropriateness-adjustment buttons, and call for legislation to require the approach. Making an entanglement legible still begins by handing someone a text. Understanding that one is entangled does not, by itself, change the actions required to accept a dispatch or price a product. Contesting a penalty is one of the actions their design is built to change. + +### 2.4 From Human–Machine Dyads to Irreducible Algorithmic Triads + +The reader in the studies above sits in a dyad. A triad is the situation the space would have to change, and explaining an allocation to one party does not take the tactical calculation off them. Mejia and Parker (2021), Li, Liu, and Sundararajan (2026), and Ehsan et al. (2021) already take up settings with more than one person and a system. In the classic cognitive engineering literature, interaction is modeled as a two-node loop: a human operator paired with a machine aid (Sheridan & Verplank, 1978; Parasuraman, Sheridan, & Wickens, 2000). The pilot monitors the autopilot; the clinician reviews a diagnostic classifier; the driver takes over from an automated steering system (Matsuo et al., 2026). In a dyad, the machine's role is auxiliary. Its explicit functional purpose is to support, inform, or substitute for the human operator in executing an external task. + +``` +Classic HCI Dyad: +[Human Operator] <====== Interaction Loop ======> [Machine / Tool] +``` + +Coordination across contemporary digital platforms does not fit this schema in every market the same way. In ride-hailing and digital labor platforms, two human counterparts coordinate through a third node: an active algorithmic mediator (Lee, Kusbit, Metsky, & Dabbish, 2015; Rosenblat & Stark, 2016; Kellogg, Valentine, & Christin, 2020; Rahman, 2019). In dealer foreign-exchange markets the dealer is a counterparty and owns finality of the deal (Cartea, Jaimungal, & Walton, 2019; Oomen, 2017). Reddit moderation in the study cited above was done by human volunteers (Jhaver et al., 2019). HCI and CSCW have studied settings with more than one person and a system, including AI-mediated communication, where suggested replies change how two correspondents write and regard each other (Hohenstein et al., 2023; Mejia & Parker, 2021; Li et al., 2026; Ehsan et al., 2021). + +``` +Irreducible Algorithmic Triad: + ┌─────────────────────────────┐ + │ Algorithmic Mediator │ + │ (Optimizing Objective J_M) │ + └──────────────┬──────────────┘ + ▲ ▲ + Mediation Probe│ │Boundary Reaction + │ │ + ┌────────────────────┴─┴────────────────────┐ + │ │ + ┌─────────┴─────────┐ ┌─────────┴─────────┐ + │ Party A │◄──Coordination Move──►│ Party B │ + │ (Worker / Driver) │ │ (Client / Rider) │ + └───────────────────┘ └───────────────────┘ +``` + +This mediator is not an advisory instrument assisting one party. It reads inputs from both sides, filters information symmetrically or asymmetrically, applies an independent objective function (J_M)—such as platform profit maximization, throughput balancing, or dispute minimization—and commits an allocation that neither party directly controls. On a ride-hailing platform the driver can decline the allocation, at a cost (Lee, Kusbit, Metsky, & Dabbish, 2015). Party A's payoff depends directly on Party B's move, yet Party A can only perceive, probe, or influence Party B through the transformation imposed by M; in that sense the triad does not decompose into two independent dyads. Transplanting dyadic transparency tools into a triadic mediator yields incoherence: explaining to Party A why the algorithm made an allocation does not eliminate the tactical necessity for Party A to second-guess what Party B will do through that same mediator. + +### 2.5 The Post Preserves Literacy + +This is the argument against the vision the reviews leave in place. Grahn et al. (2026) and Mourtzis and Tsakalos (2026) still hand a person a display, an authorization, or an instruction. Nadella's (2026) post, "Models as Insider Risks in the Super Intelligence Era," is that remedy written for the firm that deploys the model. It is a literacy solution for a triadic situation. It is a personal essay, not a Microsoft policy document, and the "we" throughout is the deploying organization. The words user, customer, employee, and people do not occur. Paragraph numbers below are the project's. + +The post opens on an attribution gap. With traditional software, "we had the tools and capability to trace behaviors to a specific code path" (P1). For frontier models, "we can't attribute model behaviors and outputs to specific inputs of training data or configurations of model weights," and organizations still give them "access to our most sensitive data" and "mission-critical actions on our behalf" (P2). The remedy he reaches for is not a change in what a person in a triad has to calculate. It is containment. "We need to separate the supply of intelligence from the authority over it" (P5, before the seven principles). Models are to be treated as insider risks because a capable actor can make mistakes or be compromised (P7), in the tradition of insider-threat practice (CERT National Insider Threat Center, 2022). + +The human task that follows is reading. Chain-of-thought transparency is "a non-negotiable," and "Neuralese" cannot excuse opaque reasoning, but the chain "alone is not sufficient or dependable" (P9). Controls over access and action "must sit outside the model" (P11), and the model is separated from its harness and action space (P12). Among the seven principles of observability, "Observe everything" requires "tamper-proof human readable evidence" for every meaningful action, so an outcome can be reconstructed without the model's own attestation (P15). Verifiability, independent controls, and independent auditability are further principles (P16–P18). Independent auditability is not the evidence sentence. The evidence is a text. Someone inside the firm has to read the chain: the chain of thought, the logs, the record that can reconstruct an outcome. + +Containment is real, and it is not a space. An authorized person must be able to pause or stop a model in mid-task, "like an emergency brake" (P19). The brake is held inside the firm. Incident disclosure tells "those affected" after a failure (P20). That is the one place a party outside the deployer appears, and the party appears as a reader of a notice. The close asks for the system that "enables us to trust the model the least" (P21). Trusting the model least, on this design, means reading it more. + +``` +The chain stays inside the firm; the notice goes out: +┌────────────────────────────────────────────────────────┐ +│ Deploying Firm │ +│ ┌──────────────────┐ ┌───────────────────┐ │ +│ │ Intelligence │ │ Authorized person │ │ +│ │ (the model) │ │ holds the brake │ │ +│ └────────┬─────────┘ └─────────▲─────────┘ │ +│ │ │ │ +│ └──────► The chain: CoT, logs, ─┘ │ +│ human-readable evidence │ +└───────────────────────────┬────────────────────────────┘ + │ Allocation the model commits + ▼ +┌────────────────────────────────────────────────────────┐ +│ Parties the model acts on │ +│ │ +│ They appear once: readers of a notice after a │ +│ failure. They do not hold the brake. │ +└────────────────────────────────────────────────────────┘ +``` + +*Scriptio continua* asked the ear to find a boundary the eye could not see. The post asks a reader to find, in a chain, what the model did. Containment and the brake do not take that task off the person and show, on the surface, that it is gone. They add readers, and a brake held inside the firm. A rider, a driver, a buyer, a seller still has to calculate through the mediator. The post does not change that calculation. The transition from oracy to literacy is the change that did, and §3 is that change. + +## 3. Re-Representation: The Scribal Precedent in Human-Factors Terms + +### 3.1 The Page Before and After Word Separation + +The reviews and the post leave the person with something to read. This section is the change they do not contain. The transition from oracy to literacy put the finding of a word's boundary on the page. The space between words was the smoking gun: one change in the representation that took that work off the voice. + +In the main Greek tradition, texts were transcribed in *scriptio continua*, words strung together in unbroken lines. Because the line gave no spatial demarcation, Saenger (1997) argues, a reader identified words by pronouncing or subvocalizing the characters. Heilmann (2020) reports that the classical consensus follows Saenger. Busch (2002) defends reading aloud as the ancient norm and allows that fluent silent readers were an exception. *Scriptio continua* made sense read aloud, because the ear did the work the eye could not. + +The device that moved the work was the space. Scribes in Ireland and Anglo-Saxon Britain learned Latin as a foreign language, not as a spoken one, and Insular copyists began grouping words on the page (Tunbridge, 1992). Saenger's statement of the effect is that word separation and related devices in Insular books "provided shortcuts for achieving the reading skills that an elite among the ancients had mastered only through a prolonged and arduous grammatical apprenticeship" (1997, p. 85). The space did not add a new word. By altering the layout, the scribes altered what a reader had to do. + +The limits are these, and they do not touch that claim. The space is not the invention of literacy. Mass literacy required centuries of institutional change, legal documentation, and administrative centralization (Clanchy, 1993; Cressy, 1980). Latin used interpuncts until about AD 100, and continuous script after that lasted about six centuries; learner and bilingual texts kept word division (Dickey, 2017; Busch, 2002). The earliest word-based separation in the Irish record is the second quarter of the seventh century, and the practice was already in some manuscripts before a start date can be fixed (Tunbridge, 1992). Insular copyists grouped prepositions with their nouns and joined *et*; fully consistent separation into individual parts of speech is not found before the twelfth century, and on Saenger's account the Continental convention was standard by the mid-eleventh, with canonical separation in the twelfth (Tunbridge, 1992; Saenger, 1997; Parkes, 1992). Silent reading is attested for letters and short texts, and for an elite, long before the space (Knox, 1968; Gavrilov, 1997; Burnyeat, 1997; Heilmann, 2020). Letter forms also changed (Tunbridge, 1992). The smoking gun is the change in what the reader had to do. + +### 3.2 Cognitive Offloading, Visual Search, and Perceptual Substitution + +What the reader no longer had to do is the content of the smoking gun. In the language of cognitive engineering, the space can be read as environmental cognitive offloading (Risko & Gilbert, 2016; Kirsh, 1995) that exploited the representational effect (Zhang & Norman, 1994). Those sources do not mention writing. The extension to the page is this paper's. Zhang and Norman showed that "different isomorphic representations of a common formal structure can cause dramatically different cognitive behaviors" (p. 88). When a constraint sits in the layout, it does not have to be held in working memory. + +Hutchins (1995) documented the same principle on the flight deck: pilots set movable speed bugs on the rim of the airspeed indicator, and the comparison of speeds is externalized into the spatial proximity between a needle and a marker. + +``` +Representational Shift on the Page: + +Scriptio Continua: +[L][E][X][I][C][A][L][B][O][U][N][D][A][R][Y] +▲ ▲ +└──────── Working Memory / Phonological ────┘ + Reconstruction Overhead + +Word Spacing: +[L][E][X][I][C][A][L] [B][O][U][N][D][A][R][Y] + ▲ ▲ ▲ + └── Visual ┴── Eye ───┘ + Saccade Targeting +``` + +*Scriptio continua* left boundary detection to the reader, who had to hold ambiguous strings of sound while testing competing segmentations. The space substituted visual segmentation for that inference. No study has measured that substitution in medieval readers. What has been measured is the modern cost of taking the space away. + +The visual system processes spatial structure across the field in parallel, within the limits of eccentricity and crowding (Wolfe, 2021). Under Wolfe's Guided Search model, bottom-up feature salience is one of five sources that guide selective attention. By carving white channels between character clusters, spacing can give the oculomotor system targets for the next saccade. Removing spaces from English changes where the eyes land as well as how fast words are identified (Rayner et al., 1998). Thai readers' initial landing positions were similar in spaced and unspaced text (Winskel, Radach, & Luksaneeyanawin, 2009), so saccade targeting does not depend on spaces in every script. Deprived of these markers, modern readers of English lose about 30 to 50 percent of their reading rate (Rayner et al., 1998; Winskel et al., 2009; Oralova & Kuperman, 2021). The geometry of the page did part of the work of finding the word. That is what a smoking gun does. + +### 3.3 Cross-Orthographic Asymmetries and the Structural Disanalogy + +The space pays where finding the boundary was the reader's job. Where a script already gives the eye another cue, inserting spaces is a different change. When spaces are removed from English, reading performance drops sharply. When spaces are inserted into Chinese or Thai, the effect is mixed (Bai et al., 2008; Winskel, Radach, & Luksaneeyanawin, 2009; Chantavarin & Leung, 2025). + +Skilled readers of Chinese read word-spaced text no faster than normal unspaced text, 5,239 against 5,221 milliseconds a sentence (Bai et al., 2008). Oralova and Kuperman (2021) replicate that null and explain it: a word-level gain is cancelled by extra saccades, while spaces placed only at boundaries that raters agree on gave a small reliable speed-up of about 2 percent. Thai readers fixate a spaced word for less time, and the sentence-level slowing is marginal (Winskel et al., 2009; Chantavarin & Leung, 2025). Learners do not reliably gain. Child learners of Chinese read word-spaced text more slowly and less accurately than text spaced between every character (Bassetti & Masterson, 2012), and for adult second-language learners the difference between spaced and unspaced text was not significant (Cui, 2023). Native readers of unspaced Thai gain at the word level (Chantavarin & Leung, 2025). Colour-marked word boundaries gave young skilled Chinese readers no significant benefit and helped older readers, whom the authors see as limited by vision (Pan, Wang, Zhang, Tsang, & Yan, 2025). An ergonomic intervention is not an abstract virtue; it must match the visual search affordances of the medium to the cognitive tasks of the user. + +The page and the platform are not the same surface, and the difference matters for what a later space would have to survive. A manuscript page does not act on its own. Once ink dries, it does not reorder its words when the reader blinks. Later readers did add punctuation and marks, and correctors' hands are routine (Tunbridge, 1992). The scribes who made the page worked under a discipline that was real and uneven. + +Carolingian rulers and church authorities ordered some books made correct. Charlemagne's letter *De litteris colendis* measures monasteries' written Latin and study against a standard of correctness (Charlemagne, trans. 1900; Boretius, 1883, no. 29). The *Admonitio generalis* of 789, chapter 72, orders books corrected and copied with care (Boretius, 1883, no. 22). Alcuin's *Carmen* XCIV instructs scribes to mark sense units for the lector (Dümmler, 1881). Only the capitulary and the poem concern books; the letter concerns written Latin and study. Westwell (2018) cautions that uniformity imposed from above can be overstated, and authority was distributed and negotiated. None of these texts concerns word separation, which the Insular scribes had adopted earlier. What they show is who stood behind some of the page: an authority with standing over the scribe, not a single court that held every copyist. + +An algorithmic intermediary is dynamic, adaptive, and governed by the commercial incentives of its owner. The platform controls the interface, observes user probes, and can rewrite its ranking rules, pricing multipliers, and allocation weights (Rosenblat & Stark, 2016; Lee et al., 2015; Rahman, 2019). The page, once dry, held the space. The platform can withdraw one. The reviews and the post, already read, still hand the person a text. The space is the change they do not make. + +## 4. Disclosure as a Literacy-Based Intervention + +### 4.1 Taxonomizing Disclosure and the Law-and-Economics Critique + +Regulation's usual remedy is the same text. When algorithmic mediation produces unfair or exploitative outcomes, public regulators and legal bodies often fall back on disclosure. Ben-Shahar and Schneider (2011) call it the most common technique in regulation generally. It is common, and it is not the whole of the law: the Platform Work Directive pairs disclosure with bans on some processing and with duties to modify a discriminatory system (Directive (EU) 2024/2831, Arts. 7, 10). The four targets below are ways of handing a person a document. Only the last is the category error. A one-line statement of a property, Target 3, can change what people do, and it is still a sentence to study (§6.2). + +``` + ┌────────────────────────────────────────────────────────┐ + │ Target 1: Presence Disclosure │ + │ ("An automated algorithm is in use") │ + └───────────────────────────┬────────────────────────────┘ + │ + ┌───────────────────────────▼────────────────────────────┐ + │ Target 2: Objective Function Disclosure │ + │ ("System maximizes driver utilization and margin") │ + └───────────────────────────┬────────────────────────────┘ + │ + ┌───────────────────────────▼────────────────────────────┐ + │ Target 3: Formal Rule Property Disclosure │ + │ ("Sincere reporting is a dominant strategy") │ + └───────────────────────────┬────────────────────────────┘ + │ + ┌───────────────────────────▼────────────────────────────┐ + │ Target 4: Full Textual Rule & Rationale │ + │ (Terms of Service, XAI feature cards, audit logs) │ + └────────────────────────────────────────────────────────┘ +``` + +1. Target 1 (Presence): Notifying users that an algorithm or bot is active (Ishowo-Oloko et al., 2019). +2. Target 2 (Objective Function): Stating the objective the mediator is optimizing (e.g., platform margin maximization vs. wait-time minimization). +3. Target 3 (Formal Rule Property): Informing users of an actionable structural property of the mechanism (e.g., that sincere bidding is a dominant strategy; Masuda et al., 2022). "Obviously strategy-proof" is Li's (2017) term for a stronger property, and the second-price sealed-bid auction, which is Masuda's case, is strategy-proof and not obviously strategy-proof. +4. Target 4 (Full Textual Rule and Rationale): Providing full terms of service, technical documentation, or post-hoc feature-attribution scores explaining an individual allocation. + +The category error named in this paper lies in Target 4. Target 3 behaves differently: a one-line statement of a property changes what people do (§6.2). In Masuda, Mikami, Sakai, Serizawa, and Wakayama (2022) that statement came with instructions and a seventeen-question quiz, and the gain in sincere bidding was concentrated among perfect scorers. It is not a signal with nothing to study. + +The legal and behavioral critique of mandated disclosure explains why a rulebook often fails, and it does not cover the text of algorithmic rules. Ben-Shahar and Schneider (2011) reviewed disclosure regimes across finance, healthcare, and consumer contracts. They write first that its failure is inevitable (p. 651) and later that it regularly, though not inevitably, fails (p. 679), because the people it addresses do not read disclosures, do not understand them when they do, and do not use them when they understand (p. 665). Fung, Graham, and Weil (2007) find that disclosure works where it is embedded in the routines of users and disclosers alike (pp. 54–55, 65), and they rate complex chemical and toxics disclosure as only moderately effective (p. 77). Psychology explains much of the gap between what is disclosed and what is done with it (Loewenstein, Sunstein, & Golman, 2014). + +The disclosures that do alter outcomes tend to be simple, low-cost signals, such as letter grades for hygiene posted in restaurant windows. They act in part on the discloser. Loewenstein, Sunstein, and Golman (2014) put the motive in the spotlight and in guilt (p. 404). Jin and Leslie (2003) put it in consumer demand: Los Angeles grade cards raised scores, made revenue sensitive to hygiene, and cut food-related hospitalisations, with inspector leniency at the A threshold. The same device can fail where the grade is not a valid measure. Ho (2012) finds San Diego grades inflated, New York scores almost uninformative, and no detectable health effect after New York grading. Dranove, Kessler, McClellan, and Satterthwaite (2003) find that cardiac-surgery report cards shifted surgery toward healthier patients and left sicker patients worse off. Ben-Shahar and Schneider hedge even the restaurant case. + +### 4.2 The Empirical Record Across Finance, XAI, and Platform Governance + +The empirical record is the same verdict. Disclosure often fails. Where a text helps, it still asks for a reader. In a meta-analysis of explainable-AI interventions, explanations yielded no reliable improvement over AI advice alone (Schemmer et al., 2022; standardized mean difference 0.07, 95% CI [−0.15, 0.30]). Against unaided humans the same paper reports a gain (0.59, 95% CI [0.39, 0.79]). Across 370 effect sizes the presence of an explanation was not a significant moderator of human–AI performance (Vaccaro et al., 2024). Heatmap explanations raised balanced accuracy by 7.7 points for 48 factory workers and by 4.7 points for 113 radiologists, over the same black-box score (Senoner et al., 2024). The heatmap re-represents the object, and radiologists overruled fewer wrong predictions when they had it. + +In consumer finance, one field experiment shows that a disclosure can backfire. A Mexican bank sent a separate mailing in a statement-like envelope, with no statement inside (Seira et al., 2017). A message stating the interest rate changed nothing in the aggregate and reduced debt by 525 pesos, about 3% of the mean, in one subgroup. A message that stated the personalized months-to-pay and advised the cardholder to pay more than the minimum raised delinquency by 0.011 points among clients who paid interest and by 0.012 among high-debt clients, 7.4% of the mean (pp. 293–295). A two-bank follow-up found effects under 5%, none significant (p. 304). A repayment disclosure under the CARD Act moved repayment slightly in the intended direction (Agarwal, Chomsisengphet, Mahoney, & Stroebel, 2015). Disclosure can backfire. It does not always. + +In machine learning interfaces, Poursabzi-Sangdeh et al. (2021) showed that a clear model, set against a black-box model, reduced users' ability to detect and correct the model's mistake on an unusual case in the first two experiments. The two-feature comparison is only the fourth experiment. The effect is a main effect of transparency, pooled over the two-feature and eight-feature clear models. The authors conjecture that the visible detail overloaded participants. + +In platform governance, Jhaver et al. (2019) surveyed 907 Reddit users whose posts had been removed. The two coefficients come from different subsamples. Among the 526 who had noticed the rules, having read them went with judging the removal less fair (β = −.314). Among 738 respondents, receiving an explanation went with judging it fairer (β = .384). Both are cross-sectional associations from self-report, and the analysis is primarily exploratory. Rule clarity also raised fairness (β = .170) and intention to post again (β = .223). The rulebook and the situated reason pulled in opposite directions for fairness. + +Nouwens, Liccardi, Veale, Karger, and Kagal (2020) showed in cookie consent interfaces that moving the reject button off the first page raised consent by 22 to 23 percentage points, in a sample of 40 participants in the United States. Users respond to the spatial architecture of the interface. The person is still choosing inside a notice. + +The reading of Nadella (2026) is in §2.5. What §4 adds is the regulatory record of the same remedy: a text handed to a person. + +## 5. Algorithmacy: Epistemic Action in Irreducible Triads + +### 5.1 Defining Algorithmacy + +Algorithmacy, as used here, is the sensibility the space would call on. It is deciding, and taking in what returns, rather than reading a record. I define it as the behavioral sensibility required to navigate irreducible algorithmic triads: the capacity to coordinate with a counterpart through an active, optimizing mediator whose internal rules shift dynamically. + +``` + ┌─────────────────────────────┐ + │ Historical Transition │ + └──────────────┬──────────────┘ + │ + ┌──────────────────────────────────────┴──────────────────────────────────────┐ + ▼ ▼ +┌─────────────────────────────────┐ ┌─────────────────────────────────┐ +│ Literacy │ │ Algorithmacy │ +├─────────────────────────────────┤ ├─────────────────────────────────┤ +│ Mode: Evaluative, textual, │ │ Mode: Enacted, exploratory, │ +│ reflective │ │ epistemic │ +│ Medium: Static pages, documents,│ │ Medium: Dynamic, optimizing, │ +│ post-hoc logs │ │ triadic interfaces │ +│ Cognitive Demands: Working │ │ Cognitive Demands: Dynamic │ +│ memory, phonological loop, │ │ calibration, boundary │ +│ semantic parsing │ │ probing, feedback integration │ +│ Stance: Reader │ │ Stance: Decision-maker │ +└─────────────────────────────────┘ └─────────────────────────────────┘ +``` + +Algorithmacy is distinct from dyadic "AI literacy." The latter typically refers to conceptual knowledge: understanding what neural networks are, how training data shapes bias, or how to compose structured text prompts in chat windows. + +Nor does it match the one earlier printed use of the word that I have found. Oberschelp (1995), writing on informatics teaching, urges that "we have to preserve logicacy and algorithmacy by continuously training simple mental activities without artificial tools" (p. 121). He does not define the term; in context it names the skill of carrying out algorithmic steps by hand. + +Algorithmacy as I use it is an active coordination competency. Just as literacy allows a reader to traverse an author's text without demanding that the reader build a printing press, algorithmacy allows an individual to formulate choices, infer hidden constraints, and protect their welfare across an algorithmic intermediary without requiring them to inspect its source code. + +### 5.2 Enacted Sensemaking and Epistemic Action + +In an irreducible triad, people do not come to understand a mediator by reading about it. They understand it by acting through it and observing what happens. This process is grounded in Weick's (1988) account of enacted sensemaking: in ambiguous, dynamic environments, "people think by acting" (p. 305). Actors take action, create consequences in the environment, and then look at the results to infer what kind of system they are dealing with. Weick adds a warning that belongs in any design that relies on this: action that clarifies a situation can also intensify it. + +In cognitive science, exploratory moves of this kind are known as epistemic actions (Kirsh & Maglio, 1994). Kirsh and Maglio showed that in Tetris, players do not rotate and translate falling pieces only to place them (pragmatic action). They also move them to learn something sooner than they could work it out in the head. Epistemic actions are "physical actions that make mental computation easier, faster, or more reliable" (p. 513). + +``` +Triadic Epistemic Probing: +[Worker / User] ──── (1) Epistemic Probe: Move / Micro-bid ────► [Algorithmic Mediator] + │ + (2) Evaluates Probe + Against Objective J_M + │ +[Worker / User] ◄─── (3) Boundary Reaction / State Update ──────────────┘ + │ +(4) Dynamic Recalibration of Strategy +``` + +Ride-hail drivers' strategies drew partly on the company's published general rules, and partly on what the system did in response to their own moves (Lee et al., 2015). They trade those theories in online forums (Lee et al., 2015; Kellogg, Valentine, & Christin, 2020). Workers also pool what they observe and push back together (Kellogg et al., 2020; Rahman, 2019). Upwork freelancers learn a rating algorithm from trial and error, forums, and documentation (Sutherland, Jarrahi, Dunn, & Nelson, 2020). When a decision rule is hidden, people infer it from the observed outcomes of others (Ghalme et al., 2021). Shen et al. (2021) already theorise that probing. What this definition adds is the name of the sensibility a space would call on: deciding, and taking in what returns. + +Learners also settle into a false rule and stop probing (Rich & Gureckis, 2018; Lee, Li, Lee, & Hayes, 2023). Taking in what returns is not a guarantee that the return was read correctly. + +### 5.3 High-Validity Learning Environments vs. Mediated Distortion + +Kahneman and Klein (2009) set two conditions for the development of skilled intuition: an environment regular enough to be learned, and an opportunity to learn it through "feedback that is both rapid and unequivocal" (p. 524). They also write that validity and uncertainty are not incompatible, and name poker and warfare. + +When feedback is delayed or noisy, calibration suffers, and not in one way. Rahmandad and Gary (2023) found that delay pushes decision-makers toward options that pay off quickly. Uncertainty was not an important predictor of bias. It cut performance in one study and not robustly in the second, and in the first study uncertainty may have helped performance when delays were long. Movement toward the middle of the strategy space comes from an increase in delay or in uncertainty, so it is not unique to noise. Learning was present across all conditions. + +Algorithmic intermediaries often fail these conditions. To analyze mediated distortion, we must separate two distinct drivers: + +1. Strategic Rule Secrecy: Some platforms conceal evaluation criteria, ranking parameters, and matching formulas (Rahman, 2019; Regulation (EU) 2019/1150, Arts. 5(1) and 5(6)). Rahman (2019) reports that operators justify this secrecy as necessary to prevent "gaming," and reads it as also keeping control of a stratified workforce with the platform. +2. Technical Noise and Latency: Algorithmic mediators can introduce jitter and stochastic allocations. Not all distortion is intended. The jitter Chen, Mislove, and Wilson (2015) measured in Uber's surge multipliers was, by Uber's own account, a bug: stale multipliers served to random clients. Deliberate noise does occur. LinkedIn added calibrated differential-privacy noise to the analytics marketers receive (Rogers et al., 2021). Yahoo researchers describe serving randomly selected articles to a randomly chosen bucket of visitors in order to learn (Li, Chu, Langford, & Schapire, 2010). Deliberate delay is not shown in the sources read here. + +Hidden rules and noisy or delayed feedback degrade learning by acting. They do not remove it. Rahmandad and Gary (2023) find that delay and noise degrade and bias learning without abolishing it, and last-round performance still beat a random benchmark. The user who probes and receives an ambiguous response later may be unable to verify whether the action contributed (Rahman, 2019). Dietvorst, Simmons, and Massey (2015) complicate the mechanism: people who watched a superior algorithm win still abandoned it for a worse human forecast. Folk theories are part of what remains. They are also what launched audits that found real bias (Shen et al., 2021). Defensive gaming is not the only residue. + +## 6. Decision Experience Design (DXD): Two Candidate Conventions + +### 6.1 Architectural Framing via Norman's Gulfs of Action + +This section is the third move. The test is the one the space supplies. A candidate is the smoking gun only if it takes a task off the person in the triad and shows, on the surface, that the task is gone, so they do not have to read a chain to find the boundary. Decision Experience Design names two candidates and holds each to that test. The vocabulary is Norman's (1988) gulfs of execution and evaluation. Those gulfs are between a person and a device. The mediator here is a third party, and the fit is an analogy: + +``` + ┌─────────────────────────────────────────┐ + │ The Transacting Human │ + └──────────────┬──────────────────▲───────┘ + │ │ + Formulate Intent │ │ Evaluate State + │ │ + ┌────────────────▼┐ ┌┴────────────────┐ + │ Gulf of │ │ Gulf of │ + │ Execution │ │ Evaluation │ + └────────────────┬┘ └▲────────────────┘ + │ │ + Bridged by: │ │ Bridged by: + Strategic │ │ Ecological + Offloading │ │ Feedback + (Convention A) │ │ Immediacy + │ │ (Convention B) + │ │ + ┌──────────────▼──────────────────┴───────┐ + │ Algorithmic Mediator │ + └─────────────────────────────────────────┘ +``` + +1. The Gulf of Execution: The cognitive distance between a user's goal and the actions permitted by the interface. Convention A (Strategic Offloading) narrows the Gulf of Execution by eliminating tactical gaming calculations from the transaction mechanism and showing on the display that they have been eliminated. +2. The Gulf of Evaluation: The cognitive distance between the system's state and the user's ability to assess that state. Convention B (Ecological Feedback Immediacy) narrows the Gulf of Evaluation by delivering valid, noise-filtered, low-latency feedback as the user formulates their choice. + +### 6.2 Convention A: Strategic Offloading (Removing a Tactical Task and Showing It Is Gone) + +Convention A passes the test only if the tactical calculation is taken off the person and the display shows that it is gone. A sentence that states the property is still a text. The evidence is about how far a menu does the first thing, and how far a description does not. + +In classical mechanism design, an allocation mechanism is strategy-proof if reporting one's true valuation or preference is a dominant strategy: no participant can gain by lying, shading their bid, or misrepresenting their constraints (Li, 2017). Yet theoretical strategy-proofness often fails to produce sincere human behavior. Non-sincere play is frequent, on the order of a fifth to a half of reports depending on the mechanism, and it is not near-universal (Masuda et al., 2022; Li, 2017; Gonczarowski, Heffetz, Ishai, & Thomas, 2024; Artemov, Che, & He, 2023). In a laboratory Vickrey auction, only about a fifth of bids were sincere without advice (Masuda et al., 2022). In field data, 17 to 35 percent of applicants make dominated reports, and only 1 to 20 percent of those mistakes change the assignment (Artemov et al., 2023). + +To address this, Li (2017) formalized obviously strategy-proof (OSP) mechanisms. A strategy is obviously dominant if, at any information set where it and a deviation first diverge, the best possible outcome under the deviation is no better than the worst possible outcome under the strategy. A mechanism is OSP if it has an equilibrium in obviously dominant strategies. The comparison is not a property of the mechanism taken apart from that information set. OSP mechanisms can be recognized as safe by agents who lack contingent reasoning capabilities, and in Li's experiments far fewer games ended off the dominant-strategy outcome under the obviously strategy-proof version of the same rule. + +Convention A embeds this property in the visual interface. It removes the tactical calculation—eliminating the need to second-guess what other participants are doing—and uses the interface to make that removal obvious. + +Recent experimental economics calibrates the convention and its limit. Gonczarowski, Heffetz, Ishai, and Thomas (2024; N = 542) tested several descriptions. The gain in understanding belongs to a property-only menu, not to a menu description of the algorithm. Understanding of strategy-proofness was 71% under the property-only menu (Menu-SP), 62% under a textbook statement of strategy-proofness, 58% under a menu description of deferred acceptance (Menu-DA), 56% under the traditional description, and 54% under a null. Straightforward play, measured as the share of rounds, was 59% under Menu-SP. Menu-DA was 50% and the textbook statement 53%. Menu-SP beat the null (p = 0.01) and beat Menu-DA (p = 0.04). It did not beat the traditional description. + +Understanding moved play by a large margin. The description moved it by a small one. Most Menu-SP participants mixed straightforward and non-straightforward rounds: 59% of rounds were straightforward, and 19% of those participants played straightforwardly in every round. Participants who understood the property, scored above 75%, played straightforwardly in 83% of rounds, against 46% for those scored below 75%. That is a 36-point difference (Gonczarowski et al., 2024). + +This is why Masuda et al. (2022) matter: advice that sincere bidding is a dominant strategy, given with instructions and a quiz, raised sincere bids from 20.6% to 46.9%, while efficiency rose by about one percentage point. The authors attribute the small efficiency effect to a symmetric design in which efficiency was already near the ceiling. Convention A must therefore pair the structural elimination of the tactical task with an interface that visibly confirms the task is gone. Designers should not treat a large change in behavior as a promise of a large change in outcomes, and they should not treat the small efficiency gap as a general law. + +``` +Convention A Interface Transformation: + +Old Gaming Interface: +┌────────────────────────────────────────────────────────┐ +│ Enter Your Hourly Rate Bid: [ $ 45.00 / hr ] │ +│ │ +│ [ Warning: Other workers may undercut your bid. ] │ +│ [ Consider platform commission (20%) and surge index. ]│ +└────────────────────────────────────────────────────────┘ +Result: Tactical anxiety; complex shade calculation. + +DXD Strategic Offloading Interface: +┌────────────────────────────────────────────────────────┐ +│ Guaranteed Clearing Menu: │ +│ │ +│ ● [ $ 52.00 / hr ] ── Guaranteed Dispatch in Tier 1 │ +│ ○ [ $ 48.00 / hr ] ── Guaranteed Dispatch in Tier 2 │ +│ │ +│ [ Strategy Notice: Posted-price menu. ]│ +│ [ Sincere preference revelation is dominant. No worker ]│ +│ [ can undercut your allocated tier. ]│ +└────────────────────────────────────────────────────────┘ +Result: Tactical calculations offloaded to mechanism display. +``` + +**Design Opening: What if...?** +What if freelance labor dispatch and procurement platforms abandoned open-ended bid-entry boxes—which force workers to calculate speculative price-shading—and replaced them with posted-price menus of guaranteed tiers, where compensation is stated and tactical undercutting is made impossible by design? An ascending clock is a different mechanism. Pycia and Troyan (2023) rank a posted-price menu, which is strongly obviously strategy-proof, above an ascending clock, which is only one-step simple, and a guaranteed-tier menu gives up efficient price discovery. The display above is a posted-price menu. It still carries a strategy notice, which is a sentence to be read. + +On the test from the history, Convention A has not passed. The measured gain is the property-only menu. Understanding moved straightforward play by 36 points; the description moved it by a few. A posted-price menu is not an ascending clock. Whether the menu takes the calculation off a person in a triad, and shows the removal without a notice to read, is what Arm 3 of §8 is for. The question stays open. + +### 6.3 Convention B: Ecological Feedback Immediacy (Valid and Plain First, Rapid Second) + +Convention B passes the test only if a valid state appears on the surface as the person acts, so they do not have to read a log afterward. Speed is not the test. The state has to be valid, and it has to be on the surface. + +Convention B addresses the Gulf of Evaluation by shortening the temporal and causal distance between an actor's exploratory probe and the mediator's response. Shneiderman (1983) states the principle that the impact on the object of interest is immediately visible, and his examples run wider than classical graphical interfaces: display editors, VisiCalc, video games, CAD, and the Xerox Star, where a document icon is dragged onto a printer icon. In algorithmic mediation, that coupling is often broken. A user submits a bid, posts a job, or positions a car, and waits while the system calculates. Uber recomputed surge on a five-minute clock (Chen et al., 2015), and Last Look gives a dealer a rejection window (Cartea et al., 2019). + +However, in designing Convention B, ecological validity strictly precedes speed. Frequent feedback can damage decisions when it makes losses salient. Larson, List, and Metcalfe (2016) found that professional traders shown prices second by second invested about 25% less in the risky asset and earned about 35% less than traders shown prices every four hours, a pattern they attribute to myopic loss aversion on an asset with a positive drift. The harm is from seeing losses often, not from noisy or unreliable data. With real funds over a year, and with realistic return distributions, the frequency effect vanished (Beshears, Choi, Laibson, & Madrian, 2017). + +Kluger and DeNisi (1996) showed in a meta-analysis that more than one-third of feedback interventions reduce performance. Harm grew as attention moved toward the self and away from the task. Wisniewski, Zierer, and Hattie (2020) put the harm in low-information feedback. Validity comes first for that reason. + +Convention B therefore specifies feedback that is valid and noise-filtered first, plain second, and low-latency third. It would give participants a state on the surface before they commit. The closest test is dyadic: a what-if console that let people drag inputs and watch a prediction change did not improve decisions and raised over-reliance in one task (Liu, Lai, & Tan, 2021). + +``` +Convention B: Pre-Commit Boundary Feedback + +Opaque Batched Flow: +[ Formulate Move ] ──► [ Commit Action ] ──► [ Silent Batching Delay ] ──► [ Irreversible Outcome ] + (Zero visibility) (Post-hoc shock) + +DXD Pre-Commit Sandbox Flow: +[ Formulate Move ] + │ + ▼ (Exploratory Probe) +┌────────────────────────────────────────────────────────┐ +│ Pre-Commit Counterfactual Contour Sandbox │ +│ │ +│ Candidate Parameter: [ X = 42 ] │ +│ Mediator Boundary Reaction: [ Cleared: P(Match) = 88%│ +│ Noise-Filtered Confidence Band: [ ± 3% ] │ +│ │ +│ [ Visual contour shows parameter boundary line ] │ +└────────────────────────────────────────────────────────┘ + │ + ▼ (Informed Commitment) +[ Commit Final Action ] ──► [ Deterministic Execution ] +``` + +**Design Opening: What if...?** +What if gig drivers and marketplace sellers were provided with noise-filtered contour maps that showed, as they dragged a price or radius slider, how the platform's clearing mechanism would allocate orders across candidate locations before they committed to a shift or contract? + +On the test from the history, Convention B has not passed. It is untested in a triad. The nearest sandbox, a what-if console, did not improve decisions and raised over-reliance (Liu, Lai, & Tan, 2021). A contour that has to be read as a log is another chain. Arms 4a and 4b of §8 are how the test would be run. The question stays open. + +## 7. Institutional Custody: Who Can Hold a Convention? + +### 7.1 The Unilateral-Modification Custody Test + +A space the operator can withdraw is not a space. That is the only condition this section adds. A candidate that takes a task off the person still fails if the party that operates the interface can take the representation back. + +To evaluate institutional stability, I formulate the Unilateral-Modification Custody Test: + +*Can the party that operates the intermediary unilaterally alter, suspend, or distort the representational convention without the informed consent of the transacting parties?* + +This test is an analogy to the economic theory of agency costs (Jensen & Meckling, 1976). Their page on continuing incentives concerns divergence between managers and shareholders, and their remedy is monitoring and bonding. Applying the point to a platform and its users is this paper's step. Farronato, Fradkin, and MacKay (2023) record one measured tilt: Amazon's own brands ranked higher than observably similar products. The authors do not claim surplus capture or consumer harm. + +### 7.2 Comparative Analysis Across Three Institutional Realms + +To evaluate how custody functions in practice, I examine three institutional environments: + +| Institutional Realm | Primary Enforcement Authority | Structural & Procedural Mechanism | Unilateral-Modification Vulnerability | +| --- | --- | --- | --- | +| Courts of Law | Constitutional oversight & statutory civil procedure | Rules Enabling Act (28 U.S.C. §§ 2072–2074); judicial disqualification (*Tumey v. Ohio*, 1927); procedural due process (*Mathews v. Eldridge*, 1976) | Low, and not airtight: a judge cannot rewrite the rules mid-trial, and failure to follow the enabling procedure does not itself invalidate a rule | +| Dealer FX Markets | Voluntary industry code; regulators after the fact | FX Global Code (GFXC, 2024): Principle 17 on last look; Principle 9 on disclosure to clients; Principle 36 on keeping reasons for rejections; Principle 40 on agreeing substantial changes to legal terms | Moderate (soft law): the Code is voluntary; control failures in G10 spot FX were fined by the FCA (2014); a 17 November 2015 NYDFS order concerns Last Look, and a 20 May 2015 order concerns benchmarks | +| Digital Platforms | Private corporate ownership | Unilateral Terms of Service (Suzor, 2018; Klonick, 2018); proprietary interface updates; advisory boards (Oversight Board, 2022) | High (unilateral): the owner sets and changes rankings, algorithms, and interface conventions, subject to statute where one applies | + +In judicial systems, procedural conventions are partly insulated from unilateral tampering. The Supreme Court prescribes the federal rules (28 U.S.C. § 2072), Judicial Conference committees recommend them and notice their meetings (§ 2073), and Congress reviews them (§ 2074). Failure to comply with the committee procedure does not invalidate a rule (§ 2073(e)), and a court may make an immediate-need local rule first and take comment afterwards (§ 2071(e)). The insulation is procedural and partly unenforceable. The Supreme Court ruled in *Tumey v. Ohio* (1927) that a direct, personal, substantial pecuniary interest, in a criminal case, violates due process. The holding is not a general ban on any pecuniary interest. Repeat litigants hold systemic advantages (Galanter, 1974). Employees fare worse in arbitration than in court in Colvin's (2011) comparison, which is unmatched and draws litigation rates from other studies. A judge cannot rewrite the federal rules in the middle of a trial. The insulation is the statute, and it is partial. + +In foreign-exchange dealer markets, banks act as both liquidity providers and counterparties on their own platforms. "Last look" lets a dealer hold a client's order briefly and decide whether to accept it. Paragraphs 8, 13, and 24 of the 17 November 2015 NYDFS consent order, which are stipulated settlement text and not adjudicated findings, state that one bank used the practice to reject trades that had moved against it and did not tell clients why (NYDFS, 2015). Cartea, Jaimungal, and Walton (2019) model Last Look as a defence against latency arbitrage that tightens quoted spreads, with the cost to ordinary clients being the rejection of their favourable fills. The consent order is not the whole of the practice. + +The FCA's press release of 12 November 2014 records fines of £1,114,918,000 against five banks for control failures in G10 spot foreign exchange, and its finding is attempted manipulation (FCA, 2014). The Financial Stability Board's 2014 report does not report those fines. The industry's code followed. The FX Global Code asks dealers to disclose last-look practice (Principle 17). Principle 36 is a record-keeping duty, to keep reasons for rejections. Principle 40, under legal risk, says substantial changes to standard legal terms between counterparties should be agreed before any transaction. It is not a client-disclosure duty of the same kind as Principles 9 and 17. Adherence is voluntary. A soft-law convention holds as far as reputation and the threat of a regulator carry it. + +Digital platforms fare worst on the custody test in the terms Suzor examined. In all fourteen sets of terms that Suzor (2018) examined, the terms "provided broad, unfettered discretion to platform owners." The fourteen are the largest English-language social media platforms by traffic, a sample Suzor calls highly western- and US-centric. It includes no matching, pricing, or dispatch platform, so the sentence does not by itself cover ride-hailing or labor dispatch. Appeals were thin at Facebook in 2016–17 when Klonick (2018) studied them, and broader at YouTube and Twitter. Case decisions of the Meta Oversight Board (2022) are binding and its policy recommendations are advisory. The narrow remit that does not reach matching, pricing, or dispatch is the point made by Wong and Floridi (2023) and by Douek (2024). + +Technical standards bodies cannot fill the gap on their own. Voluntary adherence is the rule for Internet Standards in RFC 2026. End users take little part in the work Nottingham (2020) describes. Participation in the IETF from 2001 to 2023 is concentrated in a few organisations, and that dataset cannot observe end users (Zhang et al., 2025). If a platform can rewrite its matching conventions with a software update, a designer's promise of "strategic offloading" can be withdrawn whenever margins compress. + +### 7.3 External Compulsion and Statutes Beyond Disclosure + +Because platforms fail the custody test, Decision Experience Design conventions cannot survive through voluntary designer ethics alone. They require external legal compulsion. + +First-generation platform regulations relied largely on disclosure and procedure, and not only on those. The EU Platform-to-Business Regulation (Regulation (EU) 2019/1150) requires fifteen days' notice of changes to terms (Art. 3) and a description of the main ranking parameters (Art. 5), and it states that providers are not required to disclose their algorithms (Art. 5(6)). Article 3(3) makes non-compliant terms and changes null and void. The Digital Services Act (Regulation (EU) 2022/2065) requires a statement of reasons for restrictions (Art. 17) and an internal complaint-handling system (Art. 20), gives users a complaint to a regulator and a claim for damages (Arts. 53–54), and constrains interface design (Arts. 25 and 38). Of 131 million statements of reasons under the Act, 99.8% cite terms of service, and most facts fields are templates (Kaushal et al., 2024). These measures give notice, reasons, and some recourse. They do not, taken alone, leave every owner free to set every convention. The EU Unfair Contract Terms Directive has, since 1993, treated unilateral change-of-terms clauses as presumptively unfair (Loos & Luzak, 2016). The gap is enforcement as well as the text of the rule. + +Later statutes add structural duties beside disclosure. They do not replace it. + +1. The EU Platform Work Directive (Directive (EU) 2024/2831; in force since 1 December 2024; transposition due by 2 December 2026): bans certain processing of personal data (Art. 7); requires a system to be modified or discontinued where appropriate when there is a high risk of discrimination or of infringed rights (Art. 10(3)); requires that decisions to restrict, suspend, or terminate an account be taken by a human (Art. 10(5)); requires rectification within two weeks of the decision, and compensation only where rectification is not possible, where a decision infringes a worker's rights (Art. 11(3)). Article 13 extends an existing consultation right so that it covers decisions to introduce or substantially change automated systems, and it says representatives may be assisted by an expert. It does not establish the consultation right. The Directive also requires disclosure of main parameters (Art. 9). It gives workers no veto. It does not reach non-employee sellers. +2. The German Works Constitution Act is not recent. It dates from 1972. Section 87(1) no. 6 grants works councils co-determination over technical systems designed to monitor employee behavior or performance. Section 87(2) provides that the conciliation committee's award replaces an agreement. Later amendments add information and consultation duties that name artificial intelligence (§§ 80(3), 90(1) no. 3, 95(2a)). Those are not a veto. The Act does not reach consumer-facing platforms or non-employee sellers. + +These statutes are part of the institutional setting in which a DXD convention would have to be held. They are not, by themselves, the framework that sustains it, and they do not cover the sellers in this paper's own examples. Carolingian scriptoria were not uniformly directed by one royal and ecclesiastical authority to produce correct books (Charlemagne, trans. 1900; Boretius, 1883; Westwell, 2018). Digital platforms will maintain a representational convention only where some external rule removes the power to alter it at will, and which rule can do that is not settled by the two statutes above. + +## 8. A Research Agenda for Triadic Decision Design + +### 8.1 Four Falsifiable Claims + +The four claims and the five-arm protocol are how we would know a candidate is the gun. A result counts only if the task is gone from the person in the triad, on the surface. A result that hands them another text does not. In the manner of a framework and research agenda (Terveen & McDonald, 2005), I set out four claims, each stated so that data could count against it: + +- **Claim 1 (Irreducibility).** In a triad comprising Party A, Party B, and Mediator M, a proxy measure of integrated information computed on the joint behavior of all three will exceed the same measure computed on each person–mediator pair, and the excess will shrink when M's determinations are made independent of one party's moves. + - *Operationalization.* Integrated information (Φ) has so far been applied to human groups only by proxy: Engel and Malone (2018) computed a proxy on who was active when in a study of 68 groups, with the reported correlation (r = .37) on 61 of them, in groups with no mediator. Barrett et al. (2026) write that Φ is not well-defined for real physical systems and has not been computed on any real physical system. Niizato et al. (2020) did compute IIT 3.0 Φ on fish schools of two to five, and Φ rose with group size even with zero coupling, so a triad-versus-pairs comparison needs a size-matched uncoupled null. On human triads the study can test a proxy, not the untested quantity. I therefore fix the measure, the coding of behavior into discrete states, the time lag, and the partition in advance, validate them on small Boolean models where exact values can be computed, and treat the application to human triads as untested. +- **Claim 2 (Representational conventions outperform textual disclosure).** In a triadic allocation game where Mediator M optimizes an independent objective (J_M ≠ J_A, J_B), representing the system through Convention A (menu-based strategic offloading) produces higher joint surplus and less tactical distortion than full textual rule disclosure. + - *Operationalization.* The closest precedent is Casella, Friedman, and Perez Archila (2026), where two participants bargain through a computer mediator that maximizes the chance of agreement and shows its rule every round. I give M a margin of its own and compare joint surplus between rulebook arms and menu arms. +- **Claim 3 (Validity governs the sign of feedback immediacy).** The effect of feedback immediacy on error correction depends on mediator noise. Under low noise, continuous feedback speeds convergence; under high noise, continuous feedback degrades choices, and discrete filtered feedback yields higher joint surplus. + - *Operationalization.* Larson, List, and Metcalfe (2016) test how often traders see prices, under myopic loss aversion, not mediator noise. "Jitter" is this paper's variable. I compare continuous boundary feedback against discrete-round filtered feedback across high-jitter and low-jitter mediator regimes, and I do not treat a failure of the Larson result to travel as a failure of Claim 3. Beshears et al. (2017) already found that the frequency effect vanished with real funds over a year. +- **Claim 4 (Custodial drift degrades sincere play).** Conventions subject to unannounced mediator parameter drift decay into defensive, manipulative play, whereas conventions the mediator has committed not to change within a session sustain sincere coordination over repeated rounds. + +### 8.2 Five-Arm Multi-Agent Experimental Protocol + +To test these claims, I specify a controlled laboratory experiment. The protocol follows experimental-economics norms: no false statements to participants (Charness, Samek, & van de Ven, 2022), power calculated at the triad level (the independent unit of observation is the three-node triad, not the individual participant), and joint surplus preregistered as the primary outcome variable. + +``` +Experimental Architecture (5-Arm Triadic Protocol): + + [ Participant A ] <═══════> [ Mediator M ] <═══════> [ Participant B ] + ║ + Varying Interface / Representation Condition: + ═════════════════════════════════╬═════════════════════════════════ + Arm 1: Control (Black-Box Interface; Terminal Payout Only) + Arm 2: Disclosure (Full Written Rulebook, Margin Formulas, Audit Logs) + Arm 3: DXD Convention A (Posted-Price Menu / Strategic Offloading) + Arm 4a: DXD Convention B (Continuous Low-Latency Boundary Contours) + Arm 4b: Discrete Control (Identical Filtered Feedback Between Rounds) +``` + +1. Arm 1 (Control — Black-Box): Participants submit raw bids or demands through an unadorned interface, receiving only terminal round payoffs with no explanatory rules. +2. Arm 2 (Literacy — Full Disclosure): Participants receive complete documentation of M's objective function, matching formulas, and an itemized post-round audit log explaining how both inputs were scored. Claim 4's unannounced drift cannot be added to this arm. If the documentation says what M optimizes and M then changes without notice, the documentation becomes a false statement, which the norm cited above forbids (Charness et al., 2022). Drift without notice is defensible in arms that did not promise a complete account of M. +3. Arm 3 (DXD Convention A — Strategic Offloading): Participants interact with personalized menus where dominant strategies are pre-cleared, visually confirming that strategic undercutting has been removed. +4. Arm 4a (DXD Convention B — Continuous Immediacy): Participants receive real-time, pre-commit visual boundary contours that update as they adjust their sliders, showing clearing odds before commitment. +5. Arm 4b (Discrete Control for Convention B): Participants receive identical noise-filtered boundary data, delivered only at discrete intervals between rounds, isolating the effect of temporal immediacy from informational content. + +## 9. Conclusion and Scope Boundaries + +### 9.1 Synthesis + +The argument has three moves. Grahn et al. (2026) and Mourtzis and Tsakalos (2026) state the field, and the person in both reviews still takes something in. Nadella's post keeps the chain and the brake, and leaves the person in the triad to read. The space between words was the change the transition from oracy to literacy already contains, because it changed what the reader had to do (Zhang & Norman, 1994; Hutchins, 1995; Tunbridge, 1992). Vicente and Rasmussen (1992) write that no level of representation is globally superior. The two conventions are candidates for a change of the same kind. Neither has passed the test. + +Faced with opaque algorithmic intermediaries, a common response has been to demand more literacy: more documentation, more transparency dashboards, more explainability cards, and more post-hoc audit logs (Ben-Shahar & Schneider, 2011; Regulation (EU) 2019/1150, Art. 5; Regulation (EU) 2022/2065, Arts. 14 and 27). It is not the only response. The Digital Services Act already constrains interface design (Arts. 25, 38, and 40). + +This is a category error. Rulebooks and rationales treat human beings as readers of algorithms rather than deciders within mediated environments. Navigating irreducible triads requires algorithmacy—an enacted sensibility that operates through epistemic probing and the integration of valid feedback. + +Decision Experience Design (DXD) is the name of the search. Strategic offloading and ecological feedback immediacy are two candidates for the space between the words. Either is that moment only if it takes a task off the person in the triad and shows, on the surface, that the task is gone. Whether either one does that is untested. The paper's question stands: what could the spaces-between-words moment be? + +### 9.2 Explicit Scope Boundaries + +To ensure theoretical precision, this framework establishes five explicit scope boundaries: + +1. Analytical Guide, Not Historical Determinism: Word separation is used as an ergonomic guide for one change in what a reader had to do, and not as an explanation of historical literacy (Street, 2003). +2. Asymmetric Empirical Maturity: Convention A rests on experiments in mechanism design that show a real but partial effect (Li, 2017; Gonczarowski et al., 2024). Convention B and the measurement of triadic integrated information remain hypotheses awaiting the five-arm protocol. +3. Choice Architecture Boundary: DXD is a neighbor of nudging and boosting and must be placed against them (Hertwig & Grüne-Yanoff, 2017). A nudge alters how choices are presented within an unchanged game; DXD changes what the mechanism requires a person to calculate and shows that it has. The nudge literature classes re-presentation of information with informational interventions, the weaker kind (Mertens, Herberz, Hahnel, & Brosch, 2022), and DXD's claim to differ rests on Claim 2. +4. Theoretical Status of Working Disclosures: The argument that working disclosures succeed insofar as they function as simple perceptual signals is this paper's theoretical claim, not an established consensus in prior literature. +5. Institutional Boundary: Interaction conventions cannot withstand commercial pressure without some external constraint, and the two statutes in §7.3 do not yet supply it for the sellers and riders in the examples. The Carolingian page was not sustained by one church and one court (Westwell, 2018). What a modern convention requires is a rule that removes the operator's power to alter the representation at will. Which rule can do that is open. + +## References + +Agarwal, S., Chomsisengphet, S., Mahoney, N., & Stroebel, J. (2015). Regulating consumer financial products: Evidence from credit cards. *The Quarterly Journal of Economics, 130*(1), 111–164. https://doi.org/10.1093/qje/qju037 + +Artemov, G., Che, Y.-K., & He, Y. (2023). Stable matching with mistaken agents. *Journal of Political Economy Microeconomics, 1*(2), 270–320. https://doi.org/10.1086/722978 + +Bai, X., Yan, G., Liversedge, S. P., Zang, C., & Rayner, K. (2008). Reading spaced and unspaced Chinese text: Evidence from eye movements. *Journal of Experimental Psychology: Human Perception and Performance, 34*(5), 1277–1287. https://doi.org/10.1037/0096-1523.34.5.1277 + +Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. (2021). Does the whole exceed its parts? The effect of AI explanations on complementary team performance. In *Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems* (pp. 1–16). ACM. https://doi.org/10.1145/3411764.3445717 + +Barrett, A. B., Milinkovic, B., Mediano, P. A. M., Rosas, F. E., Bor, D., Barnett, L., & Seth, A. K. (2026). *Integrated information theory: The good, the bad and the misunderstood* [Preprint]. + +Bassetti, B., & Masterson, J. (2012). Effects of removing morphemic information and adding interword spacing on reading in Chinese experienced and inexperienced readers. *Reading and Writing, 25*(9), 2291–2314. https://doi.org/10.1007/s11145-012-9358-7 + +Ben-Shahar, O., & Schneider, C. E. (2011). The failure of mandated disclosure. *University of Pennsylvania Law Review, 159*, 647–749. + +Beshears, J., Choi, J. J., Laibson, D., & Madrian, B. C. (2017). Does aggregated returns disclosure increase portfolio risk taking? *The Review of Financial Studies, 30*(6), 1971–2005. https://doi.org/10.1093/rfs/hhw086 + +Boretius, A. (Ed.). (1883). *Capitularia regum Francorum* (Vol. 1; Monumenta Germaniae Historica, Legum sectio II). Hahn. + +Bundesministerium der Justiz. (n.d.). *Betriebsverfassungsgesetz (BetrVG)* [Works Constitution Act], § 87. + +Burnyeat, M. F. (1997). Postscript on silent reading. *The Classical Quarterly, 47*(1), 74–76. https://doi.org/10.1093/cq/47.1.74 + +Busch, S. (2002). Lautes und leises Lesen in der Antike [Reading aloud and silently in antiquity]. *Rheinisches Museum für Philologie, 145*(1), 1–45. + +Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. *Proceedings of the ACM on Human-Computer Interaction, 5*(CSCW1), Article 188. https://doi.org/10.1145/3449287 + +Cabrera, Á. A., Perer, A., & Hong, J. I. (2023). Improving human-AI collaboration with descriptions of AI behavior. *Proceedings of the ACM on Human-Computer Interaction, 7*(CSCW1), 1–21. https://doi.org/10.1145/3579612 + +Cartea, Á., Jaimungal, S., & Walton, J. (2019). Foreign exchange markets with Last Look. *Mathematics and Financial Economics, 13*(1), 1–30. https://doi.org/10.1007/s11579-018-0218-3 + +Casella, A., Friedman, E., & Perez Archila, M. (2026). On the fragility of mediation: Theory and experimental evidence. *Journal of the European Economic Association, 24*(3), 801–850. https://doi.org/10.1093/jeea/jvaf049 + +Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W. (2014). The retention of manual flying skills in the automated cockpit. *Human Factors, 56*(8), 1506–1516. https://doi.org/10.1177/0018720814535628 + +CERT National Insider Threat Center. (2022). *Common sense guide to mitigating insider threats* (7th ed.). Software Engineering Institute, Carnegie Mellon University. + +Chantavarin, S., & Leung, T. T.-C. (2025). The effects of interword spacing and morphological complexity in reading Thai: An eye-tracking study. *Language and Cognition, 17*. https://doi.org/10.1017/langcog.2025.10028 + +Charlemagne. (1900). Letter to Abbot Baugulf of Fulda (*De litteris colendis*), c. 780–800 (D. C. Munro, Ed. & Trans.). In *Laws of Charles the Great*. *Translations and Reprints from the Original Sources of European History, 6*(5), 12–14. + +Charness, G., Samek, A., & van de Ven, J. (2022). What is considered deception in experimental economics? *Experimental Economics, 25*(2), 385–412. https://doi.org/10.1007/s10683-021-09726-7 + +Chen, L., Mislove, A., & Wilson, C. (2015). Peeking beneath the hood of Uber. In *Proceedings of the 2015 Internet Measurement Conference* (pp. 495–508). ACM. https://doi.org/10.1145/2815675.2815681 + +Chen, V., Liao, Q. V., Wortman Vaughan, J., & Bansal, G. (2023). Understanding the role of human intuition on reliance in human-AI decision-making with explanations. *Proceedings of the ACM on Human-Computer Interaction, 7*(CSCW2), 1–32. https://doi.org/10.1145/3610219 + +Clanchy, M. T. (1993). *From memory to written record: England 1066–1307* (2nd ed.). Blackwell. + +Colvin, A. J. S. (2011). An empirical study of employment arbitration: Case outcomes and processes. *Journal of Empirical Legal Studies, 8*(1), 1–23. https://doi.org/10.1111/j.1740-1461.2010.01200.x + +Cressy, D. (1980). *Literacy and the social order: Reading and writing in Tudor and Stuart England*. Cambridge University Press. + +Cui, Y. (2023). Eye movements of second language learners when reading spaced and unspaced Chinese texts. *Frontiers in Psychology, 14*, Article 783960. https://doi.org/10.3389/fpsyg.2023.783960 + +Dickey, E. (2017). Word division in bilingual texts. In G. Nocchi Macedo & M. C. Scappaticcio (Eds.), *Signes dans les textes, textes sur les signes* (pp. 159–175). Presses universitaires de Liège. + +Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. *Journal of Experimental Psychology: General, 144*(1), 114–126. https://doi.org/10.1037/xge0000033 + +Directive (EU) 2024/2831 of the European Parliament and of the Council of 23 October 2024 on improving working conditions in platform work. *Official Journal of the European Union*. + +Disley, M., & Khan, M. (2026). Designing cognitive friction with concept suppression in large language models. *Human–Computer Interaction*. Advance online publication. https://doi.org/10.1080/07370024.2026.2708617 + +Douek, E. (2024). The Meta Oversight Board and the empty promise of legitimacy. *Harvard Journal of Law & Technology, 37*(2), 373–445. + +Dranove, D., Kessler, D., McClellan, M., & Satterthwaite, M. (2003). Is more information better? The effects of "report cards" on health care providers. *Journal of Political Economy, 111*(3), 555–588. https://doi.org/10.1086/374180 + +Dümmler, E. (Ed.). (1881). *Poetae Latini aevi Carolini* (Vol. 1). Weidmann. Alcuin, *Carmina* XCIV, p. 320. + +Ehsan, U., Liao, Q. V., Muller, M., Riedl, M. O., & Weisz, J. D. (2021). Expanding explainability: Towards social transparency in AI systems. In *Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems* (pp. 1–19). ACM. https://doi.org/10.1145/3411764.3445188 + +Eisenstein, E. L. (1969). The advent of printing and the problem of the Renaissance. *Past & Present, 45*(1), 19–89. https://doi.org/10.1093/past/45.1.19 + +Engel, D., & Malone, T. W. (2018). Integrated information as a metric for group interaction. *PLOS ONE, 13*(10), Article e0205335. https://doi.org/10.1371/journal.pone.0205335 + +Farronato, C., Fradkin, A., & MacKay, A. (2023). Self-preferencing at Amazon: Evidence from search rankings. *AEA Papers and Proceedings, 113*, 239–243. https://doi.org/10.1257/pandp.20231068 + +Financial Conduct Authority. (2014, November 12). *FCA fines five banks £1.1 billion for FX failings and announces industry-wide remediation programme* [Press release]. + +Financial Stability Board. (2014). *Foreign exchange benchmarks: Final report*. + +Fung, A., Graham, M., & Weil, D. (2007). *Full disclosure: The perils and promise of transparency*. Cambridge University Press. + +Galanter, M. (1974). Why the "haves" come out ahead: Speculations on the limits of legal change. *Law & Society Review, 9*(1), 95–160. https://doi.org/10.2307/3053023 + +Gavrilov, A. K. (1997). Techniques of reading in classical antiquity. *The Classical Quarterly, 47*(1), 56–73. https://doi.org/10.1093/cq/47.1.56 + +Ghalme, G., Nair, V., Eilat, I., Talgam-Cohen, I., & Rosenfeld, N. (2021). *Strategic classification in the dark* [Preprint]. arXiv. + +Global Foreign Exchange Committee. (2024). *FX Global Code: A set of global principles of good practice in the foreign exchange market* (December 2024 update). + +Gonczarowski, Y. A., Heffetz, O., Ishai, G., & Thomas, C. (2024). *Describing deferred acceptance and strategyproofness to participants: Experimental analysis* (Working Paper No. 33020). National Bureau of Economic Research. https://doi.org/10.3386/w33020 + +Grahn, S., Wang, T., Wang, X. V., & Wang, L. (2026). From human–machine interaction to integration: A review of paradigms, principles, and prospects for a human-centric future. *Human-Machine Interaction, 1*(1), 1. https://www.sciltp.com/journals/hmi + +Heilmann, J. (2020). Reading early New Testament manuscripts: Scriptio continua, "reading aids", and other characteristic features. In *Material aspects of reading in ancient and medieval cultures* (pp. 177–196). De Gruyter. https://doi.org/10.1515/9783110639247-011 + +Herlocker, J. L., Konstan, J. A., & Riedl, J. (2000). Explaining collaborative filtering recommendations. In *Proceedings of the 2000 ACM Conference on Computer Supported Cooperative Work* (pp. 241–250). ACM. https://doi.org/10.1145/358916.358995 + +Hertwig, R., & Grüne-Yanoff, T. (2017). Nudging and boosting: Steering or empowering good decisions. *Perspectives on Psychological Science, 12*(6), 973–986. https://doi.org/10.1177/1745691617702496 + +Ho, D. E. (2012). Fudging the nudge: Information disclosure and restaurant grading. *The Yale Law Journal, 122*(3), 574–688. + +Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J., & Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. *Scientific Reports, 13*, Article 5487. https://doi.org/10.1038/s41598-023-30938-9 + +Horvath, A., Liu, M. X., Hendriksen, R., Shannon, C., Paterson, E., Jawad, K., Macvean, A., & Myers, B. A. (2022). Understanding how programmers can use annotations on documentation. In *CHI Conference on Human Factors in Computing Systems* (pp. 1–16). ACM. https://doi.org/10.1145/3491102.3502095 + +Hutchins, E. (1995). How a cockpit remembers its speeds. *Cognitive Science, 19*(3), 265–288. https://doi.org/10.1207/s15516709cog1903_1 + +Ishowo-Oloko, F., Bonnefon, J.-F., Soroye, Z., Crandall, J., Rahwan, I., & Rahwan, T. (2019). Behavioural evidence for a transparency–efficiency tradeoff in human–machine cooperation. *Nature Machine Intelligence, 1*(11), 517–521. https://doi.org/10.1038/s42256-019-0113-5 + +Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs and ownership structure. *Journal of Financial Economics, 3*(4), 305–360. https://doi.org/10.1016/0304-405X(76)90026-X + +Jhaver, S., Appling, D. S., Gilbert, E., & Bruckman, A. (2019). "Did you suspect the post would be removed?": Understanding user reactions to content removals on Reddit. *Proceedings of the ACM on Human-Computer Interaction, 3*(CSCW), Article 192. https://doi.org/10.1145/3359294 + +Jin, G. Z., & Leslie, P. (2003). The effect of information on product quality: Evidence from restaurant hygiene grade cards. *The Quarterly Journal of Economics, 118*(2), 409–451. https://doi.org/10.1162/003355303321675428 + +Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. *American Psychologist, 64*(6), 515–526. https://doi.org/10.1037/a0016755 + +Kaushal, R., van de Kerkhof, J., Goanta, C., Spanakis, G., & Iamnitchi, A. (2024). Automated transparency: A legal and empirical analysis of the Digital Services Act Transparency Database. In *Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency* (pp. 1121–1132). ACM. https://doi.org/10.1145/3630106.3658960 + +Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. *Academy of Management Annals, 14*(1), 366–410. https://doi.org/10.5465/annals.2018.0174 + +Kirsh, D. (1995). The intelligent use of space. *Artificial Intelligence, 73*(1–2), 31–68. https://doi.org/10.1016/0004-3702(94)00017-U + +Kirsh, D., & Maglio, P. (1994). On distinguishing epistemic from pragmatic action. *Cognitive Science, 18*(4), 513–549. https://doi.org/10.1207/s15516709cog1804_1 + +Klonick, K. (2018). The new governors: The people, rules, and processes governing online speech. *Harvard Law Review, 131*, 1598–1670. + +Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. *Psychological Bulletin, 119*(2), 254–284. https://doi.org/10.1037/0033-2909.119.2.254 + +Kluger, A. N., & DeNisi, A. (1998). Feedback interventions: Toward the understanding of a double-edged sword. *Current Directions in Psychological Science, 7*(3), 67–72. https://doi.org/10.1111/1467-8721.ep10772989 + +Knijnenburg, B. P., Willemsen, M. C., Gantner, Z., Soncu, H., & Newell, C. (2012). Explaining the user experience of recommender systems. *User Modeling and User-Adapted Interaction, 22*(4–5), 441–504. https://doi.org/10.1007/s11257-011-9118-4 + +Knox, B. M. W. (1968). Silent reading in antiquity. *Greek, Roman, and Byzantine Studies, 9*(4), 421–435. + +Körber, M. (2018). *Theoretical considerations and development of a questionnaire to measure trust in automation* [Preprint]. PsyArXiv. https://doi.org/10.31234/osf.io/nfc45 + +Larson, F., List, J. A., & Metcalfe, R. D. (2016). *Can myopic loss aversion explain the equity premium puzzle? Evidence from a natural field experiment with professional traders* (Working Paper No. 22605). National Bureau of Economic Research. https://doi.org/10.3386/w22605 + +Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In *Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems* (pp. 1–22). ACM. https://doi.org/10.1145/3706598.3713778 + +Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. *Human Factors, 46*(1), 50–80. https://doi.org/10.1518/hfes.46.1.50.30392 + +Lee, M. K., Kusbit, D., Metsky, E., & Dabbish, L. (2015). Working with machines: The impact of algorithmic and data-driven management on human workers. In *Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems* (pp. 1603–1612). ACM. https://doi.org/10.1145/2702123.2702548 + +Lee, W. J., Li, A. X., Lee, J. E., & Hayes, B. K. (2023). Learning traps lead to change blindness in dynamic environments. In M. Goldwater, F. K. Anggoro, B. K. Hayes, & D. C. Ong (Eds.), *Proceedings of the 45th Annual Conference of the Cognitive Science Society* (pp. 2318–2324). Cognitive Science Society. + +Li, L., Chu, W., Langford, J., & Schapire, R. E. (2010). A contextual-bandit approach to personalized news article recommendation. In *Proceedings of the 19th International Conference on World Wide Web* (pp. 661–670). ACM. https://doi.org/10.1145/1772690.1772758 + +Li, R., Liu, X., & Sundararajan, A. (2026). *Platform design, earnings transparency and minimum wage policies: Evidence from a natural experiment on Lyft* [Preprint]. arXiv. https://arxiv.org/abs/2602.08955 + +Li, S. (2017). Obviously strategy-proof mechanisms. *American Economic Review, 107*(11), 3257–3287. https://doi.org/10.1257/aer.20160425 + +Liu, H., Lai, V., & Tan, C. (2021). Understanding the effect of out-of-distribution examples and interactive explanations on human-AI decision making. *Proceedings of the ACM on Human-Computer Interaction, 5*(CSCW2), Article 408. https://doi.org/10.1145/3479552 + +Loewenstein, G., Sunstein, C. R., & Golman, R. (2014). Disclosure: Psychology changes everything. *Annual Review of Economics, 6*, 391–419. https://doi.org/10.1146/annurev-economics-080213-041341 + +Loos, M., & Luzak, J. (2016). Wanted: A bigger stick. On unfair terms in consumer contracts with online service providers. *Journal of Consumer Policy, 39*(1), 63–90. https://doi.org/10.1007/s10603-015-9303-7 + +Masuda, T., Mikami, R., Sakai, T., Serizawa, S., & Wakayama, T. (2022). The net effect of advice on strategy-proof mechanisms: An experiment for the Vickrey auction. *Experimental Economics, 25*(3), 902–941. https://doi.org/10.1007/s10683-021-09736-5 + +*Mathews v. Eldridge*, 424 U.S. 319 (1976). + +Matsuo, R., Liu, H., Hiraoka, T., & Wada, T. (2026). An educational human machine interface providing request-to-intervene trigger and reason explanation for enhancing the driver's comprehension of ADS's system limitations. *IEEE Transactions on Human-Machine Systems, 56*(3), 650–661. https://doi.org/10.1109/thms.2026.3679748 + +McGowan, V. A., White, S. J., Jordan, T. R., & Paterson, K. B. (2014). Aging and the use of interword spaces during reading: Evidence from eye movements. *Psychonomic Bulletin & Review, 21*(3), 740–747. + +Mejia, J., & Parker, C. (2021). When transparency fails: Bias and financial incentives in ridesharing platforms. *Management Science, 67*(1), 166–184. + +Mertens, S., Herberz, M., Hahnel, U. J. J., & Brosch, T. (2022). The effectiveness of nudging: A meta-analysis of choice architecture interventions across behavioral domains. *Proceedings of the National Academy of Sciences, 119*(1), Article e2107346118. https://doi.org/10.1073/pnas.2107346118 (Correction published 2022.) + +Metzger, U., & Parasuraman, R. (2001). The role of the air traffic controller in future air traffic management: An empirical study of active control versus passive monitoring. *Human Factors, 43*(4), 519–528. https://doi.org/10.1518/001872001775870421 + +Mourtzis, D., & Tsakalos, D. (2026). Emerging technologies in human-machine interaction towards Industry 5.0 and industrial metaverse: Challenges and opportunities. *Human-Machine Interaction, 1*(1), 4. https://www.sciltp.com/journals/hmi + +Mozannar, H., Bansal, G., Fourney, A., & Horvitz, E. (2024). Reading between the lines: Modeling user behavior and costs in AI-assisted programming. In *Proceedings of the CHI Conference on Human Factors in Computing Systems* (pp. 1–16). ACM. https://doi.org/10.1145/3613904.3641936 + +Nadella, S. (2026, October 10). *Models as insider risks in the super intelligence era* [Post]. X. https://x.com/satyanadella/status/2108931348857827686 + +New York State Department of Financial Services. (2015, November 17). *In the matter of Barclays Bank PLC, Barclays Bank PLC, New York Branch: Consent order under New York Banking Law § 44*. + +Niizato, T., Sakamoto, K., Mototake, Y., Murakami, H., Tomaru, T., Hoshika, T., & Fukushima, T. (2020). Finding continuity and discontinuity in fish schools via integrated information theory. *PLOS ONE, 15*(2), Article e0229573. https://doi.org/10.1371/journal.pone.0229573 + +Norman, D. A. (1988). *The psychology of everyday things*. Basic Books. + +Nottingham, M. (2020). *The Internet is for end users* (RFC 8890). Internet Architecture Board. + +Nouwens, M., Liccardi, I., Veale, M., Karger, D., & Kagal, L. (2020). Dark patterns after the GDPR: Scraping consent pop-ups and demonstrating their influence. In *Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems* (pp. 1–13). ACM. https://doi.org/10.1145/3313831.3376321 + +Oberschelp, W. (1995). Developing analytical competency through informatics teaching. In D. Watson & D. Tinsley (Eds.), *Integrating information technology into education* (pp. 115–126). Springer. https://doi.org/10.1007/978-0-387-34842-1_12 + +Oomen, R. (2017). Last look. *Quantitative Finance, 17*(7), 1057–1070. + +Oralova, G., & Kuperman, V. (2021). Effects of spacing on sentence reading in Chinese. *Frontiers in Psychology, 12*, Article 765335. + +Oversight Board. (2022). *Q1 2022 transparency report*. + +Özçetin, S., & Wiltse, H. (2025). Terms of entanglement: A posthumanist reading of Terms of Service. *Human–Computer Interaction, 40*(1–4), 171–194. https://doi.org/10.1080/07370024.2023.2281928 (Published online 29 November 2023.) + +Pan, J., Wang, A., Zhang, M., Tsang, Y.-K., & Yan, M. (2025). Printing words in alternating colors facilitates eye movements among young and older Chinese adults. *Psychonomic Bulletin & Review, 32*(2), 855–865. https://doi.org/10.3758/s13423-024-02581-6 + +Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. *Human Factors, 52*(3), 381–410. https://doi.org/10.1177/0018720810376055 + +Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. *IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 30*(3), 286–297. https://doi.org/10.1109/3468.844354 + +Parkes, M. B. (1992). *Pause and effect: An introduction to the history of punctuation in the West*. Scolar Press. + +Poursabzi-Sangdeh, F., Goldstein, D. G., Hofman, J. M., Wortman Vaughan, J., & Wallach, H. (2021). Manipulating and measuring model interpretability. In *Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems* (Article 237). ACM. https://doi.org/10.1145/3411764.3445315 + +Pucci, E., Piro, L., & Matera, M. (2027). From tabbing to talking: LLM-driven conversational agents for web form interaction. *International Journal of Human-Computer Studies, 218*, Article 103925. https://doi.org/10.1016/j.ijhcs.2026.103925 + +Pycia, M., & Troyan, P. (2023). A theory of simplicity in games and mechanism design. *Econometrica, 91*(4), 1495–1526. https://doi.org/10.3982/ECTA16310 + +Rahman, H. A. (2019). *Invisible cages: Algorithmic evaluations in online labor markets* [Doctoral dissertation, Stanford University]. + +Rahmandad, H., & Gary, M. S. (2023). Delays impair learning and can drive convergence to inefficient strategies. *Organization Science, 34*(6), 2392–2414. https://doi.org/10.1287/orsc.2020.1405 + +Rayner, K., Fischer, M. H., & Pollatsek, A. (1998). Unspaced text interferes with both word identification and eye movement control. *Vision Research, 38*(8), 1129–1144. https://doi.org/10.1016/S0042-6989(97)00274-5 + +Regulation (EU) 2019/1150 of the European Parliament and of the Council of 20 June 2019 on promoting fairness and transparency for business users of online intermediation services. *Official Journal of the European Union*. + +Regulation (EU) 2022/2065 of the European Parliament and of the Council of 19 October 2022 on a Single Market For Digital Services (Digital Services Act). *Official Journal of the European Union*. + +Rich, A. S., & Gureckis, T. M. (2018). The limits of learning: Exploration, generalization, and the development of learning traps. *Journal of Experimental Psychology: General, 147*(11), 1553–1570. https://doi.org/10.1037/xge0000466 + +Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. *Trends in Cognitive Sciences, 20*(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002 + +Rogers, R., Subramaniam, S., Peng, S., Durfee, D., Lee, S., Kancha, S. K., Sahay, S., & Ahammad, P. (2021). LinkedIn's Audience Engagements API: A privacy preserving data analytics system at scale. *Journal of Privacy and Confidentiality, 11*(3). https://doi.org/10.29012/jpc.782 + +Rosenblat, A., & Stark, L. (2016). Algorithmic labor and information asymmetries: A case study of Uber's drivers. *International Journal of Communication, 10*, 3758–3784. + +Rules Enabling Act, 28 U.S.C. § 2073. + +Saenger, P. (1997). *Space between words: The origins of silent reading*. Stanford University Press. + +Schemmer, M., Hemmer, P., Nitsche, M., Kühl, N., & Vössing, M. (2022). A meta-analysis of the utility of explainable artificial intelligence in human-AI decision-making. In *Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society* (pp. 617–626). ACM. https://doi.org/10.1145/3514094.3534128 + +Seira, E., Elizondo, A., & Laguna-Müggenburg, E. (2017). Are information disclosures effective? Evidence from the credit card market. *American Economic Journal: Economic Policy, 9*(1), 277–307. https://doi.org/10.1257/pol.20140404 + +Senoner, J., Schallmoser, S., Kratzwald, B., Feuerriegel, S., & Netland, T. (2024). Explainable AI improves task performance in human–AI collaboration. *Scientific Reports, 14*, Article 31150. + +Shen, H., DeVos, A., Eslami, M., & Holstein, K. (2021). Everyday algorithm auditing: Understanding the power of everyday users in surfacing harmful algorithmic behaviors. *Proceedings of the ACM on Human-Computer Interaction, 5*(CSCW2), Article 433. https://doi.org/10.1145/3479577 + +Sheridan, T. B., & Verplank, W. L. (1978). *Human and computer control of undersea teleoperators* [Technical report]. Massachusetts Institute of Technology, Man-Machine Systems Laboratory. https://doi.org/10.21236/ada057655 + +Shneiderman, B. (1983). Direct manipulation: A step beyond programming languages. *Computer, 16*(8), 57–69. https://doi.org/10.1109/mc.1983.1654471 + +Simkute, A., Tankelevitch, L., Kewenig, V., Scott, A. E., Sellen, A., & Rintel, S. (2024). Ironies of generative AI: Understanding and mitigating productivity loss in human-AI interaction. *International Journal of Human–Computer Interaction*. Advance online publication. https://doi.org/10.1080/10447318.2024.2405782 + +Street, B. V. (2003). What's "new" in New Literacy Studies? Critical approaches to literacy in theory and practice. *Current Issues in Comparative Education, 5*(2), 77–91. + +Sutherland, W., Jarrahi, M. H., Dunn, M., & Nelson, S. B. (2020). Work precarity and gig literacies in online freelancing. *Work, Employment and Society, 34*(3), 457–475. https://doi.org/10.1177/0950017019886511 + +Suzor, N. (2018). Digital constitutionalism: Using the rule of law to evaluate the legitimacy of governance by platforms. *Social Media + Society, 4*(3). https://doi.org/10.1177/2056305118787812 + +Terveen, L., & McDonald, D. W. (2005). Social matching: A framework and research agenda. *ACM Transactions on Computer-Human Interaction, 12*(3), 401–434. https://doi.org/10.1145/1096737.1096740 + +Tunbridge, G. L. (1992). *A study of scribal practices in early Irish and Anglo-Saxon manuscripts* [Doctoral dissertation, University of Oxford]. + +*Tumey v. Ohio*, 273 U.S. 510 (1927). + +Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. *Nature Human Behaviour, 8*(12), 2293–2303. https://doi.org/10.1038/s41562-024-02024-1 + +Vasconcelos, H., Jörke, M., Grunde-McLaughlin, M., Gerstenberg, T., Bernstein, M. S., & Krishna, R. (2023). Explanations can reduce overreliance on AI systems during decision-making. *Proceedings of the ACM on Human-Computer Interaction, 7*(CSCW1), Article 129. https://doi.org/10.1145/3579605 + +Vicente, K. J., & Rasmussen, J. (1992). Ecological interface design: Theoretical foundations. *IEEE Transactions on Systems, Man, and Cybernetics, 22*(4), 589–606. + +Weick, K. E. (1988). Enacted sensemaking in crisis situations. *Journal of Management Studies, 25*(4), 305–317. https://doi.org/10.1111/j.1467-6486.1988.tb00039.x + +Westwell, A. (2018). *The dissemination and reception of the Ordines Romani in the Carolingian church, c. 750–900* [Doctoral dissertation, University of Cambridge]. https://doi.org/10.17863/cam.23255 + +Winskel, H., Radach, R., & Luksaneeyanawin, S. (2009). Eye movements when reading spaced and unspaced Thai and English: A comparison of Thai–English bilinguals and English monolinguals. *Journal of Memory and Language, 61*(3), 339–351. https://doi.org/10.1016/j.jml.2009.07.002 + +Wischnewski, M., Krämer, N., & Müller, E. (2023). Measuring and understanding trust calibrations for automated systems: A survey of the state-of-the-art and future directions. In *Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems*. ACM. + +Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. *Frontiers in Psychology, 10*, Article 3087. https://doi.org/10.3389/fpsyg.2019.03087 + +Wolfe, J. M. (2021). Guided Search 6.0: An updated model of visual search. *Psychonomic Bulletin & Review, 28*(4), 1060–1092. https://doi.org/10.3758/s13423-020-01859-9 + +Wong, D., & Floridi, L. (2023). Meta's Oversight Board: A review and critical assessment. *Minds and Machines, 33*(2), 261–284. + +Zhang, J., & Norman, D. A. (1994). Representations in distributed cognitive tasks. *Cognitive Science, 18*(1), 87–122. https://doi.org/10.1207/s15516709cog1801_3 + +Zhang, Y., McQuistin, S., Karan, V., Ramirez-Centeno, H., Perkins, C., Tyson, G., & Castro, I. (2025). Two decades of IETF affiliations: Evolution and impact. In *Proceedings of the Applied Networking Research Workshop* (pp. 17–23). ACM. https://doi.org/10.1145/3744200.3744757 + +## Changes from version 1 (Claude's notes; not part of the paper) + +**Reference list.** Rebuilt. Thirty-nine entries were replaced with the published record (the full table +is in `REVIEW_v2_2026-10-13.md` §2a). Entries the text never cited were removed (Arteaga et al.; Ashlagi +& Gonczarowski; Balkin; Bar-Gill et al.; Bechavod et al.; Brehmer & Nählinder; Crabtree; Dietvorst et +al.; Dix; Douek; Hardt et al.; Lee & Moray; Rees-Jones & Skowronek; Scaife & Rogers; Thorgeirsson et al.; +Vicente & Rasmussen; Wickens & Carswell). Entries were added for every work the text cites. Three +citations were removed because no record could be found or the record did not match the claim: +Kosmyna et al. (2025), Thorgeirsson et al. (2026), and Sáenz & De Russis (2027), whose only Crossref +match is a paper with a different title. + +**Statements corrected against their sources.** + +| Section | Version 1 | Version 2 | +| --- | --- | --- | +| 1, 4.2 | Repayment disclosures "increased delinquency among borrowers"; "Mexican credit card companies"; "made the debt feel insurmountable" | One bank's messages; delinquency rose for two client groups by 0.011 points (7.4% of the mean); the invented mechanism is removed | +| 1, 4.2 | "High-interpretability models impaired users' ability to correct anomalous model mistakes"; "an illusion of competence" | A clear model left users less able to correct the mistake on an unusual case, in three of four experiments; the authors' own conjecture (overload) is given | +| 4.2 | Explained removals judged less fair than unexplained ones | Rule-readers judged removal less fair (β = −.314); those given an explanation judged it fairer (β = .384) | +| 4.2 | Layout shifted consent "by orders of magnitude"; text "produced no detectable effect" | Removing the opt-out button raised consent by 22 to 23 points; the claim about text is removed | +| 1, 4.2 | "Explanations do not improve decision performance on average (Vaccaro…)" | Schemmer et al. for the average effect; Vaccaro et al. for explanation not moderating performance | +| 3.1 | Reading "confined to an elite, highly trained scribal caste"; Insular languages "shared no linguistic roots with Latin"; distinctiones and "codex page indexing" standardized by the eleventh century | Saenger's claims attributed to him and marked as disputed; Insular scribes learned Latin as a foreign language; the unsourced details are removed; his p. 85 sentence is quoted | +| 3.2 | Boundary detection in "the phonological loop"; saccade claims unsourced | Marked as this paper's interpretation; landing-position effect cited to Rayner et al. | +| 3.3 | Spaces help children and learners; spaces slow skilled Chinese readers (Bai) | Bai: no difference; Thai: faster words, slower sentences; child learners worse (Bassetti & Masterson); adult learners not significant (Cui) | +| 3.3, 7.3 | Carolingian edicts cited to Westwell; scribes "held accountable to readers" | The three primary texts named; Westwell's caution added; a sentence states that none concerns word separation | +| 4.1 | A quotation from Ben-Shahar & Schneider not found at p. 665; "a bright caloric label" | Their mechanism paraphrased from p. 665; their own example (restaurant grades) with its hedge | +| 4.1 | Only Target 4 said to fail, while the thesis condemned all disclosure | The abstract, introduction and 4.1 now locate the category error in Target 4 and say why Target 3 differs | +| 4.3 | "Three pillars"; "halt switches" in quotation marks; affected people "handed the… audit log" | Seven principles, three of which bear on a person; his emergency-brake image; affected people are promised "timely disclosure" | +| 5.1 | Oberschelp "treated algorithmacy as elementary procedural programming literacy" | His one sentence quoted (p. 121); undefined; a hand-computation skill | +| 5.2 | Drivers with a second phone driving "two blocks north"; Ghalme et al. cited for sharing screenshots | What Lee et al. and Kellogg et al. report; Ghalme et al. cited for inferring a hidden rule from others' outcomes | +| 5.3 | Kahneman & Klein quoted beyond the verified words; platforms "deliberately disrupt ecological validity"; "an engineering synchronization bug in their geospatial grid" | Only "feedback that is both rapid and unequivocal" is quoted; intent is claimed for secrecy and denied for noise; the bug is Uber's own account | +| 5.3 | Rahmandad & Gary: failure "even after hundreds of trials… superstitious heuristics" | Delay and noise have different effects, as the paper reports | +| 6.2 | "Comprehension from 56% to 71%. Sincere play increased from 48% to 59%" | Understanding 71% against 58% and 56%; play 59% against 56% (not significant) and 48% (control) | +| 6.2 | Masuda et al. "pairing a structural representation with a clear confirmation" | A statement that sincere bidding is dominant; 20.6% to 46.9%; efficiency up about one point | +| 6.3 | Larson et al. in the *Journal of Finance*; "financial traders… real-time price telemetry" | An NBER working paper on professional traders; second-by-second against four-hourly prices | +| 7.2 | Code followed by fines; Principles 7, 9, 36, 39, 40 "governing last-look and pre-hedging" | Fines in 2014–15, Code afterward; Principle 17 for last look and the others for what they cover | +| 7.2 | RFC 8890 "explicitly declare[s]" standards do not bind platforms | Standards are voluntary and end users take little part; cited to Nottingham (2020) | +| 7.3 | P2B and DSA "merely" require plain-language description, "changing nothing" | Their notice, reasons, complaint and damages provisions stated; P2B's no-algorithm-disclosure clause added | +| 7.3 | Directive article summaries without their qualifiers | "Where appropriate" (Art. 10(3)); rights infringed (Art. 11(3)); no veto | +| 8.1 | Claim 1: "strictly positive integrated information (Φ > 0) that cannot be factored" | A comparison the data could fail, with the proxy status of the measure stated | +| 8.1 | Claim 4: "verifiable cryptographic or co-governance lock-in" | A commitment not to change the convention within a session | +| 9.1 | "where reading stood before the eighth century" | "before word separation" | +| 9.2 | Convention A "rests on established empirical mechanism design"; one sentence on nudging | "A real but partial effect"; nudge and boost named as the neighbor, with Mertens et al.'s classification | + +**Structure and voice.** The roadmap is prose. Where "we" meant the author it is now "I"; the inclusive "we" that addresses the reader is kept. Convention B is "ecological +feedback immediacy" everywhere. The claim that HCI is dyadic is narrowed to transparency research, with +AI-mediated communication acknowledged. "Importantly," and "Notably," are removed. The diagrams are kept +as they were, inside code blocks; they still need redrawing as figures. + +**Not done.** The paper is not lengthened. The body runs to about 7,600 words excluding diagrams and references, against 9,000 to 17,000 in the four recent articles read from the journal. Sections +2, 5 and 7 are the ones to grow. + +## Joins added to engage the 2026 HMI reviews (not part of the paper) + +Written 2026-10-14 after a full read of Grahn et al. and Mourtzis and Tsakalos, both in *Human-Machine Interaction* 1(1). Cards: `literature/library/grahn2026integration.md`, `literature/library/mourtzis2026emerging.md`. + +- Abstract and §1 now open on the two reviews, then the argument against Nadella (2026), then the oracy-to-literacy change. +- New §2.1 is the engagement. Old §2.1–2.3 are now §2.2–2.4. +- The Nadella reading moved from §4.3 to §2.5, so it follows the reviews and precedes the history. A pointer remains where §4.3 stood. +- §3 now opens as the change the reviews and the post do not contain. + +The sourced limits inside those joins are the reviews' own: Level 4b and "work allocation on labor platforms"; the external-display sentence; civic AI literacy as one transferred cell; the Section 5.4 sentence about information for decisions; 87 publications; sensor readiness levels 3 to 4. The claim that these remedies are still literacy, and that a rider and a driver sit in Level 4b, is this paper's application. + +## Joins added for the three-move argument (not part of the paper) + +These sentences were written on 14 October so each section names the move it serves. Sourced findings under them are the author's, as corrected by the claim check. + +- Introduction: the smoking-gun definition; the one limit paragraph; the three-move roadmap. The date-stack and the study-by-study disclosure list were moved out of the opening. +- §2: the openings of 2.1, 2.2, and 2.3, which state that HCI still hands people a text, and that the reader in those studies sits in a dyad. +- §3: the opening of 3.1 ("this section is the first move"); the single limit paragraph; the sentence that what the reader no longer had to do is the content of the gun; the close that points at the second move. +- §4.1–4.2: the openings that regulation's remedy is the same text, and that the empirical verdict is failure plus a reader. +- §4.3: the reading of Nadella (2026) by the project's paragraph numbers, and the diagram that puts the chain inside the firm and a notice outside it. +- §5.1: algorithmacy as the sensibility the space would call on, deciding and taking in what returns. +- §6: the test taken from the history, and the two paragraphs that say neither convention has passed. +- §7: the opening condition that a space the operator can withdraw is not a space. +- §8: the opening that the protocol is how we would know. +- §9.1: the three-move synthesis. The closing question is the one already drafted. + +Sentences removed from the body because they only recorded the audit, and the fact they were guarding is either still in the paragraph or in the claim-check note above: "the book was not opened"; "the abstract is all this paper has seen"; "that clause is dropped"; "the citation is removed"; "version 2 understated" and its kin; "the pages were not read in full"; "Scribner and Cole (1981) was not opened"; "Routinely and the batching wait are not established"; "Every day and consumers are not established"; the transparency-paradox citation does not say what it was asked to say (Parasuraman & Manzey, 2010, is reported for what it does say); Zuboff is not a source and the behavioral-surplus claim is dropped; movable type is dropped. Saenger (1997), Parkes (1992), Norman (1988), and Rayner et al. (1998) beyond the abstract remain unverified, as the note below says. + +## Applied from the 14 October claim check (not part of the paper) + +Version 3 changes sentences where `literature/audit/claims/SUMMARY.md` marked a claim contradicted, miscited, overstated, missing from its source, or narrower than the sentence. Supported claims were left. Claims the check could not verify, chiefly Saenger (1997), Parkes (1992), Norman (1988), and Rayner et al. (1998) beyond the abstract, are hedged in the text rather than rewritten as if the page had been opened. New citations are sources the check read in full on 14 October. `DRAFT_v2.md` is unchanged, so the ledgers' line numbers still match it. + +The largest reversals: word separation is not the oldest case and not a blank between every word; the Schemmer null is against AI advice without an explanation, not against unaided judgment, and Senoner et al. (2024) cut against a blanket null; the Seira message was a separate mailing that also gave advice; Nadella's seven principles are relabeled to the post; platforms do add noise on purpose (Rogers et al., 2021; Li et al., 2010); Gonczarowski et al.'s gain is the property-only menu, and understanding moved play by 36 points; the posted-price menu is not an ascending clock; the FCA fine was for control failures, not a finding of completed benchmark manipulation reported by the FSB. + +## Still unverified (Claude's notes; not part of the paper) + +- **Records confirmed at Crossref, texts not read.** The sentences that cite these rest on the earlier + draft's description of them: Herlocker et al. (2000); Knijnenburg et al. (2012); Wischnewski et al. + (2023; Crossref shows the preprint, and the CHI venue is from the draft); Ehsan et al. (2021); Lee & + See (2004); Körber (2018); Bansal et al. (2021); Chen et al. (2023); Metzger & Parasuraman (2001); + Casner et al. (2014); Mozannar et al. (2024); Simkute et al. (2024); Lee et al. (2025); Horvath et al. + (2022); Pucci et al. (2027); Sheridan & Verplank (1978); Wolfe (2021); Hohenstein et al. (2023, carded + for a different point); Eisenstein (1969). +- **Terveen & McDonald (2005)** is cited as the form the agenda follows and has not been read. +- **Norman (1988)** is cited for the gulfs of execution and evaluation and has not been opened. +- **Rayner et al. (1998)** was read as an abstract only. +- **Saenger (1997) and Parkes (1992).** Saenger's quoted wording is confirmed; his pages need a print + check. Parkes is known from snippets. +- **Rahman (2019)** is the dissertation; the published article (2021) was not opened. +- **Gonczarowski et al. (2024), Larson et al. (2016)** were read as working papers. +- **Barrett et al. (2026)** is a preprint; the bibliography holds no DOI or archive number for it. +- **The interface mock-ups and their numbers** (rates, tiers, "P(Match) = 88%") are illustrations. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v4.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v4.md new file mode 100644 index 00000000..7b185db1 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_DRAFT_v4.md @@ -0,0 +1,553 @@ +# The Space Between Decisions: Word Separation as a Guide to Designing for Algorithmacy + +> **Version 4, 2026-10-15. Claude/Fable-drafted, not the author's text, except the abstract, which is the author's, verbatim.** The author's read-aloud is the last gate. + +## Abstract + +Human–machine interaction has moved from signals and screens to extended reality (Mourtzis & Tsakalos, 2026) and human–machine integration (Grahn, Wang, Wang, & Wang, 2026). Yet, whether on a factory headset or in an enterprise audit log (Nadella, 2026), the system still sends a person something to read so they can check the machine. That approach assumes a dyad: an operator paired with a tool (Parasuraman, Sheridan, & Wickens, 2000). On digital platforms, though, counterparts—a driver and a rider, or a buyer and a seller—coordinate through an algorithmic mediator that binds and optimizes, forming an irreducible algorithmic triad. Mandated disclosure (Ben-Shahar & Schneider, 2011) does not spare them from having to game the mediator to coordinate with the other person. What these settings beg for instead is the kind of cognitive offloading (Risko & Gilbert, 2016) seventh-century Irish scribes achieved when they broke up scriptio continua by adding spaces between words (Tunbridge, 1992; Saenger, 1997). Unspaced text forced a reader to sound out the line just to hear where each word ended, scaffolding literacy with oracy; word separation freed literacy by moving boundary detection from the voice to the page, exploiting a representational effect (Zhang & Norman, 1994) that enabled readers to internalize text. I argue we should look for a similar "smoking gun" design change that would free algorithmacy from literacy—enabling people to decide within irreducible triads through epistemic action (Kirsh & Maglio, 1994), strategic offloading (Li, 2017), and ecological feedback immediacy (Shneiderman, 1983; Kahneman & Klein, 2009). + +## 1. Introduction + +Human–machine interaction keeps changing what it hands a person and keeps the person's job the same. Mourtzis and Tsakalos (2026) tell the industrial version of the story: levers, then buttons, then graphical interfaces that moved the human "from operators to overseers," then dashboards, augmented reality and decision support "designed to assist humans in understanding machine behavior," and now extended reality, whose interface "provides the human with the necessary information to perform tasks and make decisions in real-time" (pp. 2–3, 10). Grahn et al. (2026) tell the same story at a different scale, four levels of coupling ordered by how much it costs to leave, and concede that the classical models "were developed for supervisory control settings in which a human operator, clearly separated from the machine, allocates functions and monitors discrete levels of automation through external displays" (p. 2). The enterprise version is Nadella's (2026) proposal to treat frontier models as insider risks: every meaningful model action is to leave "tamper-proof human readable evidence" (P15), and "an authorized person" is to hold a brake (P19). A headset, a dashboard and an audit log differ in everything except the task they set. Each sends a person something to read so the person can check the machine. + +That task comes from a particular picture of who is in the room. Parasuraman et al. (2000) allocate functions between "the human operator" and "automation," which they define as a device that does something "previously, or conceivably could be, carried out ... by a human operator" (p. 287). The automation is an aid to the operator's task, and the operator reads it to confirm that the aid is working. The pair is the wrong picture for a digital platform. A rider and a driver, a buyer and a seller, a host and a guest do not each operate a tool. They coordinate with each other through a third party that matches them, prices them, ranks them, and does so to serve an objective of its own. Stark and Pais (2020) call the relation "typically three-sided" (p. 53); Vallas and Schor (2020) call the platform Simmel's *tertius gaudens*, the third who rejoices. The abstract calls it an irreducible algorithmic triad. The design question it raises is not how a person checks a machine but how two people decide through one. + +The standing answer is disclosure: publish the parameters, explain the decision, show the log. Ben-Shahar and Schneider (2011) showed that mandated disclosure fails its recipients across consumer law, and the platform statutes that followed copy its form while limiting its content. What the platform studies show is that a published rule changes how people strategize and does not remove the strategizing, because the task the rule leaves in place is reaching the other person through the mediator. Disclosure hands the person a text. It is, in the terms this paper adopts, a literacy-based intervention, and the capacity the triad calls for is not literacy. + +The paper's move is to look for a design change of a different kind, and to take its model from the history of reading. Seventh-century Irish scribes, copying Latin that was not their language, put space between its words (Tunbridge, 1992; Saenger, 1997). The change disclosed nothing about the text; it parsed it, marking boundaries with space and with points and altered letter forms (Tunbridge, 1992, pp. 165–175). It trained no reader to parse, though it presupposed a trained scribe and a convention reader and scribe shared. It moved a piece of work, finding where one word ends and the next begins, from the reader to the page: offloading "built-in" to the task environment (Risko & Gilbert, 2016, Box 3), and a representational effect, the same formal structure in a different external form with "dramatically different cognitive behaviors" (Zhang & Norman, 1994, p. 88). I argue that designing for algorithmacy, the capacity to decide well inside an algorithmic triad, should look for a change of that kind, and I name three places to look: epistemic action (Kirsh & Maglio, 1994), the removal of strategic work by the mediator's own rule (Li, 2017), and feedback that is both valid and immediate (Shneiderman, 1983; Kahneman & Klein, 2009). + +This is a framework paper in the sense this journal names: it reviews literatures that HCI has mostly left to others, builds a set of orienting claims, and ends in a research agenda. It reports no study. It makes three contributions. It states, with the field's own sources, why the read-to-check role survives every change of medium and cannot be the role of a counterpart in a triad. It gives the scribal precedent a human-factors reading that survives the historians' dispute, by separating what word separation did to the page from what it has been claimed to do to literacy. And it specifies what a "smoking gun" design change for the triad would have to do and what evidence would count as having found one. The sections follow the abstract's order, from the read-to-check role (§2) and the dyad (§3) through the triad (§4), disclosure (§5), the precedent (§6) and the three means (§7) to candidate changes and their test (§8), limits (§9) and a conclusion (§10). + +## 2. Reading to Check + +### 2.1 From signals and screens to headsets and integration + +The 2026 reviews agree about where human–machine interaction has gone and about what the human still does there. Mourtzis and Tsakalos (2026) reviewed 87 engineering publications from 2021 to 2026 and found most of the sensor work still at technology readiness levels 3 to 4, "with only a single case among forty showing TRL 5" (p. 8); in their collaborative model the robot senses and acts and the XR interface "provide[s] instructions and support[s] decision-making for humans" (p. 11). Grahn et al. (2026) rank four levels of integration by exit cost and place "work allocation on labor platforms" at Level 4b, where a person "cannot decline the system without forgoing an economic or social role" (Table 2, p. 8). Beside override and opt-out rights (p. 13), their remedy at most levels is a reader's remedy: outputs that are "transparent and interpretable to human operators" (Table 1, p. 5), explainable AI as "essential for maintaining the principle of human agency" (p. 16), and "civic AI literacy" in the education row (p. 13). +The factory headset is the clearest case, and the case where the abstract's claim needs its narrowest form. De Moura Costa et al. (2022) reviewed 32 papers on augmented reality for human–robot collaboration and found that the headset's job is to put information into the operator's field of view; safety features were the most common application, most were "for informational purposes only," 13 of 20 safety systems were passive, and most were "not certified" (pp. 27–29). Quandt and Freitag (2021) read 33 papers on industrial AR acceptance and found "no widespread adoption" (p. 14). What the headset sends is mostly instructions, warnings and status; checking the robot, through a trajectory preview or a digital twin, is one use among several. The abstract's "something to read so they can check the machine" therefore describes the audit log better than the headset, where the field's own purpose is to tell the worker what to do next. In both cases the worker reads. + +### 2.2 The audit log and the overseer + +The enterprise audit log is where the read-to-check role is written into law and into the leading statement of human-centered AI. Shneiderman (2020) prescribes "high levels of human control AND high levels of automation" (p. 1), rests reliability on "audit trails and analysis tools to review failures & near misses" (pp. 3–4), and rejects the teammate metaphor outright (p. 5). The European Union's Article 14 asks the person assigned to oversight to understand, monitor, "correctly interpret" the system's output, override it and stop it (Regulation (EU) 2024/1689). Enqvist (2023) reads that duty as one performed "through human intermediaries as interpreters of the data produced by the operating algorithms," observes that without transparency "the human overseers would have nothing substantive to review" (pp. 517–518), and notes that the Act leaves "a great deal of room for the providers to determine the detail on what human overseers will be presented with" (p. 525). Koulu (2020) explains why the role is adopted at all: "through digital traces and metadata we can easily ascertain that the oversight requirement has been fulfilled," at the risk of "turning human decision-makers into rubberstamps" (False promise; Conclusion). + +Nadella's (2026) post belongs here, read as an enterprise-security proposal rather than as a statement about people who meet a platform's mediator. "We" in the post is the deploying organization throughout; the words "user," "customer" and "employee" do not occur. The remedy is containment: "separate the supply of intelligence from the authority over it" (P5), keep every control outside the model (P11), and surround it with seven principles of observability, of which two give a person anything to do. Observe everything, with "tamper-proof human readable evidence" for every meaningful action (P15), gives a person a record to read; containment, with "an authorized person" able to stop a model "like an emergency brake" (P19), gives a person a brake. Model diversity (P14) and independent controls (P17) disclose nothing, and chain-of-thought transparency is called "not sufficient or dependable" (P9), so the post is not a disclosure proposal in the narrow sense. What it shares with the headset and the dashboard is the human seat: a reader of records after the fact and a holder of a stop. The people the model acts upon appear once, as recipients of incident disclosure (P20). + +### 2.3 What the dyadic literature already knows about checking by reading + +Human factors has measured this seat for forty years and has not liked what it found. Bainbridge (1983) gave the problem its name: the operator "may be expected to monitor that the automatic system is operating correctly," yet attention to a source "on which very little happens" lasts about half an hour, and "the human monitor has been given an impossible task" (pp. 775–776). Parasuraman and Manzey (2010) report detection of automation failures at 33% under constant reliability against 82% under variable reliability in multitask conditions, and 75% of pilots following a wrong automated checklist against 25% with paper. Dismukes and Berman (2010), observing 60 airline flights, logged 899 deviations and found crews trapped 35.4% of deviations in primary procedures, 14.4% in checklists and 5.6% in monitoring; pilots "do not receive feedback on the effectiveness of their monitoring" (pp. 23–24). +The same literature already knows the one lever that helps, and it is not more to read. Onnasch et al. (2014) meta-analyzed 18 experiments and found that return-to-manual performance fell with the degree of automation (tau −.34) and collapsed once action selection was automated (tau −.90 across six studies); they attribute the protection of keeping the choice to the generation effect and recommend that operators "should be kept involved at least to some extent in decision and action selection" (p. 486). O'Neill et al. (2022) reviewed 76 human–autonomy teaming studies and report that transparency, more of the agent's reasoning to read, is "mixed" and "may create a level of complacency that leads to less vigilance." Lai et al. (2023) surveyed over 100 human–AI decision-making studies and found the dominant paradigm "to have models providing predictions, then users can choose to follow or ignore" (§4.2.4, p. 20), with the agency-preserving alternatives in a minority. The protective factor the dyad has found is that the person keep deciding. That finding is the abstract's "epistemic action" in the dyad's own terms, and §7 picks it up. Before that, the paper has to say what the dyad is and where it stops. + +## 3. The Dyadic Assumption + +### 3.1 One operator, one aid + +The supervisory-control lineage models two parties from its first page. Sheridan and Verplank's (1978) scale and the ten levels that Parasuraman et al. (2000) adopt describe the computer's autonomy "over human action" (p. 287): at level 4 "the computer suggests one decision alternative, but the human retains authority"; at level 6 "the system gives the human only a limited time for a veto." Wiener and Curry's (1980) Figure 1 has two dimensions, the operator monitoring the computer and the computer monitoring the operator (p. 13). Every one of Onnasch et al.'s (2014) 18 experiments is an operator and an aid. Even where the model's own example has several people in it, the unit of analysis stays the pair: Parasuraman et al.'s air-traffic case has a controller issue "a verbal clearance to a pilot" that "may be a computer choice" uplinked to the aircraft (p. 293), and the model analyzes each operator's relation to the automation separately. The automation has no objective. It helps the operator do the operator's task, and the operator's reading of it is a check that the help is sound. + +### 3.2 Where the lineage admits teams, and why the machine stays a tool + +Three branches of the field admit more than two parties, and in none of them does the machine acquire an interest against the people it faces. Cognitive systems engineering takes the joint system as the unit: Rayo et al. (2023) write that cognition "resides not only in the head of an individual, but in components of the system and their interactions" (p. 184), and Woods calls the levels-of-automation proposals "obsolete, narrow, and/or counterproductive" (p. 185). Klein et al. (2004) ask automation to be a "team player" that enters a Basic Compact, "a commitment to some degree of goal alignment" (p. 91); the machine has goals, and the design aim is to align them with the team's. O'Neill et al. (2022) require of a human–autonomy team "a common goal as a collective"; of their 76 studies 45 pair one human with one agent, and the authors describe the evidence base as "dyads." Oversight law adds a chain, provider to deployer to overseer to system, with affected persons as rights-holders who have no interface (Enqvist, 2023, pp. 518–530); even there, the act of oversight is one reader facing one system's output, and the subject coordinates with no one through the system. The lineage admits teams and chains. It treats the machine as a tool, a component or an aligned teammate. The interested mediator between two counterparts is absent from it. + +### 3.3 Shneiderman's choice + +One qualification to the abstract's wording belongs here. For the leading statement of human-centered AI, the dyad is not an assumption but a position. Shneiderman (2020) keeps the pair on purpose: computers "are not teammates" (p. 5), the person should operate a controllable device, and organizations should read the records. The argument against the pair therefore has to be made against a choice, and the ground for making it is empirical: there are settings in which the machine is not the person's device at all. A rider cannot operate Uber's dispatch; a seller cannot operate Amazon's Buy Box. The person is on one side of a system whose other side is another person and whose middle has a purpose of its own. + +## 4. The Irreducible Algorithmic Triad + +### 4.1 Three parties, one of them optimizing + +Three literatures model platform coordination as three parties, with the platform as a party of its own. Economic sociology states it directly. Stark and Pais (2020) write that the relationship "is typically three-sided: on one side, the platform owner; on the second side, the provider ...; and on the third side, the user-customer" (p. 53), and that what the algorithmic-management literature misses "is that this triangular geometry ... and the power asymmetries along its various axes are key" (p. 54). Their mechanism is co-optation: the owner enrolls the other two sides' ratings "without managerial authority having been delegated to them" (p. 47), so that feedback loops are "twisted by the triadic relations of owner-provider-user" (p. 54). Vallas and Schor (2020) supply the ancestor: platforms "govern the service triangle that links employers, workers, and customers" and play "the role of tertius gaudens—'the third who rejoices'—as in Simmel's classic theory of the triad" (16.10). Simmel (1950) is the source of the concept. The third member makes possible mediation and arbitration (pp. 145–154), exploitation as *tertius gaudens* (pp. 154–162) and *divide et impera* (pp. 162–169), forms that "are impossible if there are only two elements" (p. 145). The abstract's "binds and optimizes" fuses Simmel's arbitrator, who "ends up by taking sides" (p. 148), with his self-interested third (p. 154): the platform both settles the match and gains from how it settles it. + +The economics of two-sided markets makes the platform a profit maximizer by construction. Rochet and Tirole (2006) define a market as two-sided when "the volume of transactions between end-users depends on the structure and not only on the overall level of the fees charged by the platform" (pp. 2–3), and derive the platform's price level and structure from its own profit (pp. 17, 24). Hagiu and Wright (2015) define a multi-sided platform by "direct interactions" in which the sides "retain control over the key terms" plus "affiliation," and set the platform's fees to maximize its profit (pp. 5, 12–13). The management literature dissents. Kellogg et al. (2020) read algorithms at work through labor-process theory, with the algorithm as the employer's instrument and customers as raters, so their frame is a dyad with an instrument. Stark and Pais themselves bound the triad: where the platform employs directly, as in "Amazon's enormous warehouses," there is "no triangular relationship" (p. 54, n. 13). The triad is a property of the marketplace form, not of algorithms as such, and the paper's claim sides with the sociologists against the labor-process frame for marketplaces only. + +### 4.2 Binds and optimizes + +The documented cases show both verbs, and show that the strength of the binding varies. On the rider side of Uber, Cohen et al. (2016) describe a take-it-or-leave-it surge multiplier, rounded to 1.2x, 1.3x and upward with no 1.1x and never below 1.0x, whose stated goal is to clear local supply and demand and allocate rides to the riders who value them most (p. 7); business rules override the algorithm in 11.5% of sessions, and rider purchase rates fall from 62% at base price to 39% at high surge. Castillo et al. (2025) model the dispatch: an idle driver is sent on every request, the price is set per geofence every two minutes (p. 18), and a static price would sit at the 92nd percentile of dynamic prices (p. 32). Garg and Nazerzadeh (2022) show that the platform sets the rider's price and the driver's payment separately, "decoupled pricing" (p. 8), that multiplicative surge makes it optimal for a driver to reject short surge trips and long non-surge trips, and that additive surge removes much of the gain, worth about $15 per hour for short trips at high surge in their RideAustin estimate (p. 22). +On Amazon, Chen et al. (2016) found that the Buy Box, through which 82% of sales pass, chooses the default seller by undocumented weights, and that 543 sellers repricing by algorithm won it more often at every rank but the top (pp. 2, 7, 10). Farronato et al. (2023) found Amazon's own brands ranked higher in search. On Airbnb, Haldar et al. (2019) describe a ranking model trained to put booked listings high and shipped on A/B conversion gains, whose engineers could not recover stable feature importances (pp. 1, 7–8). On remote freelance platforms, Wood et al. (2019) found that clients rate and "the platforms' algorithmic ranking of workers within search results" allocates, with control sitting "at the end of the labour process rather than during it" (Findings). + +Two boundaries follow. First, binding varies by platform: Uber sets the match and both prices; Amazon picks a default seller; Airbnb decides visibility while the host still sets the price; freelance platforms rank and filter. By Hagiu and Wright's (2015) own definition, a mediator that takes control of the key terms has moved from a pure platform toward a reseller or an integrated firm (pp. 5, 21). The triad this paper means is that hybrid, not the textbook two-sided market. Second, the binding is not total for everyone. Vallas and Schor (2020) note that supplemental earners "can refuse low-paying tasks" (16.8), and Cruz and Gameiro (2023) describe workers who "agree to allocate tasks directly without the need for platform intermediation" (p. 3). Where the match survives the mediator, the triad was reducible for that pair. + +### 4.3 What "irreducible" can mean + +The word carries the argument and needs a definition the sources can bear. Three senses came out of the reading, and the paper adopts a fourth as its gauge. Simmel's sense is formal: the triad's forms belong to the configuration and not to any pair, and adding a third "completely changes" a dyad while a fourth does not change the triad (1950, pp. 138, 141, 145). Rochet and Tirole's sense is price-structure non-neutrality: the market is two-sided exactly when the two end-users cannot bargain the mediator's choice away (2006, pp. 6, 8, 11). Castillo et al.'s sense is system feedback: trip supply depends on both sides through pickup time, a rider's request removes an idle driver and lengthens every other rider's wait, and the resulting failure mode, the "wild goose chase," is one no party in either dyad can see or cause (2025, pp. 5–8, 13–14). Against all three, Hagiu and Wright's (2015) model decomposes into N professional–platform relations plus a spillover term (pp. 9–14), which is how an economist resists the word. + +The fourth sense is the one the author's program uses: integrated information, Φ, as the gauge of whether a given triad's causal structure can be cut into parts without loss (Hunt, 2026). Φ is the right kind of instrument for the question, because it is defined over the whole system's transition structure rather than over any pair's payoffs, and it returns a quantity rather than a verdict by example. What the record shows is that it has not yet been applied to a mediated triad. Engel and Malone (2018) computed observational estimates of Φ on who-is-active data from four-person peer groups and found a correlation with collective intelligence of r = .370 (p. 11); their groups had no mediator with an objective of its own, their lag was tuned to maximize Φ, and they use the number as a predictor of performance rather than as a test of reducibility. No source outside the author's program proposes Φ for a platform triad, and none computes it. The paper therefore says "irreducible" in the senses the platform cases support, system feedback and non-neutrality, and points to Φ as the instrument that would make the word measurable on a modeled triad, with its measure, coding, lag and partition fixed before computation. §8 returns to it. + +### 4.4 Why HCI has not modeled it + +The nearest HCI field defines its mediator so that the triad falls outside it. Hancock et al. (2020) define AI-mediated communication as communication "in which an intelligent agent operates on behalf of a communicator by modifying, augmenting, or generating messages" (p. 1); the agent's autonomy "is similar to the principal–agent relationship," and "content ranking, recommendation and classification algorithms ... we therefore exclude" (p. 4). The experiments in that tradition randomize a delegate's availability to one or both partners (Hohenstein et al., 2023), and what moves is the partner's suspicion, not any property of a third party with interests. AI-MC is a dyad with a delegate. Terveen and McDonald (2005), whose framework-and-agenda form this paper borrows, did describe a system that stands between two people, matching them, introducing them and taking feedback on the match; their matching system has no objective beyond good matches, and their own caution is that "it has not been established that social network diagrams help users select others with whom to interact" (p. 418). Özçetin and Wiltse (2025), in this journal, come closest to the problem and propose the remedy the paper argues against: "revealing design," which would make platform relations "legible and actionable at the level of interface" (§6) and which the authors expect "would need to be mandated at the level of legislation" (§6). They agree that terms of service fail, judging the usability fixes "notably limited" (§2.1); the disagreement is whether the fix is more for the person to see and understand. + +## 5. Why Disclosure Does Not Spare the Counterparts + +### 5.1 Mandated disclosure and its platform descendants + +Ben-Shahar and Schneider (2011) defined their target narrowly, a technique "in which a lawmaker requires a discloser to give a disclosee a standard disclosure" (p. 651), and found it failing in three steps: disclosees "do not read disclosed information, do not understand it when they read it, and do not use it even if they understand it" (p. 665). Two features of their evidence limit what the abstract can take from them. Their cases are consumer and patient disclosures to individuals, before platforms and outside any triad, and their own remedies, symbols instead of sentences, advice and ratings instead of data, defaults and prohibition (pp. 743–749), all sit on the discloser's side of the relation. Loewenstein et al. (2014) add that disclosure "may have little effect on recipients but large effects on providers" (p. 398). What the article establishes is that handing a person a text is a weak way to change what the person does. It does not by itself establish anything about gaming a mediator. + +Platform regulation copied the form and limited the content. Regulation (EU) 2019/1150 requires "the main parameters determining ranking and the reasons for the relative importance" (Art. 5(1)) and says providers are "not required to disclose algorithms" (Art. 5(6)); the Digital Services Act (Regulation (EU) 2022/2065, Art. 27) and the Platform Work Directive (Directive (EU) 2024/2831, Art. 9) use the same formula. The Commission's guidelines on ranking transparency say what the disclosure is for: it "aims to improve predictability and help users improve the presentation of their goods and services" (European Commission, 2020, pts 5, 15), weights and algorithms need not be disclosed (pt 46), "an excess of information can mean that, in effect, no meaningful information is provided" (pt 25), and the point is to let businesses decide "how to invest in 'ranking strategies'" (pt 94). The regulator's own statement of purpose is the strongest support the abstract has: the disclosure exists so that the disclosee can strategize better. + +### 5.2 A disclosed rule redirects strategizing + +The founding models of strategic classification assume disclosure and ask the designer to survive it. Hardt et al. (2016) note that "concerns of gaming and manipulation are often used as a reason for keeping classification mechanisms secret," and reply that "secrecy is not a robust solution" because outsiders learn the rule "from classification outcomes" (p. 2). Kleinberg and Raghavan (2020) add that secrecy "can create inequalities between insiders ... and outsiders" (p. 2). Perdomo et al. (2020) show that the rule is itself a response to the population it scores (§5), so the loop runs whether or not the rule is disclosed, and Ghalme et al. (2021) show that hiding the rule does not stop gaming; people game a guessed rule at a "price of opacity." + +Where disclosure was manipulated, strategizing rose or moved. Björkegren et al. (2020) showed participants the scoring rule and found them moving "closer to the direction incentivized by the algorithm," at a cost of about 23% of predictive performance for a standard rule (pp. 5, 29). Tasnim et al. (2025) found truthful ranking fall from 80.14% with no probabilities shown to 57.45% and 40.43% when probabilities were shown, and aggressive manipulation rise from 5.67% to 39.01% when it was shown to succeed 90% of the time (Table 2, p. 2474). Eslami et al. (2019) told Yelp users that their reviews had been filtered; six of fifteen said they would "write for the algorithm," seven that they would use Yelp less or leave (p. 11). Danz et al. (2022), in a belief-elicitation rule that was truthful by construction, found that showing the exact payoff table doubled false reports, 41.5% against 20.6% with no quantitative information (Table 2, pp. 12–17). Hakimov and Kübler (2021) report that 85% truthful play under top trading cycles fell to 31% once subjects were told others might misreport (pp. 444–449). + +The models disagree about what the extra strategizing does, and the disagreement narrows the claim. Wang et al. (2023) find "a broad set of conditions under which making the algorithm transparent is beneficial for the firm," because gaming a disclosed cheap feature pushes people onto the costly causal one, while "the agents are not always better off" (p. 4). Cofone and Strandburg (2019) argue that gaming requires weak proxies and cheap modification and that "blanket claims that disclosure will lead to gaming are over-blown" (p. 623; the prerequisites are on pp. 662–663). The abstract's claim therefore holds for ranking rules built on cheap, mutable signals, reviews, engagement, response time, acceptance rate, which is the platform case, and not for every disclosed rule. + +### 5.3 Gaming as the cost of reaching the other person + +Platforms give the anti-gaming reason for opacity in their own words, and partial disclosure then produces gaming by guesswork. Upwork's help page said that revealing the calculation "would make it easier for some users to artificially boost their scores" (Rahman, 2019, p. 89); Airbnb's researchers wrote that "there needs to be some degree of uncertainty ... or otherwise, some hosts would inevitably game the system" (Jhaver et al., 2018, Discussion); the Commission's guidelines license it, since fraud mechanisms "cannot be revealed in detail" (European Commission, 2020, pt 73). The result is what Rahman (2019) calls "superstitious reactivity": a score that dropped "from 100% to 82%" after closing an idle contract (p. 93), updates at unknown times (p. 94), and freelancers "trying various methods to take apart and understand the algorithm" (p. 117). Bucher et al. (2020) record Upwork workers' "20/1 rule" and avoidance of "gray area words"; Bishop (2019) records YouTube creators' "algorithmic gossip" behind $75–$500 guild fees; Jhaver et al. (2018) record a host who changed his title daily because "lots of blogs" said to (P-15). + +Every platform study reads this work as aimed at the other person. Curchod et al. (2020) found 41 of 77 high-rated eBay sellers describing "working around the algorithm" to prevent buyer negatives, in a relation the authors call triadic. Bucher et al. (2020) describe freelancers forgoing proposals and underpricing work to keep a score that lets them reach clients, "extra work" in a "triadic relationship between algorithm, client and worker." Jhaver et al. (2018) call the host's position a "double negotiation," an appeal to guests and to algorithms at once that conflicts when declining short bookings costs acceptance rate. Luca and Zervas (2016) found about 16% of 316,415 Boston Yelp reviews filtered as suspicious, with businesses caught soliciting fake reviews averaging 2.6 stars and 18 reviews against 3.5 and 86 (pp. 6, 19, accepted manuscript). Lee et al.'s (2015) driver toggled the app at traffic lights to avoid distant requests (p. 1607). Bambauer and Zarsky (2018) name the cost: "third-order inefficiencies—the costs of the system itself," including "the persistent worry and vigilance" (p. 32). + +Three qualifications keep this claim honest. None of these studies isolates a coordination cost from a competition or reputation cost, and none measures what the other party loses, so "to coordinate with the other person" is supported by description and not by a test. "Gaming" is the platform's moral label: Petre et al. (2019) find the line between manipulation and legitimate strategy "nebulous and largely reflective of their material interests," and record Facebook advising calls-to-action and then demoting them as "engagement bait" in the same year. And the counterparts can ally against the mediator: Rahman's freelancers formed "online alliances with clients to game the performance evaluation system" (Vallas & Schor, 2020, 16.7), which means coordinating with the other person can itself be the way the mediator is gamed. + +### 5.4 Where disclosure helped + +The counter-evidence is real and is of a particular shape. Kizilcec (2016) found that a procedural explanation of an adjusted peer grade removed the trust gap between learners whose expectations were violated and met (3.72 against 3.70), where no explanation (3.02 against 4.18) and full outcome data (2.77 against 3.89) did not; the outcome is self-reported trust in 103 learners (p. 2392). Masuda et al. (2022) raised truth-telling by a net 23.8 points by announcing strategy-proofness in a Vickrey auction, and Gonczarowski et al. (2024) found that a short statement of a mechanism's property helped where teaching the algorithm did not. Matsuo et al. (2026) cut simulator collisions to 1 of 15 with a warning cue and its reason, against 7 of 15 with no cue, in a person-and-machine setting. Tasnim et al. (2025) read their own data as a case for disclosing risk, since parents shifted to aggressive strategies only at a 90% stated success chance while simulations put real gains at 1–2% of students (p. 2477). + +Two things are true of this list together. Where disclosure helped, it was a short statement of a property, a reason attached to a cue, or advice computed for the person, not a published rule or a log; and none of it is a triad with a behavioral outcome. The author's ruling that disclosure is a literacy-based intervention stands in this form: handing a counterpart a description of the mediator has not been shown to change how well the counterpart reaches the other person, and in the measured cases it changed how the counterpart strategized. The two documented changes that removed a strategic task altogether were not disclosures. Uber's move from multiplicative to additive surge removed the incentive to cherry-pick, and Garg and Nazerzadeh (2022) record Uber disclosing destinations only afterwards: "Incentive compatible pricing is an important stepping stone to showing this information" (p. 9, n. 7). Compelled pooling in US insurance and credit-card law, Bambauer and Zarsky (2018) write, "renders gaming efforts ... unnecessary" (p. 47). Both changed the rule rather than the reader. The question is what kind of change that is, and the history of reading offers a model. + +## 6. The Scribal Precedent + +### 6.1 What the Insular scribes did + +The earliest surviving Irish manuscripts "already contain some kind of separation" (Tunbridge, 1992, p. 193). Tunbridge's table runs from erratic wax tablets and syntactic-unit spacing through word-based separation in the St Gall Isidore of the second quarter of the seventh century to the Bangor Antiphonary (c. 680–691) and the Schaffhausen Adomnán (before 713) (pp. 214–215). Her explanation is linguistic: "Irish scribes when copying Latin began to abandon scriptio continua in favour of a layout which clarified the syntax and morphology of the language and hence met the needs of the non-native Latin speaker" (Long Abstract, p. 4). Saenger (1997) gives the chronology the field uses: Insular origin in the seventh and eighth centuries, Continental adoption from the mid-tenth century, standard and ubiquitous by the mid-eleventh (chs. 4, 6, 9; pp. 83, 100, 176–177). Three facts narrow the picture. The same scribes wrote Old Irish in stress-groups rather than words: Thurneysen's rule, quoted by Doyle and McCrae (2025), is that "words which are grouped round a single chief stress and have a close syntactic connexion with each other are written as one" (p. 24). Separation came as a bundle with points, hyphens, tags on final letters and variant letter forms (Tunbridge, 1992, pp. 165–175), and Saenger's own sentence names four devices, "word separation, the diastole, the hyphen, and eventually the trait d'union" (p. 85). And Latin had used interpuncts before the second century AD and dropped them (Heilmann, 2020, pp. 182–183), so separation was a return rather than an invention. The scribes' move was a learner's analysis of a foreign language, carried out on the page, and it was the central move in a complex of conventions. + +### 6.2 What the voice did and did not do + +The abstract's account of the reader before spacing is Saenger's, and it is contested at every load-bearing point. Saenger's (1997) thesis correlates oral reading with the "threshold in the duration of cognitive activity needed to achieve lexical access" in a script (p. 2), and holds that the Insular devices "provided shortcuts for achieving the reading skills that an elite among the ancients had mastered only through a prolonged and arduous grammatical apprenticeship" (p. 85). O'Donnell (2000), a sympathetic reviewer, called the stretch from the physiology of reading to the theory of reading practice "the most venturesome and vulnerable part of the book" (¶4). The classicists then tested it. Knox (1968) had already shown that ancient books were "normally read aloud" while silent reading of letters was "taken completely for granted" in fifth- and fourth-century Athens (pp. 434–435), on unspaced text. Heilmann (2020) calls the reading-aloud consensus a "quasi-dogma," says Saenger's arguments "presuppose precisely what must be proved" (pp. 178–179), reads the ancient complaints about laborious reading as being about learners (p. 181, n. 28), and concludes: "We should be careful not to project our difficulties with reading scriptio continua onto ancient readers" (p. 183). The founders of the oralization thesis cut the other way too. Balogh (1927) explained the difficulty of *scriptio continua* by whole-word recognition and long practice (pp. 227–228) and credited the printing press, not the scribes, with regular word boundaries (p. 236); Hendrickson (1929) found unspaced script "only strange, not necessarily difficult" (p. 186). + +Reading science settles what the voice does. Readers of spaced scripts lose a third to a half of their speed when the spaces are removed: 33% for English monolinguals and 45% for Thai–English bilinguals reading English (Winskel et al., 2009, p. 344), and 69% for Spanish readers of one-color unspaced text (Perea et al., 2015, Table 2). Readers of unspaced scripts do not pay that cost. Forty-six Chinese readers segmented silently by lexical knowledge with the same word-based eye-movement signature as German readers (Li et al., 2014); skilled Chinese adults read word-spaced and unspaced text in the same time, 5,239 against 5,221 ms (Bai et al., 2008); Thai children target word centers from Grade 6 at 88–93% comprehension, silently (Kasisopa et al., 2016, pp. 12–13). The voice is not the boundary detector even when it is used: in reading aloud the eye leads the voice by 16.2 letters at fixation onset, about 561 ms, so words are found before they are said (Laubrock & Kliegl, 2015). Speech itself does not mark word ends; "continuous speech lacks directly accessible cues to word boundaries" and pauses "are rare and occur only inconsistently" (Matzinger et al., 2021, Introduction). The word boundary is partly a literate idea: 60 American adults disagreed with written convention about where words end in proportion to their reading level (Barton & Hamilton, 1980, Tables 1–3). Nor did the voice leave reading when spaces arrived. Leinenger's (2014) review shows skilled silent readers of spaced text generating phonological codes within 80–125 ms of a word, before fixation through preview, and later as an inner voice that rises with difficulty. + +What survives of the abstract's sentence is a narrower claim, and it is the claim the design argument needs. For an unpracticed reader of a foreign language, voicing an unspaced line gave a second channel: a sounded string could be matched against words already known by ear (Busch, 2002, p. 20), as Scribner and Cole (1981) describe Vai readers resolving an unspaced syllabary by trial pronunciation. That is a learner's scaffold, not a physiological necessity, and "oracy," which none of the sources read here uses for voiced reading, is a new sense of the word for it; the sources say "reading aloud," "voiced reading" or "oralization." Word separation put the boundary on the page, where any salient mark will serve: alternating color recovers all but 4.7% of the cost of removing spaces (Perea et al., 2015, p. 12). It removed a segmentation task from the reader. It did not remove the inner voice from reading, and the sentence "enabled readers to internalize text" has no source behind it in either sense the phrase might carry. + +### 6.3 Offloading by design and the representational effect + +Read as a change to the task rather than to the reader, the scribes' move has an exact description in human factors. Risko and Gilbert (2016) define cognitive offloading as "the use of physical action to alter the information processing requirements of a task in order to reduce cognitive demand" (p. 2), which names the reader's own act, and then extend it: "In many cases the decision to offload is not made by the individual. Rather, offloading is 'built-in' to the task environment by design" (Box 3, p. 33). Word separation is offloading built in by design: another person's act on the page before the reader acts. Hutchins's (1995) speed bugs are the same move. Without bugs a pilot reads, remembers and compares speed values; with them "a memory and scale reading task is transformed into a judgment of spatial adjacency" (p. 283), and "individual pilot memory has not been enhanced; rather, the memory function has now become a property of a larger system" (p. 282). Larkin and Simon (1987) give the formal version: two representations can be informationally equivalent and computationally unequal, so that the pulley problem costs 138 element-considerations in sentences and essentially the seven inferences themselves in a diagram (Table 2, p. 77), and whoever draws the diagram "makes these new inferences which are then displayed explicitly" (p. 70). +Zhang and Norman (1994) supply the construct the abstract names. The representational effect holds when "different isomorphic representations of a common formal structure can cause dramatically different cognitive behaviors" (p. 88); in their Tower of Hanoi isomorphs, solution times differed by "as large as a ratio of one to ten" (p. 107), and rules carried by the physical setup were "error proof," since no subject broke one (p. 110). Two conditions in their experiments bear on a page. The external rule is enforced either by physics or by shared cultural knowledge, the coffee-cup rule being "an external rule grounded in cultural knowledge" (p. 106), and Zhang and Patel (2006) later name that case a "cognitive affordance ... provided by cultural conventions" (p. 339). A word space is a cultural affordance of this kind: it works because readers share the convention, and it can be withheld by whoever controls the page. And the effect is defined over isomorphs that carry identical information. Where an unspaced string has one lawful parse, spacing is isomorphic and the effect is representational; where it has two, the space adds information as well. +The precedent also carries the costs that re-representation brings. What is offloaded is practiced less and retained less: more offloading meant faster copying, 49.3 against 55.1 s, and worse memory for the offloaded layout, 5.91 against 7.07 bindings of 12 (Grinschgl et al., 2021, Table 1); statements believed saved were recalled at 0.19–0.22 against 0.29–0.31 for statements believed erased (Sparrow et al., 2011, pp. 777–778). People over-rely on the store, d > 0.8 (Gilbert et al., 2020), and "adapt by becoming dependent on those tools" (Kirsh, 2010, p. 450). Against that, the same group reports forgetting of about 45% unaided against about 5% with reminders and calls fears of long-term harm "overstated" (Gilbert et al., 2023). The reader of spaced text who can no longer read *scriptio continua* has paid the dependence cost; Winskel et al.'s 33–45% is its measure. + +### 6.4 What the precedent teaches, and what it does not + +The precedent teaches the shape of an intervention that is neither disclosure nor training. The letters were the same before and after, so nothing was revealed. The reader was the same before and after, so nobody was taught. A piece of inferential work moved from the person to the artifact, and it stayed there because the convention was shared. That shape is the model this paper carries forward. + +The precedent does not teach that word separation spread literacy, and the paper does not claim it. Saenger's book "is not a general history of literacy" (1997, p. x), and no historian of literacy read for this paper credits spacing with raising the number of readers; Balogh (1927, p. 236) and Hendrickson (1929, pp. 193–194) date the end of voiced reading to the press, half a millennium later. Nor does the precedent teach that one move did the work alone. The author's ruling is that spacing was the critical move within a complex, and the record supports a complex (Tunbridge, 1992; O'Donnell, 2000, ¶4) with one device, the visible word boundary, carrying the segmentation cost that modern readers still pay without it. Finally, the precedent is a guide to the move from one capacity to another and no proof of how the next move will go. The capacity it helped was literacy, a person and a page; the capacity the triad calls for is something else. + +## 7. Freeing Algorithmacy from Literacy + +### 7.1 Algorithmacy + +I define algorithmacy as the capacity to decide well within an algorithmic triad: to reach the other party through a mediator that binds and optimizes, without first reading and understanding the mediator. The word has a prior printed use in another sense: Oberschelp (1995) wrote that "we have to preserve logicacy and algorithmacy by continuously training simple mental activities without artificial tools" (p. 121), meaning a skill of carrying out algorithmic steps by hand. The sense here is the lab's (Hunt, 2026): not executing algorithms but coordinating through one. Literacy takes in a fixed text and acts afterward; the read-to-check role of §2 is literacy applied to a machine. Algorithmacy acts and takes in the result, because the mediator is not a text, it answers, and it answers differently as the other party and the mediator's own objective move. The author's ruling is that understanding in these settings comes through deciding and quickly ingesting the results. The sources support a conditional form of that ruling, and the conditions are where the design work lies. + +Two cautions set the limits of the contrast. First, the ruling's own example, that pilots do not read the situation but act, is contradicted by the aviation record: both pilots in Hutchins's (1995) cockpit keep monitoring airspeed (pp. 283, 285), and monitoring generated the largest share of the deviations Dismukes and Berman (2010) observed, 391 of 899, and the least caught, 5.6%. What that record shows is narrower and still useful: people correct what they do far more often than what they watch. Second, acting as a way of knowing has a cost that Weick (1988) stated first. "People think by acting" (p. 305), but there is "a delicate tradeoff between dangerous action which produces understanding and safe inaction which produces confusion" (p. 305), and the acts that build understanding change the thing understood. On a platform the person's acts feed an adaptive mediator. Each of the three means below answers that problem in part, and none answers it alone. + +### 7.2 Epistemic action in a triad + +Kirsh and Maglio (1994) distinguished pragmatic actions, which bring an agent "closer to his or her physical goal" (p. 515), from epistemic actions, "physical actions that make mental computation easier, faster, or more reliable" (p. 513). Their Tetris players rotated falling pieces in as few as 100 ms where mental rotation took 800 to 1200 ms (pp. 530–531), and "experts keep their costs lower, in part, by performing more epistemic actions" (p. 547). This is the dyadic literature's own lever stated as a theory: keep the person acting, because acting is how the person comes to know. Onnasch et al.'s (2014) generation effect and Shneiderman's (2020) "rapid, incremental, and reversible actions" (p. 13) are the same point in human factors and HCAI. + +The 1994 paper also states the condition under which it works, and the condition excludes the triad as written. Epistemic rotation pays because "the local effects of an action are totally determinate. There are no hidden states, exogenous influences, or other agents to change the result of hitting the rotate key" (p. 543). A platform has all three. Kirsh (2010) supplies the bridge: for outcomes that depend on "the voting of n participants" or "certain market equilibrium problems," "it is necessary to run the process, and it is best to run the process repeatedly" (p. 453). In a triad, epistemic action is not acting to see a fixed answer but acting to see a moving one, and the one experiment that put a shared display in front of several agents and one provider shows what moving means. Kayongo et al. (2022) ran a congestion game in which players saw a chart of prior play. Players followed the chart two-thirds as often when it was public as when it was private (odds 0.67), 25–40% did not follow it even in private, and the "visualization equilibrium," at which the chart predicts the behavior it induces, lay at 0.34–0.35, between the Nash outcome (0.22) and the optimum (0.44) (pp. 1, 5–7). The provider's convention is constrained to a fixed point, and the agent's act of reading and responding moves the thing read. The defensible claim is therefore "act to know," not "act and the answer holds," and what makes acting informative in a triad is the subject of the third means. + +### 7.3 Strategic offloading + +The abstract's second means is the removal of strategic work from the person by the mediator's own rule. The label "strategic offloading" is the author's. Li (2017) does not use it; his term is *obviously strategy-proof*, and the phrase has a prior printed meaning in cognitive psychology, Gilbert's (2015) title for a person's own choice to put a delayed intention into the environment, in four online experiments with 1,196 participants. The paper keeps the author's label for the mechanism and attributes the mechanism, not the phrase, to Li. The idea itself runs through market design in other words. Li's own rationale is cognitive and about trust: "Participants have to see for themselves that it is incentive-compatible; the mechanism has to be simple. Simplicity bypasses the need for participants to trust the mechanism designer" (2024, p. 3). Boston's superintendent, recommending a strategy-proof school assignment in 2005, wanted "to enable families to list their true choices of schools without jeopardizing their chances," because the old rule gave "an advantage to families who have the time, resources and knowledge to conduct the necessary research" (Abdulkadiroğlu et al., 2006, p. 24). + +The mechanism is real and its size is measured. Li (2017) defines a strategy as obviously dominant if, "for any deviation, at any information set where both strategies first diverge, the best outcome under the deviation is no better than the worst outcome under the dominant strategy," and shows that an ascending clock auction is obviously strategy-proof while the second-price sealed-bid auction, with the same outcome rule, is not (§1). In his serial-dictatorship experiment, "36% of games do not end in the dominant strategy outcome" when subjects submit ranked lists and 7% when they take turns picking (§5.3, p. 29). Chen and Sönmez (2006) found roughly 80% of participants misrepresenting under the Boston mechanism, 53% under top trading cycles and 36% under Gale–Shapley (Table 4, p. 217). Martínez-Marquina et al. (2019) isolated the piece of work that goes: removing uncertainty about a counterpart while keeping every contingency payoff-relevant doubled payoff-maximizing play, 19.7% to 41.5%, N = 371, and made advice effective, 72.5% against 41.5% under certainty (Table 3, pp. 15, 19, 26). The clock auction works, they argue, by resolving the lottery over counterparts before the bidder acts (pp. 30–31). That is the scribal shape: contingent reasoning, the segmentation task of a strategic setting, moved from the person into the rule. + +Two findings narrow the means and point the search elsewhere. First, the obviously-strategy-proof label has not survived as the active ingredient. Breitmoser and Schweighofer-Kodritsch (2022) replicated Li's auction comparison with his materials and split the move into steps: about 49% of the reduction in deviations came from showing an ascending clock after sealed bids, 12% (not significant) from live bidding on the clock, which is the step that creates obvious strategy-proofness, and almost 40% from live drop-out information; "obviousness is really a matter of what one gets to see ... not how one formally gets to decide" (pp. 3–4, 18). Bó and Hakimov (2024) found 93% truthful play inside an OSP mechanism when the dominant path only required clinching and 56% when it required passing, the latter no better than the direct mechanism (Table 4, p. 33). Li (2024) concedes both points (fnn. 5, 8). What removes the task is the shape of the decision, pick an object, clinch, and what the person sees, a clock, rivals dropping out, a step-by-step rejection. Those are the abstract's first and third means, and offloading works through them. Second, strategy-proofness removes most but not all strategizing, and the residue is systematic. In Israel's psychology match, 19.3% of 672 relevant lists ranked an unfunded contract above the funded one despite repeated advice (Hassidim et al., 2021, pp. 11–21); in Hungary, 11% of 528,172 eligible lists contained a dominated choice, more often where it was cheaper, which the authors read as ruling out cognitive limits "as a sole determinant" (Shorrer & Sóvágó, 2023, pp. 2–13); among 1,714 recent residency applicants, 23.3% submitted a suboptimal list in an incentivized replica (Rees-Jones & Skowronek, 2018). The task is removed for the person only when the person can see that it is gone, and that seeing is a representational matter. + +### 7.4 Ecological feedback and immediacy + +The abstract's third means pairs two sources that support two different things, and the paper states them as two conditions. Shneiderman's (1983) third principle is "rapid, incremental, reversible operations whose impact on the object of interest is immediately visible" (p. 64), with the benefit that "users can immediately see if their actions are furthering their goals" (p. 65). Hutchins et al. (1985) explain why: "the feeling of directness results from the commitment of fewer cognitive resources" (p. 317), and rapid feedback lets the user "make use of perceptual faculties in evaluating the outcome of actions" (pp. 333–334); their water-tank display, which shows a rate rather than a level the user would otherwise compute (p. 323), is the scribal move on the evaluation side of an interface. This is *immediacy*, a property of an interface's response to the user's own action. Kahneman and Klein (2009) name two necessary conditions for trustworthy intuition, "an environment of sufficiently high validity and adequate opportunity to practice the skill" (p. 520), where validity "describes the causal and statistical structure of the relevant environment" (p. 520), and summarize that learning needs "prolonged practice and feedback that is both rapid and unequivocal" (p. 524). This is *validity*, a property of the environment, with rapidity as one conjoined clause of the practice condition. Hogarth et al. (2015) give the fuller classification: a kind environment's feedback "links outcomes directly to the appropriate actions or judgments and is both accurate and plentiful"; in a wicked one it is "poor, misleading, or even missing" (p. 379). The two conditions coincide only where the interface is the environment, which is the dyad the abstract set aside. + +Immediacy on its own is not a learning principle, and the paper does not claim it as one. Schmidt and Bjork (1992) report that feedback after every trial gave the best practice and the worst two-day retention, because "frequent feedback comes to be a part of the task" and blocks the learner's own error detection (pp. 212–213). McKay et al. (2022), across 61 papers (k = 75, N = 2,228), found the effect of reduced feedback frequency undetectable either way, g = 0.19 at acquisition and 0.01 at immediate retention, and advised scientists to "abstain from making substantive recommendations regarding feedback frequency" (pp. 32, 35). In the field, Casas-Arce et al. (2017) randomized about 800 repair professionals to weekly or monthly, detailed or aggregate customer-score feedback: monthly-detailed did best, weekly groups did no better than the control, and the mechanism was salience of the last report (pp. 23–33). Gneezy and Potters (1997) found that subjects given feedback every round bet 50.5% of their endowment against 67.4% for those given it every three rounds, and earned less (pp. 639–640). More frequent return of a counterpart's judgment through an intermediary made things worse in the one field case that resembles a rating system. + +Validity is the condition the triad threatens, and the games literature says how. Bereby-Meyer and Roth (2006) ran ten-period prisoner's dilemmas in which the partner's action was always visible; with deterministic payoffs first-period cooperation rose from 0.61 to 0.93 over 20 supergames, and with the same payoffs paid as lotteries it stayed at 0.40 to 0.48 (pp. 1035–1037), because "the game provides a feedback loop between what players learn and what there is to learn" (p. 1040). Noise in what comes back is enough to stop two people learning to cooperate even when each can see what the other did. Burton-Chellew and West (2021) give the quantity a mediator controls: *influence*, the share of a person's payoff range under the person's own control (eq. 1, p. 4). With 43% influence, contributions in a black-box public-goods game fell from 46% to 15% in 16 rounds; with 11–12% they finished "at just over 50%, indicative of no learning on average" (p. 6); across 237 published games with 17,940 participants, influence predicted the rate of learning (p. 6). Hogarth and Soyer (2011) show that experiencing simulated outcomes lifted sophisticated participants on a Bayesian task from 17% correct analytically to 97% (pp. 15–16), and left "little or no difference in accuracy" between sophisticated and naïve participants who experienced them (p. 24). People learn by acting and reading the return, provided the return tracks the act. A mediator that pools counterparts and optimizes on its own account lowers the share of the return that the act explains, and Hogarth et al. (2015) place an environment that reacts to the person's own actions, or changes over time, among the wicked cases unless "the nature of the dynamic change can be inferred" (p. 382). "Ecological feedback" in a triad therefore means a visible correlation between the person's act and what comes back, which the mediator can raise or lower by design. Immediacy serves it and does not substitute for it. + +## 8. Candidate Design Changes and the Test for a Smoking Gun + +### 8.1 What a smoking gun would have to do + +The precedent fixes four properties. The change discloses nothing new about the mediator; the rule may stay as opaque as before. It trains nobody; the counterpart does not have to learn the mediator to benefit. It moves a piece of work from the person into the artifact, and the person can see that the work is gone. And it holds, because the convention is shared or enforced rather than granted. The three means of §7 say which work can move: contingent reasoning about the other party into the rule (§7.3), the reading of a response from inference to perception (§7.2, §7.4), and the return made to track the act closely enough to learn from (§7.4). Four candidates meet at least three of the four properties. + +**Candidate 1: a rule that removes the contingency and shows it gone.** Uber's additive surge is the documented case: under multiplicative surge a driver's best response depended on the rider-side demand process through surge transitions, and the additive rule made that dependence irrelevant to the choice (Garg & Nazerzadeh, 2022, pp. 10–13, 22). Boston's deferred acceptance is the other (Abdulkadiroğlu et al., 2006). Both change the rule, neither explains it, and both leave a residue when the person cannot see that the task is gone (Hassidim et al., 2021; Shorrer & Sóvágó, 2023). The design implication is the one the designers drew: "the advantages of strategy proofness would be lost if it is not effectively communicated" (Abdulkadiroğlu et al., 2006, p. 27), and what communicates it is not a description but the form of the decision. + +**Candidate 2: a pick-from-a-menu form with step-by-step return.** This is where the measured gains in market design sit. Dynamic deferred acceptance raised truth-telling over the static form, a gain Hakimov and Kübler (2021) attribute to "the feedback it provides after every step ... and the possibility to re-strategize" (p. 468); posting tentative cut-offs after each step raised it further, and continuous feedback on the would-be allocation raised stable outcomes from 83% to "almost 99% of markets" (p. 469). Bó and Hakimov's (2024) "pick your best feasible object and wait to see if you can keep it" (p. 3) and Breitmoser and Schweighofer-Kodritsch's (2022) clock and drop-outs are the same device. The person acts, sees the provisional result, and acts again, with the contingency resolved one step at a time. This is the scribal move applied to a sequence: the mediator writes the boundary between settled and open, and the person reads it rather than inferring it. + +**Candidate 3: a return that tracks the act.** Burton-Chellew and West's (2021) influence measure names a design variable: the share of the outcome under the person's own control, which a mediator can raise by holding other inputs still long enough for the act's effect to show, or lower by pooling and re-optimizing between the act and the return. Rahman's (2019) unknown update times and hidden private inputs (pp. 94, 96–100) are the negative case; Heiland's (2025) corrected rider theories, produced by controlled variation, are the positive one. The candidate is a convention on timing and attribution, not on content: the mediator returns the result of a decision at a known time, attributable to the decision, and holds the rule still between returns. Casas-Arce et al. (2017) warn that returning it more often is not the same as returning it more attributably. + +**Candidate 4: a representation that shows the strategic structure without stating it.** Chou et al. (2009) recast a two-person guessing game as placing troops on a hill and raised dominant-strategy play in a community-college pool from 0% to 46%, with "no changes in incentives, information or subject pool," while a one-sentence hint that disclosed the structure moved Caltech subjects from 46% to 87% and college subjects only from 0% to 7% (Tables 4–5, pp. 5–7, 16). Re-representation helped those whom disclosure did not. Subjects "either perceive the game form correctly or do not" (pp. 7–8), which is the strongest evidence in the record that a single presentational change can flip behavior rather than nudge it. Against it, Crosetto and Mantovani (2018) found that making payoffs less accessible in a centipede game shifted stopping by two-thirds of a node, in a direction that depends on the baseline game (pp. 5–9), and Hollan et al. (2000) record an airspeed-tape redesign that "inadvertently removed" the perceptual property pilots used (p. 7). The same kind of change can run either way, and it can be lost without notice. + +### 8.2 What evidence would count + +No convention of any of these four kinds has been tested in a triad with a behavioral outcome. Every measured effect above comes from a person and a system, from two players with no mediator, or from a clearinghouse with no objective of its own. That gap is the paper's main finding about the literature and the first item of its agenda. In the manner of Terveen and McDonald (2005), whose claims were "intuitions about what is most likely to be true" with research questions attached (p. 414), I state four claims so that data could count against them. + +*Claim 1. In a modeled triad with an optimizing mediator, a convention that removes a contingency from the counterpart's decision (Candidate 1 or 2) will reduce the share of dominated choices more than a full description of the mediator's rule will, and the description will raise strategizing in a subsample.* The dyadic and clearinghouse evidence predicts both halves (Danz et al., 2022; Hakimov & Kübler, 2021; Chou et al., 2009). Data that would count against it: a triad in which the disclosed rule lowers dominated choices and the rule change does not. The design is a laboratory market with two populations and an experimenter-set mediator objective, the convention varied between arms and the rule held fixed (cf. Hohenstein et al., 2023). + +*Claim 2. The gain from a step-by-step return (Candidate 2) will be carried by what the counterpart sees, not by the incentive property of the rule.* Breitmoser and Schweighofer-Kodritsch (2022) and Bó and Hakimov (2024) predict it. Against it: an arm in which the same visible steps are shown under a manipulable rule and the gain vanishes. + +*Claim 3. Learning in the triad will track the counterpart's influence over the return (Candidate 3) and not the frequency of the return.* Burton-Chellew and West (2021), Bereby-Meyer and Roth (2006) and Casas-Arce et al. (2017) predict it. Against it: a triad in which more frequent returns at fixed influence raise learning, or higher influence at fixed frequency does not. + +*Claim 4. A triad's irreducibility, gauged by Φ on the modeled system with the measure, coding, lag and partition fixed in advance, will predict which conventions help: where the mediator's objective makes the system irreducible, conventions that raise influence will help more than conventions that remove contingency, because the contingency cannot be removed without changing what the mediator optimizes.* This claim joins the lab's instrument to the design question. Engel and Malone (2018) show that a Φ-family measure can be computed on coded human interaction and can track an outcome, and also show the choices a pre-specification has to close (lag chosen by maximum Φ, p. 5; low-variance nodes dropped until a valid value appeared, p. 13). Against it: a triad with high Φ in which removal of contingency helps as much as, or more than, raised influence. Φ here classifies the triad; it does not by itself show that a convention helps the people in it, and the two measures are complements. + +### 8.3 Who would make the change + +The last property, that the convention holds, is the one the triad makes hardest. Every strategy-proof or obviously-strategy-proof result above comes from a clearinghouse or an experimenter with no stake. Where the mediator has a stake, the picture changes. A strategy-proof direct mechanism hands an auctioneer the information to cheat, so "the first-price auction is the unique credible static mechanism" and "the ascending auction is the unique credible strategy-proof mechanism" (Akbarpour & Li, 2020, pp. 425, 428). eBay told users to "bid the absolute maximum they are willing to pay" and let proxy bidding work (Ockenfels & Roth, 2006, p. 298, n. 1), and users sniped anyway: 13% of last bids on eBay came in the last five minutes against 1% on Amazon, whose soft close removed the advantage, and experience raised late bidding on eBay and lowered it on Amazon (pp. 313–315). Google moved its display inventory to a unified first-price auction in 2019 because layered auctions had "driven our publishers and app developers to implement increasingly complicated ad monetization strategies," and told buyers and sellers that they "will need to ... adjust how they bid" (Cox, 2019). The mediator's interest decided the rule each time. A platform's own terms gave "broad, unfettered discretion" in all fourteen that Suzor (2018) read, and a European business user gets fifteen days' notice of a change (Regulation (EU) 2019/1150, Art. 3). In courts an interested judge is barred (*Tumey v. Ohio*, 1927); in dealer foreign exchange the Global FX Committee asks that substantial changes "be agreed before any transaction" (Global Foreign Exchange Committee, 2024, Principle 40), voluntarily. The paper does not name who should hold a triad's convention. What the record shows is mixed: courts and the European regulation moved the convention toward the weaker party under outside force, while Google's first-price auction, Uber's additive surge and Boston's change came on the mediator's or its designers' own initiative; the author's ruling leaves the holder open. The agenda therefore includes a second question beside the laboratory one: under what condition does an optimizing mediator adopt a convention that lowers its counterparts' strategic load, and whether the condition is credibility, as Akbarpour and Li's result suggests, statute, or collective action by the counterparts themselves. + +## 9. Scope and Limits + +The paper is a framework and reports no data. Its claims about the triad rest on sociological, economic and HCI studies of marketplace platforms, and on laboratory results from dyads and clearinghouses; the step from those to a triad with an optimizing mediator is the inference §8.2 is written to test. Several of the sources behind the three means were read in working-paper or preprint versions, among them Li (2017), Martínez-Marquina et al. (2019), Breitmoser and Schweighofer-Kodritsch (2022) and Hogarth and Soyer (2011), and their locators are to those versions. Saenger (1997) was read through its chapter records, two reviews and earlier extracts rather than in full, and the one sentence the paper quotes from it (p. 85) should be checked against a library copy before submission. The historical argument stands on a narrower base than the abstract states: the Insular dating and the learner's motive are supported; the claim that unspaced text forced reading aloud is contested by classicists and contradicted by reading science for fluent readers; and no source credits word separation with spreading literacy. The paper therefore carries the precedent as a guide to the shape of an intervention and not as a causal history. "Irreducible" is used in the senses the platform studies support, and Φ is named as the instrument that would make the word measurable on a modeled triad; it has not been computed on one. Two sources that bear on the argument could not be opened and are listed in the notes; both should be read before submission. + +## 10. Conclusion + +Human–machine interaction has carried the same human role through every change of medium: the person reads what the system sends and checks the machine. The role comes from a pair, an operator and an aid, and the field's own evidence says it degrades with time, with reliability and with the amount handed over, while the one protective factor is that the person keep deciding. On a platform the pair is the wrong picture. Two people coordinate through a mediator that settles their match and gains from how it settles it. The remedy that law and HCI reach for, disclosure, hands the counterpart a text. The platform studies show that a disclosed rule redirects strategizing and does not remove it, because the work the rule leaves in place is reaching the other person; the regulator's own guidelines say the disclosure exists so that businesses can invest in ranking strategies. + +Seventh-century Irish scribes made a change of a different kind. They put the word boundary on the page, where any visible mark will serve, and moved a segmentation task from the reader into the artifact without disclosing anything and without training anyone. Read in human-factors terms, that is offloading built in by design and a representational effect, and it carries the costs of dependence that every such move carries. The paper takes the shape and not the history: the record does not support the claim that the change freed literacy or ended voiced reading, and the paper does not need it to. + +The capacity a triad calls for is algorithmacy, deciding well through a mediator one has not read. Three means could free it from literacy, and the evidence sorts them. Epistemic action is the dyad's own lever, and in a triad it is acting to see a moving answer. Removing strategic work by rule is real and measured, and what does the removing is the shape of the decision and what the person sees, not the incentive label. Feedback helps when it is valid, when the return tracks the act, and immediacy serves that condition rather than substituting for it. Four candidate conventions follow, and none has been tested where it matters. The next step is a modeled triad with an optimizing mediator, the conventions varied between arms, the rule held still, and Φ fixed in advance as the gauge of what the mediator has made irreducible. The field will know it has found a smoking gun when a counterpart who has never read the mediator reaches the other person better, and can see on the surface that the work has gone. + +## References + +Abdulkadiroğlu, A., Pathak, P. A., Roth, A. E., & Sönmez, T. (2006). *Changing the Boston School Choice Mechanism* (NBER Working Paper No. 11965). National Bureau of Economic Research. https://doi.org/10.3386/w11965 + +Akbarpour, M., & Li, S. (2020). Credible Auctions: A Trilemma. *Econometrica*, *88*(2), 425–467. https://doi.org/10.3982/ecta15925 + +Bai, X., Yan, G., Liversedge, S. P., Zang, C., & Rayner, K. (2008). Reading spaced and unspaced Chinese text: Evidence from eye movements. *Journal of Experimental Psychology: Human Perception and Performance*, *34*(5), 1277–1287. https://doi.org/10.1037/0096-1523.34.5.1277 + +Bainbridge, L. (1983). Ironies of automation. *Automatica*, *19*(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8 + +Balogh, J. (1927). „Voces paginarum": Beiträge zur Geschichte des lauten Lesens und Schreibens. *Philologus*, *82*, 84–109, 202–240. https://doi.org/10.1524/phil.1927.82.14.84 + +Bambauer, J., & Zarsky, T. (2018). The Algorithm Game. *Notre Dame Law Review*, *94*(1), 1–48. https://ndlawreview.org/wp-content/uploads/2018/11/1-Bambauer-Zarksy.pdf + +Barton, D., & Hamilton, M. E. (1980). *Awareness of the segmental structure of English in adults of various literacy levels* (ERIC ED222858). Stanford University, Department of Linguistics. + +Ben-Shahar, O., & Schneider, C. E. (2011). The Failure of Mandated Disclosure. *University of Pennsylvania Law Review*, *159*, 647–749. https://chicagounbound.uchicago.edu/journal_articles/1067 + +Bereby-Meyer, Y., & Roth, A. E. (2006). The Speed of Learning in Noisy Games: Partial Reinforcement and the Sustainability of Cooperation. *American Economic Review*, *96*(4), 1029–1042. https://doi.org/10.1257/aer.96.4.1029 + +Bishop, S. (2019). Managing visibility on YouTube through algorithmic gossip. *New Media & Society*, *21*(11–12), 2589–2606. https://doi.org/10.1177/1461444819854731 + +Björkegren, D., Blumenstock, J. E., & Knight, S. (2020). *Manipulation-Proof Machine Learning* [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2004.03865 + +Bó, I., & Hakimov, R. (2024). Pick-an-Object Mechanisms. *Management Science*, *70*(7), 4693–4721. https://doi.org/10.1287/mnsc.2023.4908 + +Breitmoser, Y., & Schweighofer-Kodritsch, S. (2022). Obviousness around the clock. *Experimental Economics*, *25*(2), 483–513. https://doi.org/10.1007/s10683-021-09720-z + +Bucher, E. L., Schou, P. K., & Waldkirch, M. (2020). Pacifying the algorithm – Anticipatory compliance in the face of algorithmic management in the gig economy. *Organization*, *28*(1), 44–67. https://doi.org/10.1177/1350508420961531 + +Burton-Chellew, M. N., & West, S. A. (2021). Payoff-based learning best explains the rate of decline in cooperation across 237 public-goods games. *Nature Human Behaviour*, *5*(10), 1330–1338. https://doi.org/10.1038/s41562-021-01107-7 + +Busch, S. (2002). Lautes und leises Lesen in der Antike. *Rheinisches Museum für Philologie*, *145*(1), 1–45. http://www.rhm.uni-koeln.de/145/Busch.pdf + +Casas-Arce, P., Lourenço, S. M., & Martínez-Jerez, F. A. (2017). The Performance Effect of Feedback Frequency and Detail: Evidence from a Field Experiment in Customer Satisfaction. *Journal of Accounting Research*, *55*(5), 1051–1088. https://doi.org/10.1111/1475-679X.12184 + +Castillo, J. C., Knoepfle, D., & Weyl, E. G. (2025). Matching and Pricing in Ride Hailing: Wild Goose Chases and How to Solve Them. *Management Science*, *71*(5), 4377–4395. https://doi.org/10.1287/mnsc.2022.00096 + +Chen, L., Mislove, A., & Wilson, C. (2016). An Empirical Analysis of Algorithmic Pricing on Amazon Marketplace. In *Proceedings of the 25th International Conference on World Wide Web* (pp. 1339–1349). International World Wide Web Conferences Steering Committee. https://doi.org/10.1145/2872427.2883089 + +Chen, Y., & Sönmez, T. (2006). School choice: an experimental study. *Journal of Economic Theory*, *127*(1), 202–231. https://doi.org/10.1016/j.jet.2004.10.006 + +Chou, E., McConnell, M., Nagel, R., & Plott, C. R. (2009). The control of game form recognition in experiments: understanding dominant strategy failures in a simple two person "guessing" game. *Experimental Economics*, *12*(2), 159–179. https://doi.org/10.1007/s10683-008-9206-4 + +Cofone, I. N., & Strandburg, K. J. (2019). Strategic Games and Algorithmic Secrecy. *McGill Law Journal*, *64*(4), 623–663. https://doi.org/10.7202/1074151ar + +Cohen, P., Hahn, R., Hall, J., Levitt, S., & Metcalfe, R. (2016). *Using Big Data to Estimate Consumer Surplus: The Case of Uber* (NBER Working Paper No. 22627). National Bureau of Economic Research. https://doi.org/10.3386/w22627 + +Cox, S. (2019). *Simplifying programmatic: first price auctions for Google Ad Manager*. Google Ad Manager blog. https://blog.google/products/admanager/simplifying-programmatic-first-price-auctions-google-ad-manager/ + +Crosetto, P., & Mantovani, M. (2018). Representation effects in the centipede game. *PLOS ONE*, *13*(10), e0204422. https://doi.org/10.1371/journal.pone.0204422 + +Cruz, S. A., & Gameiro, A. (2023). Digital work platform: Understanding platforms, workers, clients in a service relation. *Frontiers in Sociology*, *7*, 1075808. https://doi.org/10.3389/fsoc.2022.1075808 + +Curchod, C., Patriotta, G., Cohen, L., & Neysen, N. (2020). Working for an Algorithm: Power Asymmetries and Agency in Online Work Settings. *Administrative Science Quarterly*, *65*(3), 644–676. https://doi.org/10.1177/0001839219867024 + +Danz, D., Vesterlund, L., & Wilson, A. J. (2022). Belief Elicitation and Behavioral Incentive Compatibility. *American Economic Review*, *112*(9), 2851–2883. https://doi.org/10.1257/aer.20201248 + +de Moura Costa, G., Petry, M. R., & Moreira, A. P. (2022). Augmented Reality for Human–Robot Collaboration and Cooperation in Industrial Applications: A Systematic Literature Review. *Sensors*, *22*(7), 2725. https://doi.org/10.3390/s22072725 + +Directive (EU) 2024/2831 of the European Parliament and of the Council on improving working conditions in platform work. (2024). OJ L, 11.11.2024. http://publications.europa.eu/resource/celex/32024L2831 + +Dismukes, R. K., & Berman, B. (2010). *Checklists and Monitoring in the Cockpit: Why Crucial Defenses Sometimes Fail* (NASA Technical Memorandum NASA/TM-2010-216396). NASA Ames Research Center. https://ntrs.nasa.gov/citations/20110011145 + +Doyle, A., & McCrae, J. P. (2025). An assessment of word separation practices in Old Irish text resources and a universal method for tokenising Old Irish text. In B. Davis, T. Fransen, E. Uí Dhonnchadha, & A. Walsh (Eds.), *Proceedings of the 5th Celtic Language Technology Workshop (CLTW 5)* (pp. 1–11). International Committee on Computational Linguistics. https://aclanthology.org/2025.cltw-1.1/ + +Engel, D., & Malone, T. W. (2018). Integrated information as a metric for group interaction. *PLOS ONE*, *13*(10), e0205335. https://doi.org/10.1371/journal.pone.0205335 + +Enqvist, L. (2023). 'Human oversight' in the EU artificial intelligence act: what, when and by whom? *Law, Innovation and Technology*, *15*(2), 508–535. https://doi.org/10.1080/17579961.2023.2245683 + +Eslami, M., Vaccaro, K., Lee, M. K., Elazari Bar On, A., Gilbert, E., & Karahalios, K. (2019). User Attitudes towards Algorithmic Opacity and Transparency in Online Reviewing Platforms. In *Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems* (pp. 1–14). ACM. https://doi.org/10.1145/3290605.3300724 + +European Commission. (2020). *Commission Notice: Guidelines on ranking transparency pursuant to Regulation (EU) 2019/1150 of the European Parliament and of the Council (2020/C 424/01)*. Official Journal of the European Union, C 424, 8.12.2020, p. 1. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52020XC1208(01) + +Farronato, C., Fradkin, A., & MacKay, A. (2023). Self-Preferencing at Amazon: Evidence from Search Rankings. *AEA Papers and Proceedings*, *113*, 239–243. https://doi.org/10.1257/pandp.20231068 + +Garg, N., & Nazerzadeh, H. (2022). Driver Surge Pricing. *Management Science*, *68*(5), 3219–3235. https://doi.org/10.1287/mnsc.2021.4058 + +Ghalme, G., Nair, V., Eilat, I., Talgam-Cohen, I., & Rosenfeld, N. (2021). Strategic classification in the dark. In *Proceedings of the 38th International Conference on Machine Learning* (Proceedings of Machine Learning Research, Vol. 139, pp. 3672–3681). https://proceedings.mlr.press/v139/ghalme21a.html + +Gilbert, S. J. (2015). Strategic offloading of delayed intentions into the external environment. *Quarterly Journal of Experimental Psychology*, *68*(5), 971–992. https://doi.org/10.1080/17470218.2014.972963 + +Gilbert, S. J., Bird, A., Carpenter, J. M., Fleming, S. M., Sachdeva, C., & Tsai, P.-C. (2020). Optimal use of reminders: Metacognition, effort, and cognitive offloading. *Journal of Experimental Psychology: General*, *149*(3), 501–517. https://doi.org/10.1037/xge0000652 + +Gilbert, S. J., Boldt, A., Sachdeva, C., Scarampi, C., & Tsai, P.-C. (2023). Outsourcing Memory to External Tools: A Review of 'Intention Offloading'. *Psychonomic Bulletin & Review*, *30*(1), 60–76. https://doi.org/10.3758/s13423-022-02139-4 + +Global Foreign Exchange Committee. (2024). *FX Global Code: A set of global principles of good practice in the foreign exchange market*. https://www.globalfxc.org/docs/fx_global.pdf + +Gneezy, U., & Potters, J. (1997). An Experiment on Risk Taking and Evaluation Periods. *The Quarterly Journal of Economics*, *112*(2), 631–645. https://doi.org/10.1162/003355397555217 + +Gonczarowski, Y., Heffetz, O., Ishai, G., & Thomas, C. (2024). *Describing Deferred Acceptance and Strategyproofness to Participants: Experimental Analysis* (NBER Working Paper No. 33020). National Bureau of Economic Research. https://doi.org/10.3386/w33020 + +Grahn, S., Wang, T., Wang, X. V., & Wang, L. (2026). From Human–Machine Interaction to Integration: A Review of Paradigms, Principles, and Prospects for a Human-Centric Future. *Human-Machine Interaction*, *1*(1), 1. https://www.sciltp.com/journals/hmi + +Grinschgl, S., Papenmeier, F., & Meyerhoff, H. S. (2021). Consequences of cognitive offloading: Boosting performance but diminishing memory. *Quarterly Journal of Experimental Psychology*, *74*(9), 1477–1496. https://doi.org/10.1177/17470218211008060 + +Hagiu, A., & Wright, J. (2015). Multi-sided platforms. *International Journal of Industrial Organization*, *43*, 162–174. https://doi.org/10.1016/j.ijindorg.2015.03.003 + +Hakimov, R., & Kübler, D. (2021). Experiments on centralized school choice and college admissions: a survey. *Experimental Economics*, *24*(2), 434–488. https://doi.org/10.1007/s10683-020-09667-7 + +Haldar, M., Abdool, M., Ramanathan, P., Xu, T., Yang, S., Duan, H., Zhang, Q., Barrow-Williams, N., Turnbull, B. C., Collins, B. M., & Legrand, T. (2019). Applying Deep Learning to Airbnb Search. In *Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining* (pp. 1927–1935). ACM. https://doi.org/10.1145/3292500.3330658 + +Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-Mediated Communication: Definition, Research Agenda, and Ethical Considerations. *Journal of Computer-Mediated Communication*, *25*(1), 89–100. https://doi.org/10.1093/jcmc/zmz022 + +Hardt, M., Megiddo, N., Papadimitriou, C., & Wootters, M. (2016). Strategic Classification. In *Proceedings of the 2016 ACM Conference on Innovations in Theoretical Computer Science* (pp. 111–122). https://doi.org/10.1145/2840728.2840730 + +Hassidim, A., Romm, A., & Shorrer, R. I. (2021). The Limits of Incentives in Economic Matching Procedures. *Management Science*, *67*(2), 951–963. https://doi.org/10.1287/mnsc.2020.3591 + +Heiland, H. (2025). The Social Construction of Algorithms: A Reassessment of Algorithmic Management in Food Delivery Gig Work. *New Technology, Work and Employment*, *40*(1), 1–19. https://doi.org/10.1111/ntwe.12282 + +Heilmann, J. (2020). Reading Early New Testament Manuscripts: Scriptio continua, "Reading Aids", and Other Characteristic Features. In *Material Aspects of Reading in Ancient and Medieval Cultures* (pp. 177–196). De Gruyter. https://doi.org/10.1515/9783110639247-011 + +Hendrickson, G. L. (1929). Ancient reading. *The Classical Journal*, *25*(3), 182–196. + +Hogarth, R. M., Lejarraga, T., & Soyer, E. (2015). The Two Settings of Kind and Wicked Learning Environments. *Current Directions in Psychological Science*, *24*(5), 379–385. https://doi.org/10.1177/0963721415591878 + +Hogarth, R. M., & Soyer, E. (2011). Sequentially simulated outcomes: Kind experience versus nontransparent description. *Journal of Experimental Psychology: General*, *140*(3), 434–463. https://doi.org/10.1037/a0023265 + +Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J., & Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. *Scientific Reports*, *13*(1), 5487. https://doi.org/10.1038/s41598-023-30938-9 + +Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed cognition: Toward a new foundation for human-computer interaction research. *ACM Transactions on Computer-Human Interaction*, *7*(2), 174–196. https://doi.org/10.1145/353485.353487 + +Hunt, R. (2026, September 30). *Decision experience design: Yes, another XD* [Substack post]. https://rogerhuntphdcand.substack.com/p/decision-experience-design + +Hutchins, E. (1995). How a Cockpit Remembers Its Speeds. *Cognitive Science*, *19*(3), 265–288. https://doi.org/10.1207/s15516709cog1903_1 + +Hutchins, E. L., Hollan, J. D., & Norman, D. A. (1985). Direct manipulation interfaces. *Human–Computer Interaction*, *1*(4), 311–338. https://doi.org/10.1207/s15327051hci0104_2 + +Jhaver, S., Karpfen, Y., & Antin, J. (2018). Algorithmic Anxiety and Coping Strategies of Airbnb Hosts. In *Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems* (pp. 1–12). ACM. https://doi.org/10.1145/3173574.3173995 + +Kahneman, D., & Klein, G. (2009). Conditions for Intuitive Expertise: A Failure to Disagree. *American Psychologist*, *64*(6), 515–526. https://doi.org/10.1037/a0016755 + +Kasisopa, B., Reilly, R. G., Luksaneeyanawin, S., & Burnham, D. (2016). Child readers' eye movements in reading Thai. *Vision Research*, *123*, 8–19. https://doi.org/10.1016/j.visres.2015.07.009 + +Kayongo, P., Sun, G., Hartline, J., & Hullman, J. (2022). Visualization Equilibrium. *IEEE Transactions on Visualization and Computer Graphics*, *28*(1), 465–474. https://doi.org/10.1109/TVCG.2021.3114842 + +Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at Work: The New Contested Terrain of Control. *Academy of Management Annals*, *14*(1), 366–410. https://doi.org/10.5465/annals.2018.0174 + +Kirsh, D. (2010). Thinking with external representations. *AI & Society*, *25*(4), 441–454. https://doi.org/10.1007/s00146-010-0272-8 + +Kirsh, D., & Maglio, P. (1994). On Distinguishing Epistemic from Pragmatic Action. *Cognitive Science*, *18*(4), 513–549. https://doi.org/10.1207/s15516709cog1804_1 + +Kizilcec, R. F. (2016). How Much Information? Effects of Transparency on Trust in an Algorithmic Interface. In *Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems* (pp. 2390–2395). ACM. https://doi.org/10.1145/2858036.2858402 + +Klein, G., Woods, D. D., Bradshaw, J. M., Hoffman, R. R., & Feltovich, P. J. (2004). Ten Challenges for Making Automation a "Team Player" in Joint Human-Agent Activity. *IEEE Intelligent Systems*, *19*(6), 91–95. https://doi.org/10.1109/MIS.2004.74 + +Kleinberg, J., & Raghavan, M. (2020). How Do Classifiers Induce Agents to Invest Effort Strategically? *ACM Transactions on Economics and Computation*, *8*(4), 1–23. https://doi.org/10.1145/3417742 + +Knox, B. M. W. (1968). Silent Reading in Antiquity. *Greek, Roman, and Byzantine Studies*, *9*(4), 421–435. https://grbs.library.duke.edu/index.php/grbs/article/view/10731 + +Koulu, R. (2020). Proceduralizing control and discretion: Human oversight in artificial intelligence policy. *Maastricht Journal of European and Comparative Law*, *27*(6), 720–735. https://doi.org/10.1177/1023263X20978649 + +Lai, V., Chen, C., Smith-Renner, A., Liao, Q. V., & Tan, C. (2023). Towards a Science of Human-AI Decision Making: An Overview of Design Space in Empirical Human-Subject Studies. In *Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency* (pp. 1369–1385). ACM. https://doi.org/10.1145/3593013.3594087 + +Larkin, J. H., & Simon, H. A. (1987). Why a Diagram is (Sometimes) Worth Ten Thousand Words. *Cognitive Science*, *11*(1), 65–100. https://doi.org/10.1111/j.1551-6708.1987.tb00863.x + +Laubrock, J., & Kliegl, R. (2015). The eye-voice span during reading aloud. *Frontiers in Psychology*, *6*, 1432. https://doi.org/10.3389/fpsyg.2015.01432 + +Lee, M. K., Kusbit, D., Metsky, E., & Dabbish, L. (2015). Working with Machines: The Impact of Algorithmic and Data-Driven Management on Human Workers. In *Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems* (pp. 1603–1612). https://doi.org/10.1145/2702123.2702548 + +Leinenger, M. (2014). Phonological coding during reading. *Psychological Bulletin*, *140*(6), 1534–1555. https://doi.org/10.1037/a0037830 + +Li, S. (2017). Obviously Strategy-Proof Mechanisms. *American Economic Review*, *107*(11), 3257–3287. https://doi.org/10.1257/aer.20160425 + +Li, S. (2024). Designing Simple Mechanisms. *Journal of Economic Perspectives*, *38*(4), 175–192. https://doi.org/10.1257/jep.38.4.175 + +Li, X., Bicknell, K., Liu, P., Wei, W., & Rayner, K. (2014). Reading is fundamentally similar across disparate writing systems: A systematic characterization of how words and characters influence eye movements in Chinese reading. *Journal of Experimental Psychology: General*, *143*(2), 895–913. https://doi.org/10.1037/a0033580 + +Loewenstein, G., Sunstein, C. R., & Golman, R. (2014). Disclosure: Psychology Changes Everything. *Annual Review of Economics*, *6*, 391–419. https://doi.org/10.1146/annurev-economics-080213-041341 + +Luca, M., & Zervas, G. (2016). Fake It Till You Make It: Reputation, Competition, and Yelp Review Fraud. *Management Science*, *62*(12), 3412–3427. https://doi.org/10.1287/mnsc.2015.2304 + +Martínez-Marquina, A., Niederle, M., & Vespa, E. (2019). Failures in Contingent Reasoning: The Role of Uncertainty. *American Economic Review*, *109*(10), 3437–3474. https://doi.org/10.1257/aer.20171764 + +Masuda, T., Mikami, R., Sakai, T., Serizawa, S., & Wakayama, T. (2022). The net effect of advice on strategy-proof mechanisms: an experiment for the Vickrey auction. *Experimental Economics*, *25*(3), 902–941. https://doi.org/10.1007/s10683-021-09736-5 + +Matsuo, R., Liu, H., Hiraoka, T., & Wada, T. (2026). An Educational Human Machine Interface Providing Request-to-Intervene Trigger and Reason Explanation for Enhancing the Driver's Comprehension of ADS's System Limitations. *IEEE Transactions on Human-Machine Systems*, *56*(3), 650–661. https://doi.org/10.1109/thms.2026.3679748 + +Matzinger, T., Ritt, N., & Fitch, W. T. (2021). The influence of different prosodic cues on word segmentation. *Frontiers in Psychology*, *12*, 622042. https://doi.org/10.3389/fpsyg.2021.622042 + +McKay, B., Hussien, J., Vinh, M.-A., Mir-Orefice, A., Brooks, H., & Ste-Marie, D. M. (2022). Meta-analysis of the reduced relative feedback frequency effect on motor learning and performance. *Psychology of Sport and Exercise*, *61*, 102165. https://doi.org/10.1016/j.psychsport.2022.102165 + +Mourtzis, D., & Tsakalos, D. (2026). Emerging Technologies in Human-Machine Interaction towards Industry 5.0 and Industrial Metaverse: Challenges and Opportunities. *Human-Machine Interaction*, *1*(1), 4. https://www.sciltp.com/journals/hmi + +Nadella, S. (2026, October 10). *Models as insider risks in the super intelligence era* [Blog post]. sn scratchpad. https://snscratchpad.com/posts/models-as-insider-risks/ + +Oberschelp, W. (1995). Developing analytical competency through informatics teaching. In D. Watson & D. Tinsley (Eds.), *Integrating Information Technology into Education* (pp. 115–126). Springer. https://doi.org/10.1007/978-0-387-34842-1_12 + +Ockenfels, A., & Roth, A. E. (2006). Late and multiple bidding in second price Internet auctions: Theory and evidence concerning different rules for ending an auction. *Games and Economic Behavior*, *55*(2), 297–320. https://doi.org/10.1016/j.geb.2005.02.010 + +O'Donnell, J. J. (2000). Review of Paul Saenger, Space Between Words: The Origins of Silent Reading. *The Medieval Review*, 00.02.24. https://scholarworks.iu.edu/journals/index.php/tmr/article/view/14886 + +O'Neill, T., McNeese, N., Barron, A., & Schelble, B. (2022). Human–Autonomy Teaming: A Review and Analysis of the Empirical Literature. *Human Factors: The Journal of the Human Factors and Ergonomics Society*, *64*(5), 904–938. https://doi.org/10.1177/0018720820960865 + +Onnasch, L., Wickens, C. D., Li, H., & Manzey, D. (2014). Human Performance Consequences of Stages and Levels of Automation: An Integrated Meta-Analysis. *Human Factors: The Journal of the Human Factors and Ergonomics Society*, *56*(3), 476–488. https://doi.org/10.1177/0018720813501549 + +Özçetin, S., & Wiltse, H. (2025). Terms of entanglement: a posthumanist reading of Terms of Service. *Human–Computer Interaction*, *40*(1–4), 171–194. https://doi.org/10.1080/07370024.2023.2281928 + +Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. *Human Factors*, *52*(3), 381–410. https://doi.org/10.1177/0018720810376055 + +Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A Model for Types and Levels of Human Interaction with Automation. *IEEE Transactions on Systems, Man, and Cybernetics, Part A*, *30*(3), 286–297. https://doi.org/10.1109/3468.844354 + +Perdomo, J., Zrnic, T., Mendler-Dünner, C., & Hardt, M. (2020). Performative Prediction. In *Proceedings of the 37th International Conference on Machine Learning* (pp. 7599–7609). PMLR. https://proceedings.mlr.press/v119/perdomo20a.html + +Perea, M., Tejero, P., & Winskel, H. (2015). Can colours be used to segment words when reading? *Acta Psychologica*, *159*, 8–13. https://doi.org/10.1016/j.actpsy.2015.05.005 + +Petre, C., Duffy, B. E., & Hund, E. (2019). "Gaming the System": Platform Paternalism and the Politics of Algorithmic Visibility. *Social Media + Society*, *5*(4). https://doi.org/10.1177/2056305119879995 + +Quandt, M., & Freitag, M. (2021). A Systematic Review of User Acceptance in Industrial Augmented Reality. *Frontiers in Education*, *6*, 700760. https://doi.org/10.3389/feduc.2021.700760 + +Rahman, H. A. (2019). *Invisible Cages: Algorithmic Evaluations in Online Labor Markets* [Doctoral dissertation, Stanford University]. https://purl.stanford.edu/fc567cc8525 + +Rayo, M. F., Cooke, N. J., Flach, J. M., Hoffman, R. R., Smith, P. J., & Woods, D. D. (2023). Cognitive Systems Engineering at 40, Part I: Deriving Theory from Practice. *Proceedings of the Human Factors and Ergonomics Society Annual Meeting*, *67*(1), 183–186. https://doi.org/10.1177/21695067231193648 + +Rees-Jones, A., & Skowronek, S. (2018). An experimental investigation of preference misrepresentation in the residency match. *Proceedings of the National Academy of Sciences*, *115*(45), 11471–11476. https://doi.org/10.1073/pnas.1803212115 + +Regulation (EU) 2019/1150 of the European Parliament and of the Council on promoting fairness and transparency for business users of online intermediation services. (2019). OJ L 186, 11.7.2019, p. 57. https://eur-lex.europa.eu/eli/reg/2019/1150/oj + +Regulation (EU) 2022/2065 of the European Parliament and of the Council on a Single Market For Digital Services (Digital Services Act). (2022). OJ L 277, 27.10.2022, p. 1. https://eur-lex.europa.eu/eli/reg/2022/2065/oj + +Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Article 14 (Human oversight). (2024). Official Journal of the European Union, 12 Jul 2024. https://artificialintelligenceact.eu/article/14/ + +Risko, E. F., & Gilbert, S. J. (2016). Cognitive Offloading. *Trends in Cognitive Sciences*, *20*(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002 + +Rochet, J.-C., & Tirole, J. (2006). Two-sided markets: a progress report. *The RAND Journal of Economics*, *37*(3), 645–667. https://doi.org/10.1111/j.1756-2171.2006.tb00036.x + +Saenger, P. (1997). *Space Between Words: The Origins of Silent Reading*. Stanford University Press. https://doi.org/10.1515/9781503619081 + +Schmidt, R. A., & Bjork, R. A. (1992). New Conceptualizations of Practice: Common Principles in Three Paradigms Suggest New Concepts for Training. *Psychological Science*, *3*(4), 207–218. https://doi.org/10.1111/j.1467-9280.1992.tb00029.x + +Scribner, S., & Cole, M. (1981). *The Psychology of Literacy*. Harvard University Press. https://archive.org/details/psychologyoflite00sylv + +Sheridan, T. B., & Verplank, W. L. (1978). *Human and Computer Control of Undersea Teleoperators* (Technical report). Massachusetts Institute of Technology, Man-Machine Systems Laboratory. https://doi.org/10.21236/ADA057655 + +Shneiderman, B. (1983). Direct manipulation: A step beyond programming languages. *Computer*, *16*(8), 57–69. https://doi.org/10.1109/mc.1983.1654471 + +Shneiderman, B. (2020). Human-Centered Artificial Intelligence: Reliable, Safe & Trustworthy. *International Journal of Human–Computer Interaction*, *36*(6), 495–504. https://doi.org/10.1080/10447318.2020.1741118 + +Shorrer, R. I., & Sóvágó, S. (2023). Dominated Choices in a Strategically Simple College Admissions Environment. *Journal of Political Economy Microeconomics*, *1*(4), 781–807. https://doi.org/10.1086/726226 + +Simmel, G. (1950). *The Sociology of Georg Simmel* (K. H. Wolff, Ed. & Trans.). The Free Press. + +Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. *Science*, *333*(6043), 776–778. https://doi.org/10.1126/science.1207745 + +Stark, D., & Pais, I. (2020). Algorithmic Management in the Platform Economy. *Sociologica*, *14*(3), 47–72. https://doi.org/10.6092/issn.1971-8853/12221 + +Suzor, N. (2018). Digital Constitutionalism: Using the Rule of Law to Evaluate the Legitimacy of Governance by Platforms. *Social Media + Society*, *4*(3). https://doi.org/10.1177/2056305118787812 + +Tasnim, M., Verhagen, P., Blanke, T., Acar, E., & Ghebreab, S. (2025). Modeling Strategic Risk in School Choice: A Case for Transparent Design. *Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society*, *8*(3), 2470–2479. https://doi.org/10.1609/aies.v8i3.36731 + +Terveen, L., & McDonald, D. W. (2005). Social Matching: A Framework and Research Agenda. *ACM Transactions on Computer-Human Interaction*, *12*(3), 401–434. https://doi.org/10.1145/1096737.1096740 + +*Tumey v. Ohio*, 273 U.S. 510 (1927). https://www.law.cornell.edu/supremecourt/text/273/510 + +Tunbridge, G. L. (1992). *A Study of Scribal Practices in Early Irish and Anglo-Saxon Manuscripts* [Doctoral dissertation, University of Oxford]. https://doi.org/10.5287/ora-ave2z6gqd + +Vallas, S., & Schor, J. B. (2020). What Do Platforms Do? Understanding the Gig Economy. *Annual Review of Sociology*, *46*(1), 273–294. https://doi.org/10.1146/annurev-soc-121919-054857 + +Wang, Q., Huang, Y., Jasin, S., & Singh, P. V. (2023). Algorithmic Transparency with Strategic Users. *Management Science*, *69*(4), 2297–2317. https://doi.org/10.1287/mnsc.2022.4475 + +Weick, K. E. (1988). Enacted Sensemaking in Crisis Situations. *Journal of Management Studies*, *25*(4), 305–317. https://doi.org/10.1111/j.1467-6486.1988.tb00039.x + +Wiener, E. L., & Curry, R. E. (1980). Flight-deck automation: promises and problems. *Ergonomics*, *23*(10), 995–1011. https://doi.org/10.1080/00140138008924809 + +Winskel, H., Radach, R., & Luksaneeyanawin, S. (2009). Eye movements when reading spaced and unspaced Thai and English: A comparison of Thai–English bilinguals and English monolinguals. *Journal of Memory and Language*, *61*(3), 339–351. https://doi.org/10.1016/j.jml.2009.07.002 + +Wood, A. J., Graham, M., Lehdonvirta, V., & Hjorth, I. (2019). Good Gig, Bad Gig: Autonomy and Algorithmic Control in the Global Gig Economy. *Work, Employment and Society*, *33*(1), 56–75. https://doi.org/10.1177/0950017018785616 + +Zhang, J., & Norman, D. A. (1994). Representations in Distributed Cognitive Tasks. *Cognitive Science*, *18*(1), 87–122. https://doi.org/10.1207/s15516709cog1801_3 + +Zhang, J., & Patel, V. L. (2006). Distributed cognition, representation, and affordance. *Pragmatics & Cognition*, *14*(2), 333–341. https://doi.org/10.1075/pc.14.2.12zha + +## Notes (Claude/Fable; not part of the paper) + +*Written by Claude (Fable) on 2026-10-15 after the round-5 facets (S, T, U, V, W, Y, AA) and their S2 verification. The abstract above is the author's and was inserted by shell from `ABSTRACT_v4.md`; the body does not alter it. Every source cited in the body has a card in `literature/library/`; the reference list is built from the `.bib` entries for those keys, with the round-5 facet entries preferred where a key appears in more than one file. Title casing in the references follows the bib fields (Crossref) and has not been converted to APA sentence case.* + +### (1) Fact flags on the abstract, each with the correct value, its source and locator + +1. **"the system still sends a person something to read so they can check the machine" (factory headset).** Narrower than the field's own purpose. On the industrial headset the system mostly sends instructions, warnings and status; checking the robot is one use among several. De Moura Costa et al. (2022), pp. 27–29, 45 (`demouracosta2022augmented`); Quandt & Freitag (2021), abstract and p. 14 (`quandt2021systematic`). The audit-log half of the sentence is supported (Shneiderman 2020, pp. 3–4; Enqvist 2023, pp. 517–525; Koulu 2020; Nadella P15). Body §2.1 states the narrower form. +2. **"That approach assumes a dyad" (Parasuraman, Sheridan, & Wickens, 2000).** Supported for the supervisory lineage and the HCI decision-support studies (`parasuraman2000types` p. 287; `wiener1980flight` p. 1; `onnasch2014human`; `lai2023towards` §4.2.4; 45 of 76 studies in `oneill2022human`). Two caveats: the model's own ATC example involves a controller and a pilot (p. 293); and for Shneiderman (2020, p. 5) the dyad is a chosen position ("not teammates"), not an unexamined assumption. Body §3.1 and §3.3. +3. **"irreducible algorithmic triad."** "Three parties, one with its own objective" is supported (`stark2020algorithmic` pp. 53–54; `vallas2020what` 16.10; `rochet2006two`; `hagiu2015multi`). "Irreducible" has no measured sense in the sources; the defensible senses are Simmel's formal one (pp. 138, 141, 145), Rochet–Tirole's non-neutrality (pp. 6, 8, 11) and Castillo et al.'s system feedback (2017 draft, pp. 5–8, 13–14). No source computes or proposes Φ on a mediated triad; Engel & Malone (2018) estimated observational Φ on peer groups with no mediator (p. 11). Counterparts can ally against or exit the mediator (`vallas2020what` 16.7; `cruz2023digital` p. 3). The triad is a property of the marketplace form, not of algorithms as such (`stark2020algorithmic` p. 54 n. 13). Body §4.1–4.3. +4. **"binds and optimizes."** "Optimizes" supported; "binds" supported in a narrower form, since binding strength varies (Uber sets match and both prices; Amazon picks a default seller; Airbnb decides visibility while the host prices; freelance platforms rank and filter), and by Hagiu & Wright's definition (pp. 5, 21) a mediator controlling key terms is a hybrid, not a pure platform. Body §4.2. +5. **"Mandated disclosure (Ben-Shahar & Schneider, 2011) does not spare them from having to game the mediator to coordinate with the other person."** Ben-Shahar & Schneider's evidence is consumer and patient disclosure to individuals, pre-platform, with no triad (`benshahar2011failure` pp. 651, 665, 743–749). Disclosure evidence is mixed in direction: null or harm in `jhaver2019suspect`, `seira2017are`, `poursabzi2021manipulating`, `danz2022belief`; gains in `kizilcec2016how` (trust, self-report), `masuda2022net` (+23.8 points), `gonczarowski2024describing`, `matsuo2026educational` (1 of 15 vs 7 of 15), `gilbert2020optimal` (p. 28); `wang2023algorithmic` finds a region where transparency Pareto-dominates; `cofone2019strategic` (pp. 655–663) shows disclosed strong proxies are not gamed. "To coordinate with the other person" is supported by description, not by a test; no study isolates a coordination cost or measures the other party's loss. "Game" is the platform's moral label (`petre2019gaming`). Counterparts can game the mediator *by* coordinating with each other (`vallas2020what` 16.7). Body §5. +6. **"seventh-century Irish scribes ... broke up scriptio continua by adding spaces between words (Tunbridge, 1992; Saenger, 1997)."** Supported, with narrowings: the same scribes wrote Old Irish in stress-groups (Thurneysen 1946 p. 24, via `doyle2025assessment`; `tunbridge1992study` pp. 89–92); separation was a learner's analysis of Latin and came in a bundle with points, hyphens and letter forms (Tunbridge pp. 165–175; Saenger p. 85 names four devices); Latin had used interpuncts before the second century AD (`heilmann2020reading` pp. 182–183; `busch2002lautes` p. 19). Tunbridge was supervised by Parkes and is no independent test of Saenger's chronology. Body §6.1. +7. **"Unspaced text forced a reader to sound out the line just to hear where each word ended."** Cuts against as stated. Silent reading of unspaced text is attested in classical Athens (`knox1968silent` pp. 432–435); Heilmann (2020) calls the consensus a "quasi-dogma" (p. 179) and reads the laboured-reading passages as about learners (p. 181 n. 28); Balogh (1927) locates the difficulty in whole-word recognition and practice (pp. 227–228); Hendrickson (1929): "only strange, not necessarily difficult" (p. 186). Reading science: skilled readers of unspaced Chinese and Thai segment silently (`li2014reading`; `bai2008reading` 5,239 vs 5,221 ms; `kasisopa2016child` pp. 12–13); the eye leads the voice by 9.7–16.2 letters, ~561 ms (`laubrock2015eyevoice`); speech does not mark word ends (`matzinger2021influence`); word boundaries are partly a literate convention (`barton1980awareness`). Saenger's oralization thesis is contested at every load-bearing point (`odonnell2000saenger` ¶4; Heilmann pp. 178–183). Narrower form supported: for an unpractised reader of a foreign language, voicing gave a second channel for matching a sounded string to a known word (`busch2002lautes` p. 20; Vai readers via `scribner1981psychology`). Body §6.2. +8. **"scaffolding literacy with oracy."** "Oracy" is an education term for speaking-and-listening competence; none of the sources read uses it for voiced reading (facet W Q4). The sources say "reading aloud," "voiced reading" or "oralization." The term's origin (Wilkinson 1965) was reached at abstract level only and is therefore not cited in the body. +9. **"word separation freed literacy."** No historian of literacy read for the paper credits spacing with raising the number of readers; Saenger p. x: the book "is not a general history of literacy"; Balogh (p. 236) and Hendrickson (pp. 193–194) credit the printing press with the end of voiced reading; facet C records spacing standard c. 1050 and English male literacy about 20% in 1558. Narrower form: separation removed a segmentation cost for readers of Latin as a foreign language and for any reader of a spaced script deprived of the cue (33–69%: `winskel2009eye` p. 344; `perea2015colours` Table 2). Body §6.2 and §6.4. +10. **"moving boundary detection from the voice to the page."** "To the page" supported, with the narrowing that any salient mark serves (alternating colour recovers all but 4.7% of the cost: `perea2015colours` p. 12). "From the voice" cuts against for fluent readers (sources in flag 7) and is untested for learners. Body §6.2. +11. **"exploiting a representational effect (Zhang & Norman, 1994)."** Supported as an analogy; the construct is defined over isomorphs that carry identical information (p. 88) and over rules enforced by physics or shared cultural knowledge (p. 106). Where an unspaced string has two lawful parses, the space adds information. Nothing in Zhang & Norman concerns text. Body §6.3. +12. **"that enabled readers to internalize text."** Cuts against if it means reading without inner speech: skilled silent readers of spaced text generate phonological codes within 80–125 ms and move the articulators on hard text (`leinenger2014phonological`). Cuts against if it means retention: what is offloaded is practised and retained less (`grinschgl2021consequences` Table 1; `sparrow2011google` pp. 777–778). No source defines a third sense. Body §6.2 says the sentence has no source behind it. +13. **"the kind of cognitive offloading (Risko & Gilbert, 2016)."** The definition (p. 2) requires the person's own physical action; only Box 3 (p. 33) extends the term to offloading "built-in" by design. The body cites Box 3. Body §6.3. +14. **"epistemic action (Kirsh & Maglio, 1994)."** The 1994 paper ties the gain to actions whose local effects are "totally determinate," with no "other agents" (p. 543), which excludes a triad as written. `kirsh2010thinking` p. 453 supplies the bridge (outcomes depending on n participants must be run to be known); `kayongo2022visualization` shows that following a shared chart fell by a third once public (odds 0.67). Defensible form: "act to know," not "act and the answer holds." Body §7.2. +15. **"strategic offloading (Li, 2017)."** The phrase is the author's. Li's term is *obviously strategy-proof*; he does not use "offloading" (`li2017obviously`; `li2024designing` p. 3). "Strategic offloading" has a prior printed meaning in cognitive psychology, a person's own reminder-setting (`gilbert2015strategic`, N = 1,196). Further: the OSP label has not survived as the active ingredient (Breitmoser & Schweighofer-Kodritsch 2022: 49% / 12% n.s. / almost 40%, pp. 3–4; Bó & Hakimov 2024 Table 4: 93% vs 56%; Li 2024 fnn. 5, 8); 11–23% keep strategising under strategy-proof rules in the field (`hassidim2021limits`; `shorrer2023dominated`; `reesjones2018experimental`); every OSP result comes from a mediator with no stake, and interested mediators chose otherwise (`akbarpour2020credible` pp. 425, 428; `ockenfels2006late` pp. 298, 313–315; `cox2019simplifying`). Body §7.3 and §8.3 attribute the mechanism, not the phrase, to Li. +16. **"ecological feedback immediacy (Shneiderman, 1983; Kahneman & Klein, 2009)."** The two sources support two different things: Shneiderman's "immediately visible" is a property of an interface's response to the user's own action (pp. 64–65; `hutchins1985direct` pp. 317, 333–334), with predictability as a stated condition (1983 p. 62; `shneiderman1997direct` pp. 35–36); Kahneman & Klein's condition is environmental validity (p. 520) with feedback "both rapid and unequivocal" as one conjoined clause (p. 524). Immediacy as such is not a learning principle: frequent feedback helped practice and hurt retention (`schmidt1992new` pp. 212–213), the meta-analytic effect is undetectable (`mckay2022meta`, g ≈ 0.0–0.2), immediate vs delayed feedback across 38 classes was 0.002 (`fyfe2021manyclasses`), weekly feedback did worse than monthly-detailed (`casasarce2017performance` pp. 23–33), and more frequent feedback lowered bets and earnings (`gneezy1997experiment` pp. 639–640). Hogarth et al. (2015) place a reactive, nonstationary environment among the wicked cases (pp. 380, 382). The design tradition the abstract cites proposed disclosure of the model as its remedy for adaptive systems (`shneiderman1997maes` p. 58), which the author's D3 classifies as a literacy intervention. Body §7.4 states two conditions. +17. **"Human–machine integration (Grahn, Wang, Wang, & Wang, 2026)."** Supported; the review's own remedies are interpretable outputs, explicability and "civic AI literacy" (pp. 5, 13, 16), which puts it on the literacy side of the paper's distinction. Body §2.1. +18. **Nadella (2026).** Read as an enterprise-security proposal: "we" is the deployer throughout; "user," "customer," "employee" do not occur; model diversity (P14), independent controls (P17) and containment (P19) disclose nothing, and P9 calls chain-of-thought transparency "not sufficient or dependable," so "his only move is disclosure" (ruling D4) does not match the text. The critique in the body rests on the human seat (reader of records, holder of a brake) and on the post's silence about the people the model acts upon, not on a claim that people review records badly (CERT practice endorses such review; `kutasov2025shade` gives one AUC 0.65 measure). Body §2.2. + +### (2) Where the body narrows a claim the abstract makes + +- §2.1: the headset sends instructions, warnings and status; "to check the machine" fits the audit log better. +- §3.1–3.3: the dyad is the supervisory lineage's unit of analysis and, for Shneiderman (2020), a chosen position; the ATC example has several people. +- §4.1: the triad is a property of the marketplace form, not of algorithms as such; the labor-process frame (Kellogg et al. 2020) dissents. +- §4.2: binding varies by platform; the triad meant is the hybrid; counterparts can exit or ally. +- §4.3: "irreducible" is used in the senses the platform cases support (system feedback, non-neutrality); Φ is named as the instrument, not reported as applied. +- §5.1: Ben-Shahar & Schneider establish that handing a person a text is weak; they establish nothing about gaming a mediator. +- §5.2: the claim holds for ranking rules built on cheap, mutable signals, not for every disclosed rule (Cofone & Strandburg; Wang et al.). +- §5.3: "to coordinate with the other person" is supported by description, not by a test; "gaming" is the platform's label; coordinating with the counterpart can be the way the mediator is gamed. +- §5.4: disclosure helped where it was a short property statement, a reason attached to a cue, or computed advice; none in a triad with a behavioural outcome. +- §6.1: the scribes' move was a learner's analysis of a foreign language, one device in a bundle, and a return to an older Latin practice. +- §6.2: the "sound out the line to hear where each word ended" account is contested by classicists and contradicted by reading science for fluent readers; the narrower learner's-scaffold form is stated; "oracy" is flagged as a new sense; "internalize text" is stated to have no source behind it. +- §6.3: the representational effect is applied as an analogy with its isomorphism and cultural-convention conditions stated; offloading is cited through Box 3; the costs of offloading are reported. +- §6.4: the precedent is carried as a guide to the shape of an intervention, not as a causal history of literacy or of silent reading. +- §7.1: "pilots don't read the situation" is replaced by "people correct what they do far more often than what they watch"; Weick's trade-off is kept beside "think by acting." +- §7.2: epistemic action in a triad is "act to know," not "act and the answer holds." +- §7.3: "strategic offloading" is attributed to the author as a label; the OSP label is not the active ingredient; the residue of strategising is reported. +- §7.4: "ecological feedback immediacy" is split into validity (Kahneman & Klein; Hogarth et al.) and immediacy (Shneiderman; Hutchins et al.); immediacy is denied as a learning principle on its own; the triad is placed among the wicked cases unless the dynamics can be inferred. +- §8.2: no convention has been tested in a triad with a behavioural outcome; the claims are stated so data can count against them. +- §8.3: interested mediators have not adopted conventions that lower counterparts' strategic load; the holder of the convention is left open (ruling D8). +- §9: working-paper versions, the unread Saenger, and the two unreached sources are stated. + +### (3) Sources wanted that had no card + +- Parkes, M. B. (1997). Rædan, areccan, smeagan: how the Anglo-Saxons read. *Anglo-Saxon England, 26*, 1–22. doi:10.1017/s0263675100002088. Abstract/extract only (facet W). The closest source to "sound out the line"; it frames *lectio* as a schooled discipline of identifying letters, syllables and words in order to read aloud. +- Xu, L., Xie, Q., & Burtch, G. (2025). An empirical study of strategic opacity in crowdsourced evaluations. *MIS Quarterly, 49*(3). doi:10.25300/misq/2024/17441. Abstract only (facet U). Field experiments at 99designs reportedly found opacity reduced vote manipulation with no effect on participation; it cuts toward the platform's own rationale for opacity. +- Wilkinson, A. (1965). The concept of oracy. *Educational Review, 17*(4), 11–15. Abstract only (facet W). The origin of "oracy"; the body therefore describes the term's sense without citing him. +- Parkes, M. B. (1987/1991). The contribution of Insular scribes of the seventh and eighth centuries to the "grammar of legibility". No open copy (facet W). +- Bronner, D., Busch, N., Fleischer, J., & Poppe, E. (2018). (Non-)separation of words in early medieval Irish and German manuscripts. De Gruyter. doi:10.1515/9783110542899-003. Abstract only (facet W). +- Rahman, H. A. (2021). The invisible cage. *Administrative Science Quarterly*. Bib key `rahman2021invisible`; not opened. The body cites the 2019 dissertation (`rahman2019invisible`) instead. +- Möhlmann, M., & Zalmanson, L. (2017). Hands on the wheel. ICIS 2017; and Möhlmann et al. (2021), *MIS Quarterly, 45*(4). Abstract only (facet U); the press claim that drivers log off en masse to trigger surge is therefore not in the body. +- Mackworth, N. H. (1948); Endsley (1995, 2017, 2023); Parasuraman & Riley (1997); Hollnagel & Woods (1983); Goddard, Roudsari & Wyatt (2012). Not accessible or abstract only (facet S); Bainbridge's report of Mackworth's half-hour figure is cited through Bainbridge. +- Pathak & Sönmez (2008); Zhang & Levin (2017); Katuščák & Kittsteiner (2025); Troyan & Morrill (2020); Klijn, Pais & Vorsatz (2019). Abstract only (facet Y); the dynamic-DA and cut-off results are cited through `hakimov2021experiments`. +- Karelaia & Hogarth (2008), read in part with no card (facet AA): its lens-model finding that task information teaches and outcome feedback does not would strengthen §7.4 and is not cited. +- Krackhardt (1999), Monderer & Tennenholtz (2009), Caplow (1956), Yoon, Thye & Lawler (2013), Obstfeld (2005): not carded (facet T). +- Zhang (1997), Dunn & Risko (2016), Risko & Dunn (2015), Cason & Plott (2014), Gode & Sunder (1993) (facet V): abstract or record only. +- Scribner & Cole (1981): carded as partial through the authors' 1978 article; the Vai "click into meaningful units" wording is from the 1978 article (p. 456), so the body paraphrases without a page. + +### Bib-field discrepancies for the author + +- Heiland: the card header gives 2023 (accepted 5 Nov 2023), the bib gives 2025 (volume 40(1) is dated 2025). The body and references follow the bib (2025). +- Bucher, Schou & Waldkirch: the card header gives 2021 (volume 28(1), 2021), the bib gives 2020 (online 27 Nov 2020). The body and references follow the bib (2020). +- Özçetin & Wiltse: bib key `ozcetin2023terms`, bib year 2025 (volume 40). The body cites 2025. +- Oberschelp: the facet-Z bib lacks pages, editors and publisher; the entry above takes pp. 115–126, Watson & Tinsley (Eds.) and Springer from the card header. +- Hunt (2026): the bib entry has no URL or subtitle; the entry above takes both from the card header. +- Legal instruments (the four EU acts and *Tumey v. Ohio*) are listed by title, as the body cites them, rather than under the corporate author the bib records. + +## Applied from the claim check, 2026-10-15 (Claude; not part of the paper) + +The ledger is `literature/audit/claims/CLAIMS_v4.md`. Fixed in the body: the two contradicted claims (§1 "the letters were the same … It trained nobody", now the Tunbridge pp. 165–175 bundle and the trained-scribe narrowing; §8.3 "the record is uniform … under outside force", now mixed, with Google, Uber and Boston as mediator- or designer-initiated); the four miscitations (Onnasch p. 486 wording; Cofone & Strandburg p. 623; Luca & Zervas pp. 6, 19; Hogarth & Soyer's 17%/97% belongs to the sophisticated group, naïve parity p. 24); and three of the four overstatements (de Moura Costa "most … not certified"; Grahn et al.'s override and opt-out remedies; Φ "outside the author's program"). Ghalme et al. (2021) is now cited as PMLR 139:3672–3681 (checked on the PMLR page). Not changed: the abstract (the author's; its flags are in Notes 1), about 35 references whose body locators are preprint or working-paper pages while the list gives journal pagination (ledger, cross-cutting finding 2), and the six body claims that rest on the unopened Saenger (1997) and Scribner & Cole (1981) cards. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_OUTLINE.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_OUTLINE.md new file mode 100644 index 00000000..73e9e9f0 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_OUTLINE.md @@ -0,0 +1,287 @@ +# Outline — the foundation paper for Decision Experience Design + +*Written by Claude on 2026-10-12. This is a plan for the author to accept, change or reject, and he writes +the paper. Text in quotation marks under "Author's sentence" is his, copied verbatim. Everything else is +Claude's.* + +**Marks.** `[ok]` audited against the source. `[abs]` abstract only: paraphrase, do not quote. `[lib]` +needs a library copy. `[inf]` an inference with no source. `[D n]` waits on decision *n* below. + +## (a) Ten decisions, and the author's rulings of 2026-10-13 + +The author ruled on all ten on 2026-10-13. His answers are verbatim in `SOURCE.md` ("The author's ten +rulings"); the third column below is Claude's short reading of each. Sections (b) and (c) still follow +Claude's recommendations of 2026-10-12 and will be revised to the rulings once facets I, J and K are in. + +| # | Decision | Ruling, as Claude reads it | Still open | +| --- | --- | --- | --- | +| D1 | History as design precedent, or as cause of literacy? | Design precedent; the causal analysis is reviewed, not asserted | Facet K | +| D2 | `draft.md:263`: "no longer required the training" and "populace" | Rests on other scholars; review the debate first | Facet K | +| D3 | "Algorithmacy is not legibility" | Restated on disclosure: disclosure is a literacy-based intervention and does not change the ability to navigate | "Doesn't affect" against "has not been shown to change" | +| D4 | The contrast with the insider post | He argues against neuralese, which is not algorithmacy, and answers with disclosure and transparency, which are literacy-based | "His only move" does not match P11–P12, P14, P17, P19 | +| D5 | The "enterprise, not user" reading | Adopted | None | +| D6 | One move or a complex | A complex, with a search for the most critical convention in it | Which convention; facets H and I | +| D7 | Engagement layer wording | From reading and writing to deciding: understanding through deciding and taking in the results | The claim about pilots; Part 1 line 66 on disrupted feedback; facet I | +| D8 | Who builds the convention | Not known yet; compare courts, foreign-exchange markets and platforms | Facet J | +| D9 | The missing experiment | The triad is irreducible, and Φ gauges the irreducibility of any given triad | Whether the paper proposes a behavioural experiment; whether this arm computes Φ or cites the lab | +| D10 | Novice and expert support | The spacing history is a guide to the move from one -acy to another, and may or may not hold | Whether the course statement's claim is carried; Claude reads the answer as no | + +**D3, dressed in the language of Part 1, as the author asked.** *Drafted by Claude for the author to +accept, change or discard. The last sentence is the author's own, from Part 1, line 140.* + +> Disclosing an algorithm does not change a person's ability to navigate it. Disclosure hands the person +> a text—model weights, a saliency map, an activity log, a notice of change—and so asks for a reader. +> That makes it an intervention in literacy. Algorithmacy is exercised elsewhere, in the decisions a +> person makes inside a coordination that a third party keeps adjusting, and no document about the third +> party makes those decisions for them. A transparent triad remains a triad. + +The first sentence states the ruling as the author gave it. What the evidence in hand supports is +narrower: Part 1 (line 138) reports that FeedVis and Instawareness raised awareness without producing +critical evaluation or agency, and facet H finds weaker and less consistent effects for aids that hand +the person something to read (table (i)). A first sentence the sources carry: "Disclosing an algorithm +has not been shown to change how well a person navigates it." + +**What the rulings do to the argument (Claude's inferences, none sourced).** + +- D3 settles the legibility problem better than Claude's recommendation did. Word separation disclosed + nothing: the letters were the same before and after. It changed the form of the page, which is Zhang + and Norman's representational effect (1994, p. 89). The paper can oppose disclosure and keep its + precedent. +- Word separation is still an intervention on the literacy side. D10 answers this: the history guides + the search for an algorithmacy-side convention and does not supply one. +- D7 and Part 1 line 66 point at the same place. If people come to understand a mediator by deciding + and taking in the results, and the mediator returns results that are "partial, delayed, and mediated + through platform metrics" (Part 1, line 66), then a convention that returns the result of a decision + quickly and plainly is a candidate for the critical move of D6. Facet I tests whether the literature + supports this. +- D4 needs the post's own split. Chain-of-thought transparency (P9) and two of the seven principles + are disclosure outright: observe everything (P15) and incident disclosure (P20). Two more are testing + and audit, which produce evidence for an auditor to read (P16, P18). Three disclose nothing: model + diversity (P14), independent controls (P17) and containment (P19). The paper can say that everything + the post offers a *person* is something to read, and that its other controls are held by the + organisation. That keeps the author's contrast and matches the text. + +Also open: the due date, length and rubric for the course paper, and whether the instructor must agree to +the change of topic. + +## (b) Terms + +| Term | Recommendation | Note | +| --- | --- | --- | +| *scriptura continua* / *scriptio continua* | Either; pick one and say so | Saenger writes *scriptura continua* (p. 69), as the author does; Parkes and the classicists write *scriptio continua* | +| word separation | Keep | Saenger's term. "Spaces between words" on first use | +| aerated script | Define on first use or omit | Saenger's term for partly spaced Continental script | +| legibility | Use only with a qualifier | Graphic legibility (Parkes) against interpretability of a system (the ruling) | +| re-representation; representational effect | Adopt | Zhang and Norman 1994, p. 88 (card `zhang1994representations`; the earlier p. 89 was wrong) | +| task transformation | Adopt | Hutchins 1995, "How a cockpit remembers its speeds", p. 283 `[ok]`. Cite the article, not *Cognition in the Wild* | +| cognitive offloading | Available | Risko and Gilbert count offloading designed into the task environment (Box 3; card `risko2016cognitive`). *Corrected 2026-10-13; the earlier advice was to avoid the term* | +| function allocation | Avoid for the scribes' move | A page is not an agent | +| the move | Replace with "a convention" or "a complex of conventions" | `[D6]` | +| observability | Treat as a form of transparency | Rieder and Hofmann say it is not "a radical departure" from transparency (p. 21; card `rieder2020towards`). *Corrected 2026-10-13* | +| disclosure | Adopt as the paper's term for what it opposes | The author's ruling D3: disclosure is a literacy-based intervention | +| irreducible triad | Use in place of "mediated triad" | The author's ruling D9 | +| algorithmacy | Say "in this sense" where coinage is claimed | Oberschelp 1995, p. 121, prints the word in another sense (card `oberschelp1995developing`) | +| the insider post | Use in place of "Nadella" when the post is meant | His other 2026 posts say different things | +| engagement layer | Define from `ENGAGEMENT_LAYER.md:9` | `[D7]` | +| ignited literacy | Drop | `[D1]` | + +## (c) Section plan + +> **Superseded in part (2026-10-13).** The plan below is Claude's first-round plan and predates the +> author's rulings. Read it with `literature/facets/T_adjudication.md` §9 and the revised +> `ARGUMENT_MAP.md`, which are the current state. The main changes the rulings and the full readings +> make to it: section 3 opposes *disclosure*, not legibility, and may use "offloading"; section 4 drops +> "most for learners" and carries Saenger's p. 85 sentence under his name; section 5 moves from reading +> and rewriting to deciding (D7), with the pilot analogy narrowed; section 6 reads the post as offering a +> person something to read, takes up the enterprise reading (D5), and no longer rests on people reviewing +> records badly; section 7 carries two candidate conventions and the courts, foreign-exchange and +> platform comparison (D8). Claude has not redrafted the section plan, because the author writes the +> paper and the plan is his to re-cut. + +Word budgets are shares of the body, since the course length is not recorded. + +### 1. The standing answer (8%) + +- Open on the insider post as the current answer to opaque systems: external controls, records, audit, + a brake, authority with the deploying organisation. Post, P5, P11, P15, P17, P19 `[ok]`. +- State it fairly, in its own scope: enterprise security for agentic systems `[D5]`. +- Pose the question the post leaves open: what is asked of a person, and what of the people the model + acts upon? P20 is the only place they appear `[ok]`. +- State the paper's claim and its plan in two sentences. +- *Author's sentence:* "how should we design for 'algorithmacy,' which is similar to asking how we should + design for 'literacy.'" (`phd1750_topic_and_rationale.md:15-16`). + +### 2. The page before and after (18%) + +- Unspaced writing and who read it. Knox 1968, p. 435: "Ancient books were normally read aloud, but there + is nothing to show that silent reading of books was anything extraordinary…" `[ok]`; carry the whole + sentence. Saenger 1997, p. 85: reading rested on a "prolonged and arduous grammatical apprenticeship" + `[ok]`. +- Romans had word dividers and gave them up. Heilmann 2020, p. 182 `[ok]`. Saenger, p. 11: the ancient + world "did not possess the desire, characteristic of the modern age, to make reading easier and + swifter" `[ok]`. The device was known and set aside. +- Who separated words and for whom. Saenger, p. 83: the Book of Mulling (before 692) and Adomnán's *Vita + Columbae* (before 713) `[ok]`. Parkes 1992, p. 1: learners of Latin as a foreign language `[lib]`. +- The slow spread. Saenger, p. 100: "not adopted by Continental scribes until the mid-tenth century"; + pp. 176–177: standard by the mid-eleventh `[ok]`. Three and a half centuries from Ireland to the + Continent's norm (before 692 to c. 1050). +- What Saenger claims and what he does not. Speed and reference reading (p. 13); scholasticism (p. 122); + "not a general history of literacy" (p. x) `[ok]`. +- The dispute, told as a dispute. Vatri: "Saenger's model must be rejected" (2012, p. 639, via Heilmann + 2020, p. 181) `[ok]` for the sentence; Vatri's article `[lib]`. O'Donnell 2000 on the "necessary and + sufficient" claim `[ok]`. +- A complex of conventions: word order, punctuation, capitals, later tables and running heads `[D6]`. +- *Author's sentence:* "Word separation did not train the populace to read *scriptio continua*; it changed + the interface so that reading no longer required the training." (`draft.md:263`) `[D2]`. + +### 3. What the scribes did, in human-factors terms (18%) + +Organised by the course's three headings. + +**Perception.** +- Removing spaces costs practised readers of English a third of their reading rate (Winskel et al. 2009, + p. 344: "an average of 33% decrement") `[ok]`; about half in Rayner et al. 1998 `[abs]`. +- The cost belongs to the missing boundary: alternating bold recovers most of it (Perea and Acha 2009) + `[abs]`. +- Fluent readers of unspaced Chinese read no faster with spaces: 5,221 ms against 5,239 ms, 16 readers + (Bai et al. 2008, Table 1) `[ok]`. Thai: "marginally significant", F(1, 7) = 4.21, p < .08, n = 8 + (Kohsom and Gobet 1997) `[ok]`. + +**Cognition.** +- The representational effect: "different isomorphic representations of a common formal structure can + cause dramatically different cognitive behaviors" (Zhang and Norman 1994, p. 89) `[ok]`. +- Task transformation: "a memory and scale reading task is transformed into a judgment of spatial + adjacency"; "Individual pilot memory has not been enhanced" (Hutchins 1995, pp. 283, 282) `[ok]`. +- The work moved to the scribe; it was not removed. Hutchins 1995, p. 280 `[ok]`. `[D2]` + +**Human factors and design.** +- Who gains. Learners most, on mixed evidence: Shen et al. 2012 `[abs]`, Blythe et al. 2012 `[abs]`; + against, Bassetti and Masterson 2012: "no facilitative effects" `[ok]`. `[D10]` +- Costs of an artefact absorbing a task: Risko and Gilbert 2016, Box 3 `[ok]`; Bainbridge's remedy of + deliberate pacing (1983) `[ok]`. +- The contrast case, set up for section 6: "The human monitor has been given an impossible task" + (Bainbridge 1983, p. 776) `[ok]`. + +### 4. What the precedent does not show (10%) + +- It did not spread literacy. Ten histories of literacy name schools, religion, administration, commerce + and print; one indexes word separation, once (Clanchy, p. 130) (facet C; searched, not read). +- Timing. Cressy 1980: "a reasonable guess might place male illiteracy around 80%" in 1558, five + centuries after spacing was standard `[ok]`, page `[lib]`. +- An unspaced script learned informally by adults: the Vai (Scribner and Cole 1981) `[ok]`, pages `[lib]`. +- One convention among several, in every parallel case: Guido's staff ("untenable" as invention, Haines + 2008, p. 331 `[ok]`); the 1299 Florentine rule ("a matter of speculation", Nothaft 2020, p. 15 `[ok]`); + Hangul, marginal for four centuries (facet F; second-tier sources). +- A measured modern parallel from the numeracy line the author already uses: easier formats help + low-numeracy people decide (Peters et al. 2007 `[abs]`; Hibbard and Peters 2003). +- Street's critique stands: literacy is "a social practice, not simply a technical and neutral skill" + (Street 2003, p. 77) `[ok]`. `[D1]` + +### 5. From the page to the mediator (14%) + +- What algorithmacy asks of a person, in the author's current definition + (`algorithmacy_ot/manuscript/PAPER.md:19-21`): interpret, specify intent, keep track. +- Where the precedent carries: work that a convention could carry is being done by each person, by folk + theory (Eslami et al. 2016; repo cards). +- Where it breaks, stated in the paper's own voice: + - Media hold still; mediators adapt. Hutchins 1995, p. 277; Scaife and Rogers 1996, p. 210; Vicente + and Rasmussen 1992, V.B `[ok]`. Part 1, line 118. + - The scribe shared the reader's interest; the interface's owner does not. Jensen and Meckling 1976, + p. 308 `[ok]`; Part 1, line 60. `[D8]` + - No trained elite exists to learn from. Part 1, lines 66, 138, 156. + - Algorithmacy is a sensibility of persons; the remedy sits in artefacts. The paper says how the two + relate. +- *Author's sentence:* "The design object is the engagement layer — where reading and rewriting the shared + mediator is the fundamental mode rather than a product layer that hides the mediator behind a + distribution network." (`phd1750_topic_and_rationale.md:52-54`) `[D7]`. + +### 6. The insider post, read closely (14%) + +- What it gets right, conceded in the body: the model cannot be read (P2, P4); chain-of-thought + transparency is "not sufficient or dependable" (P9); controls belong outside the model (P11); + disclosure "to those affected" (P20) `[ok]`. +- Its lineage: the reference monitor (Anderson 1972) and Saltzer and Schroeder 1975, whose list includes + "psychological acceptability", a human-factors principle the post leaves unused `[ok]`. +- The difference, on three counts `[D4]`: + 1. *Authority.* It rests with "Orgs" (P5, P17). The brake belongs to "an authorized person" (P19), so + it resembles structural refusal in form only. + 2. *The residual human task.* A person reads "human readable evidence" (P15) and decides when to halt + (P19). Sustained monitoring is the task people do worst: Bainbridge 1983, p. 776 `[ok]`; Green 2022 + `[ok]`; alert-review studies `[abs]`. Limit: the evidence does not reach after-the-fact + reconstruction by trained auditors. + 3. *Who is absent.* No user, customer or counterpart appears; "those affected" appear once, after a + failure (P20). +- Containment as the floor under the paper's proposal, not its opposite. +- The two tests that separate the paper from the post, what task the person keeps and whom the move + serves, are the paper's own `[inf]`. Say so. + +### 7. Candidates for the convention, and the experiment nobody has run (12%) + +- Sort candidates by what they do with the work: remove a task by fixing the mediator's rule; hand over + a finished comparison; hand the person something to read or tune; hand the task to the person's agent; + remove exposure. +- Rules that remove a strategic task: under a formally equivalent rule, "36% of games do not end in the + dominant strategy outcome" against 7% when the rule's form makes the dominance obvious (Li 2017, §5.2, + 144 students) `[ok]`. In the field 23% still misreport (Rees-Jones and Skowronek 2018, n = 1,714) + `[ok]`. Presentation decides who benefits, which is the spacing lesson. +- A finished comparison: 28% against 17% switched plans (Kling et al. 2012, n = 406) `[ok]`. +- Reading aids: mostly satisfaction, several nulls (facet H, table (i)) `[abs]`. +- The task relocates: Arteaga et al. on Chilean school choice `[ok]`. +- The missing experiment `[D9]`: two human parties and a mediator that reads both; one arm fixes the + mediator's rule, one adds a reading aid, one gives a delegate; novices and experienced users sampled + separately; outcomes are joint surplus and each party's regret. + +### 8. What Decision Experience Design is, on this footing (4%) + +- Its object: conventions that carry the work of coordination through an algorithmic third party. +- Its expertise: Part 1, line 150 ("UX designers did not become vision scientists…"). +- Its constraint: conventions the mediator's owner cannot withdraw alone `[D8]`. + +### 9. Limits (2%) + +- A historical case and a set of experiments from neighbouring settings; no test in a mediated triad. +- The history rests on Saenger's chronology and a disputed mechanism. +- The post was days old at the time of writing. +- The gap statement, in the only form the searches support: "I found no work that presents the move + from continuous script to word separation as a design precedent for AI interfaces, algorithmic + intermediaries or human-factors practice in [the sources and queries listed in + `literature/facets/Z_gap_search.md`]." Cite Dembeck (2023) as the nearest precedent, and Illich, Mak + and Drucker for the page as an interface `[lib]`. + +## (d) Course-concept crosswalk + +| Course heading | Section | Sources | +| --- | --- | --- | +| Perception | 3, first part | Winskel et al. 2009; Bai et al. 2008; Kohsom and Gobet 1997; Rayner et al. 1998; Perea and Acha 2009; Larkin and Simon 1987 | +| Cognition | 3, second part; 5 | Zhang and Norman 1994; Hutchins 1995; Scaife and Rogers 1996; Risko and Gilbert 2016 | +| Human factors and design | 3, third part; 6; 7 | Bainbridge 1983; Vicente and Rasmussen 1992; Saltzer and Schroeder 1975; Li 2017; Kling et al. 2012 | + +## (e) Flags on the author's existing texts + +Each gives the correct value and its source. Nothing has been changed. + +| Where | As written | What the source says | +| --- | --- | --- | +| `dial_response_algorithmacy/draft.md:339` | Ehsan et al. (2024) ran a process "that measurably improved their sense of agency over it" | The paper reports participant comments and no agency measure (facet H, (vi)) | +| `dial_response_algorithmacy/draft.md:337` | Inman and Ribes (2019) on what the word-separation precedent redistributes | Their paper is on seamful design; nobody has confirmed it mentions the precedent | +| Part 1, line 138 | Moon et al. (2025): literacy "did not benefit users unless both explainability and direct control mechanisms were present simultaneously" | Their abstract says literacy mattered when at least one feature was present, not only when both were (facet H, (vi)) | +| Part 1, line 118 | "Traditional media hold still" | Street (2003, p. 77): literacy is "a social practice, not simply a technical and neutral skill" | +| Part 2 facet E | "the ancient world had no desire to make reading easier and swifter", as a quotation | Saenger's words: "did not possess the desire, characteristic of the modern age, to make reading easier and swifter" (p. 11) | +| The three Gemini essays | N400 as an fMRI result; word superiority effect credited to Cattell (1886) and to lowercase against capitals | The N400 is an event-related potential (Kutas and Federmeier 2011). Reicher (1969) and Wheeler (1970) established the effect; it is not about letter case | + +## (f) Sources to fetch before drafting, in order + +1. Saenger, *Space Between Words* (1997): a print copy, to confirm page numbers and read chapters 1–4 + and 14 whole. +2. Parkes, *Pause and Effect* (1992), chapters 1–3. +3. Vatri (2012), *Classical Quarterly* 62; Johnson (2000), *American Journal of Philology* 121. +4. Dekker and Woods (2002), "MABA-MABA or abracadabra?"; Norman (1991), "Cognitive artifacts"; Zhang and + Norman (1995), on numeration systems. +5. Rayner, Fischer and Pollatsek (1998); Perea and Acha (2009); Shen et al. (2012); Blythe et al. (2012). +6. Wingo (1972), *Latin Punctuation in the Classical Age*; Gavrilov (1997); Burnyeat (1997). +7. Peters et al. (2006, 2007), for the numeracy parallel. +8. Illich (1993), *In the Vineyard of the Text*; Mak (2011), *How the Page Matters*. + +A text of Saenger's book that an agent downloaded from an unrestricted archive.org copy of doubtful +provenance sits in `~/iit-playground/dxd_foundation_local/pdfs/A/`, outside the repo. The audit did not +rely on it for confirmation. Keep or delete it as you see fit. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_README.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_README.md new file mode 100644 index 00000000..dd87314d --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_README.md @@ -0,0 +1,68 @@ +# The foundation paper for Decision Experience Design + +**Status (2026-10-13):** researched in three rounds, audited, carded and outlined. The author ruled on the +ten decisions on 2026-10-13 and then asked Claude for an abstract, an introduction and a section outline: +Claude drafted one, the author replaced it with a Gemini-written version in a Google Doc, and +[`DRAFT.md`](DRAFT.md) is a snapshot of that doc with Claude's edits and citation bullets marked. The +author then supplied a full draft; [`REVIEW_v2_2026-10-13.md`](REVIEW_v2_2026-10-13.md) reviews it and +[`DRAFT_v2.md`](DRAFT_v2.md) is the repaired version, with a rebuilt reference list and a table of every +change. [`DRAFT_v3.md`](DRAFT_v3.md) (2026-10-14) applies the claim check in `literature/audit/claims/` +and reties the essay to one argument in three moves. Version 2 is left in place so those line numbers still match. The target journal is *Human–Computer Interaction* ([`venue/`](venue/)). Claude wrote every file in this arm, and +none of the wording is the author's unless quoted. + +**Where the research stands.** [`literature/facets/T_adjudication.md`](literature/facets/T_adjudication.md) +§9 and §10 are the current statement, and [`ARGUMENT_MAP.md`](ARGUMENT_MAP.md) follows it. In brief: readers used +to spaces lose a third to a half of their speed without them, and that dependence is the robust fact; +no source credits word separation with spreading literacy, and the nearest support for a shortened +training is one sentence of Saenger's (1997, p. 85). Word separation re-represented the page without +disclosing anything, which is what the paper needs from it. The author's rulings oppose *disclosure* as +a literacy-based intervention and move the engagement layer from reading and rewriting to deciding. The +evidence on disclosure is mixed in direction; no convention has been tested in a triad; and in courts, +dealer foreign exchange and platforms the third party yielded a convention only under outside force. +Nadella's insider post offers a person a record to read and a brake, and is silent on the people the +model acts upon. + +A paper for PHD1750-3 / HF700-1, Foundations in Human Factors (Bentley, Fall 2026). It replaces the DIAL +response ([`../dial_response_algorithmacy/`](../dial_response_algorithmacy/)) as the course paper and is +meant to found the Decision Experience Design series +([`../decision_experience_design/`](../decision_experience_design/)). + +**The argument under test.** Scribes moved from *scriptio continua* to spaces between words and so took +the work of finding words off the reader. Algorithmacy needs a move of the same kind. Satya Nadella's +post of 10 October 2026, "Models as Insider Risks in the Super Intelligence Era", proposes the opposite: +observation and containment by the organisation. See [`CLAIM.md`](CLAIM.md). + +## Files + +| File | Use | +| --- | --- | +| [`SOURCE.md`](SOURCE.md) | The author's instruction, answers and ten rulings, the course facts, the Nadella post record, the author's own sentences the paper keeps | +| [`CLAIM.md`](CLAIM.md) | The provisional claim, the two hypotheses on the history, and what not to claim | +| [`RESEARCH_PLAN.md`](RESEARCH_PLAN.md) | Terms, standing rulings, the four problems, the ten facets, the facet file format and rules | +| [`literature/REUSE.md`](literature/REUSE.md) | What the lab's other arms already hold, with recorded read depths | +| `literature/facets/` | Ten research facets (A–H, X, Z), three from the second round (I, J, K) and five from the third (L–P), one `.md` and one `.bib` each, then [`T_adjudication.md`](literature/facets/T_adjudication.md) | +| `literature/references.bib` | Built by `python3 literature/merge_bibs.py` from the facet bibs (first facet wins) | +| `literature/audit/` | `AUDIT_01`–`11.md` (every quotation and number in the facet syntheses, re-checked by independent agents), `CARD_FACET_CHECKS.md` (where a full reading disagreed with a facet), `DOI_CHECK.md` (written by `check_dois.py`), `BIB_CHANGES.md` | +| [`literature/library/`](literature/library/INDEX.md) | The contextual bibliography: 251 cards, one per source read from full text, each re-checked by a second agent. [`INDEX.md`](literature/library/INDEX.md) lists them by facet; [`README.md`](literature/library/README.md) gives the card format | +| [`ARGUMENT_MAP.md`](ARGUMENT_MAP.md) | Twelve premises to the conclusion, each with its source, depth mark and weakest point | +| [`OUTLINE.md`](OUTLINE.md) | Decisions, terms, section plan, course crosswalk, flags on existing texts, sources to fetch | + +**How it was made.** Ten Sonnet agents ran one facet each on 2026-10-12; three more audited the +syntheses and three applied the audits' factual fixes. On 2026-10-13, after the author's rulings, three +agents ran facets I–K and three audited them. An open-access chain then fetched 134 full texts, agents +found 117 more by hand from legitimate open copies, and one agent read and carded each source; a second +agent re-opened every source and corrected 151 of the 251 cards. Claude (Opus) led, adjudicated and wrote +the argument map and outline. No paywall, bot wall or captcha was bypassed, and no full text is in this +repo. Monographs were read as snippets or not at all. + +## Open for the author + +- The points his rulings left open: last column of [`OUTLINE.md`](OUTLINE.md) (a), and + `T_adjudication.md` §9.3 (a 1995 printed use of "algorithmacy"; a seventh flag on Part 1). +- The course paper's due date, length and rubric, and whether the change of topic from the approved + DIAL proposal needs the instructor's agreement. +- The library list in [`OUTLINE.md`](OUTLINE.md) (f): Saenger and Parkes first. About 40 further + sources have open copies behind bot walls that a person can open in a browser + (`T_adjudication.md` §9.5). + +This arm borrows lab constructs. It does not re-run Φ. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_RESEARCH_PLAN.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_RESEARCH_PLAN.md new file mode 100644 index 00000000..b223532c --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_RESEARCH_PLAN.md @@ -0,0 +1,190 @@ +# Research plan — the foundation paper for Decision Experience Design + +*Written by Claude on 2026-10-12 from the author's instruction of the same day and his three answers +(see `SOURCE.md`). The author approved the plan; none of the wording below is his unless quoted.* + +Every facet agent reads this file first. It states the thesis under test, the constraints, the facet +questions, the file format and the tool rules. + +## 1. What the paper is + +A paper for PHD1750-3 / HF700-1, Foundations in Human Factors (Bentley, Fall 2026), by Roger Hunt. It +replaces the DIAL response (`../dial_response_algorithmacy/`) as the course paper and is meant to found +Decision Experience Design (DXD; `../decision_experience_design/`). + +The author's argument, in his words (2026-10-12): "We need to be reviewing the hisotry of the critical UX +move converting scriptura continua to spaces between words which ignited literacy. we will need a +similar move here." He contrasts this with Satya Nadella's post of 10 Oct 2026, "Models as Insider Risks +in the Super Intelligence Era", which he reads as "the exact opposite of the soltion and what +algorthmacy proposes". + +His earlier statement of the same idea, which the paper keeps verbatim +(`../dial_response_algorithmacy/draft.md:263`): "Word separation did not train the populace to read +*scriptio continua*; it changed the interface so that reading no longer required the training." + +**This run produces research and an outline. Nobody drafts the paper; the author writes it.** + +## 2. Terms + +- **Algorithmacy.** The author's current wording (`../algorithmacy_ot/manuscript/PAPER.md:19-21`): "the + sensibility to navigate opaque triadic coordinative cooptation", with three competencies: interpret a + variable counterpart's intentions from an otherwise innocuous output; specify intent so meaning + survives the opaque, adaptive mediator; keep track as governance shifts. The triad is two human + parties and an algorithmic third party that reads both and commits determinations neither controls. +- **Decision Experience Design.** Published Part 1 + (`../decision_experience_design/published/2026-09-30_decision-experience-design.md`, lines 22, 34, + 142, 150): design for choices brokered through an adaptive, opaque algorithmic intermediary; its + expertise is structural intervention (counter-delegation, structural refusal, bounded operating + envelopes), not turning users into decoders. +- **Engagement layer.** `../dial_response_algorithmacy/corpus/ENGAGEMENT_LAYER.md:9`: the layer at + which people meet a system that is still being made, where steering is the ordinary mode of use. +- **Scriptio continua.** Writing without word boundaries. Both spellings are in scholarly use: Saenger + writes *scriptura continua* (1997, p. 69), as the author does; Parkes, Heilmann and most classicists + write *scriptio continua*. *(Corrected 2026-10-12: an earlier version of this line called the + author's spelling non-standard.)* + +## 3. Standing rulings that bind the research + +| Ruling | Locator | +| --- | --- | +| Algorithmacy is not legibility. DXD designs for counter-delegation, structural refusal and bounded, predictable outputs, "not for a user who reads the algorithm more skilfully" | `../decision_experience_design/README.md:38-43` | +| The numeracy framing, not Ong's. No claim that algorithmic mediation restructures cognition as Ong said writing did; Street's critique stands | `../decision_experience_design/README.md:45-52` | +| The history of reading is an analogy, not a mechanism (Part 2) | `../decision_experience_design/part2_ethics/README.md:8` | +| A source is quotable only at full-text depth | `../algorithmacy_ot/README.md:29-41` | +| "Coordinative co-optation" is the author's coinage; do not attribute it to Stark and Vanden Broeck | `../algorithmacy_ot/manuscript/FLAGS_2026-10-07.md:21` | + +**Author's answer of 2026-10-12 on the history.** "Let the research decide." The third ruling is +therefore open for this paper: the research carries two hypotheses and the author rules afterwards. + +- **H-cost (design precedent).** Word separation moved parsing work from the reader to the page. It + lowered the cost of reading and widened the group who could read without long training. It did not + create silent reading and did not by itself spread literacy. +- **H-cause (ignition).** Word separation was a cause of the spread of silent reading and, through it, + of wider literacy. + +**Author's answer on Nadella.** The paper opposes his post on three counts: +1. *Legibility paradigm.* He answers opacity with more reading: observe, log, audit. +2. *Keeps the dyad.* He treats the model as a contained insider under the organisation's authority. +3. *Containment versus engagement.* He separates the supply of intelligence from authority over it; + algorithmacy proposes an engagement layer where the user reads and rewrites the shared mediator. + +## 4. Four problems the research must settle + +1. **"Ignited literacy" outruns the evidence in hand.** `../decision_experience_design/part2_ethics/ + literature/facets/E_reading_history_meaning.md` concludes that spacing "lowered a cost" and "widened + the group". Readers of unspaced scripts read fluently. Continental adoption ran to the late tenth + century; mass literacy came much later. +2. **Word spacing is itself a legibility remedy** (Parkes's "grammar of legibility"), yet the first + ruling rejects legibility and the paper opposes Nadella as the legibility paradigm. Candidate + distinction: spacing moved work from reader to artefact; observability asks the human to read more. + Does that distinction hold in the human-factors literature? +3. **Nadella is not a clean opposite.** He says chain-of-thought transparency "alone is not sufficient + or dependable", puts controls outside the model (close to "bounded outputs" and to structural + refusal), and treats the model as an actor. Steelman, then convergences, then divergences. +4. **"One similar move" may be the wrong shape.** Parkes describes a complex of conventions over + centuries. Part 1 line 118 says algorithmacy engages a coordination, not a medium. + +## 5. Facets + +| Facet | File stem | Question | +| --- | --- | --- | +| A | `A_the_record` | What changed on the page, where, when, by whom and why: Saenger 1997 and 1982, Parkes 1992, Wingo 1972, the Insular scribes, Carolingian and later diffusion | +| B | `B_the_dispute` | Knox, Gavrilov, Burnyeat, Johnson, Vatri, McCutcheon and the published reviews of Saenger. What of his thesis survives | +| C | `C_spacing_to_literacy` | Evidence linking page conventions to who could read: Clanchy, Stock, Graff, Houston, Eisenstein, Street, Havelock. Tests H-cause directly | +| D | `D_reading_science` | Eye-movement and unspaced-script studies (Rayner, Mirault, Bai on Chinese, Thai and Japanese studies), crowding, the visual word form area | +| E | `E_human_factors_frame` | Spacing as reallocation of work from operator to artefact: function allocation, cognitive artefacts, the representational effect, cognitive offloading, ecological interface design. Maps onto the course's perception, cognition and design headings | +| F | `F_parallel_moves` | Other single conventions credited with opening a competence: place-value numerals and zero, staff notation, the index, double-entry, direct manipulation | +| G | `G_nadella_lineage` | The post; the reference monitor and Saltzer and Schroeder; insider-threat and zero-trust practice; AI control and chain-of-thought faithfulness; what the oversight literature says about humans monitoring logs | +| H | `H_candidate_moves` | What could be the space between words for algorithmacy, limited to interventions with measured effects, in settings with two humans and a mediator | +| X | `X_counter_case` | The strongest case against the thesis | +| Z | `Z_gap_search` | Whether anyone has drawn the scriptio continua analogy for AI or HCI; current uses of "decision experience design" | + +The lead writes `T_adjudication.md` after the facets and the audit. + +### Second round (2026-10-13), from the author's rulings in `SOURCE.md` + +| Facet | File stem | Question | +| --- | --- | --- | +| I | `I_understanding_through_action` | The author's ruling D7: "we understand THROUGH deciding and quickly ingesting the results. pilots dont spend time reading the situation, they act and through the action the gauge and adjust in the moment." What does the human-factors and cognitive-science literature say about understanding gained by acting: manual control and tracking, skill-based behaviour (Rasmussen), epistemic action (Kirsh and Maglio), recognition-primed decision (Klein), direct perception and ecological interface design, sensemaking by enactment (Weick)? How much of a pilot's task is monitoring (Endsley; Sarter and Woods; instrument-scan studies)? What must feedback be like for learning by acting to work (Kahneman and Klein 2009; Hogarth's kind and wicked environments), and what do studies of platform users (drivers, sellers, creators) show about learning by trial under delayed, noisy or personalised feedback? | +| J | `J_who_holds_the_convention` | The author's ruling D8: "we can look at how courts and forex markets work, compare that platforms (which i think take a highly draconian position), and look for where the convention bends towards". In each of three settings where a third party stands between two parties (courts and arbitration; foreign-exchange and other dealer or clearing markets; digital platforms), who writes the procedural conventions, who can change them, what notice and appeal the parties get, and what the evidence says about outcomes: procedural justice (Thibaut and Walker; Tyler; Lind), repeat players (Galanter 1974), the FX Global Code, the WM/Reuters fix and last-look cases, CLS settlement, platform terms of service, deactivation, appeals, the EU Platform-to-Business Regulation and Digital Services Act. Test "draconian" against the record; report where it fails | +| K | `K_training_and_cause` | The author's rulings D1 and D2: review the historical analysis for cause, and review the debate on whether word separation changed the training reading required. How were readers taught before and after separation (Quintilian; Dionysius of Halicarnassus; Cribiore; Marrou; Bonner; Reynolds, *Medieval Reading*; Orme; Black; Grendler; Saenger's chapters on novices and the syllabary method)? Does any scholar claim, test or deny that separation shortened or eased learning to read? Does any scholar make a causal claim from separation to anything (silent reading, reference reading, scholasticism, private devotion, authorship), and how have reviewers received each claim? Start from facets A, B and C and do not repeat them | + +Facets I, J and K follow the same format and rules (§7, §8). + +### Third round (2026-10-13), from the author's outline + +The author adopted a nine-section outline (the Google Doc "Human Factors Paper"; sections: 2 the scribal +precedent, 3 the human factors of re-representation, 4 the failure of disclosure, 5 from literacy to +algorithmacy, 6 Decision Experience Design conventions, 7 institutional custody, 8 testing the +irreducible triad, 9 limits) and asked for "deep research to identify any other aspects of that +literature that we should include". Each facet below looks for literatures the first two rounds did not +cover. Read `literature/library/INDEX.md` and the earlier facets first and do not repeat them. + +| Facet | File stem | Outline sections | Literatures to search | +| --- | --- | --- | --- | +| L | `L_disclosure_and_choice` | 4 | The law-and-economics and behavioural critique of mandated disclosure (Ben-Shahar and Schneider; Loewenstein, Sunstein and Golman; Bar-Gill; Fung, Graham and Weil on targeted transparency); privacy notices and consent; nudging against boosting (Hertwig and Grüne-Yanoff) and choice architecture as the nearest rival to "design, don't inform"; meta-analyses and reviews of explainable-AI human-subject studies and of human–AI decision performance; any disclosure study in a two-sided or mediated setting with a behavioural outcome | +| M | `M_acting_on_algorithms` | 5 | Strategic classification and performative prediction in machine learning; algorithmic recourse; folk theories, algorithmic imaginaries, algorithmic gossip and "gaming the algorithm" among creators, sellers and workers; algoactivism and resistance (Kellogg, Valentine and Christin; Cameron and Rahman); whether platforms keep rules opaque or noisy on purpose (Rahman's "invisible cage"; obfuscation and shrouded attributes in economics, Ellison and Ellison, Gabaix and Laibson; sludge and dark patterns); experimentation and learning by doing on platforms | +| N | `N_feedback_and_presentation` | 3, 6 | Interventions that change feedback timing or form with a behavioural outcome (real-time feedback field experiments, feedforward, previews, upfront pricing and information shown to gig workers); experiments on how a strategy-proof or simple mechanism is presented (menu descriptions; "the mechanism is truthful, why aren't you?"; obvious strategy-proofness follow-ups); display-design principles the outline has not used (proximity compatibility, stimulus–response compatibility, affordances and signifiers, gulfs of execution and evaluation, cognitive load theory) | +| O | `O_custody_of_conventions` | 2, 7 | Who has held interface and procedural conventions and how they were enforced: the Carolingian script reform and monastic and court standardisation of writing; standards bodies for technical and interface conventions (IETF, W3C, accessibility standards, the browser padlock); platform governance scholarship; information fiduciaries; collective bargaining and statute on algorithmic management (EU Platform Work Directive, Spain's Rider Law, co-determination); commons and polycentric governance (Ostrom); worker data collectives and audits | +| P | `P_triad_experiments` | 8 | Laboratory and field experiments with two human parties and an algorithmic intermediary that has its own objective (experimental markets with intermediaries, platform experiments, human–AI-mediated communication and negotiation); integrated information, synergy and partial information decomposition applied to groups, teams or social systems; design and preregistration standards for such experiments | + +Facets L–P follow the same format and rules (§7, §8), with three additions. Card-level reading is the +aim: open the full text wherever an open copy exists and say which pages you read. End the synthesis +with a section "What the outline should add", keyed to the outline's section numbers, giving for each +proposed addition the claim, the two or three sources that carry it, and their read status. Never send +any e-mail address to any service (so no Unpaywall); use OpenAlex for open-access locations. + +## 6. What the repo already holds + +Start from `literature/REUSE.md`. Each row there is a lead: reopen the source and record your own read +status before relying on it. The Gemini essays in `../dial_response_algorithmacy/sources/` and +`../triadic_reduction/sources/G7_tablets_to_algorithms.md` are leads only. They carry no checkable +citations, at least one error (N400 reported as an fMRI result) and an "HIAG framework" nobody has +confirmed. Never cite them and never repeat a figure from them without finding its source. + +## 7. Facet file format + +Two files per facet in `literature/facets/`: `.md` and `.bib`. Model: +`../decision_experience_design/part2_ethics/literature/facets/E_reading_history_meaning.md`. + +`.md` has, in order: +1. A header line: facet, date, "Agent-drafted. The author has not reviewed it, and none of the text is + the author's." +2. **What this facet covers** and **the question the synthesis answers**. +3. **Legend for read status**: *read in full*, *read in part* (say which pages), *abstract only*, + *not accessible*. +4. **(a) Search log**: a numbered table of source, exact query, hits, notes. Log failures. +5. **(b) Entries**, one per source: full reference, bib key, how you verified it exists, read status, + what it argues, quotations with page or section locators, and its bearing on H-cost, H-cause or the + four problems (supports, complicates, refutes). +6. **(c) Synthesis** by question, ending with a list of claims the sources contradict. +7. **(d) Gaps and unverified leads**, including every source you could not open and where a human with + library access would find it. + +`.bib`: one entry per source, key in the form `surnameYEARfirstword` (reuse an existing key when +the source is already in a repo bib), with a DOI, arXiv ID or stable URL, and a `note` giving read +status. + +## 8. Rules + +- Report what the sources say, including when they sink the thesis. A facet that confirms everything is + suspect. +- Quote exactly, with a locator. Paraphrase anything you read only as an abstract and say so. +- Never invent a DOI, a page number, a quotation or a statistic. Write "not found" instead. +- Resolve every DOI against Crossref (`https://api.crossref.org/works/`) and compare the title. +- Write only inside `submissions/dxd_foundation/literature/facets/`, and only your own two files. Do + not edit any other arm. Do not run git. +- Do not use the Chrome browser tools; several agents run at once and would collide. List sources that + need a browser in section (d) and the lead will open them. +- Do not try to get around a paywall or a bot wall. Record the block. +- Tools that have worked: Crossref, OpenAlex (`https://api.openalex.org/works?search=`), Unpaywall + (`https://api.unpaywall.org/v2/?email=` with the address in `org_frontier/research/ + oa_acquire.py`), Europe PMC, arXiv, HAL, author copies by `curl` and `pdftotext`, archive.org + (including its full-text search and borrowable books), Open Library, Perseus, Consensus (at most + three calls at a time; on a rate-limit error wait 30 seconds), WebSearch and WebFetch. +- Scholar Gateway failed in every earlier run. Try it once and log the result. +- Put downloaded PDFs under `~/iit-playground/dxd_foundation_local/pdfs//`, never in + the repo, which is public. +- The Nadella post is captured at `~/iit-playground/dxd_foundation_local/ + nadella_2026-10-10_models_as_insider_risks.txt`, with paragraph numbers. Quote from it sparingly (one + or two sentences at a time) and cite the paragraph number. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_REVIEW_v2_2026-10-13.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_REVIEW_v2_2026-10-13.md new file mode 100644 index 00000000..9fd11acb --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_REVIEW_v2_2026-10-13.md @@ -0,0 +1,208 @@ +# Review of the full draft received 2026-10-13 ("new version") + +*Written by Claude on 2026-10-13 at the author's request ("i need a full review"). One reader, one pass. +Every factual check below was made against a source record this project holds: the verified card +library (`literature/library/`), the audits (`literature/audit/`), the merged bibliography +(`literature/references.bib`, where each DOI was resolved at Crossref), and the journal pages opened on +2026-10-13 (`venue/HCI_STRUCTURE.md`). Where the project holds no record, the item is marked "not +checked". Nothing in the draft was changed.* + +## Verdict + +The architecture is right for *Human–Computer Interaction* and the argument is stronger than the last +version. The draft cannot be submitted as it stands, for one reason above all: **the reference list is +unreliable**. At least 39 of about 105 entries do not match the published record, and seven statements +in the text reverse or misstate what their sources found. A reviewer who checks three references will +find one that does not exist as cited. + +## 1. What works + +- The introduction is continuous prose, states its contribution outright, and runs in the order the + journal's recent papers use. +- Section 2 now reviews HCI's own transparency literature and answers Özçetin et al. (2023) directly. + That reply (2.2, last paragraph) is the best passage in the draft. +- The four-target taxonomy of disclosure (4.1) resolves the tension the research kept raising: a stated + property of a rule (Target 3) is a different thing from a rulebook (Target 4). +- Convention B is restated as "valid and noise-filtered first, plain second, and low-latency third". +- The unilateral-modification test (7.1) gives section 7 an argument of its own. +- Section 8 is a research agenda with a discrete-time control arm, and 9.2 says which claims are the + paper's own. +- Each convention ends in a "Design opening", the form the journal's own papers use. + +## 2. The reference list + +### 2a. Entries that do not match the record (39) + +| Entry in the draft | What the record shows | +| --- | --- | +| Hutchins (1995), *Cognition in the Wild* | The quotations at p. 282 are from "How a cockpit remembers its speeds", *Cognitive Science* 19(3), 265–288 | +| Jhaver, Bruckman, & Gilbert (2019), "Does transparency in moderation improve behavior?" | The finding cited is from Jhaver, Appling, Gilbert, & Bruckman (2019), "'Did you suspect the post would be removed?'", *PACM HCI* 3(CSCW), Article 192 | +| Rayner, Fischer, & Pollatsek (1998), "Punctuation and space between words", *Perception & Psychophysics* | "Unspaced text interferes with both word identification and eye movement control", *Vision Research* 38(8), 1129–1144 | +| Vaccaro et al. (2024), "Explanations do not improve…", pp. 1102–1118 | "When combinations of humans and AI are useful: A systematic review and meta-analysis", *Nature Human Behaviour* 8, 2293–2303 | +| Schemmer, Hemmer, Kühl, Benz, & Satzger (2022), *PACM HCI* | Schemmer, Hemmer, Nitsche, Kühl, & Vössing, "A meta-analysis of the utility of explainable artificial intelligence in human-AI decision-making", AIES 2022, 617–626 | +| Gonczarowski et al. (2024), "Menu representations…", *AER* 114(6) | "Describing deferred acceptance and strategyproofness to participants: Experimental analysis"; held as a working paper | +| Masuda, Okano, & Stern (2022), *Games and Economic Behavior* | Masuda, Mikami, Sakai, Serizawa, & Wakayama, "The net effect of advice on strategy-proof mechanisms: an experiment for the Vickrey auction", *Experimental Economics* 25, 902–941 | +| Larson, List, & Metcalfe (2016), *Journal of Finance* 71(5) | "Can myopic loss aversion explain the equity premium puzzle? Evidence from a natural field experiment with professional traders"; an NBER working paper | +| Rahmandad & Gary (2023), "Why human learning fails…", 34(1), 12–35 | "Delays impair learning and can drive convergence to inefficient strategies", *Organization Science* 34(6), 2392–2414 | +| Rahman (2019), *The Invisible Leash* | *Invisible Cages: Algorithmic Evaluations in Online Labor Markets* (Stanford dissertation) | +| Cameron & Rahman (2022), *Academy of Management Annals* 16(1) | *Organization Science* 33, 38–58; the subtitle is "…the co-constitution of control and resistance in the gig economy" | +| Ghalme, Nair, Eilat, Ingel, & Rosenfeld (2021) | Fourth author is Talgam-Cohen | +| Bechavod, Ligett, Wu, & Ziani (2022), "Information design for strategic classification", NeurIPS | Bechavod, Podimata, Wu, & Ziani, "Information discrepancy in strategic learning"; held as an arXiv preprint | +| Engel & Malone (2018), "…metric of collective intelligence", *Cognitive Science* | "Integrated information as a metric for group interaction", *PLOS ONE* 13, e0205335 | +| Barrett, Mediano, Rosas, & Seth (2026), *Neural Computation* | Barrett, Milinkovic, Mediano, Rosas, Bor, & Barnett, "Integrated information theory: the good, the bad and the misunderstood"; a preprint | +| Casella et al. (2024), "Mediated bargaining", *JET* 215 | "On the fragility of mediation: Theory and experimental evidence", *JEEA* 24, 801–850 (2026); read as a 2024 working paper | +| Charness, Samek, & van de Ven (2022), "The design of laboratory experiments…", 1–29 | "What is considered deception in experimental economics?", *Experimental Economics* 25, 385–412 | +| Matsuo, Tanaka, & Yamada (2026), *Human–Computer Interaction* | Matsuo, Liu, Hiraoka, & Wada, "An educational human machine interface providing request-to-intervene trigger and reason explanation…", *IEEE Transactions on Human-Machine Systems* 56(3), 650–661 | +| Crabtree (2025), "Human-computer interaction as a design science" | "H is for human and how (not) to evaluate qualitative research in HCI"; posted online, no volume in Crossref | +| Disley & Khan (2026), 41(2), 112–148 | Posted online September 2026 as pp. 1–42 with no volume; co-authors not checked | +| Dix (2024), 39(3), 205–244 | 41(2), 126–153 | +| Özçetin et al. (2023), 38(5–6), 389–422 | 40(1–4), 171–194 | +| Suzor (2018), *Lawless* | The quotation is from "Digital constitutionalism: Using the rule of law to evaluate the legitimacy of governance by platforms", *Social Media + Society* 4(3) | +| Nadella (2026), "Governing frontier models as insider risks… Microsoft Executive Research Whitepaper" | "Models as insider risks in the super intelligence era", a post on X and on his blog, 10 October 2026. No whitepaper | +| Oberschelp (1995), "Algorithmacy: Der Stiefbruder der Literacy…", *Informatik Spektrum* | "Developing analytical competency through informatics teaching", in Watson & Tinsley (Eds.), *Integrating Information Technology into Education*, pp. 115–126; the word is on p. 121 | +| Westwell (2018), *Early Medieval Europe* 26(3) | *The Dissemination and Reception of the Ordines Romani in the Carolingian Church, c. 750–900* (Cambridge dissertation) | +| Chantavarin & Leung (2025), *Journal of Memory and Language* | "The effects of interword spacing and morphological complexity in reading Thai: An eye-tracking study", *Language and Cognition* 17 | +| Winskel et al. (2009), *Vision Research* 49(24) | "Eye movements when reading spaced and unspaced Thai and English…", *Journal of Memory and Language* 61(3), 339–351 | +| Cabrera et al. (2023), "…with behavioral feedback", *TOCHI* | "Improving human-AI collaboration with descriptions of AI behavior", *PACM HCI* 7(CSCW1) | +| Alfrink et al. (2023), *TOCHI* 30(4) | *Minds and Machines* 33, 613–639 | +| Vasconcelos et al. (2023), "Generation mutualism…", CHI | "Explanations can reduce overreliance on AI systems during decision-making", *PACM HCI* 7(CSCW1), Article 129; last author Krishna | +| Douek (2024), "The limits of platform constitutionalism", *Harvard Law Review* | "The Meta Oversight Board and the empty promise of legitimacy", *Harvard Journal of Law & Technology* 37(2), 373–445 | +| Brehmer & Nählinder (2004), *Human Factors* 46(1) | "Achieving what cannot be done: Coping with the time constants in a dynamic decision task", conference proceedings | +| Bar-Gill et al. (2024), "Algorithmic recommendations and merchant performance" | "Helping small businesses become more data-driven: A field experiment on eBay", *Management Science* 70, 7345–7372 | +| Rees-Jones & Skowronek (2018), *JET* 178 | "An experimental investigation of preference misrepresentation in the residency match", *PNAS* 115, 11471–11476 | +| Ashlagi & Gonczarowski (2018), 564–602 | "Stable matching mechanisms are not obviously strategy-proof", *JET* 177, 405–425 | +| Heilmann, M. (2020), *Classical Philology* 115(2) | Heilmann, J., "Reading early New Testament manuscripts: scriptio continua, 'reading aids', and other characteristic features", in *Material Aspects of Reading in Ancient and Medieval Cultures*, pp. 177–196 | +| Nouwens, Liccardi, Veale, Karger, & Mackay (2020) | Fifth author is Kagal | +| Arteaga, Kapor, & Neilson (2022) | Fourth author Zimmerman; "…in centralized school choice" | + +### 2b. Not checked + +The project holds no record for these, so they are unverified: Bansal et al. (2021), Ehsan et al. +(2021), Kosmyna et al. (2025), Körber (2019), Wischnewski et al. (2023), Wolfe (2021), Mozannar et al. +(2024), Casner et al. (2014), Metzger & Parasuraman (2001), Herlocker et al. (2000), Lee & Moray (1994), +Eisenstein (1969), Thorgeirsson et al. (2026), Terveen & McDonald (2005), and the two entries dated +**2027** (Pucci, Piro, & Matera; Sáenz & De Russis). Given the rate in 2a, each needs checking against +Crossref before it stays. The 2027 dates are in the future. + +### 2c. Cited in the text with no entry in the list + +Knijnenburg et al. (2012); Simkute et al. (2024); Lee et al. (2025); Horvath et al. (2022); Sheridan & +Verplank (1978); Ishowo-Oloko et al. (2019); Lyons et al. (2021); Kluger & DeNisi (1998); Mertens et al. +(2022); Street (2003); Scribner & Cole (1981); FCA (2014); FSB (2014); NYDFS (2015); Oversight Board +(2022); CERT (2022); RFC 8890; the EU regulations and directive other than 2024/2831; the two cases and +the statutes. + +### 2d. In the list and never cited + +Arteaga et al.; Ashlagi & Gonczarowski; Balkin; Bar-Gill et al.; Bechavod et al.; Brehmer & Nählinder; +Crabtree; Dietvorst et al.; Dix; Douek; Hardt et al.; Lee & Moray; Rees-Jones & Skowronek; Scaife & +Rogers; Thorgeirsson et al.; Vicente & Rasmussen; Wickens & Carswell. + +## 3. Statements that reverse or misstate their sources + +| # | Where | The draft says | The source says | +| --- | --- | --- | --- | +| 1 | 4.2 | When communities "posted detailed explanations for moderation removals, sanctioned participants reported lower perceived procedural fairness than when removals were unexplained" | The reverse. Users who received an explanation judged the removal fairer (β = .384). Users who had *read the rules* judged it less fair (β = −.314). Survey, not an experiment (Jhaver et al., 2019, Table 5). The introduction has it right | +| 2 | 3.3 | For experienced readers of Chinese, inserting spaces "slows fixation rates and disrupts comprehension (Bai et al., 2008)" | Bai et al. found no difference between word-spaced and normal text (5,239 against 5,221 ms). Character spacing was slower | +| 3 | 3.3 | "For children or non-native learners… spaces provide a helpful scaffold" | Child learners of Chinese read spaced text more slowly and less accurately (Bassetti & Masterson, 2012); the learner contrast in Cui et al. (2023) is not significant | +| 4 | 6.2 | The menu "raised comprehension from 56% to 71%. Sincere play increased from 48% to 59%" | Two different baselines. Understanding: 71% against 58% and 56%. Straightforward play: 59%, against 56% for the traditional description (p = 0.47, no difference) and 48% for a near-zero-information control (Gonczarowski et al., 2024) | +| 5 | 4.2 | Layout "shifted consent rates by orders of magnitude, while changing the textual disclosure language produced no detectable effect" | Removing the opt-out button raised consent by 22 to 23 percentage points (Nouwens et al., 2020). The claim about text was not found | +| 6 | 7.2 | The industry "attempted self-regulation through the FX Global Code… But voluntary soft-law adherence failed… until public financial regulators intervened" (2014–15) | The chronology is reversed. The fines came in 2014–15; the Code followed in 2017. Principles 7, 9, 36, 39 and 40 concern complaints, disclosure and terms, not last look or pre-hedging | +| 7 | 7.3 | The P2B Regulation and the DSA "merely mandate that platforms describe their ranking parameters and terms in plain language… changing nothing about operational mechanics" | P2B Art. 3 requires 15 days' notice of changes to terms. DSA Arts. 17 and 20 require a statement of reasons and an internal complaint system; Arts. 53–54 give a complaint to a regulator and damages | + +Other statements that go beyond their sources: + +- **Seira et al. (4.2, intro).** One bank's field experiment with messages, not "Mexican credit card + companies" under a mandate. Delinquency rose for two client groups by 0.011 points, 7.4% of the mean. + "The disclosure made the debt feel insurmountable" is not in the paper. +- **Poursabzi-Sangdeh et al. (4.2).** "An illusion of competence" is the draft's mechanism. The authors + conjecture information overload, and the effect holds in three of four experiments on one unusual case. +- **Ben-Shahar & Schneider (4.1).** The quotation "the most common and least effective regulatory + technique in modern law" (p. 665) is not the wording the audit found on that page. The "bright caloric + label" is not their example; theirs is restaurant letter grades (p. 743), which they hedge. +- **Vasconcelos et al. (2.1)** is cited for misplaced confidence. The paper shows explanations *reducing* + overreliance in a hard task. **Parasuraman & Manzey** do not use "transparency paradox". +- **Nadella (4.3).** The post has seven principles, not "three pillars". "Halt switches" is in quotation + marks and is not his phrase; he writes of an emergency brake. The people affected are not "handed the + tamper-proof human-readable audit log"; P20 promises them disclosure of incidents. +- **Saenger (3.1).** "Confined to an elite, highly trained scribal caste" and "impossible to parse + phonologically" are not his. Celtic and Old English do share roots with Latin; his point is that the + Irish and English learned Latin as a foreign language. The distinctiones are an ancient system, and + "codex page indexing by the eleventh century" has no source here. The conclusion says "before the + eighth century"; the body says late seventh. +- **Carolingian sources (3.3, 7.3).** None of them mentions word separation, and Insular spacing predates + them. "Held accountable to readers by external sovereign authority" is the draft's reading. +- **Chen, Mislove, & Wilson (5.3).** The bug is Uber's own account; "an engineering synchronization bug + in their geospatial grid" adds detail the paper does not give. +- **Ghalme et al. (5.2)** is a machine-learning theory paper and cannot be cited for gig workers sharing + screenshots. The driver vignette (second phone, "two blocks north") is not in Lee et al. or Kellogg et + al. as described. +- **Oberschelp (5.1)** did not treat algorithmacy as "procedural programming literacy for + schoolchildren". He used the word once, undefined, for mental algorithmic skill without tools. +- **Rahmandad & Gary (5.3).** "Even after hundreds of trials" and "superstitious heuristics" are not + theirs; delay and noise had different effects. +- **RFC 8890 (7.2)** is from the IAB and does not "explicitly declare that internet standards do not bind + proprietary platforms". +- **Alfrink et al. and Lyons et al. (2.3)** are about contestability and do not describe a mediator that + reads both sides and commits a binding allocation. + +## 4. The argument + +1. **The abstract and introduction still condemn all disclosure; section 4.1 does not.** The taxonomy + says Target 4 fails, and 6.2 relies on Target 3 working (Masuda et al.). The thesis sentence should + say which disclosure is the category error. +2. **"HCI remains oriented around dyadic interaction" (2.3) will be contested.** CSCW is the study of + multi-party systems, and AI-mediated communication is an established HCI topic. The project holds + Hohenstein et al. (2023) on exactly this. The claim needs narrowing to transparency and explanation + research, with that literature cited. +3. **The second introduction paragraph says "In none of these settings does the algorithm issue a + determination binding both parties".** Jhaver et al. is a platform setting. The sentence refers to the + two helpful cases and should say so. +4. **Claim 1 is not falsifiable as written.** "Strictly positive integrated information (Φ > 0) that + cannot be factored" is a definition, and the measure proposed is a proxy. Engel and Malone had no + mediator. Barrett et al. say Φ has only ever been computed by proxy. State the claim as a comparison + the data could fail, and say the application to human triads is untested. This also keeps the paper + inside the lab's own rule that irreducibility is explored, not established. +5. **Claim 4 introduces "verifiable cryptographic… lock-in"** with no source and no arm in the protocol. +6. **Scope boundary 2 overstates Convention A.** "Rests on established empirical mechanism design" sits + beside a result in which the best presentation did not beat the traditional one on play. +7. **Scope boundary 3 says nudging "alters choice presentation within an unchanged game".** The nudge and + boost literature is the nearest rival and needs more than one sentence (Hertwig & Grüne-Yanoff, 2017). + Mertens et al. also class re-representation of information with informational interventions. +8. **"Following the model established by Terveen and McDonald (2005)".** Nobody on this project has read + it. +9. **Section 5.3 says platforms "deliberately disrupt ecological validity"**, then correctly separates + secrecy from noise. The opening sentence should match the distinction that follows. + +## 5. Fit to *Human–Computer Interaction* + +- **Length.** By my estimate the body runs to about 6,500 words, roughly half the journal's norm (not counted by script; the draft arrived as pasted text). Its articles average 35 printed pages; the + four read ran 9,000 to 17,000 words. Sections 2, 5 and 7 are thin against their models. +- **Roadmap.** A numbered list with bold labels. The journal's introductions end in prose. +- **Diagrams.** Eight ASCII boxes. They need to be real figures, and three (the dyad, the page shift, the + literacy/algorithmacy contrast) add nothing the prose does not say. +- **Voice.** "I" in the introduction, "we" in sections 3, 7 and 8. Sole authors in this journal write "I". +- **Convention B has three names**: "valid, low-latency feedback" (abstract), "ecological feedback + immediacy" (body), and "Radical" survives nowhere, which is right. Pick one. +- **Evidence.** The journal says "every paper should have empirical evidence that validates its results". + Convention B has none from a triad, and the draft says so only in 9.2. Say it in section 6. +- **Prior posting.** Still to raise with the editor. + +## 6. Prose + +The author's own style rules ban emphasis openers. The draft opens sentences with "Importantly," and +"Notably,"; it uses the "not X; it is Y" frame repeatedly ("The failure is not an empirical anomaly; it +is a category error"); and it ends paragraphs on epigrams ("an audit log is an +autopsy report"). "Zenith", "apex", "shatters", "paranoia", "victim" and "heroic" overstate. The best +paragraphs are the plain ones: the reply to Özçetin et al. and the definition of the custody test. + +## 7. Order of repair + +1. Rebuild the reference list from `literature/references.bib`, which holds the verified record for + nearly every source in 2a. Check the sixteen entries in 2b against Crossref. Reconcile 2c and 2d. +2. Correct the seven reversals in section 3 above, then the overstated attributions. +3. Narrow the thesis to Target 4 disclosure and fix the dyad claim in 2.3. +4. Rewrite Claim 1 and drop or source Claim 4's cryptographic clause. +5. Read Terveen and McDonald (2005) before citing it as the model. +6. Lengthen sections 2, 5 and 7; turn the roadmap into prose; settle "I"; name Convention B once. +7. The author's read-aloud pass. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_SOURCE.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_SOURCE.md new file mode 100644 index 00000000..fb488095 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_SOURCE.md @@ -0,0 +1,98 @@ +# Source record — the foundation paper for Decision Experience Design + +*Compiled by Claude on 2026-10-12. Text inside quotation marks is the author's or a cited source's; the +rest is Claude's.* + +## The author's instruction (2026-10-12, verbatim) + +> "yea the dial response paper. BUT i want to develop this into something that is the foundation for +> decision experience design. We need to be reviewing the hisotry of the critical UX move converting +> scriptura continua to spaces between words which ignited literacy. we will need a similar move here. +> i also want to contrast this view with Satya Nadellas recent tweet, which is the exact opposite of the +> soltion and what algorthmacy proposes. here is his tweet. +> https://x.com/satyanadella/status/2108931348857827686 - lets create a new paper for the "HUman +> factors" course that works this out. review all our content thoruhg the repo so you have a slid +> background and lan out brand new deep research to get evefyhitng we need to make this arguemnt" + +## The author's three answers (2026-10-12) + +| Question | Answer | +| --- | --- | +| How should the paper carry the history, given the earlier rulings (analogy, not mechanism; no Ong-style restructuring)? | "Let the research decide." Both readings stay live; the author rules once the evidence on Saenger and his critics is in. | +| Which reading of the Nadella post does the paper oppose? | Three of the four offered: the legibility paradigm; that it keeps the dyad; containment versus engagement. Not selected: "enterprise, not user". | +| What does this run deliver? | Research and an outline. The author writes the prose. | + +## The author's ten rulings (2026-10-13) + +Claude put the ten decisions of `OUTLINE.md` (a) to the author as questions with options. The second +column is his answer, verbatim where it is in quotation marks. The third column is Claude's reading of +what the answer changes, and the flags in it are Claude's. + +| # | The author's answer | What it changes; flags | +| --- | --- | --- | +| D1 | "Design precedent, and lets review the historical analysis for cause" | The paper argues the precedent. The causal question is reviewed, not asserted: facets B and C hold the review so far, and facet K extends it | +| D2 | "this is a question we rely on other scholars for, we need to reveiw teh debate, if any" | The sentence at `draft.md:263` waits on the scholars. Facets B and C found a debate on whether unspaced script was hard for trained readers, and found nobody who argues that separation shortened the training of readers. Facet K searches the history of reading instruction for that claim | +| D3 | "disclosure of the algorithm doesnt affect our ability to navigate it. disclosure is a literacy based intervention, not an algorithmacy based one. you should dress this up in the lagauge of the ssay" | Replaces "not legibility" with a ruling on disclosure. Claude's draft of the statement is in `OUTLINE.md` (a), marked as Claude's. **Flag:** the evidence in facet H and in Part 1 (lines 136–140) is that disclosure raises awareness and has not been shown to change what people do; "doesn't affect" is stronger than that | +| D4 | "he is arguing aginst neuralese, which is NOT algorithmacy, and his only move is disclosure design, and transparency, which are literacy based interventiones." | The contrast becomes literacy-based against algorithmacy-based intervention. The post does name "Neuralese" (P9). **Flag:** "his only move" does not match the post. Of the seven principles, model diversity (P14), independent controls (P17) and containment (P19) disclose nothing, and P11–P12 put the controls outside the model; P9 itself calls chain-of-thought transparency "not sufficient or dependable" | +| D5 | "Adopt it (Recommended)" | The paper takes up the "enterprise, not user" reading | +| D6 | "its a complex, BUT we are looking for the most critical, if not the critical move. spacing between workds wasnt the ONLY move, but it was certainly a huge move" | A complex of conventions, with a search for the one that does most of the work | +| D7 | "we are going from reading and writing to deciding...we understand THROUGH deciding and quickly ingesting the results. pilots dont spend time reading the situation, they act and through the action the gauge and adjust in the moment. \"Understanding\" is becoming something similar." | The engagement layer is restated as deciding and taking in results, not reading and rewriting. **Flags:** the claim about pilots needs the aviation literature, which treats monitoring as a large part of the pilot's task; and Part 1 (line 66) says an adaptive intermediary disrupts the rapid, unequivocal feedback that learning by acting needs. Facet I tests both | +| D8 | "we just dont know yet. we can look at how courts and forex markets work, compare that platforms (which i think take a highly draconian position), and look for where the convention bends towards" | The paper does not name a holder of the convention. Facet J compares courts, foreign-exchange markets and platforms | +| D9 | "this is an irreducible traid, so different than statistical mediation. we use IIT and phi to gauge the irreducibiltiy of any given triad" | The paper says "irreducible triad" and points to Φ as the gauge. "Mediated triad" in the question was the lab's own term and did not mean statistical mediation. **Open:** the answer does not say whether the paper proposes a behavioural experiment, and this arm's README says it does not re-run Φ | +| D10 | "the spacing intervention is something we are pointing to to explore the transiton from one -acy to another -acy...going from literacy to algorithmacy might be similar, it might not be. but its a guie" | The history is a guide, which agrees with the Part 2 ruling (analogy, not mechanism). **Open:** the answer does not rule on the course statement's novice and expert claim; Claude's reading is that the paper does not carry it | + +## The course + +PHD1750-3, Independent Research Project, mapped to HF700-1, Foundations in Human Factors (Bentley +University, Fall 2026; instructor named Ericson on the proposal branch). The course statement the author +submitted is "Algorithmacy as a Human Factors Problem: Re-coding Digitally Induced Altered States +(DIAS)", snapshot at `../dial_response_algorithmacy/sources/phd1750_topic_and_rationale.md`. Its three +course-concept headings are Perception, Cognition, and Human factors and design. A proposal was due +2026-10-06 (unmerged branch `cursor/phd1750-paper-proposal`). + +**Not recorded anywhere in the repo:** the final paper's due date, length, format or rubric. The author +supplies these. + +This paper replaces the DIAL response as the course paper. The approved topic was the DIAL re-coding; +whether the change of topic needs the instructor's agreement is the author's call. + +## The Nadella post + +- **Reference.** Nadella, S. [@satyanadella]. (2026, October 10). *Models as insider risks in the super + intelligence era* [Post]. X. https://x.com/satyanadella/status/2108931348857827686 +- **Capture.** Read from the live page on 2026-10-12 (13M views at capture). The full text is kept in a + local file outside this public repo, with paragraph numbers P1–P21 added for locators. Facet G quotes + it by paragraph. +- **Summary (Claude's).** Organisations cannot trace a frontier model's behaviour to its causes, yet + give such models sensitive data and the power to act. Nadella proposes treating models as insider + risks: surround them with deterministic system design, human controls and operating procedures, and + keep every control outside the model. He lists seven principles of "observability": model diversity, + observe everything, verifiability, independent controls, independent auditability, containment and + incident disclosure. +- **Short quotations.** Authority: "separate the supply of intelligence from the authority over it" (P5). + Records: every meaningful model action must leave "tamper-proof human readable evidence" (P15). On + chain-of-thought transparency: "not sufficient or dependable" (P9). His closing line says the most + trustworthy system is the one that lets us trust the model least (P21, paraphrased). + +## The author's own text this paper builds on + +| Text | Locator | +| --- | --- | +| "Word separation did not train the populace to read *scriptio continua*; it changed the interface so that reading no longer required the training." | `../dial_response_algorithmacy/draft.md:263` | +| "the remedy was never to reserve reading for those who could bear the load but to change the interface until the load fell" | `../dial_response_algorithmacy/draft.md:337` | +| "how should we design for 'algorithmacy,' which is similar to asking how we should design for 'literacy.'" | `../dial_response_algorithmacy/sources/phd1750_topic_and_rationale.md:15-16` | +| "The design object is the engagement layer — where reading and rewriting the shared mediator is the fundamental mode rather than a product layer that hides the mediator behind a distribution network." | same file, lines 52-54 | +| The seed prompt for the reading-history line: "From a user experience perspective, we are in the infancy of developing ways to navigate triadic cooptive coordination and developing algorithmacy. I want to look at the development of texts and how users actually pick them up and read them…" | `../triadic_reduction/sources/G7_tablets_to_algorithms.md` (Turn 1) | +| The published Part 1 of the DXD series | `../decision_experience_design/published/2026-09-30_decision-experience-design.md` | +| The current definition of algorithmacy | `../algorithmacy_ot/manuscript/PAPER.md:19-21` | + +## Standing rulings + +Listed with locators in `RESEARCH_PLAN.md` §3. + +## Process notes + +- The four Gemini essays that develop the scriptio continua analogy are process, not citable sources + (`RESEARCH_PLAN.md` §6). +- The repo is public. No publisher PDF, extracted full text or full copy of the post is committed. +- Research facets were run by Sonnet agents on 2026-10-12, one per facet, with Claude (Opus) as lead. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_A_the_record.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_A_the_record.bib new file mode 100644 index 00000000..594b4efa --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_A_the_record.bib @@ -0,0 +1,80 @@ +@book{saenger1997space, + author = {Saenger, Paul}, + title = {Space Between Words: The Origins of Silent Reading}, + publisher = {Stanford University Press}, + address = {Stanford, CA}, + year = {1997}, + doi = {10.1515/9781503619081}, + note = {Read in part at full-text depth, 2026-10-12: OCR text of an unrestricted archive.org copy (filenames show a shadow-library origin; lead to decide whether to cite from it). Read: Preface, ch. 1, ch. 2 pp. 18-41, ch. 3 pp. 69-79, ch. 4, ch. 6 pp. 120-126, ch. 14 pp. 243-251 and 259-261, ch. 15 pp. 264-277, endnotes to pp. 9-13; samples of pp. 100, 177, 185. Not read: chs. 7-8, most of chs. 9-13, Appendix, most endnotes. Page numbers from OCR running heads; confirm against a library copy.} +} + +@article{saenger1982silent, + author = {Saenger, Paul}, + title = {Silent Reading: Its Impact on Late Medieval Script and Society}, + journal = {Viator}, + volume = {13}, + pages = {367--414}, + year = {1982}, + doi = {10.1484/J.VIATOR.2.301476}, + note = {Not accessible: Brepols bot wall (Unpaywall lists it as bronze OA). Crossref pages 367-414; Saenger 1997 Preface gives 366-414. Thesis known only through the 1997 Preface, pp. ix-x.} +} + +@book{parkes1992pause, + author = {Parkes, M. B.}, + title = {Pause and Effect: An Introduction to the History of Punctuation in the {W}est}, + publisher = {Scolar Press}, + address = {Aldershot}, + year = {1992}, + note = {Read in part at snippet depth (Google Books windows and Crossref chapter abstracts), carried over from facet E; no new text reached 2026-10-12. Routledge reprint 2016, doi 10.4324/9781315247243. archive.org lending copy pauseeffectintro0000park (guest search-inside refused).} +} + +@book{wingo1972latin, + author = {Wingo, E. Otha}, + title = {Latin Punctuation in the Classical Age}, + publisher = {Mouton}, + address = {The Hague}, + year = {1972}, + doi = {10.1515/9783110805215}, + note = {Not accessible: De Gruyter closed; archive.org lending copy latinpunctuation0000wing (guest search-inside refused). Cited second-hand by Saenger 1997, endnotes to pp. 9-11.} +} + +@incollection{cavallo1999history, + editor = {Cavallo, Guglielmo and Chartier, Roger}, + title = {A History of Reading in the {W}est}, + publisher = {Polity Press}, + address = {Cambridge}, + year = {1999}, + note = {Not accessible (archive.org lending copy historyofreading0000cava; the item cavallo-a-history-of-reading-in-the-west is mislabelled and holds Graff, Undisciplining Knowledge). Contains Parkes, Reading, copying and interpreting a text in the early Middle Ages, and Saenger, Reading in the later Middle Ages; chapter page ranges not verified.} +} + +@phdthesis{tunbridge1992study, + author = {Tunbridge, Genny Louise}, + title = {A Study of Scribal Practices in Early {I}rish and {A}nglo-{S}axon Manuscripts}, + school = {University of Oxford}, + year = {1992}, + doi = {10.5287/ora-ave2z6gqd}, + url = {https://ora.ox.ac.uk/objects/uuid:8fc8f8de-0229-4f89-9816-e53bf8c6cc7e}, + note = {Read in part (card: library/tunbridge1992study.md); abstract first read via DataCite 2026-10-12. Rights: all rights reserved.} +} + +@incollection{saenger2020orality, + author = {Saenger, Paul}, + title = {Orality and Visible Language}, + booktitle = {The Oxford Handbook of Latin Palaeography}, + publisher = {Oxford University Press}, + pages = {692--704}, + year = {2020}, + doi = {10.1093/oxfordhb/9780195336948.013.112}, + note = {Abstract only (Crossref). Closed.} +} + +@incollection{bronner2018nonseparation, + author = {Bronner and Busch and Fleischer and Poppe}, + title = {(Non-)separation of words in early medieval {I}rish and {G}erman manuscripts and the concept ``word''}, + booktitle = {Empirical Approaches to the Phonological Structure of Words}, + publisher = {De Gruyter}, + pages = {45--70}, + year = {2018}, + doi = {10.1515/9783110542899-003}, + note = {Not accessible: Crossref record has no abstract; Unpaywall reports no open copy. Given names not checked (Crossref family names only read).} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_A_the_record.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_A_the_record.md new file mode 100644 index 00000000..3fd87dc0 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_A_the_record.md @@ -0,0 +1,254 @@ +# A — The record: what changed on the page, where, when, by whom and why + +*Corrected on 2026-10-12 against `../audit/AUDIT_01.md`: 9 factual fixes.* + +*Facet A of the foundation-paper literature review. Agent-drafted on 2026-10-12. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** The documentary record of the move from *scriptio continua* to word separation: interpuncts in early Greek and Latin; the Roman abandonment of them; the aids that existed meanwhile; Insular (Irish, then Anglo-Saxon) separation in the seventh and eighth centuries; "aerated" script; the Continental spread; the vernaculars; Greek and Byzantine practice; what else changed on the page alongside spacing; and who read before and after. It also records exactly what Saenger claims about consequences, and how strongly. It supersedes the snippet-level work in `decision_experience_design/part2_ethics/literature/facets/E_reading_history_meaning.md`, section B1 and E1. + +**The question the synthesis answers.** What does the record show about who changed the page, for whom and why; was spacing one move or one of a bundle; and what does that leave standing of H-cost (design precedent) and H-cause (ignition)? + +**Legend for read status.** *Read in full* = whole text read this run. *Read in part* = text obtained and the stated pages read; unread parts are listed. *Abstract only* = database abstract or record only. *Not accessible* = no text. A further caution for the main source: the only full text I could reach of Saenger (1997) is an OCR text of a scan, not the print edition (see entry A1). Page numbers are print pages recovered from the running heads in that OCR; a few pages have garbled heads, and I flag each locator that rests on a merged or missing head. The audit (`../audit/AUDIT_01.md`) confirmed 19 of 25 checked Saenger quotations, wording and page, through Google Books search-inside and the archive.org lending scan. + +--- + +## (a) Search log + +All searches ran on 2026-10-12. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | repo | Read `RESEARCH_PLAN.md`, `REUSE.md`, facet E (lines 1–236, 305–330, 379–414) | — | Starting claims, locators and gaps. | +| 2 | archive.org advancedsearch | `title:(space between words) AND creator:saenger` | 0 | The creator field is empty on the item that exists. | +| 3 | archive.org advancedsearch | `title:(space between words)` | 4 | One relevant: `space-between-words-the-origins-of-silent-reading`, unrestricted, with a `_djvu.txt` OCR file and a PDF. The filenames carry a "z-lib.org" tag, i.e. an upload from a shadow library. | +| 4 | archive.org advancedsearch | `title:(pause and effect) AND creator:parkes`; `title:(latin punctuation classical age)`; `title:(history of reading in the west)` | 1; 1; 2 | Found `pauseeffectintro0000park`, `latinpunctuation0000wing`, `historyofreading0000cava`: all `access-restricted-item: true` (lending only). `cavallo-a-history-of-reading-in-the-west` is unrestricted but mislabelled: its text is Graff, *Undisciplining Knowledge* (checked the first 800 characters). `pause-and-effect` is an unrelated video. | +| 5 | archive.org download | `_djvu.txt` and `_text.pdf` of the Saenger item | 1.17 MB text; 253-page PDF of two-page spreads | Saved under `dxd_foundation_local/pdfs/A/`. Used for all Saenger quotations. | +| 6 | archive.org search-inside | `fulltext/inside.php?item_id=…&q=grammar of legibility` for Parkes and `q=interpunct` for Wingo | "Item not available" ×2 | Guest users cannot search inside lending items. Not circumvented. | +| 7 | archive.org full-text search | `services/search/beta/page_production/?service_backend=fts&user_query="grammar of legibility"` | error | Parameter rejected; guest session. Not pursued. | +| 8 | Brepols (Saenger 1982) | `curl` of `brepolsonline.com/doi/pdf/10.1484/J.VIATOR.2.301476` | HTTP 200, 114-byte redirect page to `/lander` | Bot wall. Unpaywall lists the article as "bronze" open access (`brepolsonline.net/doi/epdf/…`). I did not go around the wall. | +| 9 | Google Books search-inside | Parkes 1992 (id `X11nAAAAMAAJ`), 12 queries: *grammar of legibility; foreign language; Irish; word separation; separation of words; scriptio continua; Carolingian scribes; inexperienced readers; per cola et commata; distinctiones; praelectio; Insular* | 0 snippets | The endpoint answered `searchable: true` with zero results for every query, including a control query ("legibility"). The index is unreliable (facet E found the same). No new Parkes text this run. | +| 10 | WebSearch | `Parkes "Pause and Effect" punctuation pdf "grammar of legibility" Irish scribes word separation` | 10 | No PDF. Confirms the chapter list and that "grammar of legibility" is the title of Parkes's ch. 2 subtitle. Blog and review hits are secondary and not used. | +| 11 | WebSearch | `Tunbridge "A study of scribal practices in early Irish and Anglo-Saxon manuscripts" ora.ox.ac.uk` | 9 | No hit on the thesis. | +| 12 | DataCite | `api.datacite.org/dois/10.5287/ora-ave2z6gqd` | 1 | Record confirmed (Tunbridge, Genny Louise; University of Oxford; 1992). Full abstract read. Rights: all rights reserved; I found no download link on the ORA landing page. | +| 13 | OpenAlex | six searches: *Stutzmann word separation…; word separation early medieval Latin manuscripts Irish scribes…; scriptio continua Latin papyri interpunct…; Saenger Space between words review Speculum; history of word spacing Insular manuscripts non-native readers Latin; Tunbridge scribal practices…* | 6; 6; 0; 6; 6; 6 | Noise only. No Stutzmann (2020) item identified; no dated-manuscript test of Saenger's chronology; no papyrological study of the interpunct decline. | +| 14 | Crossref | DOIs 10.1515/9781503619081 (Saenger 1997); 10.4324/9781315247243 (Parkes reprint); 10.1515/9783110805215 (Wingo); 10.1484/J.VIATOR.2.301476 (Saenger 1982); 10.1093/oxfordhb/9780195336948.013.112 (Saenger 2020); 10.1515/9783110542899-003 (Bronner et al.) | 6/6 resolved | Titles, authors, years matched. Viator pages in Crossref: 367–414 (Saenger's own preface gives 366–414). Bronner et al. has no abstract. | +| 15 | Crossref | `query.bibliographic` for Stutzmann 2020 and for Parkes's Cavallo–Chartier essay | 5; 4 | Stutzmann not found. Nearest hits are Coulson, "Punctuation" (Oxford Handbook of Latin Palaeography, 10.1093/oxfordhb/9780195336948.013.115) and Galambos, "Premodern punctuation and layout" (10.7312/chen19552-017); not read, not Stutzmann. Parkes's essay is not a Crossref record. | +| 16 | Unpaywall | Bronner et al.; Wingo | closed ×2 | No open copy. | +| 17 | Scholar Gateway | `semanticSearch`: *word separation Insular scribes scriptio continua Parkes Saenger* | `ACCESS_DENIED` | Failed again; not retried. | +| 18 | local grep of the Saenger text | *literacy; Wingo; foliation; index; running head; middlebrow; paragraphos* | — | `literacy` occurs four times in the main text (Preface p. x; p. 11; p. 177; p. 185). `paragraphos` does not occur. | + +--- + +## (b) Entries + +### A1. Saenger, *Space Between Words* (1997) + +**saenger1997space** — Saenger, P. (1997). *Space between words: The origins of silent reading*. Stanford University Press. https://doi.org/10.1515/9781503619081 +- *Verification:* Crossref record (row 14). Print extent not independently checked. +- *Read status:* **read in part, at full-text depth.** I read the OCR text of the archive.org copy (row 3–5). Read in full: Preface (pp. ix–xi), ch. 1 Introduction (pp. 1–17), ch. 2 pp. 18–41 (Nomenclature; Separated Script; Aerated Script), ch. 3 pp. 69–79 (punctuation, foliation, résumé notes), ch. 4 pp. 83–99 (Insular culture), ch. 6 pp. 120–126 (Protoscholasticism, Arabic influence), ch. 14 pp. 243–251 and 259–261, ch. 15 pp. 264–277, and the endnotes to pp. 9–13 (pp. 300–301). Read by sample only: pp. 177 and 185 (ch. 9, ch. 10), p. 100 (ch. 5 opening), pp. 241–242. **Not read:** ch. 5 beyond its opening, ch. 7 (numbers and music), ch. 8 (Gerbert, Abbo, Heriger), most of chs. 9–13 (the abbey-by-abbey diffusion), the Appendix tables and most endnotes. The diffusion chapters hold the dated-manuscript evidence for the Continental chronology; I have Saenger's own summaries of them but not the manuscript descriptions. The audit (`../audit/AUDIT_01.md`) confirmed 19 of 25 checked Saenger quotations, wording and page, through Google Books search-inside and the archive.org lending scan. +- *Provenance and reliability of the copy.* The archive.org item is unrestricted but its filenames show a shadow-library origin ("z-lib.org"), so it may be an unauthorised copy. I read it because the lead asked for full text from archive.org and no wall stood in the way; the lead should decide whether to cite from it. The OCR drops characters at the edges of some pages (for example pp. 85–89, 94–95, 98–99) and merges running heads on pp. 86–87, 94–95 and 98–99. I quote only lines whose OCR is clean. Every locator below should be confirmed against a library copy before it enters the paper. Page numbers rest on running heads; where a head is missing or merged I say so. +- *What it argues.* A single thesis about the physiology of reading, stated at p. 2: "It is a premise of this book that given such differences in transcription, there is a correlation between a propensity to read orally in both past and contemporary cultures and the threshold in the duration of cognitive activity needed to achieve lexical access in that culture's script." Word separation lowers that threshold; the book then narrates, manuscript by manuscript, when and where it arrived. The author says what the book is not: "this book, which, I would stress, is not a general history of literacy" (Preface, p. x). + +**Chronology, with locators.** +- *Interpuncts.* "While the very earliest Greek inscriptions were written with separation by interpuncts, points placed at midlevel between words, Greece soon thereafter became the first ancient civilization to employ scriptura continua" (pp. 9–10). Before vowels were added to the Phoenician alphabet, "all the ancient languages of the Mediterranean world … were written with word separation by either space, points, or both in conjunction" (p. 9). Saenger links scriptio continua to vocalic alphabets: "the adoption of vowels and of scriptura continua went hand in hand" (p. 9). +- *Roman abandonment.* The Romans "maintained the earlier Mediterranean tradition of separating words by points far longer than the Greeks, but they, too, after a scantily documented period of six centuries, discarded word separation as superfluous and substituted scriptura continua for interpunct-separated script in the second century A.D." (p. 10). Elsewhere: "A century after Seneca's death in A.D. 65, the Roman page had become increasingly similar to that of Greek, as the Romans gave up separation by interpuncts and adopted the Greek convention of scriptura continua" (p. 12); "Separation by interpuncts alone was primarily a phenomenon of Latin antiquity prior to the end of the second century A.D." (p. 29). The endnotes to pp. 9–11 (p. 300) cite Wingo (1972) for Latin separation "before the end of the second century"; the OCR interleaves the note text, so I cannot read Wingo's page. **Saenger's own dates differ by passage** (second century; a century after 65; and "the late second or third century A.D." at p. 69, where the subject is the Greek hyphen entering Latin "when the Romans adopted scriptura continua in emulation of the Greeks"). Facet E took p. 69 as a date for the adoption of scriptio continua in general. The audit finds that the dependent clause at p. 69 does date the Roman adoption of scriptura continua "in emulation of the Greeks" to "the late second or third century A.D.", so E's use is not a misreading; the sentence's subject is the hyphen, and pp. 10 and 29 give "the second century A.D.". +- *Interpunct page, not spaced page.* Roman interpuncts "were not usually accompanied by quantities of space any greater than that ordinarily placed between adjacent letters within a word" (p. 26); the largest space he recorded with an interpunct is "0.67 unities of space", against a minimum of 1.5 for rapid perception (p. 28); interpuncts and hederae "were ineffective aids to the physiological process of reading when compared with the interword spaces familiar in late medieval manuscripts and modern printed books" (p. 28) [page to check in a library copy]. He also notes a hedera in second-century inscriptions (p. 26). +- *Aids during the scriptio continua centuries.* In the imperial period "punctuation notes (which grammarians termed positurae) were tools intended primarily for the inexperienced reader, who employed them when mastering a text by copying it onto wax tablets, and they were not generally considered appropriate for a formal book" (p. 72). Such marks "were added by the reader rather than the scribe" (p. 72). "At the end of antiquity, the only 'punctuation' sometimes present in formal Roman books was spaces setting off paragraphs" (p. 72). Cassiodorus is the first he sees imagining the scribe inserting positurae (p. 72). *Per cola et commata* is Jerome's format for the Vulgate, "creating contextual space in which each line represented either a phrase, a clause, or a sentence" (p. 16); Saenger reads it as "a level of legibility in Latin appropriate to the middlebrow Christian audience of the late Roman world" (p. 124, with the remark that he has no textual evidence for the Semitic-influence theory). Foliation and pagination: attested in Greek codices from the third century, "never employed by the Romans" (p. 12; p. 77: "No evidence of foliation exists on any surviving Latin codex"). He mentions the *distinctiones* only as the sentence-divisions an Irish grammarian wanted readers to recognise (p. 83); I found no passage in the parts I read that discusses *paragraphos*. +- *Irish origin.* "The first separated Latin manuscript books in western Europe were Irish, and their text format was probably inspired by Syriac Gospel books" (p. 83). The first datable word-separated manuscripts: Dublin, Trinity College 60 (the Book of Mulling, "before 692") and Schaffhausen, Stadtsbibliothek Generalia 1 (Adomnán's *Vita Columbae*, "before 713") (p. 83). Earlier Irish manuscripts were aerated: "In Ireland, aerated manuscripts were uniquely a transient phenomenon of the seventh century, after which all Latin writing was separated" (p. 34). Within "three-quarters of a century of their initial experimentation" Irish scribes had moved from letter blocks to syllable blocks and word-and-morpheme spacing (p. 34). The Anglo-Saxon side: "Indeed, space for word separation was regularly inserted in the earliest surviving English books, and the first English documents, which date from the end of the seventh century, were all separated" (p. 87 per the audit's Google Books check; the earlier p. 86 came from OCR with merged page heads). Anglo-Saxon grammarians trained by Irish masters (Aldhelm, Bede, Tatwine, Boniface) wrote in separated script and wrote about it (pp. 83–89). +- *Aerated script on the Continent.* Aerated Latin codices "were common from the seventh century through the eleventh" in France, the Low Countries, Germany and Iberia, and "prevailed in central and southern Italy even into the thirteenth century"; formal Roman-law texts were aerated in central Italy in the fourteenth century (p. 34). "The Insular graphic innovation of word separation was not adopted by Continental scribes until the mid-tenth century" (p. 100). Continental aerated script is a hybrid of scriptio continua and Insular practice, arranged in "hierarchical word blocks" (pp. 35–36, 100). Insular scholars on the Continent (Alcuin, Sedulius Scottus) brought true separation, "Its general acceptance there, however, had to await a range of further developments" (p. 100). +- *Why the Continent waited, and what ended the wait.* "Continental monks perceived no need to make the reading of this corpus easier. Indeed, an ambiguous text format enhanced the mystery and power of clerics, who were the sole dispensers of the Divine Word to a laity almost entirely illiterate" (p. 120). After 950 a "new corpus of protoscholastic texts" demanded technical reading: "The desirability of making reading easier became apparent to Continental scholars, and to this end the word separation by space of Insular manuscripts became an essential linguistic tool" (pp. 120–121). A second model was Arabic: Latin translations of Arabic science "brought word separation with them and formed the earliest body of writings to circulate invariably in word-separated text format" (p. 124). +- *Not yet ubiquitous before the middle decades of the eleventh century.* "Word separation prior to the middle decades of the eleventh century was still not a ubiquitous practice in western Europe"; authors of unseparated texts wrote Latin "in no way inferior to that of the authors who observed word separation" (p. 177). Yet "By the mid-eleventh century, word separation was standard and ubiquitous" (p. 176, under Sigebert of Gembloux) [page to check in a library copy]. On Adémar of Chabannes: "in the eleventh century, the absence of regular word separation was not a sign of a low level of literacy or defective Latinity" (p. 185). "Canonical separation" (a regular space between every word, bar monosyllabic prepositions) "became the norm" in the twelfth century (p. 243) [page to check in a library copy]. +- *Vernaculars.* Earliest vernacular texts in Latin letters were "aerated" and written "in the same geographical regions and at approximately the same time as the acceptance of the separation of words by space", Ireland and England first, Italy last, "in the twelfth and thirteenth centuries" (p. 97). "Writing in the vernacular, like word separation, was a way of accommodating reading" (p. 97). Vernacular spacing was looser: reading one's own tongue "required fewer graphic aids" (p. 97). "Separation of function words normal for Latin, had still not been realized in 1500 for the Romance vernacular tongues" (p. 266) [page to check in a library copy]. Regions with Celtic or Teutonic vernaculars show more space than Romance-speaking ones, "where the vernacular … was not consciously recognized to be a distinct language" (p. 41). +- *Greek, Byzantine, other scripts.* Separation of Greek "originated in manuscripts copied by Irish scribes in the early Middle Ages and became a frequent practice in Renaissance Italy and France well before it became normal in Byzantium in the second half of the sixteenth century"; Cyrillic by "the beginning of the seventeenth century"; for Sanskrit-derived scripts, "only in the modern period" (p. 13) [page to check in a library copy]. "In Byzantium, aerated Greek manuscripts were normal from the ninth century to the sixteenth" (p. 34). Hebrew: the Masoretic script, with its added vowels, was "probably modeled, directly or indirectly, on" Carolingian aerated writing (p. 10). Semitic scripts without vowels were separated throughout (p. 10). +- *Universal when?* No single date. Latin: normal from the twelfth to thirteenth century in northern Europe (p. 243 [page to check in a library copy]; "the norm of scholastic culture", p. 13), later in Italy (p. 34). Romance vernaculars: separation of function words unfinished in 1500 (p. 266 [page to check in a library copy]). Greek: second half of the sixteenth century (p. 13). Cyrillic: c. 1600 (p. 13 [page to check in a library copy]). + +**Who did it and why (Saenger's account).** +- *Irish and Anglo-Saxon scribes and teachers.* "The origins of rapid, silent reading lie in the scribal techniques and grammatical teachings that developed in Ireland and England in the seventh and eighth centuries" (p. 83). Aldhelm, Bede, Tatwine and Boniface turn the *figura* of a word from an abstraction into "the outline of the block of letters on the page, where space separated one pars, or part, from another" (p. 85). The Irish context: "In contrast to the way in which young Romans learned Greek, Irish- [text lost; the sentence says Irish and Anglo-Saxon pupils learned Latin from word-separated grammars and glossaries]" (p. 91, OCR lost the middle of the sentence; I quote only the clean fragment and paraphrase the rest). Alphabetical glossaries were "an aid to those readers of Latin for whom the vernacular offered no phonic clues for comprehension" (p. 91); the earliest substantial alphabetical glossaries were bilingual (p. 91, paraphrased because the OCR is damaged there). So the foreign-language-learner motive is Saenger's too, with a second strand, explicit grammatical instruction about words (pp. 83–89). +- *Intention.* For the Irish lettered tradition he writes: "Irish scribes were not seeking to establish a text free of barbarisms, but rather … one that was unambiguous and easy to read" (pp. 94–95; merged page head) [page to check in a library copy]. Aeration he reads as "an unconscious quest for a mode of textual display that enhanced the reader's exploitation of his hitherto untapped neurophysiological faculties" (p. 33). So the intention to ease reading is stated for the Insular scribes, and inferred (not documented) for the earliest aerated work. +- *Who the aid was for.* Bede's diastole corrects "a scribe's omission of the word space that ordinarily guided the inexperienced reader" (p. 86). Separation gave "shortcuts for achieving the reading skills that an elite among the ancients had mastered only through a prolonged and arduous grammatical apprenticeship" (p. 85) [page to check in a library copy]. +- *Why Rome did not.* Saenger gives a social explanation (pp. 10–12): "the ancient world did not possess the desire, characteristic of the modern age, to make reading easier and swifter because the advantages that modern readers perceive as accruing from ease of reading were seldom viewed as advantages by the ancients" (p. 11). Those advantages "include the effective retrieval of information in reference consultation, the ability to read with minimum difficulty a great many technical, logical, and scientific texts, and the greater diffusion of literacy throughout all social strata of the population" (p. 11). Reading was "oral and rhetorical by physiological necessity as well as by taste", the canon was small, and oralisation compensated; "the notion that the greater portion of the population should be autonomous and self-motivated readers was entirely foreign to the elitist literate mentality of the ancient world"; the labour of reading went to "skilled slaves, who acted as professional readers and scribes" (p. 11). Seneca (*Epist.* 40.11) reads the interpunct as a sign of Latin's "slower cadence", not a speed aid (p. 12). +- *Why the Continent did.* The protoscholastic and Arabic-science demand described above (pp. 120–124), plus a heresy argument: "errors in reading could have potentially grave consequences by breaching the subtle distinctions separating orthodoxy from heterodoxy" (p. 123). + +**What Saenger claims about consequences, and how strongly (verbatim).** + +| Claim | Quotation and locator | Strength | +|---|---|---| +| Silent reading | "The importance of word separation by space is unquestionable, for it freed the intellectual faculties of the reader, permitting all texts to be read silently, that is, with eyes alone" (p. 13). Preface: separation was "the crucial element in the change to silent reading" (p. x). | Unhedged. But silent reading as such is not the claim: "when Ambrose read silently, given the constraints of ancient writing, he could not have read rapidly in the manner of the modern silent reader, even if he had wanted to do so" (p. 8), and "the reading of letters and documents, whether silently in the strict sense, or more probably, quietly with suppressed voice" is conceded (pp. 8–9). The claim is about *rapid* silent reading and *reference* reading: "no classical author described rapid, silent reference consultation as it exists in the modern world" (p. 9). | +| Speed and the less able reader | "As a consequence, even readers of modest intellectual capacity could read more swiftly, and they could understand an increasing number of inherently more difficult texts" (p. 13). | Unhedged, and the one sentence closest to "widened the group". "Modest intellectual capacity" is still a literate reader of Latin. | +| Not instant | Highly aerated Continental script: "Certainly, rapid, silent reference reading was not possible even within highly aerated early medieval Continental minuscule scripts" (p. 39). On devotional silent reading by Carolingian monks he writes that the sources are "too ambiguous for us to be certain" (p. 39). | Explicit lag: the Insular innovation (c. 690) does not yield rapid silent reading on the Continent until after the eleventh century. | +| Adaptation concession | "Psychologists concede that readers who habitually read unseparated writing would adapt and improve their reading rates over time. However, they maintain that these readers' brains would always compensate for the extra cognitive burden by more numerous ocular fixations and regressions" (p. 7) [page to check in a library copy]. | Saenger states the concession himself. | +| Reference reading | The skills of reference reading "emerged first in the British Isles in the early Middle Ages and on the Continent beginning in the late tenth century" (p. 4). Alphabetical glossaries, cartularies, tables, running heads: pp. 90–96, 259–260. | Strong, tied to new genres. | +| Scholasticism | "Separated script was an indispensable medium for this new learning, and intrinsic to it" (p. 122); authors "relied on the separation of words, combined with conventions for word order, to formulate increasingly subtle and complex thoughts, which without these tools would not have been possible" (p. 122); "Just as separated written Latin had facilitated the birth of scholasticism, separated vernacular writing allowed for the transference of the subtleties of fully developed scholastic thought to a new lay audience" (p. 273). | The strongest causal wording in the book. The direction is two-way: the demand of the new texts explains adoption (p. 120) and the new script then makes the texts possible (p. 122). | +| Private reading and composition | Hugh of St Victor's three modes of reading, including "reading to self by gazing (inspicere), that is, silent, private reading" (pp. 244–245). "The adoption of separated writing sparked an interest in autograph composition, creating the possibility of more intimate expression on the part of the author" (p. 249). Silent reading and writing gave "a new dimension of privacy" (p. 264). | Strong. "Sparked" is causal; the evidence is individual authors (Othloh, Guibert of Nogent). | +| Heresy and dissent | "Psychologically, silent reading emboldened the reader because it placed the source of his curiosity completely under personal control"; "Private, visual reading and private composition thus encouraged individual critical thinking and contributed ultimately to the development of skepticism and intellectual heresy" (pp. 264–265) [page to check in a library copy]. Subversive politics: "private reading provided a medium for expressing subversive political thoughts" (p. 274). | "Encouraged … contributed" is moderate wording; "emboldened" is a psychological assertion with no measurement. | +| Erotic reading | "Private reading stimulated a revival of the antique genre of erotic art" (p. 274); "the existence of twelfth-century erotica was possible because of the new intimacy between author, reader, and book" (p. 250). | Causal and stated as fact; evidence is elite books (Guibert, Burgundian court). | +| Private devotion | The Harleian, Royal and Nunnaminster prayer books, with "prayers in the first-person singular intended for intense private devotion; no equivalent can be found in unseparated manuscripts"; "The legibility of these books no doubt encouraged the development of private religious devotion" (p. 96). Vernacular: silent reading "gave lay readers the means of pursuing the individual relationship to God" (p. 275). | "No doubt encouraged" is hedged-in-form but confident. | +| Reformation | "Isolated, private reading and prayer as the pathway to salvation, in turn, may have fostered insecurities about the worthiness of each individual's faith and devotion and stimulated zeal for religious reform" (p. 276); the receptivity of Europe's elite "owed much to a long evolution that began in the late seventh century and culminated in the fifteenth century" (p. 276, per the audit's Google Books check; the OCR page head was ambiguous between pp. 276 and 277). | "May have" — the most hedged claim in the book. | +| Literacy in general | Disclaimed: the book "is not a general history of literacy" (Preface, p. x). In the body, "literacy" appears as a modern value the ancients did not share (p. 11) and as a point against a theory (pp. 177, 185). | Saenger calls his book "not a general history of literacy" (p. x). On p. 11 he lists "the greater diffusion of literacy throughout all social strata of the population" among the "advantages that modern readers perceive as accruing from ease of reading", which the ancients "seldom viewed as advantages". He attaches this to ease of reading in general and presents it as a modern value; he offers no evidence that spacing raised the number of readers. His detailed claims concern *how already-literate people read*. | +| Direction of the aim | His 1978 paper, described in the Preface: "the separation of words … altered the physiological process of reading and by the fourteenth century enabled the common practice of silent reading as we know it today" (p. ix). | The date for "common" silent reading is the fourteenth century, about 650 years after the Irish innovation. | + +**Other changes on the page alongside spacing (Saenger).** +- *Word order.* "Word order in medieval Latin affected the neurophysiology of reading in a way that was complementary to word separation, and together they laid the foundation for rapid, silent reading" (p. 14). Insular authors imposed fixed word orders from the sixth century (p. 89); Tatwine made Donatus's order of the parts of speech the order of a sentence (p. 90). +- *Punctuation.* The restructuring of written language "was a necessary prerequisite for the introduction of consistent signs of syntactic punctuation" (p. 15). New emblematic punctuation "emerged in the British Isles in manuscripts written with word separation" (p. 73). The question mark, "unknown in antiquity", appears in aerated Continental script (p. 36). Construction notes (Insular, ninth century), tie notes, résumé notes (pp. 69–79). +- *Accents and small signs.* Apices for vowel length, the diastole, the hyphen and the *trait d'union*, terminal forms, monolexic abbreviations (pp. 32, 40–41, 84–86). +- *Script and capitals.* Minuscule matters "insofar as it contributed, in conjunction with word separation, to giving each word a distinct image" (the "Bouma shape", pp. 18–19). Without an upper/lower-case distinction, Roman interpuncts lost "much of the potential difference in the Bouma shape" (pp. 28–29). +- *Musical notation.* Neumes spread in ninth- and tenth-century aerated texts (pp. 36–37). +- *Genres.* Five new Insular genres: alphabetical glossaries, interlinear translations, pocket Gospel books, personal prayer books, cartularies, and vernacular books (pp. 90–98). (He says "five" and lists six headings; I count six sections.) +- *Page apparatus.* The codex: "originated in the first century A.D. in Rome", with early codices "separated by points" (p. 77). Foliation in Latin codices begins in tenth-century aerated books and becomes current "after 1200" (p. 77). Schematic diagrams from the late tenth century at Reims, Fleury and Chartres (p. 79). Chapters and *distinctiones*, "tables of chapter headings, alphabetical tables by subject, and running headings became standard features of the scholastic codex" (p. 259); the coloured paragraph mark from the early thirteenth century (p. 259); indexes and tables in lay aristocratic books after 1350 (p. 270 area; locator approximate, OCR head missing). +- *Responsibility.* "Clearly, during the course of the nine centuries following Rome's fall, the task of separating written text, which had been for half a millennium a cognitive function of the reader, became instead the task of the scribe" (p. 13); Hugh's "presumption that it was the scribe's responsibility to prepare the text for the reader exemplified the transformed mentality that the previous century had wrought" (p. 245). + +**Readers before and after (Saenger).** Before: the *lector* and the professional reader ("skilled slaves", p. 11); the elite reader trained through "a prolonged and arduous grammatical apprenticeship" (p. 85) and the *praelectio* (p. 8); "inexperienced readers" with reader-added marks (p. 72). Insular: Latin learners with a foreign mother tongue (pp. 91, 97). Carolingian Continent: monks reading a liturgical, patristic and grammatical corpus, mostly aloud (pp. 39, 120). Eleventh to thirteenth century: protoscholastic clerics, schoolmen, students (pp. 120–123, 259–260). Fourteenth and fifteenth century: princes, aristocrats, urban merchants (pp. 265–266, 273–276). The laity as a whole stays outside until late: before 1350 French kings "rarely read themselves" (p. 265). + +**Bearing.** +- H-cost: **supports**, strongly, in the form "moved parsing work from reader to scribe and page" (pp. 13, 85, 245) and "lowered the cost" (pp. 7, 13). It also supports "the audience widened step by step": Insular learners, then clerical scholars, then lay elites. +- H-cause: **complicates, and in the author's strong form refutes.** Saenger disclaims a general history of literacy (p. x), and lists "the greater diffusion of literacy throughout all social strata" among the advantages of ease of reading that the modern age values (p. 11). His strongest causal claims concern scholasticism, privacy and dissent among readers already literate in Latin, and each comes centuries after the innovation. Silent reading itself existed (p. 8). The adoption story runs from demand to script as much as the reverse (p. 120). +- Problem 4: **bundle** (word order, punctuation, notes, accents, glossaries, tables, running heads, foliation, minuscule, capitals; pp. 14–15, 36, 69–79, 90–98, 259–260). +- Problem 1: **supports** the worry. The Continent took about 260 years to adopt it (c. 690 to the mid-tenth century, p. 100: 950 − 690 = 260) and about 350 years to make it common (c. 690 to the middle decades of the eleventh century, p. 177, taking c. 1040 as the middle: 1040 − 690 = 350), and mass literacy is outside the book's scope. + +**Cautions on his evidence.** Every cognitive claim rests on 1970s–80s experiments with English-speaking readers, cited in the endnotes (not examined here); the "more than twice the normal quantity of fixations" figure (p. 7) is stated without a source in the text. Facet D should test it. His dates for individual codices come from his own 1984–85 survey (Preface, p. x), described in the chapters I did not read. + +--- + +### A2. Parkes, *Pause and Effect* (1992) + +**parkes1992pause** — Parkes, M. B. (1992). *Pause and effect: An introduction to the history of punctuation in the West*. Scolar Press. (Routledge reprint 2016: https://doi.org/10.4324/9781315247243) +- *Verification:* Crossref record for the reprint (row 14). The archive.org lending copy `pauseeffectintro0000park` exists (dated 1993; the University of California Press edition). +- *Read status:* **read in part, at snippet depth, carried over from facet E; not re-read this run.** I tried again to reach text (rows 6, 9, 10) and got none. The Crossref chapter abstracts (chs. 1–3) were read by facet E. The quotations below come from E's section B1, taken from forty-word Google Books windows; I have not re-verified them and give them only so the lead has them in one place. +- *What it argues (from E).* Ch. 2 (pp. 20–29) names the "grammar of legibility" as the decorum governing "this complex of graphic conventions and the message of a text". Ch. 3 (pp. 30–34): "Irish and Anglo-Saxon missionaries taught their scribal practices to pupils in new foundations"; Carolingian scribes "developed a complex of new graphic conventions both to improve the intelligibility of texts copied in minuscule scripts, and to facilitate access to the information transmitted in the written medium". Snippet at p. 1: "word separation, features of layout and punctuation, were developed to make it easier for readers to extract the information conveyed in the written medium", and, where Latin was "a foreign language, there was a need for conventions which made it easier to read". Snippet at p. 12: marks "linking or separating words (hyphen or diastole)", apices, *positurae*; p. 13: *codices distincti*. +- *A point in Parkes's favour for this paper.* The "grammar of legibility" is a complex: word separation sits in a list with layout and punctuation. That is Parkes's own framing of problem 4. +- *Not established:* whether Parkes dates the Insular innovation, names scribal intent beyond the p. 1 snippet, or says anything on readers' numbers. I have no Parkes text on any of these. +- *Bearing:* H-cost **supports** (ease-of-reading motive, for new readers of a foreign language); problem 4 **supports** (a complex). Read the book before quoting. + +### A3. Saenger, "Silent Reading" (1982) + +**saenger1982silent** — Saenger, P. (1982). Silent reading: Its impact on late medieval script and society. *Viator, 13*, 367–414. https://doi.org/10.1484/J.VIATOR.2.301476 +- *Verification:* Crossref (row 14). Saenger's 1997 Preface gives the pages as 366–414 (p. x); Crossref gives 367–414. +- *Read status:* **not accessible.** Brepols redirected to a bot wall (row 8), although Unpaywall marks the article open ("bronze"). Only the description in the 1997 Preface (p. ix–x): the 1978 conference paper "outlined the thesis that the separation of words, which began in the early Middle Ages, altered the physiological process of reading and by the fourteenth century enabled the common practice of silent reading as we know it today", and the Preface says the article grew out of it. A human can open the article through the Brepols link in a browser. +- *Bearing:* none claimable. The book states the thesis in the same terms, so the book is the better text to quote. + +### A4. Wingo, *Latin Punctuation in the Classical Age* (1972) + +**wingo1972latin** — Wingo, E. O. (1972). *Latin punctuation in the classical age*. Mouton. https://doi.org/10.1515/9783110805215 +- *Verification:* Crossref record (De Gruyter e-book, 1972-12-31); archive.org lending item `latinpunctuation0000wing`. +- *Read status:* **not accessible.** The archive.org copy is lending-only and a guest cannot search inside (row 6); De Gruyter is closed (row 16). Second-hand only: Saenger's endnote to pp. 9–11 (p. 300) cites it for Latin separation before the end of the second century, and a later note (p. 301, for PSI 743) cites Wingo p. 15; I cannot read the page digits cleanly. +- *Bearing:* none claimable. This is the missing primary source for "when the Romans dropped the interpunct". Saenger's own text says the period is "scantily documented" (p. 10). + +### A5. Cavallo and Chartier (eds.), *A History of Reading in the West* (1999), with Parkes's and Saenger's chapters + +**cavallo1999history** — Cavallo, G., & Chartier, R. (Eds.). (1999). *A history of reading in the West*. Polity Press. Contains Parkes, "Reading, copying and interpreting a text in the early Middle Ages", and Saenger, "Reading in the later Middle Ages". +- *Verification:* Only through facet E (table of contents) and the archive.org lending record `historyofreading0000cava` (Polity, 2004 printing, ISBN 9780745630540). Chapter page ranges not verified. +- *Read status:* **not accessible.** Lending-only; the other archive.org item carrying the title is a mislabelled copy of Graff (row 4). +- *Bearing:* none claimable. These chapters are Parkes's and Saenger's later short statements and would show whether either moderated the thesis after the 1997 book. Priority for a library user. + +### A6. Tunbridge, DPhil thesis (1992) + +**tunbridge1992study** — Tunbridge, G. L. (1992). *A study of scribal practices in early Irish and Anglo-Saxon manuscripts* (DPhil thesis). University of Oxford. https://doi.org/10.5287/ora-ave2z6gqd +- *Verification:* DataCite record, resolving to `ora.ox.ac.uk/objects/uuid:8fc8f8de-0229-4f89-9816-e53bf8c6cc7e` (row 12). +- *Read status:* **read in part (card: library/tunbridge1992study.md)**; the facet first read the abstract only. Rights: all rights reserved. +- *What it argues (abstract):* "The presence of Latin as a non-vernacular Church language in Ireland and Anglo-Saxon England resulted in a dependence on and enthusiasm for the study of Latin grammar, and innovations in scribal practice must be seen in the context of this special linguistic environment." "Irish grammarians understood their own language in terms of syntactic groups rather than distinct parts of speech: this and other differences between Irish and Latin may have encouraged the practices of separation (and abbreviation) in the copying of Latin, as a means of making the latter easier to read." The eight-parts-of-speech teaching "underlies the practice of word separation, but a lack of explicit teaching on word boundaries accounts for the characteristic 'errors' of Insular separation." Abbreviation and separation "bring into the alphabetic system an ideogrammatic element which is symptomatic of a tendency to view Latin primarily as a graphical means of communication." +- *Bearing:* H-cost **supports** (the stated purpose is ease of reading, for a foreign-language readership). It **complicates** the "no training needed" line: the scribes needed grammatical training and made characteristic errors where teaching on word boundaries was lacking. Different method from Saenger (grammar teaching and manuscript counts), but not an independent test of Saenger's chronology: Parkes supervised the thesis (Acknowledgements) and its dating of the earliest manuscripts rests on Brown and Lowe (p. 194, n. 3). On motive the full text is firmer than the abstract's "may have": separation "must be seen as a practical outcome" of the grammarians' study (p. 98) and is "primarily a linguistic phenomenon" (Long Abstract p. 7). + +### A7. Saenger, "Orality and Visible Language" (2020) + +**saenger2020orality** — Saenger, P. (2020). Orality and visible language. In *The Oxford Handbook of Latin Palaeography* (pp. 692–704). Oxford University Press. https://doi.org/10.1093/oxfordhb/9780195336948.013.112 +- *Verification:* Crossref (row 14). +- *Read status:* **abstract only** (Crossref abstract). It restates the thesis in compressed form: "readers of codices first began to introduce visual elements to aid in reading, such as numbered sections and chapters, as well as word separation, marginal symbols, and graphic signs of various sorts"; "A shift towards visual access to the text occurred in the Middle Ages, allowing private understanding unmediated by aural cues"; the author "discusses the role of non-Latin speakers in this change". +- *Bearing:* **supports** the persistence of the bundle framing and the foreign-speaker motive in Saenger's later view; confirms nothing beyond the abstract. By this abstract, Saenger credits *readers* in late antiquity with the first aids, which sits oddly with the 1997 book's attribution to scribes (pp. 13, 72). Worth reading in full. + +### A8. Bronner et al. (2018) + +**bronner2018nonseparation** — Bronner, Busch, Fleischer, & Poppe (2018; given names not checked). (Non-)separation of words in early medieval Irish and German manuscripts and the concept "word". In *Empirical approaches to the phonological structure of words* (pp. 45–70). De Gruyter. https://doi.org/10.1515/9783110542899-003 +- *Verification:* Crossref (row 14); closed (row 16). +- *Read status:* **not accessible** (no abstract in Crossref). A recent study of how early Irish and German scribes separated words; a candidate test of Saenger's account on the manuscripts. Listed in (d). + +--- + +## (c) Synthesis + +### 1. The chronology, as far as the record I could read carries it + +| Period | What the page did | Source and depth | +|---|---|---| +| Earliest Greek inscriptions; Latin to c. 2nd c. | Words separated by mid-height points (interpuncts); later a hedera in inscriptions. Little extra space. | Saenger pp. 9–10, 26, 28–29 (full text); Wingo not read | +| Greek from soon after the earliest inscriptions; Latin from the 2nd c. (Saenger's dates vary, p. 10, p. 12, p. 69) | *Scriptio continua*, with vowel alphabet; reader-added marks; spaces for paragraphs; *per cola et commata* for the Vulgate; foliation in Greek codices from the 3rd c. | Saenger pp. 9–12, 16, 72, 77 | +| Late 7th c., Ireland | Aerated letter and syllable blocks, then word-separated Latin; first datable separated books before 692 and 713; probable Syriac model | Saenger pp. 34–35, 83 | +| 8th c., England | Separation "regularly inserted" in the earliest English books; grammarians Aldhelm, Bede, Tatwine, Boniface write on word shape | Saenger pp. 83–89 | +| 7th–11th c., Continent | Aerated "hierarchical word blocks"; Insular separation not adopted until the mid-10th c. | Saenger pp. 34, 36, 100 | +| c. 950–1050 | Protoscholastic and Arabic-science demand; separation spreads (Fleury, Reims, Chartres, Ripoll) but is, per Saenger, "still not a ubiquitous practice in western Europe" "prior to the middle decades of the eleventh century" (p. 177) and "standard and ubiquitous" "by the mid-eleventh century" (p. 176 [page to check in a library copy]) | Saenger pp. 120–126, 176–177 (partial) | +| 12th–13th c. | "Canonical separation" is the norm in northern Europe; aerated Italian books to the 13th c. | Saenger pp. 34, 243 [page to check in a library copy] | +| 14th–15th c. | Lay aristocrats read privately; Romance-vernacular function-word separation still lags (1500) | Saenger pp. 265–266 [page to check in a library copy], 273–277 | +| 16th–17th c. | Greek normal in Byzantium (second half of 16th c.); Cyrillic (c. 1600) | Saenger p. 13 | + +On the question of dated manuscripts, I found no work that tests Saenger's chronology independently of Saenger in the sources I could open (rows 13–15). Tunbridge (read in part, card) and Parkes (snippets) agree on the Insular origin and the date range, but Tunbridge was supervised by Parkes and takes the earliest dates from Brown and Lowe (p. 194, n. 3), so her thesis adds no independent test of the chronology (P2). Bronner et al. may be the nearest test. + +### 2. Who did it and why + +Insular scribes and teachers did it, in a Latin-learning setting where Latin was a foreign language. Saenger (pp. 83–97), Tunbridge (thesis, card) and Parkes's snippet agree on that. Two strands appear: an ease-of-reading motive (Parkes p. 1; Tunbridge, p. 98 and Long Abstract p. 7; Saenger pp. 94–95 [page to check in a library copy]) and a grammatical-teaching strand about word shape (Saenger pp. 83–89; Tunbridge). Parkes's "grammar of legibility" is Parkes's name for the second, wider complex. + +The intention is documented for the Anglo-Saxon grammarians, who wrote about it. For the first aerated books, "intention" is Saenger's inference ("an unconscious quest", p. 33). Whether the scribes aimed at *novices* or at *foreign-language learners* is not the same thing, and the sources I read point to learners of Latin, not to the general public. + +The Continental adoption was by a different set of actors and for a different reason: clerics and scholars who needed to read hard dialectical and scientific texts (Saenger pp. 120–124). The ancient non-adoption, in Saenger's account, was also a design choice by those who held the power to make it: an elite that handed reading to slaves and valued oral delivery (p. 11). + +### 3. Consequences, per Saenger + +See the table in A1. In short: his strongest, least hedged claims are physiological ("freed the intellectual faculties", p. 13) and scholastic ("without these tools would not have been possible", p. 122). His claims about private reading, heresy and erotic and devotional reading are causal in verb ("sparked", "encouraged", "stimulated", "emboldened") but rest on case evidence from elite individuals and books. The Reformation claim is explicitly hedged ("may have fostered", p. 276). He offers no evidence that spacing raised the number of readers; on p. 11 he lists "the greater diffusion of literacy throughout all social strata of the population" among the advantages of ease of reading that the modern age values. + +### 4. One move or a bundle? + +A bundle, on Saenger's own text and on Parkes's framing. Spacing arrived with, and in some cases after, word order conventions (pp. 14–15, 89–90), accents and small signs (pp. 84–86), syntactic punctuation (p. 15, 73), construction notes (pp. 69–71), glossaries and other genres (pp. 90–98), then tables, chapters, foliation and running heads (pp. 77, 259), and minuscule script with capitals (pp. 18–19). Saenger himself writes that "word order and word separation are thus intimately linked" and that he concentrates on separation for narrative reasons (p. 17) [page to check in a library copy]. Two further points weigh against "one move": (i) the order of arrival is mixed, since foliation and diagrams come after spacing and word-order conventions come earlier, (ii) the strongest consequence claim (scholasticism) is stated for spacing "combined with conventions for word order" (p. 122). The record cannot isolate the contribution of spacing alone. + +### 5. Readers before and after + +The readership widened in stages, and the stages are not one population. Ancient: lectors, slaves, trained elites. Insular: Latin learners. Continental monks: unchanged for three centuries. After 950: clerical scholars. After 1300: princes and urban elites. The record I read has nothing on peasants, women outside convents, or the broad urban population. That is a result: the evidence for "who could read because of spacing" stops at lay elites. + +### What the record supports + +- **H-cost: supported.** The page-level change moved word-boundary work from the reader to the scribe (Saenger pp. 13, 245). It lowered the cost (pp. 7, 13, 85) and widened the group in steps. It was motivated, at the Insular start, by ease of reading for learners of a foreign language (Parkes p. 1, Tunbridge p. 98, Saenger pp. 94–95 [page to check in a library copy]). +- **H-cause: not supported in its strong form, with one qualification.** Saenger calls his book "not a general history of literacy" (p. x). On p. 11 he lists "the greater diffusion of literacy throughout all social strata of the population" among the "advantages that modern readers perceive as accruing from ease of reading", which the ancients "seldom viewed as advantages". He attaches this to ease of reading in general and presents it as a modern value; he offers no evidence that spacing raised the number of readers. The sources I could read give no measure of numbers of readers. He does claim that spacing was a cause of rapid silent reading, reference reading, scholastic thought and private devotion among Latin readers, and these claims carry the "cause" language ("indispensable", "sparked"). Those are the claims the paper can use for H-cause, and they come with three limits: they rest on a cognitive model that facet B and D must test; the adoption story runs from demand to script as well as the reverse (p. 120); and they unfold over six hundred years. +- **Problem 4: a bundle.** See section 4. +- **Problem 1:** the record confirms that "ignited literacy" outruns the evidence in hand (adoption took about 260 to 350 years on the Continent, Saenger pp. 83, 100, 177: c. 690 to c. 950 is about 260 years, c. 690 to c. 1040 about 350; the book disclaims a general history of literacy). + +### Claims in the author's framing that the record contradicts or qualifies + +1. **"Converting scriptura continua to spaces between words … ignited literacy."** Saenger's book "is not a general history of literacy" (p. x). The nearest claim is that "even readers of modest intellectual capacity could read more swiftly" (p. 13), which is about speed among Latin readers. The record gives no evidence for a spread of literacy. (On p. 11 Saenger lists diffusion of literacy among the advantages of ease of reading that the modern age values; he gives no evidence that spacing produced it.) +2. **"A critical UX move … ignited" as a single, dated move.** The Irish innovation dates to the late seventh century; Continental adoption began in the mid-tenth century (p. 100) and was not general before the middle decades of the eleventh century (p. 177; St Gall, "middle to late eleventh century", p. 105, is a regional exception); canonical separation was the norm in the twelfth century (p. 243 [page to check in a library copy]); Romance vernaculars lagged to 1500 for function words (p. 266 [page to check in a library copy]); Greek and Cyrillic to the sixteenth and seventeenth centuries (p. 13). +3. **"Word separation did not train the populace to read *scriptio continua*; it changed the interface so that reading no longer required the training" (draft.md:263).** Partly borne out and partly not. Saenger says separation supplied "shortcuts" for skills that "an elite" had acquired through "a prolonged and arduous grammatical apprenticeship" (p. 85 [page to check in a library copy]), and that the work moved to the scribe (p. 13). But the scribes needed grammatical training and made characteristic errors where teaching on word boundaries was missing (Tunbridge, pp. 124–150); skilled readers of unspaced text adapt (Saenger p. 7, conceding, but maintaining that the cost remains [page to check in a library copy]; Mirault et al. in facet E). "Populace" does not fit: the readers were learners of Latin, then clerics, then lay elites. +4. **The move as a designed intervention for the many.** The record shows scribes serving a learner readership (Insular) and then scholars (Continental). Rome's elite chose not to simplify; there is no sign of anyone designing for "the populace". +5. **Silent reading as the product.** Saenger concedes that Ambrose read silently (p. 8) and allows that letters and documents may have been read "silently in the strict sense, or more probably, quietly with suppressed voice" (pp. 8–9). His claim is about *rapid* and *reference* silent reading. He says it was not possible on the Continent in the Carolingian aerated script (p. 39). +6. **Facet E's reading of Saenger p. 69.** E took "the late second or third century A.D." as a date for the Roman adoption of scriptio continua. The audit finds that p. 69 does date both the hyphen's entry into Latin and the Roman adoption of scriptura continua "in emulation of the Greeks" to "the late second or third century A.D.", so E's use is not a misreading; pp. 10 and 29 give "the second century A.D." and p. 12 gives "a century after Seneca's death in A.D. 65". Saenger's dates differ by passage. E's quotation "the ancient world had no desire to make reading easier and swifter" is a paraphrase; the sentence is "the ancient world did not possess the desire, characteristic of the modern age, to make reading easier and swifter" (p. 11). +7. **The scribe-as-designer intent.** For the first aerated books Saenger calls the motive "unconscious" (p. 33). Treat "scribes designed spacing to ease reading" as documented for the Anglo-Saxon grammarians and as inference for the earliest Irish scribes. + +--- + +## (d) Gaps and unverified leads + +**Sources I could not open, and where a library user finds them.** +- *Parkes (1992), Pause and Effect.* Snippets only (carried over from facet E). Library: Scolar Press 1992 or Univ. of California Press 1993; Routledge reprint 2016 (Taylor & Francis e-book). The pages that matter: pp. 1 (the purpose statement), 9–13 (antiquity, *codices distincti*), 20–29 (the grammar of legibility), 30–34 (Irish, Anglo-Saxon, Carolingian). archive.org: borrow `pauseeffectintro0000park` with a free account. +- *Wingo (1972).* Library: Mouton, or De Gruyter e-book (10.1515/9783110805215). archive.org: borrow `latinpunctuation0000wing`. Needed for the interpunct decline and the papyri. +- *Cavallo and Chartier (1999).* Parkes's and Saenger's chapters. archive.org: borrow `historyofreading0000cava`. +- *Saenger (1982), Viator 13.* Open the Brepols link in a browser (the lead may use one). +- *Tunbridge (1992).* Oxford Research Archive record above; request via the Bodleian if not downloadable. +- *Bronner et al. (2018).* De Gruyter, closed; library access. +- *Stutzmann (2020).* I could not identify the item. The lead should give the exact citation. Coulson, "Punctuation", and Galambos, "Premodern punctuation and layout" (see row 15) are candidates for related reading, not substitutes. +- *Saenger (2020), Oxford Handbook chapter.* Closed; abstract only. +- *Saenger's reviews* (Aronoff, Richter, Mulchahey, Hanna, Walker, Bäuml) are facet B. + +**Unread parts of Saenger (1997) that matter.** Chapters 8–13 (the abbey-by-abbey evidence for the Continental chronology; ch. 5's manuscript list), the Appendix tables (the surveyed manuscripts with space measures), and the endnotes (the psychology cited for the fixation figures). A library copy would also confirm every page locator above. + +**Gaps in the record itself.** +- No independent, dated-manuscript test of Saenger's chronology reached me. A specialist on Insular manuscripts (the Trinity College Dublin "Early Irish Hands" project came up in a search but I did not read it) may supply one. +- No source I could read dates the Roman shift beyond Saenger's "second century", and he calls the evidence "scantily documented" (p. 10). +- Parkes's and Saenger's later statements (Cavallo and Chartier 1999) are unread, so any softening after 1997 is unchecked. +- I found nothing on the size of the reading public at any date. +- Whether any scribe wrote that he was separating words to ease reading for novices, as opposed to for learners of Latin, I did not find in what I read (Bede's *De orthographia* and the Aldhelm passages concern the correctness of word division, pp. 85–86). + +**Methodological caveat.** The Saenger copy is an OCR of a scan from a shadow-library upload (row 3). The paper should cite the print edition only after the locators are checked against it. + +## Changes after full-text cards (2026-10-13) + +- `tunbridge1992study`: read status changed from "abstract only" to "read in part (card)" (A6, and the .bib note). +- `tunbridge1992study`: "independent of Saenger" replaced; the thesis was supervised by Parkes and its early dates come from Brown and Lowe (p. 194, n. 3), so it is not an independent test of Saenger's chronology (A6 bearing, dated-manuscripts paragraph). +- `tunbridge1992study`: motive statement made firmer than "may have" (p. 98; Long Abstract p. 7) (A6 bearing). +- `tunbridge1992study`: "Tunbridge abstract" citations in "Who did it and why", H-cost and claim 3 replaced with the card's page locators. + +Synthesis verdicts changed: none (H-cost remains "supported"; the dated-manuscript gap stands and is now stated more firmly). diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_B_the_dispute.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_B_the_dispute.bib new file mode 100644 index 00000000..a10b7aff --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_B_the_dispute.bib @@ -0,0 +1,361 @@ +% Facet B, the dispute over Saenger's thesis. Agent-drafted 2026-10-12. Every entry was checked against a database record (Crossref, OpenAlex, publisher page or journal page) unless the note says otherwise. + +@article{knox1968silent, + author = {Knox, Bernard M. W.}, + title = {Silent Reading in Antiquity}, + journal = {Greek, Roman, and Byzantine Studies}, + year = {1968}, + volume = {9}, + number = {4}, + pages = {421--435}, + url = {https://grbs.library.duke.edu/index.php/grbs/article/view/10731}, + note = {No DOI. Open access (journal PDF). Read in full 2026-10-12 (card: library/knox1968silent.md); page locators checked against the PDF.} +} + +@article{balogh1927voces, + author = {Balogh, Josef}, + title = {Voces paginarum: Beitr{\"a}ge zur Geschichte des lauten Lesens und Schreibens}, + journal = {Philologus}, + year = {1927}, + volume = {82}, + pages = {84--109, 202--240}, + note = {No DOI found. Citation confirmed in Knox 1968 p. 421 n. 1 and Heilmann 2020 bibliography. Not opened; known through Knox (read in full) and Heilmann.} +} + +@article{gavrilov1997techniques, + author = {Gavrilov, A. K.}, + title = {Techniques of Reading in Classical Antiquity}, + journal = {The Classical Quarterly}, + year = {1997}, + volume = {47}, + number = {1}, + pages = {56--73}, + doi = {10.1093/cq/47.1.56}, + note = {Crossref record verified. Abstract only (Crossref abstract, the paper's opening paragraph). Closed access (Unpaywall, OpenAlex). Not read in full; arguments known only through Johnson 2000 p. 593.} +} + +@article{burnyeat1997postscript, + author = {Burnyeat, M. F.}, + title = {Postscript on Silent Reading}, + journal = {The Classical Quarterly}, + year = {1997}, + volume = {47}, + number = {1}, + pages = {74--76}, + doi = {10.1093/cq/47.1.74}, + note = {Crossref record verified. Not accessible: no abstract, closed access. Known only through Johnson 2000 p. 593 ("Burnyeat on Ptolemy").} +} + +@article{johnson2000toward, + author = {Johnson, William A.}, + title = {Toward a Sociology of Reading in Classical Antiquity}, + journal = {American Journal of Philology}, + year = {2000}, + volume = {121}, + number = {4}, + pages = {593--627}, + doi = {10.1353/ajp.2000.0053}, + note = {Crossref record verified. Opening paragraph only (verbatim via Consensus; Project MUSE returned a verification wall). Closed access. Not read in full.} +} + +@book{johnson2010readers, + author = {Johnson, William A.}, + title = {Readers and Reading Culture in the High Roman Empire: A Study of Elite Communities}, + publisher = {Oxford University Press}, + year = {2010}, + doi = {10.1093/acprof:oso/9780195176407.001.0001}, + note = {Crossref book record verified (ch. 1, pp. 3--16; ch. 2 "Pragmatics of Reading", pp. 17--31). Not read; known only through the reviews of Ker 2012, Reis 2011 and Mannering 2012, all read.} +} + +@misc{ker2012johnson, + author = {Ker, James}, + title = {Review of William A. Johnson, Readers and Reading Culture in the High Roman Empire}, + howpublished = {Classical Journal Online 2012.02.03}, + year = {2012}, + url = {https://people.duke.edu/~wj25/pubs/CJ2012review.pdf}, + note = {Author-hosted copy on Johnson's Duke page. Read in full 2026-10-12 (card: library/ker2012johnson.md). No pagination.} +} + +@article{reis2011johnson, + author = {Reis, David M.}, + title = {Review of William A. Johnson, Readers and Reading Culture in the High Roman Empire}, + journal = {Religious Studies Review}, + year = {2011}, + volume = {37}, + number = {3}, + pages = {219--220}, + doi = {10.1111/j.1748-0922.2011.01538_4.x}, + note = {Crossref verified. Read in full (card: library/reis2011johnson.md) from the author-hosted copy (people.duke.edu/~wj25/pubs/RRC_Reis_review.pdf), about 300 words.} +} + +@misc{mannering2012johnson, + author = {Mannering, Jonathan}, + title = {Review of William A. Johnson, Readers and Reading Culture in the High Roman Empire}, + howpublished = {Bryn Mawr Classical Review 2012.11.37}, + year = {2012}, + url = {https://bmcr.brynmawr.edu/?p=38403}, + note = {Open access. Read in full (card: library/mannering2012johnson.md).} +} + +@article{vatri2012physiology, + author = {Vatri, Alessandro}, + title = {The Physiology of Ancient Greek Reading}, + journal = {The Classical Quarterly}, + year = {2012}, + volume = {62}, + number = {2}, + pages = {633--647}, + doi = {10.1017/S0009838812000213}, + note = {Crossref record verified (no abstract field). Opening paragraphs only (verbatim via Consensus) plus passages from pp. 638-639 and 646-647 quoted in Heilmann 2020 p. 181. Closed access. Not read in full. A third-party proxied PDF exists; not opened.} +} + +@article{mccutcheon2015silent, + author = {McCutcheon, R. W.}, + title = {Silent Reading in Antiquity and the Future History of the Book}, + journal = {Book History}, + year = {2015}, + volume = {18}, + number = {1}, + pages = {1--32}, + doi = {10.1353/bh.2015.0011}, + note = {Crossref record and abstract verified. Abstract only. Closed access. A copy on a personal website was found and not opened (unauthorised).} +} + +@incollection{heilmann2020reading, + author = {Heilmann, Jan}, + title = {Reading Early New Testament Manuscripts: Scriptio continua, ``Reading Aids'', and Other Characteristic Features}, + booktitle = {Material Aspects of Reading in Ancient and Medieval Cultures}, + publisher = {De Gruyter}, + year = {2020}, + pages = {177--196}, + doi = {10.1515/9783110639247-011}, + note = {Crossref verified. Open access, CC BY-NC-ND; read from the LMU Munich repository copy (epub.ub.uni-muenchen.de/75679). Read in full (card: library/heilmann2020reading.md) for the scriptio continua section (pp. 178-184) and conclusions; remainder by search.} +} + +@book{heilmann2021lesen, + author = {Heilmann, Jan}, + title = {Lesen in Antike und fr{\"u}hem Christentum: Kulturgeschichtliche, philologische sowie kognitionswissenschaftliche Perspektiven und deren Bedeutung f{\"u}r die neutestamentliche Exegese}, + publisher = {Narr Francke Attempto}, + address = {T{\"u}bingen}, + year = {2021}, + note = {Bibliographic data (707 pp., ISBN 978-3-7720-8729-5) taken from the record of Feldt 2023. No catalogue entry seen. Not accessible.} +} + +@article{feldt2023reading, + author = {Feldt, Laura}, + title = {Reading -- Silent or Aloud? Review of J. Heilmann, Lesen in Antike und fr{\"u}hem Christentum}, + journal = {The Classical Review}, + year = {2023}, + volume = {73}, + number = {2}, + pages = {688--690}, + doi = {10.1017/s0009840x23000033}, + note = {Crossref verified. First paragraph only (via Consensus, p. 688). Closed access.} +} + +@misc{flory2007silent, + author = {Flory, Stewart}, + title = {Silent Reading in Antiquity, More or Less?}, + howpublished = {Abstract, Classical Association of the Middle West and South annual meeting}, + year = {2007}, + url = {https://camws.org/meeting/2007/program/abstracts/07B5 Flory.htm}, + note = {Conference abstract. Content known only through a tool paraphrase (not verbatim); nothing quoted. No published version found.} +} + +@article{mollon2001scriptura, + author = {Mollon, J.}, + title = {Scriptura continua and the question of silent reading in antiquity}, + journal = {Perception}, + year = {2001}, + volume = {30}, + number = {Suppl. 1}, + doi = {10.1177/03010066010300s101}, + note = {ECVP 2001 abstract inside the supplement "Abstracts" (Crossref record verified for the supplement, pp. 1-123; the abstract's own page not located). Abstract text via Consensus, cut off before its end. Not peer reviewed in the usual sense.} +} + +@book{svenbro1993phrasikleia, + author = {Svenbro, Jesper}, + title = {Phrasikleia: An Anthropology of Reading in Ancient Greece}, + publisher = {Cornell University Press}, + year = {1993}, + note = {Existence confirmed by an archive.org lending record and by citations in Vatri 2012 p. 633 and Heilmann 2020. Not opened. Second-hand only (Vatri). The 1999 chapter "Archaic and Classical Greece: The Invention of Silent Reading", in Cavallo and Chartier, pp. 37-63, is in Heilmann's bibliography.} +} + +@book{cavallo1999history, + editor = {Cavallo, Guglielmo and Chartier, Roger}, + title = {A History of Reading in the West}, + publisher = {Polity Press}, + year = {1999}, + note = {Contents list seen via Consensus (Svenbro, Cavallo, Parkes, Hamesse, Saenger chapters). Not accessible; archive.org copies are lending or user uploads and were not opened.} +} + +@book{saenger1997space, + author = {Saenger, Paul}, + title = {Space Between Words: The Origins of Silent Reading}, + publisher = {Stanford University Press}, + year = {1997}, + doi = {10.1515/9781503619081}, + note = {Key reused from facet E. Facet A covers it. In this facet known through the reviews (Garrett and O'Donnell read in full; Aronoff abstract only), the publisher abstract and the critics. A z-library-style upload on archive.org was not opened.} +} + +@article{odonnell2000saenger, + author = {O'Donnell, James J.}, + title = {Review of Paul Saenger, Space Between Words: The Origins of Silent Reading}, + journal = {The Medieval Review}, + year = {2000}, + number = {00.02.24}, + url = {https://scholarworks.iu.edu/journals/index.php/tmr/article/view/14886}, + note = {Open access (JATS XML body at /view/14886/21004). Read in full 2026-10-12 (card: library/odonnell2000saenger.md); no page numbers, so cited by paragraph (4 paragraphs).} +} + +@article{odonnell1994parkes, + author = {O'Donnell, James J.}, + title = {Review of M. B. Parkes, Pause and Effect (II)}, + journal = {Bryn Mawr Classical Review}, + year = {1994}, + number = {1994.03.04}, + url = {https://bmcr.brynmawr.edu/1994/1994.03.04/}, + note = {Open access. Read in full (card: library/odonnell1994parkes.md); mentions Saenger's work as missing from Parkes; does not cite the 1982 Viator article.} +} + +@article{garrett1998saenger, + author = {Garrett, Jeffrey}, + title = {Review of Paul Henry Saenger, Space between Words: The Origins of Silent Reading}, + journal = {College \& Research Libraries}, + year = {1998}, + volume = {59}, + number = {5}, + pages = {486--488}, + doi = {10.5860/crl.59.5.486}, + note = {Crossref verified. Open access (crl.acrl.org). Read in full (card: library/garrett1998saenger.md). A review in a librarianship journal; endorses the thesis without engaging the classicists.} +} + +@article{aronoff2002saenger, + author = {Aronoff, Mark}, + title = {Review of Paul Saenger, Space between words: The origins of silent reading}, + journal = {Language in Society}, + year = {2002}, + volume = {31}, + number = {4}, + pages = {624--628}, + doi = {10.1017/s0047404502254056}, + note = {Crossref verified. Abstract only (Crossref abstract, verbatim). Closed access.} +} + +@article{hanna1999saenger, + author = {Hanna, Ralph}, + title = {Review of Paul Saenger, Space between Words: The Origins of Silent Reading}, + journal = {Speculum}, + year = {1999}, + volume = {74}, + number = {3}, + pages = {830--831}, + doi = {10.2307/2886848}, + note = {Crossref verified. Not accessible (JSTOR).} +} + +@article{richter2001saenger, + author = {Richter, Michael}, + title = {Review of Paul Saenger, Space Between Words: The Origins of Silent Reading}, + journal = {The American Historical Review}, + year = {2001}, + volume = {106}, + number = {2}, + pages = {627--628}, + doi = {10.1086/ahr/106.2.627}, + note = {Crossref verified (a second Crossref record, 10.2307/2651725, lists Richter and Saenger together). Not accessible.} +} + +@article{walker2000saenger, + author = {Walker, Thomas D.}, + title = {Review of Paul Saenger, Space between Words: The Origins of Silent Reading}, + journal = {The Library Quarterly}, + year = {2000}, + volume = {70}, + number = {1}, + pages = {166--167}, + doi = {10.1086/603182}, + note = {Crossref verified. Not accessible.} +} + +@article{mulchahey2001saenger, + author = {Mulchahey, M. Michele}, + title = {Review of Paul Saenger, Space Between Words: The Origins of Silent Reading}, + journal = {Canadian Journal of History}, + year = {2001}, + volume = {36}, + number = {2}, + pages = {323--325}, + doi = {10.3138/cjh.36.2.323}, + note = {Crossref verified (container title per Crossref; OpenAlex prints "Journal of History"). Not accessible.} +} + +@article{bauml2002saenger, + author = {B{\"a}uml, Franz H.}, + title = {Review of Saenger (1997): Space between words. The origins of silent reading}, + journal = {Written Language \& Literacy}, + year = {2002}, + volume = {5}, + number = {1}, + pages = {131--134}, + doi = {10.1075/wll.5.1.08bau}, + note = {Crossref verified. Not accessible (preview image only).} +} + +@article{lyons2001saenger, + author = {Lyons, M. C.}, + title = {Space between Words: The Origins of Silent Reading. By Paul Saenger}, + journal = {The European Legacy}, + year = {2001}, + volume = {6}, + number = {5}, + note = {OpenAlex record only; no DOI, no pages found. Existence not confirmed at Crossref. Not accessible.} +} + +@article{gilliard1993more, + author = {Gilliard, Frank D.}, + title = {More Silent Reading in Antiquity: Non omne verbum sonabat}, + journal = {Journal of Biblical Literature}, + year = {1993}, + volume = {112}, + number = {4}, + pages = {689}, + doi = {10.2307/3267404}, + annote = {Crossref registers the first page (689) only; the end page is unconfirmed (AUDIT_01). Check in print before use.}, + note = {Crossref verified. Not accessible (JSTOR paywall).} +} + +@article{slusser1992reading, + author = {Slusser, Michael}, + title = {Reading Silently in Antiquity}, + journal = {Journal of Biblical Literature}, + year = {1992}, + volume = {111}, + number = {3}, + pages = {499}, + doi = {10.2307/3267265}, + note = {Crossref verified. One-page note. Not accessible.} +} + +@article{torrensalvarez2022latin, + author = {Torrens-{\'A}lvarez, Mar{\'i}a Jes{\'u}s and Tuten, Donald}, + title = {From ``Latin'' to the Vernacular: Latin-Romance Hybridity, Scribal Competence, and Social Transformation in Medieval Castile}, + journal = {Speculum}, + year = {2022}, + volume = {97}, + number = {3}, + pages = {698--736}, + doi = {10.1086/720547}, + note = {Crossref verified. Read in part (card: library/torrensalvarez2022latin.md). Not about word separation; nothing in it bears on spacing.} +} + +@book{wingo1972latin, + author = {Wingo, E. O.}, + title = {Latin Punctuation in the Classical Age}, + series = {Janua Linguarum, Series Practica}, + number = {133}, + publisher = {Mouton}, + address = {The Hague}, + year = {1972}, + note = {Citation taken from Heilmann 2020 bibliography; the quotation from p. 15 appears at Heilmann 2020 p. 182. Book not opened. Needed for facet A and for Q4.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_B_the_dispute.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_B_the_dispute.md new file mode 100644 index 00000000..900e0310 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_B_the_dispute.md @@ -0,0 +1,355 @@ +# B — The dispute over Saenger's thesis: silent reading, word separation and what survives + +*Corrected on 2026-10-12 against `../audit/AUDIT_01.md`: 11 factual fixes.* + +*Facet B of the foundation-paper literature run. Agent-drafted on 2026-10-12. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** The reception of Paul Saenger's thesis (*Space Between Words*, 1997) by classicists, medievalists and book historians: Knox 1968, Gavrilov and Burnyeat 1997, Johnson 2000 and 2010, Vatri 2012, McCutcheon 2015, Heilmann (2020, 2021), the published reviews of Saenger's book, and the dissenters who side with Saenger (Flory, Mollon, Svenbro in part). Facet A covers what Saenger and Parkes say themselves; I quote them here only as the critics report them. + +**The question the synthesis answers.** What do the critics reject in Saenger (the claim about antiquity, the physiological mechanism, the causal link from spacing to silent reading, or the medieval story), what do they leave standing, and what does that do to H-cost and H-cause in `RESEARCH_PLAN.md` §3? + +**Legend for read status.** *Read in full* = I read the whole text this session from a legitimate open copy. *Read in part* = I read the pages named. *Abstract only* = Crossref abstract, publisher preview or database abstract. *Second-hand* = I know the work only through a named reviewer or critic who describes it; no quotation from the work itself is claimed. *Not accessible* = no text and no abstract. Page locators for the open-access PDFs I downloaded are the journal's own printed pagination (Knox, Heilmann) or the review's own pages (Garrett). Where a quotation comes from another scholar's quotation, the entry says so. + +**Headline, before the detail.** (1) The sources I could read agree that silent reading existed in antiquity, and that Knox (1968) showed it for documents and short texts; none of them says literary reading was *normally* silent. (2) The sharpest attack on Saenger is on the *mechanism* (Vatri 2012, as reported by Heilmann 2020), not on the medieval manuscript survey. (3) The one classicist review of Saenger I could read in full (O'Donnell 2000) praises the survey and flags exactly the causal step H-cause needs ("necessary and sufficient condition") as the part some would reject. (4) I found no source, pro or con, that measures when silent reading of literary texts became typical; the late-medieval dates are Saenger's own inference from manuscript layout. (5) My depth on Gavrilov, Burnyeat, Johnson (2000) and Vatri is the **opening paragraph or abstract only, plus what Johnson 2000 and Heilmann 2020 say about them**. None of those four was read in full. Every statement below about their arguments carries that limit. + +--- + +## (a) Search log + +All searches ran on 2026-10-12. Email for Unpaywall as in `org_frontier/research/oa_acquire.py`. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | Crossref `/works/{doi}` | Gavrilov 10.1093/cq/47.1.56; Burnyeat 10.1093/cq/47.1.74; Johnson 10.1353/ajp.2000.0053; Vatri 10.1017/S0009838812000213; McCutcheon 10.1353/bh.2015.0011; Gilliard 10.2307/3267404; Slusser 10.2307/3267265 | 7 of 7 resolved | Titles, containers, pages and years match the brief. Gavrilov and McCutcheon carry an abstract field; Vatri, Burnyeat, Johnson do not. | +| 2 | Unpaywall `v2/{doi}` | the same 7 DOIs | 0 open | All "closed". No repository copy. | +| 3 | OpenAlex `works/doi:` | the five main DOIs | 0 open | OpenAlex lists only the publisher DOI, JSTOR or Project MUSE landing pages. `any_repository_has_fulltext: False` for all five. | +| 4 | Crossref `query.bibliographic` and OpenAlex `search` | "Saenger Space Between Words Origins of Silent Reading review" (with and without a 1997–2005 date filter) | 9 reviews found | Hanna (*Speculum*), Richter (*AHR*), Walker (*Library Quarterly*), Mulchahey (*Canadian Journal of History*), Bäuml (*Written Language & Literacy*), Aronoff (*Language in Society*), Garrett (*College & Research Libraries*), Lyons (*The European Legacy*, OpenAlex only), a *Choice* notice (OpenAlex only). The earlier run named the first six and could read none. | +| 5 | Unpaywall on the review DOIs | 10.2307/2651725, 10.2307/2886848, 10.1086/603182, 10.3138/cjh.36.2.323, 10.1075/wll.5.1.08bau, 10.1017/s0047404502254056, 10.5860/crl.59.5.486 | 1 open | Garrett (C&RL) is open: `crl.acrl.org/index.php/crl/article/download/15238/16684`. Downloaded and read in full. | +| 6 | Medieval Review (IU ScholarWorks) site search | `…/tmr/search/search?query=Saenger+silent+reading` | 1 | "00.02.24, Saenger, Space Between Words", reviewed by James J. O'Donnell (2000). The JATS XML at `…/article/view/14886/21004` carries the full body. **Read in full.** | +| 7 | Bryn Mawr Classical Review site search | `bmcr.brynmawr.edu/?s=Saenger+silent+reading` and `?s=scriptio+continua` | 1 relevant hit | No BMCR review of *Space Between Words*. The hit is O'Donnell's BMCR 1994.03.04 review of Parkes's *Pause and Effect*, whose last paragraph says he misses Saenger's work. The `?s=scriptio+continua` page returned, but I did not parse its result list (see §d). | +| 8 | Cambridge Core and Project MUSE landing pages (WebFetch) | Gavrilov article page; Johnson 2000 via doi.org; Aronoff via doi.org; MUSE `/article/1288` and `/article/597277` | previews only | CUP pages give a one-paragraph preview. Both MUSE pages returned a "Verification required" challenge. I stopped there. | +| 9 | WebSearch | Gavrilov pdf; Vatri pdf and repositories; McCutcheon pdf; Saenger reply to critics; BMCR Saenger; BMCR Johnson 2010 | no legitimate full text | Surfaced (i) a Cambridge PDF of Vatri reachable only through a third library's EZproxy, (ii) a PDF of McCutcheon on a personal website, (iii) z-library-style uploads on archive.org (row 11). **I did not open any of these.** They are not publisher, repository or author-authorised copies, and the rule is not to get around a paywall. They are listed in §d for a library user. | +| 10 | OpenAlex `search` | "Heilmann Lesen in Antike und frühem Christentum"; "silent reading antiquity scriptio continua Saenger critique…"; "interpunct decline…"; "Hendrickson ancient reading"; "Balogh Voces paginarum"; "Svenbro Invention of Silent Reading"; Coleman; Carruthers; Jajdelska; Howley; Burnyeat | several | Found Heilmann 2020 (De Gruyter OA, hybrid), Feldt's *Classical Review* notice of Heilmann 2021 (closed; its first paragraph reachable via Consensus), Knox's OA record. No hit at all for a Saenger rejoinder or for the interpunct decline as a topic. | +| 11 | archive.org `advancedsearch` and `metadata/` | titles: "silent reading", "readers and reading culture", "phrasikleia", "space between words", "history of reading in the west", "pause and effect" | Saenger 1997, Svenbro, Cavallo & Chartier present | The Saenger item is a z-lib file upload; the Svenbro "2018 Cornell" item is a user upload with a CC license label I could not verify; two others are print-disabled lending items. I did not download or borrow any. No Johnson 2010 item. | +| 12 | Unpaywall for Heilmann 2020 | `10.1515/9783110639247-011` | open, CC BY-NC-ND | The De Gruyter PDF URL returned an empty body to curl (a bot wall). The LMU Munich repository copy (`epub.ub.uni-muenchen.de/75679/1/10.1515_9783110639247-011.pdf`, found in Unpaywall's second location) is the author's open copy. **Read in full.** | +| 13 | Johnson's own Duke page | `people.duke.edu/~wj25/pubs/` directory listing | 4 PDFs | Reviews of Johnson 2010 hosted by the author: Ker (*CJ*), Reis (*Religious Studies Review*), a *TLS* page (a blank e-paper stub, nothing readable), a *G&R* PDF (a review of someone else's book; irrelevant). Ker and Reis read in full. No copy of Johnson 2000 or 2010 itself. | +| 14 | BMCR 2012.11.37 (Mannering on Johnson 2010) | `bmcr.brynmawr.edu/?p=38403` | 1 | Read (the review is about 2,400 words; I read the opening, chapters 1–3 and the closing paragraphs through the text dump, not every line of chapters 4–9). | +| 15 | Consensus | (a) "Saenger word separation silent reading medieval critique…"; (b) "why did Romans abandon interpuncts scriptio continua…" | 20; 20 | (a) gave Vatri's and Johnson's opening paragraphs verbatim, Feldt's notice of Heilmann 2021, Mollon's ECVP abstract, the Aronoff abstract and what was presented as the book's publisher abstract [unverified: see AUDIT_01; the audit found no publisher abstract]. (b) was mostly noise (Latin phonology); useful only for Torrens-Álvarez and Tuten (a lead, abstract only) and Parkes's contents list. Two calls of the three allowed. | +| 16 | Scholar Gateway `semanticSearch` | "Did word separation make silent reading possible? Saenger critique classicists ancient silent reading" | 0 | **Failed:** `ACCESS_DENIED: Access denied`. Not retried. | +| 17 | WebFetch | CAMWS 2007 abstract for Flory | summary only | The tool returned a paraphrase, not verbatim text. I use it as a lead and do not quote it. | +| 18 | Knox 1968 (GRBS open PDF) | `grbs.library.duke.edu/index.php/grbs/article/download/10731/4297` | 15 pp. | Re-read in full this session to locate quotations by page. | + +--- + +## (b) Entries + +Entries run from the ancient evidence, through the classicists' replies and the dissenters, to the reviews of Saenger and the reception in other fields. + +### B1. Background: the old view and the first correction + +**balogh1927voces** — Balogh, J. (1927). Voces paginarum: Beiträge zur Geschichte des lauten Lesens und Schreibens. *Philologus, 82*, 84–109, 202–240. +- *Verification:* the citation is confirmed in Knox 1968, p. 421 n. 1, and in the bibliography of Heilmann 2020 (p. 192, "Philologus 82, 84–109.202–240"). No DOI. +- *Read status:* **second-hand** (Knox, in full; Heilmann). Not opened. +- *What it argues, per Knox:* Balogh's aim (p. 87 of the article) was to prove "dass das Altertum alles, was wir heute stumm für uns lesen, stets laut las" (Knox p. 421 n. 3). Knox: "Since the publication of Josef Balogh's article 'Voces Paginarum,' it has become standard doctrine that silent reading (and writing) was, if not completely unknown in the ancient world, at least so rare that whenever it was observed, it aroused astonishment, even suspicion" (p. 421). Heilmann (2020, p. 179) quotes Balogh: "Der Mensch des Altertums las und schrieb in der Regel laut; das Gegenteil war zwar nicht unerhört, doch immer eine Ausnahme" (Balogh 1927, 220). +- *Bearing:* the target of the whole classicist reply. It is the position Parkes (1992, p. 9, per facet E) still repeats. + +**knox1968silent** — Knox, B. M. W. (1968). Silent reading in antiquity. *Greek, Roman, and Byzantine Studies, 9*(4), 421–435. https://grbs.library.duke.edu/index.php/grbs/article/view/10731 +- *Verification:* journal page and OpenAlex (W record, diamond OA, 242 citations); the PDF is the journal's own and carries the original pagination 421–435. (O'Donnell 2000 cites 421–436; the PDF's last printed page is 435.) +- *Read status:* **read in full (card: library/knox1968silent.md)** (re-read this session; the card confirms the quotations, pages 421, 422, 427, 434–435, and that Knox never mentions word separation). +- *What it argues:* Balogh's evidence supports reading aloud of *literary* texts and no more. "The evidence assembled by Balogh (which should be supplemented by that of G. L. Hendrickson) does indeed make it perfectly clear that the normal way to read a literary text (non-literary texts are a different matter and will be dealt with later) was out loud, whether before an audience, in the company of friends or alone. But Balogh's insistence that silent reading was not just unusual but almost unheard of seems to go too far" (p. 421). The positive evidence: Cicero, *Tusc.* 5.116 on the deaf reading songs, with "multo maiorem percipi posse legendis his quam audiendis voluptatem" ("Cicero could not possibly have written the concluding phrase if silent reading of poetic texts had been impossible or even a nine-day wonder", p. 427); Antiphanes' *Sappho* riddle, which shows "silent reading of a letter in the presence of others was taken completely for granted" in fourth-century Athens (pp. 432–433); Euripides *Hipp.* 856ff and Aristophanes *Knights* 115ff, both silent reading on stage (p. 433); and the conclusion "These three passages … clearly demonstrate for fifth and fourth century Athens that silent reading of letters and oracles (and consequently of any short document) was taken completely for granted" (p. 434). Final: "Ancient books were normally read aloud, but there is nothing to show that silent reading of books was anything extraordinary except the famous passage from Augustine's *Confessions*, and that is countered by the phrase of Cicero" (p. 435). +- *Augustine and Ambrose:* Augustine's surprise is the surprise of "an African provincial from a poor family" at a man who had been "consular prefect" and educated at Rome; "there is no reason to suppose that Ambrose's silent reading would have excited such comment in the Italian imperial circles" (p. 422). Augustine "was quite capable of adopting it": at *Conf.* 8.12 he reads the Epistle to the Romans "in silentio" (p. 422). +- *Metaphor point:* Balogh's "acoustic" evidence (*pagina loquitur* and similar) is read literally; Theseus's letter "sings", "speaks" and "shouts aloud" in Euripides, "of course it does nothing of the sort; the audience has watched Theseus read the letter, and he did so in silence" (pp. 434–435). +- *What Knox does not do:* he never mentions word separation, spacing or *scriptio continua*; 1968 predates Saenger (1982, 1997). He also does not date when silent reading of books became ordinary. His "except the famous passage" concedes Augustine's report is one piece of evidence that silent reading of books *did* surprise someone, then discounts it by class and provenance. +- *Bearing:* **Refutes** "made possible" (supports the weak form of H-cost, not H-cause). **Complicates** any "made ordinary" wording: Knox's own conclusion is that *books* were normally read aloud (p. 435), so for literary reading "ordinary" silent reading is not shown for antiquity either way. + +### B2. The 1997 correction and its aftermath + +**gavrilov1997techniques** — Gavrilov, A. K. (1997). Techniques of reading in classical antiquity. *The Classical Quarterly, 47*(1), 56–73. https://doi.org/10.1093/cq/47.1.56 +- *Verification:* Crossref record (title, container, pages 56–73, issued May 1997); Cambridge Core landing page (WebFetch). Cited as "Gavrilov, Aleksandr K. (1997) … CQ 47, 56–73" in Heilmann 2020's bibliography. +- *Read status:* **abstract only** (Crossref abstract, which is the paper's opening). Unpaywall: closed; OpenAlex: JSTOR 639597 only. **Not read in full.** +- *What the abstract says (verbatim from Crossref):* "It has gradually become accepted among historians of ancient culture that the Greeks and Romans always, or nearly always, read aloud. They did not read to themselves, silently, save in rare and special cases. Either they were not able to read silently, or they felt no need to do so, or they did not enjoy doing it even when they were alone." The paper then attacks that view; the abstract does not state the attack. What the paper argues is known to me only from Johnson 2000, p. 593 (below): Gavrilov and Burnyeat "improved the evidential base, by refining interpretation (especially Gavrilov on Augustine) … and by adding observations from cognitive psychology." +- *Bearing:* supports silent reading existing. **I cannot say whether Gavrilov addresses word separation or Saenger.** Johnson's "observations from cognitive psychology" may bear on it; the claim is untested here. + +**burnyeat1997postscript** — Burnyeat, M. F. (1997). Postscript on silent reading. *The Classical Quarterly, 47*(1), 74–76. https://doi.org/10.1093/cq/47.1.74 +- *Verification:* Crossref (three pages, issued May 1997); Heilmann 2020 bibliography; Johnson 2000 p. 593. +- *Read status:* **not accessible** beyond Johnson's one-line characterisation. No abstract anywhere. +- *What it argues, per Johnson (2000, p. 593):* it supplies "neglected but important evidence (Burnyeat on Ptolemy)." Nothing else. +- *Bearing:* none I can verify beyond adding to the Knox–Gavrilov line. + +**johnson2000toward** — Johnson, W. A. (2000). Toward a sociology of reading in classical antiquity. *American Journal of Philology, 121*(4), 593–627. https://doi.org/10.1353/ajp.2000.0053 +- *Verification:* Crossref; MUSE record; Heilmann 2020 bibliography (593–627). +- *Read status:* **opening paragraph only** (verbatim via Consensus; the MUSE page was a verification wall and the article is closed everywhere). **Not read in full.** Heilmann 2020 names Johnson 1994, 2000, 2004 and 2009 at pp. 177 and 190 (n. 82, n. 85) as holding that reader's aids served reading aloud and that line length served performance; he argues against that position (see below). +- *What it says in the opening paragraph (p. 593; OCR hyphenation removed):* "In the last century, scholarly debate on ancient reading has largely revolved around the question 'Did the ancient Greeks and Romans read aloud or silently?' Given the recent work of Gavrilov and Burnyeat, which has set the debate on new, seemingly firmer, footing, the question is at first glance easily answered. Without hesitation we can now assert that there was no cognitive difficulty when fully literate ancient readers wished to read silently to themselves, and that the cognitive act of silent reading was neither extraordinary nor noticeably unusual in antiquity. This conclusion has been known to careful readers since at least 1968, when Bernard Knox demonstrated beyond reasonable doubt that the silent reading of ancient documentary texts, including letters, is accepted by ancient witnesses as an ordinary event. Gavrilov and Burnyeat have improved the evidential base, by refining interpretation (especially Gavrilov on Augustine), by focusing on neglected but important evidence (Burnyeat on Ptolemy), and by adding observations from cognitive psychology. The resulting clarity is salutary. Yet I suspect many will be dissatisfied with the terms in which the debate has been couched." +- *Two things the paragraph does and does not do.* (i) It is the source of the "closed" characterisation in `REUSE.md`; the word "closed" is the plan's, Johnson says "easily answered" and "seemingly firmer". (ii) It names *documentary* texts for Knox ("silent reading of ancient documentary texts, including letters"), which matches Knox's own limit. Johnson does not say literary reading was silent. +- *Johnson's pivot (second-hand, via reviewers of his 2010 book):* the question should be social, "the negotiated construction of meaning within a particular sociocultural context" (Reis 2011, quoting Johnson 2010, p. 12; Ker 2012 gives it as p. 12, emphasis in the original). +- *Bearing:* **Supports** that "impossible" is gone as a position among the scholars Johnson cites. **Complicates** the plan's "closed question": Heilmann 2020 (B3) says the opposite about the field, ten years later (see §c, Q1). + +**johnson2010readers** — Johnson, W. A. (2010). *Readers and reading culture in the High Roman Empire: A study of elite communities*. Oxford University Press. https://doi.org/10.1093/acprof:oso/9780195176407.001.0001 +- *Verification:* Crossref (book and chapter records: ch. 1 "Reading as a Sociocultural System" pp. 3–16; ch. 2 "Pragmatics of Reading" pp. 17–31); three published reviews (below). +- *Read status:* **second-hand only.** No full text and no free chapter found. Chapter 2 is the one that bears on *scriptio continua*, and I have it only through the reviewers. +- *What it argues, per the reviewers:* + - Ker (*CJ* 2012): chapter 2 shows "that certain general characteristics of ancient reading and ancient bookrolls, such as the absence of spaces between words (scriptio continua), are not signs of a technological lack, but are consistent with an elite culture in which reading from a bookroll with full articulation and comprehension is a marker of education and status, and often an act of connoisseurship." Johnson's purpose in treating Ambrose "is less to prove that silent reading is known much earlier … than to draw attention to Augustine's expectations and the reasons for his surprise at Ambrose's habit." + - Reis (*Religious Studies Review* 2011): "Johnson shows that the scriptio continua of the bookroll and its costly aesthetic features ensured that the consumers of texts were limited. Yet restricting the book's accessibility was precisely the point for elites who used reading to create and reaffirm the values of high society." + - Mannering (BMCR 2012.11.37): "The considerable effort in apprehending scriptio continua with only minimal punctuation is interpreted by Johnson as a hallmark of the exclusivity of literary culture as social institution," and the bookroll needed "years of training required for its proper use." Chapter 1 moves "beyond speculation on the neurophysiology of reading, the cognitive act having been settled by scholarly consensus as essentially the same for the ancients as for us today." +- *Bearing:* **Supports** the social half of Saenger's own line, "the ancient world did not possess the desire, characteristic of the modern age, to make reading easier and swifter" (Saenger 1997, p. 11), and so Q4's answer. It **complicates** H-cost's premise that spacing *lowered* a cost: the reviewers describe Johnson as treating the cost as real (Mannering: "considerable effort", "years of training") and valued. That is nearer the author's sentence ("the interface … required the training") than anything in the physiological critics. All of this rests on three reviewers and must be checked against chapter 2. + +**ker2012johnson** — Ker, J. (2012). Review of W. A. Johnson, *Readers and Reading Culture in the High Roman Empire*. *Classical Journal Online* 2012.02.03. Author's copy at `people.duke.edu/~wj25/pubs/CJ2012review.pdf`. **Read in full (card: library/ker2012johnson.md)** (about 940 words, no pagination). + +**reis2011johnson** — Reis, D. M. (2011). Review of W. A. Johnson, *Readers and Reading Culture in the High Roman Empire*. *Religious Studies Review, 37*. https://doi.org/10.1111/j.1748-0922.2011.01538_4.x. Author-hosted copy. **Read in full (card: library/reis2011johnson.md)** (about 300 words). + +**mannering2012johnson** — Mannering, J. (2012). Review of W. A. Johnson, *Readers and Reading Culture in the High Roman Empire*. *Bryn Mawr Classical Review* 2012.11.37. https://bmcr.brynmawr.edu/?p=38403. **Read in full (card: library/mannering2012johnson.md)** (the whole review; the facet first read the opening, chapters 1–3 and the conclusion). The full text confirms the facet's reading of chapters 1 and 2 and shows the review never mentions Saenger, spacing as a medieval change, or Knox. + +**vatri2012physiology** — Vatri, A. (2012). The physiology of ancient Greek reading. *The Classical Quarterly, 62*(2), 633–647. https://doi.org/10.1017/S0009838812000213 +- *Verification:* Crossref (no abstract field); Heilmann 2020 bibliography; JSTOR 23470128. +- *Read status:* **opening paragraphs only** (verbatim via Consensus), **plus what Heilmann 2020 quotes from pp. 638–639 and 646–647**. **Not read in full.** The one open-looking PDF is reachable only through another library's proxy; I did not use it. +- *What the opening paragraphs say (p. 633):* "In the 1980s and 1990s, the medievalist Paul Saenger argued in several publications that it was physiologically impossible for the ancient Greeks and Romans to read silently. His contributions revived an age-old controversy, which was first raised by Balogh in the early twentieth century and seemed to have been settled in 1968 by Knox, who concluded that silent reading was possible and not even unusual in antiquity. Unlike his predecessors, however, Saenger did not base his arguments on the interpretation of the literary sources for ancient reading, but tackled the question from the point of view of cognitive psychology. He developed a detailed model of the processes involved in reading ancient scriptura continua, claiming that the very lack of word separation made vocal rehearsal necessary for the readers to recognize the words. Classical sources hinting at silent reading, in his view, might rather refer to silent, or sub-vocal, oral reading – a sort of mumbling. Since their first appearance, Saenger's ideas have been met with partial acceptance or more or less harsh criticism by classicists. Svenbro, for example, acknowledges that scriptura continua hindered silent reading. However, he also maintains that the Greeks were able to practise it because of a change of attitude towards the written word." +- *What Heilmann 2020 reports Vatri as concluding (quotations are Heilmann's, p. 181, from Vatri pp. 638–639 and 646–647):* "the readers' deep-rooted habits play a major role, and this must also have applied to the ancient readers of scriptura continua. […] [T]here is no reason to assume that reading unspaced text is a particularly demanding cognitive task in itself, and Saenger's model must be rejected." And: "No physiological constraints prevent the Greeks from reading silently." Heilmann also reports that Vatri rests on "recent cross-cultural comparative studies" (Kajii et al. 2001; Sainio et al. 2007; Bai et al. 2008) showing "that the addition of spaces between words is redundant and can even disrupt the reading process" (this sentence is Heilmann's own summary of those studies, p. 181, n. 22, citing Vatri 2012, 638f. and others; it is not a quotation from Vatri), on Thai eye movements ("exactly the same as those of English readers", Winskel et al. 2009, via Vatri p. 639), and on Vatri's own frequency analysis of Greek word beginnings and endings (p. 181). +- *Bearing:* **Refutes** the physiological mechanism (the "made possible" version of H-cause, and the strong form of H-cost), on the account I can verify only second-hand. **Complicates** my own use: Vatri frames Saenger's claim as "physiologically impossible"; Saenger's book (per facet E, p. 8) says Ambrose read silently but not rapidly. Vatri's opening also states Saenger's real claim as *vocal rehearsal*, "a sort of mumbling", i.e. a claim about a mode of reading, not a prohibition. + +**mccutcheon2015silent** — McCutcheon, R. W. (2015). Silent reading in antiquity and the future history of the book. *Book History, 18*(1), 1–32. https://doi.org/10.1353/bh.2015.0011 +- *Verification:* Crossref; MUSE record. +- *Read status:* **abstract only** (Crossref). A PDF exists on a personal website (found by WebSearch); I did not open it (unauthorised copy). +- *What the abstract says (verbatim):* "Book historians have long insisted that silent reading was a rare or nonexistent practice in classical antiquity. This belief runs counter to much research in the field of Classics. In this paper, I offer a critical review of the evidence and the scholarship concerning reading in antiquity before offering an explanation for these conflicting views on the subject. I argue that this debate reveals much about the epistemology of book history as a discipline." +- *Bearing:* **Supports** the claim that the two fields never converged: the Classics rejection did not reach book history. That sets a cap on how far the plan's "closed" can be pushed. + +**heilmann2020reading** — Heilmann, J. (2020). Reading early New Testament manuscripts: *Scriptio continua*, "reading aids", and other characteristic features. In *Material aspects of reading in ancient and medieval cultures* (pp. 177–196). De Gruyter. https://doi.org/10.1515/9783110639247-011 (CC BY-NC-ND; read from the LMU Munich repository copy) +- *Verification:* Crossref (title, container, pages 177–196, 5 May 2020, De Gruyter); Unpaywall open locations; the PDF's own running heads (the part on *scriptio continua* is pp. 178–183). +- *Read status:* **read in full (card: library/heilmann2020reading.md)** for the *scriptio continua* section (pp. 178–184) and its notes, and the conclusions (pp. 190–191); the rest (diacritics, paragraphoi, nomina sacra) was read by search, not line by line. The card confirms the page locators used here (pp. 178–183, 190). Addition from the card: Heilmann relies on Vatri for the Thai result and does not cite the 2009 study's sentence-level slowing, which Chantavarin and Leung (2025, `chantavarin2025effects`) report. +- *Why it matters:* it is the only post-2012 source I could read in full that argues the case against Saenger's mechanism in print, and it supplies the Latin-side evidence (interpuncts, Vindolanda) that bears on Q4. Caveat: Heilmann is writing here on New Testament manuscripts and arguing his own thesis (that Christian papyri's reader's aids were not for public reading), so he has a stake in dismantling the "needs reading aloud" premise. His treatment of Saenger is a few pages, in service of that. +- *What it argues (all quotations are from the pages cited):* + - Against Saenger's proof structure, p. 178: "The most elaborated defense of this view can be found in P. Saenger's book 'Space between Words: The Origins of Silent Reading,' and it is a view commonly accepted in Classical Studies as well as New Testament Studies. However, the arguments of P. Saenger and others for the thesis that texts in scriptio continua were designed to be read aloud cannot carry the burden of proof for the proposed interrelationship. On the contrary, they presuppose precisely what must be proved." + - On the state of the field, p. 179: reading aloud is "still taken for granted today"; "Well-argued counterproposals (Knox 1968; Gavrilov 1997; Burnyeat 1997; Svenbro 1999, 2002, 2005; Burfeind 2002) have not been accepted against this standard view, which has become a quasi-dogma." (Counterproposals named in n. 13; the list of those who accept the standard view is n. 8, pp. 178.) + - On evidence generally, p. 180: "the sources do not allow for statistical generalizations as one can find them as the basis of the common opinion outlined above. In unmarked cases, one cannot clearly decide if someone read with or without his voice." And a distinction (p. 179) between "use of the voice" and "volume/acoustic perceptibility", so that the silent/aloud binary is "too narrow". + - On the mechanism, pp. 180–181: Heilmann concedes the premise, "average readers socialized in modern 'western' scripts read English texts in scriptio continua slower and … the PVL shifts from the middle of the word to the beginning of the word" (p. 180), then calls the inference "methodologically problematic in that this research data is based exclusively on test subjects that have been socialized in modern 'western' scripts and untrained in reading scriptio continua" (p. 180). He cites Epelboim et al. 1994/1997 as showing "one can train the eye to read scriptio continua and that spaces between words are not the primary factor controlling eye movement" (pp. 180–181). **He also lists, in n. 20, "the controversy between Epelboim/Booth/Steinman 1996 … and Rayner/Pollatsek 1996"** (Rayner and Pollatsek's reply is titled "Reading Unspaced Text Is Not Easy"). So the eye-movement literature is itself contested, and Heilmann's presentation is one side of it; facet D should settle it. + - On the ancient sources, pp. 181–182: those cited to prove difficulty "should rather be interpreted within the context of the process of learning to read and/or improving reading ability" (Dion. Hal. *Comp.* 25; Quint. *Inst.* 1.1.34; Lucian *Adv. ind.* 2; Petron. *Sat.* 75.4, n. 28), or discourses about illiteracy (Gell. 13.31); "some … refer to Latin texts, which were not written in scriptio continua until the early imperial period" (p. 182); scribal errors "do not prove that reading scriptio continua would have been particularly demanding" (p. 182). + - On the Roman change, p. 182: "Moreover, the fact that the Romans replaced their habit of writing scriptio discontinua with scriptio continua in the second century CE shows that scriptio continua is not a characteristic of deficient writing systems." He cites Oliver (1951) for noticing the late introduction in Latin and Wingo (1972), quoting Wingo p. 15: "The word-divider is regularly found on all good inscriptions, in papyri, on wax tablets, and even in graffiti from the earliest Republican times through the Golden Age and well into the Second Century." (Oliver's full reference is not in Heilmann's bibliography; I could not find it.) + - Spacing in Roman use, n. 37 (pp. 182–183): "The transition from writing with frequent word separations to scriptio continua is documented impressively by the heterogeneous evidence of the Vindolanda tablets (around the turn from 1st/2nd century). … in T. Vindol. II 297; 315; 323; 345 almost all words are separated from each other with a middle dot, in many other tablets the words are separated with spaces [!] (e.g. T. Vindol. II 296; 299; 301; 316; 343); only a few (e.g. T. Vindol. II 292) are in scriptio continua." Heilmann's own bracketed [!] marks the surprise. I could not check the tablets. + - On attitude, p. 183: Suetonius (*Aug.* 87.3) "finds it peculiar that Augustus did not separate individual words" ("non dividit verba"); Seneca (*Ep.* 40.10–11) "associates Greek not with the difficulties of deciphering scriptio continua, but with speed." Summary, p. 183: "there is no need to assume that reading unspaced scripts was any more cognitively challenging for ancient readers than reading spaced scripts is for us. We should be careful not to project our difficulties with reading scriptio continua onto ancient readers." + - Conclusion on aids, p. 190: the diacritics etc. "all functioned to assist in reading without vocalization." +- *Weak points I can see:* (i) The inference from the second-century Roman shift to "scriptio continua is not a characteristic of deficient writing systems" does not refute Saenger's *social* explanation; a community can drop a useful aid for reasons of status. Heilmann argues against a physiological deficiency, which is not the same claim. (ii) Heilmann nowhere shows that spaces would not have lowered the cost for *beginners*; his own n. 28 reads the difficulty evidence as about learners, which is the cost-for-novices claim H-cost makes. (iii) His n. 8 list of works that accept the standard view mixes authors who hold only Knox's position (literary reading mostly aloud) with those who hold Saenger's mechanism, and he does not separate them. That is my reading of the note, not his statement. +- *Bearing:* **Refutes** (on his account) the physiological premise of Saenger's thesis, so **refutes the strong form of H-cause and strong H-cost**. **Complicates** H-cost's weak form: he concedes slowing for untrained moderns and reads the ancient difficulty-sources as about learners. **Complicates Problem 1** by showing the field has not settled: the consensus Johnson announced in 2000 is not what Heilmann sees in 2020. + +**heilmann2021lesen** — Heilmann, J. (2021). *Lesen in Antike und frühem Christentum: Kulturgeschichtliche, philologische sowie kognitionswissenschaftliche Perspektiven und deren Bedeutung für die neutestamentliche Exegese*. Narr Francke Attempto. ISBN 978-3-7720-8729-5, 707 pp. (as printed in the *Classical Review* notice's heading) +- *Verification:* the title and publisher string are in the OpenAlex and Crossref record of Feldt's review (10.1017/s0009840x23000033). I have not seen a catalogue entry for the book. +- *Read status:* **not accessible**; known only through Feldt (below) and through Heilmann 2020, which says he treats the sources in detail "in my habilitation thesis" (p. 182) and which I take to be this book. +- *Bearing:* the likeliest place for the full argument against Saenger's mechanism and the fullest recent catalogue of evidence for silent reading (Feldt: lists of "the evidence for silent reading" in an appendix). A library user should fetch it first. + +**feldt2023reading** — Feldt, L. (2023). Reading – silent or aloud? Review of J. Heilmann, *Lesen in Antike und frühem Christentum*. *The Classical Review, 73*(2), 688–690. https://doi.org/10.1017/s0009840x23000033 +- *Verification:* Crossref (pages 688–690, 21 Feb 2023). +- *Read status:* **first paragraph only** (via Consensus; p. 688); the rest is behind the publisher. +- *What it says (p. 688):* the book "changes influential assumptions about reading in antiquity"; "In research on ancient reading it has been common knowledge that ancient literature was read aloud. This is the key idea that H. challenges in a set of convincing analyses. Moreover, and persuasively, he offers analyses of several other empirical sources that demonstrate that individual, quiet reading was much more common than previously thought." The book includes "a stimulating discussion of the typographical qualities of scriptio continua up against contemporary cognitive and neuroscientific studies of reading processes." +- *Bearing:* **Supports** reading Heilmann 2020 as a fair summary of his position, and confirms that in 2023 a classicist reviewer still calls reading-aloud-of-literature "common knowledge". That bears on Q1. + +### B3. Dissenters who side with Saenger, in whole or in part + +**flory2007silent** — Flory, S. (2007). Silent reading in antiquity, more or less? Abstract, annual meeting of the Classical Association of the Middle West and South. `camws.org/meeting/2007/program/abstracts/07B5 Flory.htm` +- *Read status:* **abstract only, via a tool summary, not verbatim.** The WebFetch tool declined to reproduce the abstract and returned a paraphrase. I therefore quote nothing. +- *What the paraphrase says:* Flory argues that true silent (visual) reading was not possible in Greece and Rome; reads the Knox and Johnson passages (*Hipp.*, *Conf.*) as subvocalisation; cites Saenger's lack-of-spacing argument. A conference abstract, not a published paper. I found no later published version. +- *Bearing:* shows the Saenger position still has an advocate among classicists in 2007, but with no printed argument to weigh. + +**mollon2001scriptura** — Mollon, J. (2001). Scriptura continua and the question of silent reading in antiquity. In *Abstracts, ECVP 2001*, *Perception, 30*(Suppl. 1), 1–123. https://doi.org/10.1177/03010066010300s101 +- *Verification:* Crossref record for the supplement ("Abstracts", pp. 1–123); abstract text via Consensus. +- *Read status:* **abstract only**, and the Consensus copy cuts off after "Such scripts may limit readers to a single route of lexical access." +- *What it says:* "There is scattered textual evidence that silent reading was unknown to the sophisticated and literate civilisation that built the theatre of Ephesus and indeed was not practised in the West until the mediaeval period … Can we relate the absence of silent reading to the remarkably late invention of spaces between words? An authoritative survey by Paul Saenger suggests that the first appearance of such spaces in Western Europe was in Ireland in the seventh century AD." +- *Bearing:* a vision scientist (affiliation per abstract: Experimental Psychology, Cambridge) who accepts the thesis. A conference abstract, not peer-reviewed in the usual sense. It shows that the thesis crossed into perception research as a given. + +**svenbro1993phrasikleia** — Svenbro, J. (1993). *Phrasikleia: An anthropology of reading in ancient Greece* (J. Lloyd, Trans.). Cornell University Press. Also Svenbro (1999), Archaic and Classical Greece: The invention of silent reading, in Cavallo & Chartier (Eds.), *A history of reading in the West* (pp. 37–63). +- *Verification:* archive.org lists a print-disabled lending record titled "Phrasikleia : an anthropology of reading in ancient Greece"; Vatri's opening paragraph and Heilmann's bibliography cite the 1999 chapter and the German edition (2005). +- *Read status:* **second-hand only** (Vatri p. 633; Heilmann n. 13). +- *What it argues, per Vatri:* Svenbro "acknowledges that scriptura continua hindered silent reading" but "maintains that the Greeks were able to practise it because of a change of attitude towards the written word." Vatri adds his own judgement: "Ingeniously, but questionably, he argues that silent reading implies a passive attitude for the reader, who relates to the text as the spectator in a theatre." +- *Bearing:* the one classicist who *grants* Saenger's cost premise but moves the cause from the page to attitude. **Complicates** H-cause: a causal route in antiquity that runs through a change of attitude, not spacing. Read Svenbro before the claim is used. + +### B4. The reviews of *Space Between Words* + +Only one review below reads as a classicist's: O'Donnell's (University of Pennsylvania), which engages *Confessiones* 6.3.3 and 8.12.29 directly and appeared in a medievalists' journal. The medievalists' reviews are the reception that matters for the medieval half of the thesis, and I could read only two of the nine. + +**odonnell2000saenger** — O'Donnell, J. J. (2000). Review of P. Saenger, *Space between words*. *The Medieval Review* 00.02.24. https://scholarworks.iu.edu/journals/index.php/tmr/article/view/14886 (JATS XML body at `…/view/14886/21004`) +- *Verification:* IU ScholarWorks search and article page; the XML carries the author's name, "University of Pennsylvania", year 2000 and the copyright notice. His BMCR review of Parkes (1994.03.04) is the same author. +- *Read status:* **read in full (card: library/odonnell2000saenger.md)** (about 830 words, four paragraphs; no page numbers, so I locate by paragraph, ¶1–¶4). The card confirms the ¶2 and ¶3 quotations, including that "commonplace" in ¶2 is O'Donnell's own claim. +- *What it argues (all quotations verbatim):* + - ¶2, on the book's question: "the ancient question of when and how silent reading became a common practice. … The easiest half-answer is that silent reading was invented by Ambrose of Milan and observed by Augustine of Hippo (Confessiones 6.3.3 …). But already thirty years ago, Bernard Knox … showed that the practice was attested long before Augustine. At the same time, it is clear that it did not become commonplace until long after." (The last sentence is O'Donnell's own assertion; he gives no evidence for "commonplace" here.) + - ¶3, O'Donnell's statement of "the short form of Saenger's answer", not his own endorsement: "It was when, in the later middle ages, regular spacing of words on the written page became commonplace, that the practice of a purely visual and private reading became so easy as to become widespread." + - ¶3, the key passage for H-cause: "It is possible, perhaps likely, that some will dissent from the particular argument that Saenger structures—that is, his claim that the textual practices he observes were the necessary and sufficient condition of the rise of silent reading. But that argument is, for all its power, a superstructure erected on the vast bulk of this work, like a dizzying spire looming over a richly constructed Gothic cathedral. Had the spire never been built, the cathedral would have been the poorer—but some observers might never have noticed the lack." + - ¶4, on the weak point: "Where he has left himself most vulnerable is in a brief initial chapter on the physiology of reading practices. … But the stretch from that survey to the manuscript practices he outlines and then to the theory about reading practices he constructs is the most venturesome and vulnerable part of the book. It will surely be tested." + - ¶4, a prediction: the value of the book "does not lie in the outcome of that debate (one which I expect to see resolved in Saenger's favor)". "That debate" follows his sentence that the leap from the physiology survey to the theory "will surely be tested", so the prediction concerns the physiology-to-practice step. Vatri (2012) and Heilmann (2020) (B2) reject the physiology O'Donnell expected to be vindicated; the eye-movement literature is itself contested (Heilmann n. 20). + - ¶3 also credits the Irish source: "The fons et origo in some ways is Virgilius Maro, the Irish grammarian of the seventh century, who theorized the 'word' in new ways and showed a practice compatible with that theory." +- *Bearing:* **the best single source on the shape of the reception.** The survey stands ("cathedral"); the causal claim is the "spire"; physiology is flagged as the weak foundation; and the reviewer already uses "necessary and sufficient condition" as Saenger's claim. **Supports** a version of H-cost as "the practices made silent reading easy"; **refutes** nothing about H-cause directly but names it as the contestable step. Note that the wording "so easy as to become widespread" is his summary of Saenger's answer, not his own endorsement, and is *H-cost plus diffusion*, not *H-cause for literacy*. And Q3: he gives no measurement of "widespread". + +**garrett1998saenger** — Garrett, J. (1998). Review of P. H. Saenger, *Space between words*. *College & Research Libraries, 59*(5), 486–488. https://doi.org/10.5860/crl.59.5.486 (open access) +- *Verification:* Crossref; Unpaywall (publisher version, open). +- *Read status:* **read in full (card: library/garrett1998saenger.md)** (the page and a half of review text; the card confirms the strong-form restatement, p. 487 for the co-constituent passage and p. 488 for the continuum and screens sentences). +- *What it argues:* an uncritical librarian's endorsement. It restates Saenger's thesis in strong form: "Space between words, as Saenger documents, was introduced into medieval manuscripts to serve as a co-constituent of meaning, a visual cue allowing the rapid parsing of units of meaning at the sentence, word, and morphemic level. Without this space, text must be separated by the reader at considerable cognitive expense, usually by repeated passes through the text …" (p. 487); Saenger's book is about "the emergence of (rapidly) readable text" and "the spaces between words and their revolutionary implications for modern reading, which is silent, fast, and discontinuous" (p. 487). Garrett also calls the scribal changes "technological devices improving reading efficiency" (p. 487) and credits Saenger with showing "the evolution of writing, print, and computing not as a succession of fitful revolutions but, rather, as a continuum of technological innovation," adding "Many of the issues we face today (e.g., the readability of extended texts on scrollable computer screens) had their roots in the ruminations of Irish scribes of the seventh century" (p. 488). It makes no use of the classicist debate. +- *Bearing:* **Supports** nothing as evidence; it is *reception data*: a 1998 reader outside the field takes the cost-and-efficiency framing as the book's message and extends it to screens. That is the nearest thing in this facet to the author's own use. For facet Z, it is an early (1998) analogy between spacing and interface readability, not between spacing and AI. It is also an example of the failure the dispute literature warns about: the efficiency claim is taken from the book and the mechanism left unexamined. + +**aronoff2002saenger** — Aronoff, M. (2002). Review of P. Saenger, *Space between words*. *Language in Society, 31*(4), 624–628. https://doi.org/10.1017/s0047404502254056 +- *Verification:* Crossref; Cambridge Core page. +- *Read status:* **abstract only** (the Crossref abstract; the full text is behind the publisher). +- *What the abstract says (verbatim):* "This is an impressive, fascinating, and exasperating work of scholarship, based on an astonishingly exhaustive survey of manuscript codices produced in the British Isles and western continental Europe between the 7th and the 13th centuries. Saenger traces the transition from continuous to word-divided script, which, he contends, reflects a fundamental shift in style from reading aloud to reading silently." +- *Bearing:* a linguist; "exasperating" signals a criticism I could not read. The abstract shows what the survey is *of*: manuscripts 7th–13th c., so the dating is of spacing, not of silent reading. + +**Not accessible, verified to exist (Crossref):** Hanna, R. (1999). *Speculum, 74*(3), 830–831, https://doi.org/10.2307/2886848. Richter, M. (2001). *The American Historical Review, 106*(2), 627–628, https://doi.org/10.1086/ahr/106.2.627. Walker, T. D. (2000). *The Library Quarterly, 70*(1), 166–167, https://doi.org/10.1086/603182. Mulchahey, M. M. (2001). *Canadian Journal of History, 36*(2), 323–325, https://doi.org/10.3138/cjh.36.2.323. Bäuml, F. H. (2002). *Written Language & Literacy, 5*(1), 131–134, https://doi.org/10.1075/wll.5.1.08bau (Bäuml also reviewed Coleman 1996 in the same journal, https://doi.org/10.1075/wll.1.2.10bau, which suggests a view on oral reading in the Middle Ages that I could not read). Lyons, M. C. (2001), *The European Legacy, 6*(5): OpenAlex record only, no DOI, no pages. A *Choice* notice (OpenAlex DOI 10.5860/choice.35-6061) is a one-line record. For none of these could I read even the first paragraph. + +**odonnell1994parkes** — O'Donnell, J. J. (1994). Review of M. B. Parkes, *Pause and effect*. *Bryn Mawr Classical Review* 1994.03.04. https://bmcr.brynmawr.edu/1994/1994.03.04/ +- *Read status:* **read in full (card: library/odonnell1994parkes.md)** (the whole review is short). +- *What it says (verbatim, final paragraphs):* "I miss mention of Paul Saenger's work on the growth of silent reading the later middle ages, and here the work of Richard and Mary Rouse offer the necessary bridge between punctuation studies and a wider consideration of how books were being used in the later middle ages to transmit information." Also, on the book's thesis: "earlier practice emphasized punctuation in the service of one who would read aloud … while our modern practice is to regard punctuation as a way of marking logical structures in a text for the eye of a silent reader." +- *Bearing:* **Complicates** the plan's sense that Parkes and Saenger are one voice: in 1994 a reviewer treated them as two bodies of work needing a bridge. The 1994 review does not cite Saenger's 1982 *Viator* article; that reference is in O'Donnell's 2000 review of Saenger (`odonnell2000saenger`, second paragraph). Earlier wording in this facet merged the two reviews. + +### B5. Items I could verify but not use + +**gilliard1993more** — Gilliard, F. D. (1993). More silent reading in antiquity: *Non omne verbum sonabat*. *Journal of Biblical Literature, 112*(4), 689 [Crossref gives p. 689 only; the 689–694 range is unverified: see AUDIT_01]. https://doi.org/10.2307/3267404. **Not accessible** (JSTOR; Consensus shows only a paywall page). Record verified at Crossref. Known only as a title; the earlier run flagged it as a lead. + +**slusser1992reading** — Slusser, M. (1992). Reading silently in antiquity. *Journal of Biblical Literature, 111*(3), 499. https://doi.org/10.2307/3267265. **Not accessible.** A one-page note. Record verified. + +**cavallo1999history** — Cavallo, G., & Chartier, R. (Eds.). (1999). *A history of reading in the West* (L. G. Cochrane, Trans.). Polity. **Not accessible.** Consensus shows the table of contents (Svenbro, Cavallo, Parkes, Hamesse, Saenger, …). The archive.org copies are print-disabled lending or a user upload; I did not open either. + +**torrensalvarez2022latin** — Torrens-Álvarez, M. J., & Tuten, D. (2022). From "Latin" to the vernacular: Latin-Romance hybridity, scribal competence, and social transformation in medieval Castile. *Speculum, 97*(3), 698–736. https://doi.org/10.1086/720547. **Read in part (card: library/torrensalvarez2022latin.md).** The abstract states: "we also argue against an overemphasis on ease of comprehension as a factor shaping the design and reception of written texts and oral performances of them." The full text confirms the paper has nothing on word separation, silent reading or the Insular scribes: "continuous Romance" in it means a text wholly in one register, not unspaced script, and its argument against ease of comprehension concerns Latin in the Mass and the public reading of deeds (pp. 731–732, 735). Nothing in it bears on the critics of an ease-of-reading explanation of spacing; do not cite it for word separation. Evidence is one scribe in one city, 1211–1228. + +--- + +## (c) Synthesis + +### Q1. Did ancient readers read silently, and how ordinary was it? What is the consensus and who dissents? + +**What the sources I could read support.** +- Silent reading of *short documents and letters* was taken for granted in fifth- and fourth-century Athens (Knox 1968, pp. 432–434) and is "accepted by ancient witnesses as an ordinary event" (Johnson 2000, p. 593, summarising Knox). +- Silent reading of *poetry* by at least one reader (Cicero) was thinkable (Knox, p. 427). Augustine reads silently himself at *Conf.* 8.12 (Knox, p. 422). +- Ambrose's silence is not an index of general practice (Knox, p. 422). +- Johnson, summarising the field after Gavrilov and Burnyeat, writes that they "improved the evidential base" and concludes there was "no cognitive difficulty" for fully literate readers (Johnson, p. 593; I did not read either). + +**What the sources I read do not support.** None says that literary reading was *normally* silent. Knox says the opposite: "Ancient books were normally read aloud" (p. 435). Heilmann (2020, pp. 179–180) argues the silent/aloud binary is too crude and the evidence cannot support "statistical generalizations", which pushes against *both* "usually aloud" and "usually silent". + +**Where the "consensus" claim fails.** The plan's phrase "closed the question for classicists" rests on Johnson's opening paragraph (p. 593), which says the question is "at first glance easily answered" and that the conclusion "has been known to careful readers since at least 1968." Three later or wider statements cut against a flat "closed": +1. Heilmann 2020 (p. 178–179): the reading-aloud view "is commonly accepted in Classical Studies as well as New Testament Studies", "has become a quasi-dogma", and "Well-argued counterproposals have not been accepted." His n. 8 (p. 178) lists eighteen works from 1990 to 2015 that accept it. +2. Feldt 2023 (CR 73.2, p. 688, first paragraph): "In research on ancient reading it has been common knowledge that ancient literature was read aloud." +3. McCutcheon 2015 (abstract): book historians "have long insisted" that silent reading was rare or nonexistent; this "runs counter to much research in the field of Classics." + +These are consistent once three claims are separated. (i) *Silent reading was possible and occurred*: settled among those who have engaged with Knox, Gavrilov, Burnyeat; (ii) *Literary texts were normally read aloud*: accepted by Knox himself, by Flory/Mollon-type dissenters on stronger grounds, and challenged only by Heilmann; (iii) *Unspaced script forced vocalisation*: rejected by Vatri and Heilmann, held by Saenger, partly by Svenbro. The disagreement about "the consensus" is a disagreement about which of the three is meant. + +**Dissent from the Knox line.** Flory (2007, abstract only, paraphrase) reads Knox's and Johnson's passages as subvocalisation; Mollon (2001, abstract) accepts that silent reading was "unknown" and "not practised in the West until the mediaeval period"; Svenbro grants the cost premise and gives a different cause (Vatri p. 633). O'Donnell (2000) states as "clear" that silent reading "did not become commonplace until long after" Augustine. + +**Problem 1 bearing.** The plan's "ignited literacy" has no support in this literature. The dispute is about *mode* of reading (aloud, murmured, silent) among people who could read, not about how many could read. No source here connects spacing to the size of the reading public. + +### Q2. What exactly do the critics reject in Saenger, and what do they leave standing? + +| Component of Saenger's thesis | Who rejects it | Who leaves it standing | Depth | +|---|---|---|---| +| **A. Antiquity: silent reading was impossible or very rare** | Knox 1968 (against Balogh, before Saenger), Gavrilov, Burnyeat, Johnson 2000, Vatri 2012, Heilmann 2020 | Flory 2007, Mollon 2001 (abstracts), partly Svenbro | Knox full; Johnson opening; Gavrilov, Vatri abstract; Burnyeat none | +| **B. Mechanism: unspaced script forces vocal rehearsal** | Vatri 2012 ("Saenger's model must be rejected", p. 639 via Heilmann), Heilmann 2020 (pp. 180–181) | Svenbro grants the cost; Flory, Mollon cite it; O'Donnell flags it as the vulnerable part | Vatri, via Heilmann; Heilmann full | +| **C. Cause: spacing was a (necessary and sufficient) condition of the rise of silent reading** | not directly rejected in anything I read in full; O'Donnell (2000) says some will dissent (¶3) and that this is the "spire" over the "cathedral" | O'Donnell expects the debate over the physiology-to-practice step to be resolved in Saenger's favour (¶4); Mollon, Flory accept; Garrett takes it as given | O'Donnell full; Garrett full | +| **D. Medieval story: Irish spacing 7th–8th c., Continent from late 10th c., manuscripts 7th–13th c.** | no source I read challenges it | O'Donnell ("cathedral"), Aronoff abstract ("impressive, fascinating, and exasperating … astonishingly exhaustive"), Garrett | O'Donnell full; Aronoff abstract; Hanna, Richter, Walker, Mulchahey, Bäuml unread | +| **E. Social reading: antiquity "did not possess the desire … to make reading easier and swifter" (Saenger 1997, p. 11)** | not rejected; Johnson (2010, per Ker, Reis, Mannering) supplies a parallel, independent account | Johnson 2010; Saenger (1997, p. 11) | second-hand | + +Three observations. +1. The critics' target is A and B. **The medievalist half (D) is not contested in anything I could read.** I cannot say whether the reviews I could not open (Hanna, Richter, Walker, Mulchahey, Bäuml) attack it. "Exasperating" (Aronoff) shows at least one reviewer was irritated. +2. C is the step H-cause needs, and the only reader of Saenger I could read in full who names it (O'Donnell) predicts dissent but does not supply it. Since B, its physiological bridge, is the part Vatri and Heilmann say is wrong, **C loses its mechanism**; what remains is a correlation of timing (spacing arrives; silent reading becomes common) plus social explanation. +3. Heilmann's own concession (p. 180) that untrained moderns read unspaced English slower, and his reading of the Dionysius, Quintilian and Lucian passages as about learners (n. 28, pp. 181–182), leave a smaller thesis standing: spacing cuts the cost for beginners and untrained readers. That is H-cost's weak form, and nobody in this facet's reading denies it. + +### Q3. Does anyone measure or date when silent reading of literary texts became typical? + +I found no such measurement. What exists: +- **Saenger's own dates** (7th-century Irish spacing; Continental late tenth century; 7th–13th c. survey) are dates of *word separation in manuscripts*, not of silent reading (Aronoff abstract; [unverified: see AUDIT_01] the book's publisher abstract, which the audit could not find). The inference from layout to practice is the contested step. +- **O'Donnell (2000, ¶2)** asserts that silent reading "did not become commonplace until long after" Augustine, without evidence in the review, and calls it "clear". +- **Knox (1968)** gives no date; his "normally read aloud" (p. 435) covers antiquity as a whole. +- **Johnson (2000, 2010)** reframes the question; he is not trying to date a shift. +- **McCutcheon (2015)** explains why book history insists on the late date; it does not give one (abstract only). +- Possible dating studies I could not reach: Saenger (1982), the Parkes and Saenger chapters in Cavallo & Chartier (1999), Coleman (1996) on public reading in late medieval England and France (title only; not read). + +So: **a typical-date for silent reading of literary texts is not measured in the sources I could read.** Any sentence in the paper that says "by X, silent reading was usual" would rest on Saenger's inference or on O'Donnell's unsupported "clear". + +### Q4. Why did the Romans drop interpuncts if separation helps, and what does that do to a story in which spacing is an obvious improvement waiting to be found? + +**What the sources say.** +- Heilmann 2020 (p. 182, citing Oliver 1951 and Wingo 1972, p. 15): Latin used word-dividers "from the earliest Republican times through the Golden Age and well into the Second Century", then "replaced" them by *scriptio continua* "in the second century CE". Heilmann's n. 37 (pp. 182–183) reports the Vindolanda tablets (c. 100 CE) showing three practices at once: middle dots, plain *spaces* between words, and a few in *scriptio continua*. If that report holds, spacing was in Roman use before and during the move to continuous script. I could not check the tablets. +- Heilmann (p. 183) reports Suetonius, *Aug.* 87.3, calling it odd that Augustus did not separate his words, and Seneca, *Ep.* 40, linking Greek continuous script with speed, Latin word-division with restraint. Both are attitudes to the practice, not engineering judgements. +- Saenger: the ancient world "did not possess the desire, characteristic of the modern age, to make reading easier and swifter" (1997, p. 11), and interpuncts "were ineffective aids to the physiological process of reading" when compared to spaces (p. 28 [page to check in a library copy]; as quoted in facet E, snippets). So Saenger's own answer is *social*: the improvement was available (or would have been recognised) and unwanted. +- Ker (reviewing Johnson 2010) writes that for Johnson continuous script is "consistent with an elite culture" in which fluent reading from a bookroll "is a marker of education and status"; Mannering writes that Johnson interprets the effort of reading it as "a hallmark of the exclusivity of literary culture". These quoted phrases are the reviewers' words, not Johnson's. On their account Johnson gives a social explanation with no physiological premise. +- Heilmann's own inference (p. 182): the change "shows that scriptio continua is not a characteristic of deficient writing systems." This disproves a *deficiency* story; it is silent on whether spacing would have lowered cost. + +**What this does to the "obvious improvement" story.** Three results, none of which needs the other: +1. **The improvement was not waiting to be found.** Romans used dividers for centuries and, per Heilmann's n. 37, spaces too; they then moved to continuous script at the time of the elite reading communities Johnson studies (Trajan to Marcus Aurelius, per Reis). The device was known and set aside. +2. **Both the critic and the defender point to social choice.** Saenger (no desire) and Johnson (elite exclusivity) both give a social account of why spacing was not adopted, and they differ on whether unspaced text was hard; the point survives whichever side is right about physiology. That is the strongest finding in this facet for the paper: *the history cannot be told as "a good UX move discovered late"*. It has to be told as a move that became attractive when the readers it served became a different group (Saenger's Irish scribes reading Latin as a foreign language, per facet E on Parkes p. 1). +3. **Heilmann's version supports a stronger and different claim: continuous script was not a handicap.** That sits against H-cost: if trained readers lose nothing, spacing moved work only for novices. + +**Caveats.** The interpunct was a mark within continuous lines, not space; Saenger and Parkes distinguish them and the plan's author should read their texts (facet A). The Vindolanda reading is Heilmann's and not checked. The Roman change is dated by Heilmann to "the second century CE" and by Saenger (via facet E snippets) to before the end of the second century with adoption "late second or third century"; facet A should confirm. + +### Q5. Verdict for H-cost and H-cause, in the critics' terms + +**H-cost** (word separation moved parsing work from the reader to the page; lowered the cost of reading; widened the group who could read without long training; did not create silent reading). + +- *Survives in the form:* "spacing lowers the cost of reading unspaced script for untrained readers." Supports: Heilmann concedes the slowing for untrained readers (p. 180) and reads the ancient difficulty-passages as about learners (n. 28); Mannering reports that, for Johnson, the bookroll needed "years of training"; O'Donnell's summary of Saenger's answer: spacing made visual reading "so easy as to become widespread" (¶3). The critics who reject Saenger's mechanism (Vatri, Heilmann) reject it for *trained* readers; they do not deny that learners found continuous script harder. +- *Does not survive in the form:* "unspaced script carried a cost large enough to determine how people read." Vatri via Heilmann (p. 181): "there is no reason to assume that reading unspaced text is a particularly demanding cognitive task in itself"; Heilmann (p. 181), citing Vatri 638 f. and others, writes that spaces are "redundant and can even disrupt the reading process." Facet D must say whether the eye-movement literature backs Vatri or Rayner and Pollatsek (Heilmann's n. 20 names the 1996 exchange). +- *"Widened the group … without long training"* is not tested in this facet. The critics' framing (Johnson) is that long training was kept on purpose. This is a statement about the *ancient* choice; the *medieval* widening (Irish scribes and later) rests on Saenger and Parkes and on O'Donnell's reading. +- *"Did not create silent reading"*: supported (Knox p. 434–435; Johnson p. 593; Vatri p. 633; Heilmann). **Verdict: H-cost is the safer hypothesis, in its weak form only, and only if the paper says "for untrained readers" and does not claim a mechanism.** + +**H-cause** (word separation was a cause of the spread of silent reading and, through it, of wider literacy). + +- *Silent reading:* no source I read in full proves it; the physiological bridge is rejected by Vatri and Heilmann; the dating inference is Saenger's; O'Donnell expects some dissent from the "necessary and sufficient" claim (¶3) and calls the physiology step "the most venturesome and vulnerable part of the book" (¶4). Svenbro's rival (attitude) and Johnson's rival (social setting) are non-physiological causes the field takes seriously. **H-cause for silent reading is unsupported as a mechanism and unmeasured as a date.** +- *Wider literacy:* untouched by anything in this facet. **Not supported, not tested.** `C_spacing_to_literacy` must carry it. +- **Verdict: H-cause cannot be defended from the dispute literature.** It can be stated as Saenger's thesis, dated, with the critics' rejection; it should not be adopted as the paper's claim. + +**Where this leaves the author's sentence** (`dial_response_algorithmacy/draft.md:263`: "Word separation did not train the populace to read *scriptio continua*; it changed the interface so that reading no longer required the training."). From the critics' side: the first clause is untouched; the second is a claim H-cost makes in a weak form only for untrained readers, and the training Johnson describes was an elite practice kept long on purpose. The word "populace" and "no longer required" go further than anything I found. + +### Reused repo material + +| Item | Where | How used | +|---|---|---| +| Knox 1968 | facet E (Part 2) | Re-read in full; all page locators checked again and match. | +| Gavrilov, Burnyeat, Johnson, Vatri, McCutcheon | facet E (Part 2) | Records re-verified at Crossref; Gavrilov's abstract and Johnson's opening paragraph now quoted verbatim; no deeper reading possible. | +| Saenger p. 8, 28, 29, 69, 11 snippets | facet E (Part 2) | Cited as "per facet E", not re-checked here (facet A's task). | + +### Claims the sources contradict or qualify + +1. **"Johnson (2000) closed the question for classicists."** Johnson says "easily answered" and "seemingly firmer". Heilmann (2020, pp. 178–179) says the opposite view is a "quasi-dogma"; Feldt (2023, p. 688) calls reading-aloud of literature "common knowledge"; McCutcheon's abstract says the split persists. +2. **"Critics show silent reading was common/ordinary in antiquity."** Knox's own conclusion is that books were "normally read aloud" (p. 435); his silent reading is for documents, letters and short texts, plus the Cicero inference for verse. +3. **"Saenger says the ancients could not read silently."** Vatri reports "physiologically impossible"; the book, per facet E, concedes Ambrose read silently but not rapidly. Vatri himself notes Saenger's reading of the sources as "silent, or sub-vocal, oral reading – a sort of mumbling." The strong form is the critics' framing. +4. **"Spacing is obviously better and was waiting to be found."** The Romans used dividers and, per Heilmann (n. 37), spaces, then moved to continuous script; Saenger and Johnson both give social accounts of non-adoption and differ on whether unspaced text was hard. +5. **"Continuous script was a handicap."** Vatri and Heilmann dispute it for trained readers; Heilmann concedes it for untrained modern readers. +6. **"The medievalists' reviews attack the thesis."** The two I read in full (O'Donnell, Garrett) praise it; O'Donnell calls the survey a "cathedral" and expects the debate over the physiology-to-practice step to be resolved in Saenger's favour. Vatri (2012) and Heilmann (2020) reject the physiology O'Donnell expected to be vindicated, and the eye-movement literature is itself contested (Heilmann n. 20); that does not make the survey wrong. +7. **"Spacing ignited literacy."** No source here speaks to literacy rates. +8. **"Parkes and Saenger are one view."** A 1994 reviewer (O'Donnell on Parkes) treats Saenger's work as missing and needing a bridge. Also, per facet E, Parkes (1992, p. 9) repeats the older view. + +--- + +## (d) Gaps and unverified leads + +**The main gap.** I read none of Gavrilov, Burnyeat, Johnson 2000 or Vatri in full. A library user must fetch, in this order: +1. **Vatri 2012**, *CQ* 62.2, 633–647 (doi 10.1017/S0009838812000213; JSTOR 23470128): the actual argument against Saenger's model and its experimental sources (Kajii et al. 2001; Sainio et al. 2007; Bai et al. 2008; Winskel et al. 2009); check that Heilmann's quotations at pp. 639 and 646–647 are accurate and in context. +2. **Johnson 2000**, *AJP* 121.4, 593–627 (doi 10.1353/ajp.2000.0053; MUSE 1288; JSTOR 1561728): especially pp. 609–612 (cited by Heilmann n. 82 on line length), and whatever he says about spacing and Saenger. +3. **Johnson 2010**, ch. 2 "Pragmatics of Reading", pp. 17–31 (doi 10.1093/acprof:oso/9780195176407.003.0002): the *scriptio continua* and elite-culture claim, now known only through Ker, Reis and Mannering. +4. **Gavrilov 1997**, *CQ* 47.1, 56–73 (doi 10.1093/cq/47.1.56; JSTOR 639597) and **Burnyeat 1997**, *CQ* 47.1, 74–76 (doi 10.1093/cq/47.1.74; JSTOR 639598). +5. **Heilmann 2021**, *Lesen in Antike und frühem Christentum* (Narr Francke Attempto, ISBN 978-3-7720-8729-5), the *scriptio continua* chapter and the appendix list of silent-reading evidence; and the **Feldt review** (*CR* 73.2, 688–690) in full. +6. **McCutcheon 2015**, *Book History* 18, 1–32 (MUSE 597277): the explanation of why book history keeps the old view, and any date given for a shift. + +**Reviews of Saenger not read** (fetch from the journals): Hanna, *Speculum* 74.3 (1999) 830–831 (JSTOR 2886848); Richter, *AHR* 106.2 (2001) 627–628; Walker, *Library Quarterly* 70.1 (2000) 166–167; Mulchahey, *Canadian Journal of History* 36.2 (2001) 323–325; Bäuml, *Written Language & Literacy* 5.1 (2002) 131–134; Aronoff, *Language in Society* 31.4 (2002) 624–628 (only the abstract); Lyons, *The European Legacy* 6.5 (2001); the *Choice* note. Hanna and Richter are the most likely to engage the causal claim from the medieval side. None is a classicist's review of the ancient claim, and I found no review in BMCR. + +**Searches not completed or not parsed.** The BMCR search for "scriptio continua" returned a page I did not parse (only the "Saenger silent reading" page was). A manual BMCR search for reviews of Johnson and Parker (eds.), *Ancient Literacies* (2009), Howley (2018), and Heilmann (2021) would be worth a minute. + +**Gaps on Saenger's side.** Saenger's 1982 *Viator* article (doi 10.1484/J.VIATOR.2.301476) and his 2020 *Oxford Handbook of Latin Palaeography* chapter "Orality and visible language" (doi 10.1093/oxfordhb/9780195336948.013.112) were not read (facet A). **I found no direct reply by Saenger to Knox, Gavrilov, Vatri or Johnson.** Whether he replied is unknown; his book's n. on p. 298 answers Knox on Lucian only (facet E, snippet). The 2020 chapter's abstract (Consensus) says it describes the shift "from the predominantly oral or aural reading of antiquity to the more visual act of reading that emerged in Late Antiquity along with the codex", so his current position may have moved away from "spacing alone"; facet A should check. + +**Sources I located but did not open because the only copies were unauthorised or proxied:** +- a Vatri PDF reachable only through a third institution's EZproxy; +- a McCutcheon PDF on a personal website; +- z-library-style uploads of Saenger (archive.org `space-between-words-the-origins-of-silent-reading`) and a Cavallo & Chartier upload (`cavallo-a-history-of-reading-in-the-west`) and a Svenbro upload with an unverified license label. +Print-disabled lending copies on archive.org (`phrasikleiaanthr0000sven`, `historyofreading0000cava`) need a borrower account. + +**Not found:** +- A modern edition or translation of Balogh 1927 (not on any open service I tried). +- Oliver (1951) on the late introduction of continuous script in Latin: Heilmann gives no full reference and I did not find it. +- Wingo (1972), *Latin Punctuation in the Classical Age* (Mouton, The Hague, Janua Linguarum Series Practica 133): known from Heilmann's bibliography; unread. It is the primary source for Q4 and should be fetched (also for facet A). +- Any source that explicitly argues a *deliberate* design choice for continuous script beyond Johnson (2010) and Saenger's own phrase. +- Any source that dates the typical silent reading of literary texts (Q3), other than Saenger's inference. +- Coleman (1996), *Public Reading and the Reading Public in Late Medieval England and France* (Cambridge UP): the Crossref/OpenAlex search shows five reviews of it but I read none and the book. If it argues oral reading persisted through the late Middle Ages, it contradicts the late-medieval silent-reading date; I cannot say. + +**Leads not pursued.** Gilliard 1993 and Slusser 1992 (JBL, closed); Howley (2018) *Aulus Gellius and Roman Reading Culture* and Johnson and Parker (2009) *Ancient Literacies*; Cavallo's "Between volumen and codex"; Parkes's chapter in Cavallo and Chartier 1999; Petrucci; Carruthers. Pohl (2016, *Tabularia*, diamond OA) on punctuation and performance in the Normandy *Historia*: it cites Saenger (1982, 1997) in its footnotes and I did not read it for argument. Pancotti (2025, *Interface*, diamond OA) cites Knox 1968 and Johnson 2000 in passing (silent reading "although not ruled out in antiquity, required special circumstances"): it repeats the consensus but adds no evidence. Torrens-Álvarez and Tuten (2022, *Speculum*) is listed above as an abstract-only lead. + +**Files kept under** `~/iit-playground/dxd_foundation_local/pdfs/B/` (outside the repo): `knox1968.pdf`, `heilmann2020.pdf`, `garrett1998.pdf` and text extractions; the Johnson-review PDFs from his Duke page; the BMCR and TMR HTML/XML; `pancotti2025.pdf` and `pohl2016.pdf`. + +## Changes after full-text cards (2026-10-13) + +- `odonnell1994parkes`: removed the claim that the 1994 review records Saenger's 1982 *Viator* article; that reference is in O'Donnell 2000 (`odonnell2000saenger`, ¶2). Read status now "read in full (card)". +- `torrensalvarez2022latin`: read status "abstract only" changed to "read in part (card)"; the "possible parallel" is withdrawn, since the paper has nothing on word separation (its ease-of-comprehension argument concerns the Mass and deeds). +- `heilmann2020reading`: read status updated to "(card)"; added that Heilmann relies on Vatri for the Thai result and omits Chantavarin and Leung's sentence-level slowing. +- `mannering2012johnson`: read status "read in part" changed to "read in full (card)"; the full review never mentions Saenger, spacing or Knox. +- `garrett1998saenger`, `knox1968silent`, `ker2012johnson`, `reis2011johnson`, `odonnell2000saenger`: facet confirmed; read status annotated "(card)". + +Synthesis verdicts changed: none. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_C_spacing_to_literacy.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_C_spacing_to_literacy.bib new file mode 100644 index 00000000..4df2123d --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_C_spacing_to_literacy.bib @@ -0,0 +1,373 @@ +% Facet C, spacing to literacy. Agent-drafted 2026-10-12. Read status is in each note. +% Archive.org "search-inside" = public full-text API snippets of a lending-library scan; no book borrowed or read in full. + +@book{clanchy1993from, + author = {Clanchy, M. T.}, + title = {From Memory to Written Record: England 1066--1307}, + edition = {2}, + publisher = {Blackwell}, + year = {1993}, + url = {https://archive.org/details/frommemorytowrit0000clan_e1n6_2ed}, + note = {Read in part, search-inside (archive.org item frommemorytowrit0000clan_e1n6_2ed). Index: word separation under format and layout, p. 130. Third edition (2013) is in repo card clanchy2013.} +} + +@book{stock1983implications, + author = {Stock, Brian}, + title = {The Implications of Literacy: Written Language and Models of Interpretation in the Eleventh and Twelfth Centuries}, + publisher = {Princeton University Press}, + year = {1983}, + url = {https://archive.org/details/implicationsofli0000stoc}, + note = {Read in part, search-inside. Zero hits for word separation, scriptio continua, silent reading, Saenger, punctuation.} +} + +@book{mckitterick1989carolingians, + author = {McKitterick, Rosamond}, + title = {The Carolingians and the Written Word}, + publisher = {Cambridge University Press}, + year = {1989}, + url = {https://archive.org/details/carolingianswrit0000mcki}, + note = {Read in part, search-inside. Zero hits for legibility, punctuation, word divisions, reading aloud.} +} + +@book{graff1987legacies, + author = {Graff, Harvey J.}, + title = {The Legacies of Literacy: Continuities and Contradictions in Western Culture and Society}, + publisher = {Indiana University Press}, + year = {1987}, + url = {https://archive.org/details/legaciesoflitera0000graf}, + note = {Read in part, search-inside. Mentions Latin punctuation and silent reading as supporting developments; cites Saenger 1982.} +} + +@book{graff1979literacy, + author = {Graff, Harvey J.}, + title = {The Literacy Myth: Literacy and Social Structure in the Nineteenth-Century City}, + publisher = {Academic Press}, + year = {1979}, + url = {https://archive.org/details/literacymythlite0000graf}, + note = {Read in part, search-inside (two hits on silent reading, nineteenth-century pedagogy). Not read for the general argument.} +} + +@book{houston2002literacy, + author = {Houston, R. A.}, + title = {Literacy in Early Modern Europe: Culture and Education 1500--1800}, + edition = {2}, + publisher = {Longman}, + year = {2002}, + url = {https://archive.org/details/literacyinearlym0000hous_i9l5}, + note = {Read in part, search-inside. Cites Saenger 1989, p. 142 for comprehension literacy.} +} + +@book{cressy1980literacy, + author = {Cressy, David}, + title = {Literacy and the Social Order: Reading and Writing in Tudor and Stuart England}, + publisher = {Cambridge University Press}, + year = {1980}, + url = {https://archive.org/details/literacysocialor0000davi}, + note = {Read in part, search-inside. Male illiteracy about 80 percent, female about 95 percent at Elizabeth's accession (his estimate).} +} + +@book{cipolla1969literacy, + author = {Cipolla, Carlo M.}, + title = {Literacy and Development in the West}, + publisher = {Penguin}, + year = {1969}, + url = {https://archive.org/details/literacydevelopm0000unse_r1i6}, + note = {Read in part, search-inside. Zero hits on word separation, silent reading, punctuation, minuscule.} +} + +@book{eisenstein1979printing, + author = {Eisenstein, Elizabeth L.}, + title = {The Printing Press as an Agent of Change}, + publisher = {Cambridge University Press}, + year = {1979}, + url = {https://archive.org/details/printingpressas01eise}, + note = {Read in part, search-inside; plus repo card eisenstein1979 and Johns card (secondary).} +} + +@book{goody1986logic, + author = {Goody, Jack}, + title = {The Logic of Writing and the Organization of Society}, + publisher = {Cambridge University Press}, + year = {1986}, + url = {https://archive.org/details/logicofwritingor0000good}, + note = {Read in part, search-inside. Zero hits for all spacing and silent-reading terms.} +} + +@book{street1984literacy, + author = {Street, Brian V.}, + title = {Literacy in Theory and Practice}, + publisher = {Cambridge University Press}, + year = {1984}, + url = {https://archive.org/details/unset0000unse_e6x6}, + note = {Read in part, search-inside. Notes that Vai script has no spaces between words or sentences. Repo card street1984 is secondary.} +} + +@book{scribner1981psychology, + author = {Scribner, Sylvia and Cole, Michael}, + title = {The Psychology of Literacy}, + publisher = {Harvard University Press}, + year = {1981}, + url = {https://archive.org/details/psychologyoflite00sylv}, + note = {Read in part, search-inside. Vai script runs without division into words; learned informally; no general cognitive effect without schooling.} +} + +@book{havelock1963preface, + author = {Havelock, Eric A.}, + title = {Preface to Plato}, + publisher = {Belknap Press of Harvard University Press}, + year = {1963}, + url = {https://archive.org/details/prefacetoplato0000have}, + note = {Repo card havelock1963 (pp. 38--41, pagination probable); key phrases supply of readers, reading trauma, silent revolution re-confirmed by search-inside this session.} +} + +@book{ong1982orality, + author = {Ong, Walter J.}, + title = {Orality and Literacy: The Technologizing of the Word}, + publisher = {Methuen}, + year = {1982}, + url = {https://archive.org/details/oralityliteracyt00ongw}, + note = {Repo card ong1982 (chs. 3--5) plus search-inside snippets on spacing and legibility in print.} +} + +@book{harris1989ancient, + author = {Harris, William V.}, + title = {Ancient Literacy}, + publisher = {Harvard University Press}, + year = {1989}, + doi = {10.4159/9780674038370}, + note = {Review only: figures and preconditions taken from Zanker 1991 and Sherratt 1989. No accessible copy found. Not read.} +} + +@article{zanker1991ancient, + author = {Zanker, G.}, + title = {Ancient Literacy (review)}, + journal = {Philosophy and Literature}, + volume = {15}, + pages = {173--175}, + year = {1991}, + doi = {10.1353/phl.1991.0070}, + note = {Review of Harris 1989; read via Consensus abstract field, which reproduces the review text.} +} + +@article{sherratt1989language, + author = {Sherratt, A.}, + title = {Review of Marija Gimbutas, The Language of the Goddess, with a section on Harris, Ancient Literacy}, + journal = {The Antiquaries Journal}, + year = {1989}, + doi = {10.1017/s0003581500085577}, + note = {Review text on Harris read via Consensus; the DOI belongs to the combined review of a different book. Volume and pages not retrieved.} +} + +@incollection{saenger1991separation, + author = {Saenger, Paul}, + title = {The Separation of Words and the Physiology of Reading}, + booktitle = {Literacy and Orality}, + editor = {Olson, David R. and Torrance, Nancy}, + publisher = {Cambridge University Press}, + year = {1991}, + url = {https://archive.org/details/literacyorality0000unse}, + note = {Chapter 12. Read in part, search-inside; chapter pages not seen.} +} + +@book{saenger1997space, + author = {Saenger, Paul}, + title = {Space between Words: The Origins of Silent Reading}, + publisher = {Stanford University Press}, + year = {1997}, + doi = {10.1515/9781503619081}, + url = {https://archive.org/details/spacebetweenword0000saen}, + note = {Read in part, search-inside on the lending scan. Preface: not a general history of literacy. Chronology and Part 2 facet E snippets reused.} +} + +@article{saenger1982silent, + author = {Saenger, Paul}, + title = {Silent Reading: Its Impact on Late Medieval Script and Society}, + journal = {Viator}, + volume = {13}, + pages = {367--414}, + year = {1982}, + doi = {10.1484/J.VIATOR.2.301476}, + note = {Not accessible (Brepols bot wall). DOI resolved in Crossref.} +} + +@incollection{saenger1989books, + author = {Saenger, Paul}, + title = {Books of Hours and the Reading Habits of the Later Middle Ages}, + booktitle = {The Culture of Print}, + editor = {Chartier, Roger}, + year = {1989}, + note = {Not accessible. Title and container taken from the bibliographies of Houston 2002 and Clanchy 1993 (search-inside). Publisher and page range not found. Houston cites p. 142.} +} + +@book{zieman2008singing, + author = {Zieman, Katherine}, + title = {Singing the New Song: Literacy and Liturgy in Late Medieval England}, + publisher = {University of Pennsylvania Press}, + year = {2008}, + url = {https://archive.org/details/singingnewsongli0000ziem}, + note = {Read in part, search-inside. Questions the category phonetic literacy. Needs a fuller read.} +} + +@book{thomas1992literacy, + author = {Thomas, Rosalind}, + title = {Literacy and Orality in Ancient Greece}, + publisher = {Cambridge University Press}, + year = {1992}, + url = {https://archive.org/details/literacyoralityi0000thom}, + note = {Read in part, search-inside. Applies Saenger's phonetic and comprehension literacy to antiquity.} +} + +@article{seymour2003foundation, + author = {Seymour, Philip H. K. and Aro, Mikko and Erskine, Jane M.}, + title = {Foundation Literacy Acquisition in European Orthographies}, + journal = {British Journal of Psychology}, + volume = {94}, + number = {2}, + pages = {143--174}, + year = {2003}, + doi = {10.1348/000712603321661859}, + note = {Read in full (card: library/seymour2003foundation.md); first read as abstract only (Consensus). Crossref-verified. Not open access.} +} + +@article{ziegler2005reading, + author = {Ziegler, Johannes C. and Goswami, Usha}, + title = {Reading Acquisition, Developmental Dyslexia, and Skilled Reading across Languages: A Psycholinguistic Grain Size Theory}, + journal = {Psychological Bulletin}, + volume = {131}, + number = {1}, + pages = {3--29}, + year = {2005}, + doi = {10.1037/0033-2909.131.1.3}, + note = {Not accessible; abstract not retrieved. Crossref-verified. Not open access.} +} + +@article{ziegler2010orthographic, + author = {Ziegler, Johannes C. and others}, + title = {Orthographic Depth and Its Impact on Universal Predictors of Reading}, + journal = {Psychological Science}, + volume = {21}, + number = {4}, + pages = {551--559}, + year = {2010}, + doi = {10.1177/0956797610363406}, + note = {Read in full (card: library/ziegler2010orthographic.md). Crossref-verified.} +} + +@article{wood1929latinizing, + author = {Wood, M. M.}, + title = {Latinizing the Turkish Alphabet: A Study in the Introduction of a Cultural Change}, + journal = {American Journal of Sociology}, + volume = {35}, + number = {2}, + pages = {194--203}, + year = {1929}, + doi = {10.1086/214978}, + note = {Abstract only (Consensus). Crossref-verified.} +} + +@article{yilmaz2011learning, + author = {Y{\i}lmaz, Hale}, + title = {Learning to Read (Again): The Social Experiences of Turkey's 1928 Alphabet Reform}, + journal = {International Journal of Middle East Studies}, + volume = {43}, + number = {4}, + pages = {677--697}, + year = {2011}, + doi = {10.1017/s0020743811000900}, + note = {Abstract only. Crossref-verified.} +} + +@article{filiztekin2025unification, + author = {Filiztekin, Alpay}, + title = {Unification of Education, Ethnic Diversity, and Literacy in the Early Turkish Republic}, + journal = {Economic History of Developing Regions}, + volume = {40}, + pages = {80--97}, + year = {2025}, + doi = {10.1080/20780389.2025.2450331}, + note = {Abstract only. Crossref-verified.} +} + +@article{pae2011korean, + author = {Pae, Hye K.}, + title = {Is Korean a Syllabic Alphabet or an Alphabetic Syllabary}, + journal = {Writing Systems Research}, + volume = {3}, + number = {2}, + pages = {103--115}, + year = {2011}, + doi = {10.1093/wsr/wsr002}, + note = {Abstract only. Crossref-verified.} +} + +@article{clemens2025korea, + author = {Clemens, W.}, + title = {Review of Cha and Pacheco Pardo, Korea: A New History of South and North}, + journal = {Journal of East Asian Studies}, + volume = {25}, + pages = {159--162}, + year = {2025}, + doi = {10.1017/jea.2025.6}, + note = {Read in full (card: library/clemens2025korea.md). Literacy figures (20 percent in 1912, over 60 percent by 1940) are the reviewer's own, unsourced; not the book's. Unchecked against a census.} +} + +@article{ivanova2026process, + author = {Ivanova, Yanitsa}, + title = {A Process of Implementing the Korean Alphabet on the Korean Peninsula}, + journal = {Ezikov Svyat (Orbis Linguarum)}, + volume = {24}, + number = {1}, + year = {2026}, + doi = {10.37708/ezs.swu.bg.v24i1.9}, + note = {Read in full (card: library/ivanova2026process.md). DOI not resolved in Crossref this session.} +} + +@article{becker2024conquering, + author = {Becker, Sascha O. and others}, + title = {Conquering Korea for Jesus: Protestant Missionaries, Local Churches, and Literacy in Colonial Korea}, + journal = {Labour Economics}, + volume = {91}, + pages = {102618}, + year = {2024}, + doi = {10.1016/j.labeco.2024.102618}, + note = {Read in full (card: library/becker2024conquering.md). Crossref-verified.} +} + +@article{hong2018colonization, + author = {Hong, Ji and others}, + title = {Colonization and Education: Exploring the Legacy of Local Elites in Korea}, + journal = {Economic History Review}, + volume = {71}, + number = {3}, + pages = {938--964}, + year = {2018}, + doi = {10.1111/ehr.12538}, + note = {Abstract only. Crossref-verified (online 2017).} +} + +@article{weng2023stop, + author = {Weng, Jeffrey}, + title = {Stop the Presses! Publishing Chinese Character Simplification, 1935--1936}, + journal = {Harvard Journal of Asiatic Studies}, + year = {2023}, + doi = {10.1353/jas.2023.a938222}, + note = {Abstract only (Consensus); DOI not Crossref-checked. Context for facet gap on China.} +} + +@article{kuzuoglu2022basic, + author = {Kuzuo{\u g}lu, U.}, + title = {Basic Chinese: Cognitive Management, Communication Engineering, and Mass Literacy in China}, + journal = {Modern Chinese Literature and Culture}, + year = {2022}, + doi = {10.3366/mclc.2022.0003}, + note = {Abstract only (Consensus); DOI not Crossref-checked.} +} + +@article{li2004reexamination, + author = {Li, Alan L.}, + title = {A Reexamination of Protoliteracy through an Analysis of Modern Chinese Character Use}, + journal = {Written Communication}, + year = {2004}, + doi = {10.1177/0741088303262843}, + note = {Abstract only (Consensus); DOI not Crossref-checked.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_C_spacing_to_literacy.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_C_spacing_to_literacy.md new file mode 100644 index 00000000..09e6f996 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_C_spacing_to_literacy.md @@ -0,0 +1,302 @@ +# C — Did a change in page conventions spread reading? (tests H-cause) + +*Corrected on 2026-10-12 against `../audit/AUDIT_01.md`: 9 factual fixes.* + +*Facet C of the DXD foundation-paper literature review, 2026-10-12. Agent-drafted. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** (1) What the standard histories of medieval, early modern and ancient literacy name as the drivers of the spread of reading (Clanchy, Stock, McKitterick, Graff, Houston, Cressy, Cipolla, Eisenstein/Johns, Goody, Street, Scribner and Cole, Havelock, Harris). (2) Whether any of them treats word separation, punctuation, script reform or layout as a cause or enabling condition, and with what weight; silence is reported as a finding with the search evidence. (3) Dates of spacing against dates and estimates of literacy. (4) Evidence on how novices were taught, and Saenger's "phonetic" versus "comprehension" literacy. (5) Modern cross-script evidence: orthographic depth, Hangul, the Turkish reform of 1928, Chinese simplification. (6) The strongest statements that technology or script design alone does not produce literacy. + +**The question the synthesis answers.** Does any credible scholar support the author's phrase that the move from *scriptio continua* to spaces between words "ignited literacy" (H-cause), or do the sources support the weaker claim that spacing moved parsing work from reader to page and lowered a cost (H-cost)? Section (c) gives a verdict, the most defensible sentence, and the claims in the author's framing that the sources contradict. + +**Legend for read status.** +- *Read in full*: I read the whole text this session. **No book in this facet reaches this level.** +- *Read in part, search-inside*: the book is a borrowable scan on archive.org that I did not borrow. I queried archive.org's public full-text search API (`be-api.us.archive.org/fts/v1/search`), which returns OCR snippets of about one line around each hit, plus an index of hits. Each quotation below is one such snippet, normalised for OCR hyphenation and spacing only. I did not see the surrounding paragraph, and the API gives a scan-page index, not the printed page number, so **no printed page locator is given unless the book's own index or contents supplies it**. Re-check every quotation in context before it enters a draft. A search result of "0 hits" means the term is absent from the OCR text archive.org indexed; OCR can drop words, so a zero is strong evidence of rarity, not proof of absence. +- *Abstract only*: database abstract (Consensus, Crossref) read, text not. +- *Not accessible*: record verified, no text or abstract. +- *Review only*: I read a review's account of the work. + +The snippet route is the same thing as a library user's search-inside box. The Cluster H cards in `algorithmacy_design_ethics/literature/library/` rest on secondary reading for most of these books; the entries below add search-inside evidence, not a full read. + +--- + +## (a) Search log + +All searches ran on 2026-10-12. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | Repo | Read `RESEARCH_PLAN.md`, `REUSE.md`, Part 2 facet E, library cards `clanchy2013`, `street1984`, `havelock1963`, `ong1982`, `scribner1981`, `INDEX.md` 331-384 | n/a | Reused: Havelock card (re-verified below), Part 2 facet E for Saenger/Parkes chronology. | +| 2 | archive.org `advancedsearch` | titles of 14 works (Clanchy, Stock, McKitterick, Graff x2, Houston, Cipolla, Cressy, Eisenstein, Goody, Scribner and Cole, Street, Saenger, Olson and Torrance) | identifiers found for all but Harris | Harris, *Ancient Literacy*: no archive.org item found under title or author. | +| 3 | archive.org full-text API | for each book, `"word separation"`, `"separation of words"`, `"word division"`, `"scriptio continua"`, `"scriptura continua"`, `"silent reading"`, `Saenger`, `punctuation`, `"Carolingian minuscule"`, `"spaces between"`, `"between words"`, `"word spacing"`, `"word-separation"`, each `AND identifier:` | see entries | The core of the silence evidence. Identifiers are given in each entry. | +| 4 | archive.org full-text API | follow-ups per book (e.g. `"caroline minuscule"`, `legibility`, `"lay literacy"`, `"largely a response to need"`, `"supply of readers"`, `"runs across the page"`, `"physiology of reading"`) | see entries | Used to widen snippet context. | +| 5 | archive.org full-text API (global) | `"phonetic literacy" "comprehension literacy"` | 39 items | Found Zieman 2008, Thomas 1992 and others citing or criticising Saenger's pair. | +| 6 | archive.org full-text API (global) | `"autonomous model" "ideological model" Street` | 761 | Located Street 1984 (`unset0000unse_e6x6`). | +| 7 | Crossref `/works/{doi}` | 10.1348/000712603321661859; 10.1037/0033-2909.131.1.3; 10.1484/J.VIATOR.2.301476; 10.4159/9780674038370; 10.1017/s0020743811000900; 10.1016/j.labeco.2024.102618; 10.1080/20780389.2025.2450331; 10.1093/wsr/wsr002; 10.1111/ehr.12538; 10.1086/214978; 10.1177/0956797610363406; 10.1353/phl.1991.0070; 10.1017/jea.2025.6 | 13 of 14 resolved | Titles, containers, volumes, pages matched. Hong et al. is vol. 71 (2018), online 2017. 10.1017/S0003055400210234 (a guess) returned nothing and is not used. | +| 8 | OpenAlex `works?search=` | Saenger Books of Hours; Seymour/Aro/Erskine; Hangul literacy; Turkish reform; Rawski | 0 | Daily free budget exhausted ("Rate limit exceeded"). Not retried. Rawski, *Education and Popular Literacy in Ch'ing China* (1979), is therefore **unchecked** and not cited. | +| 9 | Unpaywall | Seymour et al. 2003; Ziegler and Goswami 2005 | `is_oa: false`, no locations | Both paywalled; no author copy found. | +| 10 | Consensus | `orthographic depth and speed of literacy acquisition across European languages` | 20 | Seymour et al. abstract read; Ziegler et al. 2010 abstract read. | +| 11 | Consensus | `Turkish alphabet reform 1928 effect on literacy rates schooling` | 20 | Wood 1929, Yilmaz 2011, Filiztekin 2025 abstracts read. | +| 12 | Consensus | `Hangul design and Korean literacy: script or schooling` | 20 | Pae 2011 and others; used mainly for design claims. | +| 13 | Consensus | `Korean literacy rate rise after 1945 illiteracy campaign Hangul history schooling` | 20 | Clemens 2025 review, Becker et al. 2024, Hong et al. 2018, Ivanova 2026. | +| 14 | Consensus | `effect of Chinese character simplification on literacy rates evidence` | 20 | No causal estimate found. Weng 2023, Kuzuoglu 2022, Li 2004 are context only. | +| 15 | Consensus | `Harris Ancient Literacy rates of literacy in the Greek and Roman world estimate 10 percent` | 20 | Zanker 1991 and Sherratt 1989 reviews of Harris read (abstracts reproduce the review text). | +| 16 | WebSearch | `Saenger "Books of Hours and the Reading Habits of the Later Middle Ages" "phonetic literacy" "comprehension literacy"`; `historians criticism Saenger Space Between Words ...` | 0 useful | Returned blogs and catalogue pages only. Nothing relied on. | +| 17 | Scholar Gateway `semanticSearch` | `Did the introduction of word separation in medieval manuscripts increase literacy rates?` | error | `ACCESS_DENIED: Access denied`. Not retried. | +| 18 | archive.org metadata | each item identifier | n/a | Every book item is `access-restricted` (lending library). I did not borrow or download any of them. | + +A copy of Saenger's *Space Between Words* also sits on archive.org as an upload from a pirate-library mirror (item `space-between-words-the-origins-of-silent-reading`). Its search snippets surfaced in the global search. I **do not rely on it**: every Saenger quotation below was re-confirmed against the lending-library scan `spacebetweenword0000saen`, and I downloaded nothing from either. + +--- + +## (b) Entries + +Order: the two sources that bear most directly on H-cause (Saenger and Saenger's critics); then the histories of literacy; then the theory and the Vai evidence; then the cross-script evidence. + +### B1. Saenger and the reception of "phonetic" and "comprehension" literacy + +**saenger1997space** — Saenger, P. (1997). *Space between words: The origins of silent reading*. Stanford University Press. DOI 10.1515/9781503619081. +- *Verification:* Crossref (via Part 2 facet E); archive.org lending scan `spacebetweenword0000saen` (514 scan pages). +- *Read status:* read in part, search-inside (items 3-4 above). Nothing else was read this session. Part 2 facet E holds the snippet-level chronology. +- *What it says about literacy as a whole.* Saenger disclaims the question. Preface: "this book, which, I would stress, is not a general history of literacy. Instead, I have attempted to explain the physiological and psychological influence of word separation on the reading [process]" (preface, snippet; the object after "reading" is cut). On p. 11 (the audit corrects "front matter") he lists "the greater diffusion of literacy throughout all social strata of the population" among the "advantages that modern readers perceive as accruing from ease of reading", which the ancients "seldom viewed as advantages". He attaches this to ease of reading in general and presents it as a modern value; he offers no evidence that spacing raised the number of readers (AUDIT_01, Problems item 4). +- *Spacing as pedagogy for novices.* "The exclusive use of a syllabary method of presenting text to novice readers, necessitating the oral reconstruction..." (snippet; the sentence is cut, and it concerns the pre-separation teaching of beginners). Other snippets: "manuscripts and early printed books intended for beginning readers also followed the practice of inserting additional [marks]"; "regular word separation was not a sign of a low level of literacy or defective Latinity" (about an eleventh-century scribe; the sentence is cut). Saenger therefore treats separation both as help to the beginner and as a feature of competent scribes. +- *Why it came late.* "that modern readers perceive as accruing from ease of reading were seldom viewed as advantages by the ancients"; Part 2 facet E reports p. 11: "The notion that the greater portion of the population should be autonomous and self-motivated readers was entirely foreign to the elitist literate mentality of the ancient world." This is a **demand-side** explanation: the ancient world did not want mass reading, so it did not adopt spacing. +- *Spread.* Contents (OCR snippet, merged): "Chapter 15 Twelfth and Thirteenth Centuries: The Spread of Canonical Separation, 244". The audit finds this is chapter 14, "The Twelfth and Thirteenth Centuries" (opening p. 243), and that "The Spread of Canonical Separation" is a section at p. 244; chapter 15 is "Written Culture at the End of the Middle Ages"; "The Spread of Word Separation from England to the Low Countries"; "...to Northern France". Dates for facet A to settle. +- *Bearing:* **Supports H-cost** (separation simplified a physiological task). **Complicates H-cause**: the author most cited for the thesis calls his book "not a general history of literacy" and gives the ancient world's lack of wish for mass reading as the reason spacing was absent, which makes the demand for readers precede the page convention. + +**saenger1991separation** — Saenger, P. (1991). The separation of words and the physiology of reading. In D. R. Olson & N. Torrance (Eds.), *Literacy and orality* (ch. 12). Cambridge University Press. +- *Verification:* contents and index in `literacyorality0000unse` (archive.org lending scan; the book is Olson and Torrance 1991). +- *Read status:* read in part, search-inside. Chapter pagination not seen (the index snippet suggests the chapter runs about pp. 198-214; unconfirmed). +- *What it says:* "Word separation, by altering the physiology of reading and thereby simplifying the reading process, enabled the reader to simultaneously [...]" and "...more difficult texts with greater ease" (two adjacent snippets). The editors' introduction, summarising the chapter, says Saenger "points out that while syllabic scripts such as Hebrew have used word separation since antiquity, alphabetic scripts (Greek and Latin) abandoned [it]" and that "Japanese syllabary transcriptions contain no word separation and impose on the young reader the formidable burden of word recognition through the synthetic [method]" (the second sentence reads as Saenger's own, but the snippet's speaker is uncertain). +- *Bearing:* **Supports H-cost.** The claim is about reading efficiency for the fluent reader, not about who reads. The Hebrew sentence is a counter-datum to H-cause: a separated script from antiquity, with no mass literacy attached. + +**saenger1982silent** — Saenger, P. (1982). Silent reading: Its impact on late medieval script and society. *Viator, 13*, 367-414. DOI 10.1484/J.VIATOR.2.301476. +- *Verification:* Crossref. *Read status:* not accessible (Brepols is bot-walled; Part 2 facet E logged the same). Cited by Clanchy and Graff (see below). + +**saenger1989books** — Saenger, P. (1989). Books of Hours and the reading habits of the later Middle Ages. In R. Chartier (Ed.), *The culture of print* (page range **not found**). Princeton UP. +- *Verification:* the title and the Chartier container are confirmed by the bibliographies of Houston 2002 and Clanchy 1993 (snippets); the page range and publisher were not found. +- *Read status:* not accessible. Known through secondary use: + - Houston (2002) cites it for comprehension literacy: "This meant reading silently and with full understanding (Saenger 1989: 142)". + - Two other books paraphrase the pair: one (Margery Kempe study, item `margerykempesmed0000yosh`) writes that Saenger "categorizes reading ability into two types: phonetic literacy, defined by the ability to decode texts syllable [by syllable]... and comprehension literacy, defined as the ability to decode a written [text silently word by word and understand it]". The ancient-history volume (Thomas 1992, below) quotes the definition: "'Comprehension literacy' was 'the ability to decode a text silently word by word' and understand it fully; 'phonetic literacy' was the ability to 'decode texts syllable [by syllable] orally', close to oral rote memorization." +- *Bearing:* the pair is a classification of **reader types**, not an account of what caused readers to appear. It fits the author's interest in who can read in what way, but it ties the shift to late-medieval devotional use, with the Book of Hours as the evidence. + +**zieman2008singing** — Zieman, K. (2008). *Singing the new song: Literacy and liturgy in late medieval England*. University of Pennsylvania Press. +- *Verification:* archive.org metadata (`singingnewsongli0000ziem`). *Read status:* read in part, search-inside. +- *What it says:* a criticism of the "phonetic literacy" category. "Saenger, by contrast, identifies 'phonetic literacy' as a skill possessed by lay users of Latin..."; "modern perceptions of 'phonetic literacy' are, I suspect, based less on medieval evidence than on our own experience of, for [example...]" (snippet cut). Zieman's point (as far as the fragments show) is that medieval people did not hold a single scale on which "decoding without understanding" counts as lesser literacy. +- *Bearing:* **complicates** the use of Saenger's pair as a measure of who "could read". A fuller read is needed before this entry supports any claim stronger than "a specialist in the field disputes the category". + +**thomas1992literacy** — Thomas, R. (1992). *Literacy and orality in ancient Greece*. Cambridge University Press. +- *Verification:* archive.org metadata. *Read status:* read in part, search-inside. +- *What it says:* applies Saenger's pair to antiquity: "'Phonetic literacy' seems particularly relevant to the ancient context where reading was not done [silently...]". Also: "treat literacy as if it were a monolithic skill may be a modern fallacy." +- *Bearing:* **complicates** H-cause. An ancient historian treats unspaced reading as a different kind of literacy, not a lack of it. + +### B2. Histories of literacy that name drivers + +For each: identifiers, read status, what drivers they name, what the term search returned. All are search-inside only. + +**clanchy1993from** — Clanchy, M. T. (1993). *From memory to written record: England 1066-1307* (2nd ed.). Blackwell. (The repo card `clanchy2013` covers the 3rd edition, Wiley-Blackwell 2013.) Item `frommemorytowrit0000clan_e1n6_2ed`, 438 scan pages. +- *Read status:* read in part, search-inside. +- *Drivers named.* Administration and the habits of record. Chapter 10 is titled "Practical Literacy"; Part II is "The Literate Mentality". "The book therefore describes the growth of a literate mentality" (preface/introduction snippet). Another snippet: "Emphasis on practical literacy is a commonplace of twentieth-century culture" followed by a "not necessarily an inevitable process" clause (snippet cut). Government: "extended to nearly every sphere of government, both central and local" (ch. 10 snippet). On the pace of change: "Individuals still had few books of any size, and reading aloud persisted." +- *Word separation.* The index has "word separation" under "format and layout of writings" with a page given as 130 (index snippets: "word separation, 130"; "word separation see format and layout"; "format and layout of writings, 104-5, 131-5, 141-3, 172-4, 280-1"). Clanchy treats it as one feature of how twelfth-century documents were laid out, in a chapter on the technology of writing. Other term hits: `"scriptio continua"` 0; `"separation of words"` 0; `"word division"` 0; `"spaces between"` 0; `"silent reading"` only in notes citing Saenger (1982) and Kittay; `"Carolingian minuscule"` 1 hit (a comparison with the later script); `Saenger` only in bibliography notes (the 1982 *Viator* paper and the 1989 Chartier chapter). The audit also found a sentence beginning "Word separation was developed by Irish and Anglo-Saxon monks" (OCR-garbled), so Clanchy does state the Irish origin; the zero for "separation of words" holds for that phrasing only. Sentence and page: NEEDS LIBRARY CHECK. +- *Bearing:* **Refutes the ignition framing from the field's founding English-language case.** Clanchy's drivers are the king's government, lawyers and landholders' need for documents; spacing is a page-layout detail on p. 130. + +**stock1983implications** — Stock, B. (1983). *The implications of literacy: Written language and models of interpretation in the eleventh and twelfth centuries*. Princeton UP. Item `implicationsofli0000stoc` (626 scan pages). +- *Read status:* read in part, search-inside. +- *Drivers named.* Interpretive communities organised around a text: ch. on "Textual communities", with the heresies at Orleans (1022), Arras (1025) and Monforte (1028) as cases. "In this sense they were 'textual communities.'" Also "the notion that literacy is identical [with ...]" (snippet cut): Stock objects to equating literacy with the ability to write or read Latin. +- *Term search:* `"word separation"` 0; `"separation of words"` 0; `"word division"` 0; `"scriptio continua"` 0; `"scriptura continua"` 0; `"silent reading"` 0; `Saenger` 0 (although the Saenger 1997 acknowledgements thank Stock, and Saenger cites Stock); `punctuation` 0; `"Carolingian minuscule"` 0; `"between words"` hits are about the relation of words to things. +- *Bearing:* **Silent.** The medieval history of literacy that Saenger himself cites as an influence does not mention spacing or punctuation. + +**mckitterick1989carolingians** — McKitterick, R. (1989). *The Carolingians and the written word*. Cambridge UP. Item `carolingianswrit0000mcki` (314 scan pages). +- *Read status:* read in part, search-inside. +- *Drivers named.* The state and church: the *Admonitio generalis* of 789, which "had decreed that schools were to be [established at monasteries and cathedrals]" (snippet cut); law and the written word; the "literacy of the laity" chapter; "levels of at least pragmatic literacy were high enough for there to be some relation [between written and oral...]" (snippet cut); she follows Parkes on "pragmatic literacy". +- *Script.* The book discusses Carolingian minuscule in detail as a regional hand: "A clear St Gall version of the caroline minuscule emerged by the third decade of the ninth [century]". Term search: `legibility` 0; `legible` 0; `punctuation` 0; `"reading aloud"` 0; `"read aloud"` 0; `"word divisions"` 0; `"easier to read"` 0; `Saenger` 0 (not in the OCR, though the book predates Saenger 1997 by eight years; Saenger's 1982 article exists). +- *Bearing:* **Silent on word separation and legibility** in the one standard book on the literacy of the century in which Carolingian minuscule spread. Script appears as a palaeographic and institutional matter (scriptoria, regions), not as an access matter. + +**graff1987legacies** — Graff, H. J. (1987). *The legacies of literacy: Continuities and contradictions in western culture and society*. Indiana UP. Item `legaciesoflitera0000graf` (522 scan pages). +- *Read status:* read in part, search-inside. +- *Drivers named / thesis.* Graff's thesis is that the "literacy myth" fails: the snippets include "the 'literacy myth' no longer serves as a satisfactory [guide]" and a chapter on a "printing myth": "'printing myth,' to rest beside the 'literacy myth.'... Despite the undoubted significance of [printing]..."; "the myth of printing and the book has grown up with the literacy myth." On the quality of mass literacy: "Mass literacy was, however, never a highly skilled one; it was socially, sexually [stratified]" (snippet cut). +- *Script and silent reading.* Graff does touch the topic, briefly: "Latin punctuation developed to assist such readers" (readers who received most literature orally, snippet); "portability and readability became a greater concern. Silent reading was spreading, too."; "Personal, private, and individualized silent reading was becoming more common, but oral reading [was] not disappearing."; "The style of totally silent reading developed slowly and probably with some [difficulty]". Saenger is cited twice (1982 *Viator*; notes at "Culture, pp. 162-163, 164"). `"word separation"` 0; `"scriptio continua"` 0; `"Carolingian minuscule"` 0. +- *Bearing:* **Complicates** H-cause. Graff mentions silent reading and punctuation as developments inside the late Middle Ages, in a book whose argument is that technologies, print included, are mythologised as causes of literacy. Here is a credible historian who names the page features, and gives them a supporting, not a driving, role. + +**graff1979literacy** — Graff, H. J. (1979). *The literacy myth: Literacy and social structure in the nineteenth-century city*. Academic Press. Item `literacymythlite0000graf`. +- *Read status:* read in part, search-inside. Two hits for "silent reading" (nineteenth-century school pedagogy: "At mid-century, before silent reading was valued as a pedagogical tool...") and zero for the other terms. The book's period is the nineteenth century. The title's argument (literacy does not automatically produce the economic and moral benefits ascribed to it) is general; I did not read it here. + +**houston2002literacy** — Houston, R. A. (2002). *Literacy in early modern Europe: Culture and education 1500-1800* (2nd ed.). Longman. Item `literacyinearlym0000hous_i9l5` (326 scan pages). +- *Read status:* read in part, search-inside. +- *Drivers named.* Schooling, Reformation and Counter-Reformation, social and economic need, printing as a secondary factor. "Schooling must be seen in the context of social and economic [conditions]"; schooling "was more often an expression of social practices than it was an independent stimulator of literacy" (two snippets, the second reconstructed from "for learning as much as it was an independent stimulator of literacy"; treat the wording as approximate). "The invention and spread of printing by movable type was the most significant [of the autonomous developments]" followed by a statement that "the development of printing was episodic and its impact, as with any innovation [was ...]" (snippets cut). The Reformation's effects were "not always immediate. In Prussia and Denmark it was not the Lutheran Reformation of the [sixteenth century that mattered...]". +- *Term search:* `"word separation"` 0; `"scriptio continua"` 0; `punctuation` appears only for the quality of writers' spelling; `"silent reading"` 1 (Danish *indenadslesning* contrasted with oral reading, "first developed among the laity during the [period]"); `Saenger` 2 (the Chartier chapter, p. 142, as above). +- *Bearing:* **Refutes** a spacing-first account of the spread of reading in the period when literacy rose fastest in Europe. Houston uses Saenger only for the definition of a kind of reading. + +**cressy1980literacy** — Cressy, D. (1980). *Literacy and the social order: Reading and writing in Tudor and Stuart England*. Cambridge UP. Item `literacysocialor0000davi` (266 scan pages). +- *Read status:* read in part, search-inside. +- *Drivers named.* "The evidence indicates that the penetration of literacy in England was largely a response to need, need felt selectively in different trades [and regions]" (snippet cut after "trades"). Religion and schools: he discusses and qualifies Lockridge's thesis ("England's precocious literacy in protestantism... Intense protestantism supplied the drive to found schools"), noting that a "combination of religion and utility, rather than protestantism alone, [singled out New England]". Teaching: a chapter on the pedagogy of basic literacy under "Elizabethan and Stuart" schoolmasters (Mulcaster, Bullokar, Clement's *Petie schole*). +- *Numbers (England, signatures and marks).* "Aggregate figures for the sixteenth century are not readily available, but a reasonable guess might place male illiteracy around 80% and female illiteracy close to 95% at the time of the accession of Elizabeth" (1558). Another snippet: "The general level of literacy in England at the accession of George III would have astonished the proponents of [the old view]", with an earlier clause that a proportion "was reduced to about two out of three by the time of the accession of George III" (the subject of that proportion is unclear in the snippet; do not use as a male-literacy figure without checking). +- *Term search:* `"word separation"` 0; `"scriptio continua"` 0; `"silent reading"` 0; `Saenger` 0; `punctuation` hits are editorial notes ("Spelling and punctuation have been modernized") and one chapter reference to the *teaching* of punctuation. +- *Bearing:* **Refutes** ignition for England 1500-1750: the rise follows need and schools, 550 years after continental spacing. + +**cipolla1969literacy** — Cipolla, C. M. (1969). *Literacy and development in the West*. Penguin. Item `literacydevelopm0000unse_r1i6` (150 scan pages; Cipolla's name confirmed in the item text). +- *Read status:* read in part, search-inside. +- *Drivers named.* Economic surplus and need, state and church action, and print. "The smallness of the economic surplus remaining in hand after the satisfaction of elementary economic needs"; "The reason for this was essentially economic" (about social stratification of schooling); "On occasion, enlightened rulers such as Charlemagne and Emperor Lothar I gave a helpful hand."; on print: "All this changed - with the invention of printing. It made more sense for the common people to acquire the art of [reading]." The Reformation "indirectly... also had beneficial effects on education"; Counter-Reformation concern too. +- *Term search:* all of `"word separation"`, `"scriptio continua"`, `"silent reading"`, `Saenger`, `punctuation`, `minuscule` return 0. +- *Bearing:* **Silent**; the book is the standard economic-history account, and its drivers are incentives, religion and the state, with print as the late stimulus. + +**eisenstein1979printing** — Eisenstein, E. L. (1979). *The printing press as an agent of change*. Cambridge UP. Item `printingpressas01eise`. +- *Read status:* read in part, search-inside. `"word separation"` 0, `"scriptio continua"` 0, `punctuation` 1 (printed books supplied "regularly numbered pages, punctuation marks, section breaks, running heads, indices" as new reading aids), `Saenger` several (only the 1978 *Colard Mansion* article, on printing in Bruges, and a remark that Saenger "would have us place [certain data] in the post-Reformation [period]"). The press is her agent of change; punctuation and layout are print-era aids. Johns's critique (card `johns2002`) holds that uniformity was a social achievement, not a property of the press, and the repo card is the only treatment I rely on here. +- *Bearing:* the one thesis in this set that puts a page technology in the causal role places it in the fifteenth century, after spacing was already general, and layout (not spacing) is the affordance she names. **Complicates**: if anyone's mechanism is "interface change spreads reading", it is print's, and Johns and Graff dispute it. + +**goody1986logic** — Goody, J. (1986). *The logic of writing and the organization of society*. Cambridge UP. Item `logicofwritingor0000good`. +- *Read status:* read in part, search-inside. All 13 terms return 0, including `punctuation` and `"silent reading"`. The index scan works (other terms hit), so this is a real silence in the OCR. Goody's account of writing's effects is institutional (bureaucracy, religion, law) and about the invention of writing, not about page layout. The repo card `goody1986` is TOC-level only. + +**ong1982orality** — Ong, W. J. (1982). *Orality and literacy*. Methuen. Item `oralityliteracyt00ongw`; repo card `ong1982` (chs. 3-5 read). +- *Read status:* read in part, search-inside plus the card. One passage of interest: manuscripts "commonly ran words together or kept spaces between them minimal. Eventually, however, print [...]" (snippet; sentence cut), and "the greater legibility of print... ultimately makes for rapid, silent reading" (two snippets). Ong ties word spacing and legibility to **print**. `Saenger` 0; `"scriptio continua"` 0. +- *Bearing:* **Complicates**; in the most-cited orality theory, spacing consolidates with print. Per the standing ruling, Ong's cognitive claims are out of scope for DXD; this entry matters only for who he credits. + +### B3. Theory: literacy is not a self-acting technology + +**street1984literacy** — Street, B. V. (1984). *Literacy in theory and practice*. Cambridge UP. Item `unset0000unse_e6x6` (270 scan pages); repo card `street1984` (secondary). +- *Read status:* read in part, search-inside. +- *What it says:* the autonomous/ideological distinction: "shall characterise them as the 'ideological' model and the 'autonomous' model of literacy respectively" (snippet); the Olson, Hildyard and Goody strands are the "autonomous" targets; he criticises claims that "literacy in itself constructs superior logical functions". On writing conventions: "Recent writers (Crystal, 1976; Linell, 1983; Deuchar, 1983; Stubbs, 1980) have challenged this view" (that written conventions mirror speech). **Directly on point:** "In Vai script there are no spaces between either words or sentences and the reader has [to ...]" (snippet; the continuation is cut). Street also writes of the pre-print period that reading "was a long and laborious process; and that silent reading as we know it was very rare; until the advent of printing" (snippet, probably reporting a position he goes on to qualify; the speaker is uncertain). +- *Bearing:* **Supports** the claim that a convention of the page does not by itself define what a reader can do. Street offers no account of word separation as a cause; the Vai mention is a passing example. + +**scribner1981psychology** — Scribner, S., & Cole, M. (1981). *The psychology of literacy*. Harvard UP. Item `psychologyoflite00sylv` (366 scan pages); repo card `scribner1981` (secondary via Ong). +- *Read status:* read in part, search-inside. This is the strongest evidence in the facet. +- *What it says:* + - The Vai script, a syllabary used in Liberia, is unspaced: "A string of syllabic characters runs across the page without division into words or phrases." (ch. 3 area, near "Figure 3.2. The modern Vai syllabary"; printed page not seen). + - It was learned informally: "Vai script does not require an extended period of study and is customarily learned in adulthood"; "Unlike English schooling, Vai literacy was traditionally obtainable in the countryside"; respondents said they had begun to learn at "fourteen or fifteen years of age" (three snippets). + - Literacy without schooling had no general cognitive effect: "We could obtain no evidence that literacy in Vai script (or other [...]) [produces general cognitive effects]" (snippet cut); the question posed is "Does literacy per se (that is, literacy without schooling) [produce ...]"; "effects of schooling" were stronger: "was weak compared with effects of schooling". They name their framework "a practice account of literacy". + - A passage on English: "[conventions of English] tried to introduce word spacing and punctuation. But these attempts at standardization [...]" (snippet; the subject is a historical English scribal convention, but the sentence is cut). +- *Bearing:* Vai is a documented unspaced syllabary used by a small adult population and learned informally in adulthood by respondents who had not gone to school (Scribner and Cole). It does not show whether spacing lowers the cost for learners; Heilmann (n. 24, citing Vatri 639) notes that Saenger himself uses Vai (Saenger p. 4) and that Thai is the closer comparison for Greek. + +**havelock1963preface** — Havelock, E. A. (1963). *Preface to Plato*. Harvard UP / Belknap. Items `prefacetoplato0000have`; repo card `havelock1963`. +- *Read status:* the repo card (pp. 38-41, "probable" pagination, verified verbatim in an earlier pass) plus this session's snippet check of the key phrases. +- *Verified this session (snippets):* "upon the supply of readers, and this depended on a universalisation of [letters]"; "The reading trauma, to use a modern term, had to be imposed at [the primary level of schooling]"; "by the middle of the fourth century the silent revolution had been accomplished". Full quotations in the card. +- *Bearing:* **Supports** H-cost and the schooling driver. Havelock's decisive variable is "the supply of readers", produced by primary schooling. He names no page convention. A classicist can still be cited against the author's sentence: the "reading trauma" was imposed by an educational system, not removed by a script. + +**harris1989ancient** — Harris, W. V. (1989). *Ancient literacy*. Harvard UP. DOI 10.4159/9780674038370. +- *Verification:* Crossref. *Read status:* review only. No lending copy found (log row 2). Two reviews were read in the Consensus abstracts: Zanker (1991, *Philosophy and Literature* 15:173-175, DOI 10.1353/phl.1991.0070) and Sherratt (1989, *Antiquaries Journal*, as a section of a longer review whose DOI points to a different book; the review text on Harris is intact). +- *What the reviews report.* Zanker: Harris "convincingly shows that the ancient world lacked some of the major preconditions for widespread literacy, especially urbanization, subsidized schools, technology for the inexpensive copying of texts, a state economy complex enough to require at least a semi-literate population, and an official acceptance of the belief... that literacy might improve the quality of life of the lower social orders." Numbers: "fifth-century Attica attained a literacy level of only 5 percent to 10 percent, and certain cities of the Hellenistic age like Rhodes formed the high point at 20 percent to 30 percent (the level reached by England in the period from 1580 to 1700), while literacy in the Roman Empire never achieved more than 15 percent." Sherratt gives "15 per cent total adult literacy would be the highest likely estimate" for classical Athens and imperial Italy, and says Harris is "sceptical of the ideas of Goody... and Havelock". The 20-30 per cent figure for Rhodes is Zanker's report of Harris; the audit did not find it in Harris, whose own text uses 20-30% as the male-literacy benchmark for England 1580-1700 and puts antiquity "well below" it. +- *Bearing:* **Complicates** H-cause. Zanker reports that an unspaced antiquity reached, in some cities, the literacy level that spaced England reached in 1580-1700 [unverified: see AUDIT_01; Harris's own text puts antiquity "well below" the 20-30% England benchmark] (the equivalence is Zanker's, and Harris's estimates are disputed; see Warkentin's 2020 review of Frampton in the Consensus result, and Johnson and Parker 2009). Harris's list of preconditions contains no page convention. I could not check whether the book discusses interpuncts or word division; flag for a library user. + +### B4. Modern cross-script evidence + +**seymour2003foundation** — Seymour, P. H. K., Aro, M., & Erskine, J. M. (2003). Foundation literacy acquisition in European orthographies. *British Journal of Psychology, 94*(2), 143-174. DOI 10.1348/000712603321661859. +- *Verification:* Crossref. *Read status:* **read in full (card: library/seymour2003foundation.md)**; first read as abstract only (Consensus). Unpaywall: not open access. +- *What it says (abstract):* "children from a majority of European countries become accurate and fluent in foundation level reading before the end of the first school year. There are some exceptions, notably in French, Portuguese, Danish, and, particularly, in English." "The rate of development in English is more than twice as slow as in the shallow orthographies." The body gives the ratio as "about 2.5:1" (English readers need "2½ or more years" for what most European children reach in the first year, pp. 167–168; a regression extrapolation on reading age, not an observed time-to-mastery). Letter-sound knowledge does not differ between the languages (about 90% or better in every group, p. 151); the gap sits in word and nonword reading (Tables 5 and 6, pp. 153, 156). Cause: "fundamental linguistic differences in syllabic complexity and orthographic depth". +- *Bearing:* **Supports H-cost in the narrow sense** that orthographic design changes the **rate** of acquisition **among children in school**. It says nothing about who gets to school. + +**ziegler2005reading** — Ziegler, J. C., & Goswami, U. (2005). Reading acquisition, developmental dyslexia, and skilled reading across languages: A psycholinguistic grain size theory. *Psychological Bulletin, 131*(1), 3-29. DOI 10.1037/0033-2909.131.1.3. +- *Verification:* Crossref. *Read status:* not accessible (abstract not retrieved; Unpaywall reports no open copy). Cited by the research plan only. +- **Ziegler et al. (2010)**, *Psychological Science 21*(4), 551-559, DOI 10.1177/0956797610363406: abstract read. "Most predictors of reading performance were relatively universal across these alphabetic languages, although their precise weight varied systematically as a function of script transparency." *Read status:* **read in full (card: library/ziegler2010orthographic.md)**. Not the same bearing as Seymour: it reports predictors of reading (reading speed and accuracy) in Grade 2 in five alphabetic languages, not rates of acquisition, and the paper restricts its claim to alphabetic scripts and flags Chinese as unexplored (p. 557). + +**Turkey 1928.** +- **wood1929latinizing** — Wood, M. M. (1929). Latinizing the Turkish alphabet. *American Journal of Sociology, 35*(2), 194-203, DOI 10.1086/214978. Abstract only. A contemporary sociologist: "The old Turkish writing required nearly five hundred separate Arabic characters. Its difficulty to learn was a factor in Turkish educational lag"; but "The master-stroke in spreading knowledge of the new alphabet came with the opening of the national schools on January 1, 1929", with a four-month course for illiterates and fines for those aged sixteen to forty who could not read and had not attended. **This is the closest thing to an "interface change plus state campaign" case, and it needs both halves.** +- **yilmaz2011learning** — Yilmaz, H. (2011). Learning to read (again): The social experiences of Turkey's 1928 alphabet reform. *IJMES, 43*(4), 677-697, DOI 10.1017/s0020743811000900. Abstract only: it stresses "the actual processes, institutions, and lived experiences" and "the persistence of the Ottoman script", against a picture of silent obedience. +- **filiztekin2025unification** — Filiztekin, A. (2025). *Economic History of Developing Regions, 40*, 80-97, DOI 10.1080/20780389.2025.2450331. Abstract only: district-level data 1927-1945; "in districts where state presence was strong literacy rates increased faster, confirming the importance of state capacity"; slower in districts with a larger Kurdish minority. +- *Bearing on all three:* **Complicates** and half-supports. The script change plausibly lowered learning cost (Wood), but the measured variation in the following decades tracks state presence and schooling (Filiztekin), and the spread was uneven. +- *Not verified:* a figure circulating in a secondary article (literacy "from about 8 per cent in 1928 to over 20 per cent in 1935") appeared only in a Consensus abstract of an article about a music magazine; I did not follow it. Do not cite it. A census-based series is needed. + +**Korea and Hangul.** +- **pae2011korean** — Pae, H. K. (2011). Is Korean a syllabic alphabet or an alphabetic syllabary. *Writing Systems Research, 3*(2), 103-115, DOI 10.1093/wsr/wsr002. Abstract only: describes the script as meeting Rogers's criteria "easily acquired by the beginning reader and efficient for the skilled reader" ("arguably"). +- **clemens2025korea** — Clemens, W. (2025). Review of Cha and Pardo, *Korea: A new history of South and North*. *Journal of East Asian Studies, 25*, 159-162, DOI 10.1017/jea.2025.6. Read in full (card: library/clemens2025korea.md). The reviewer says: Hangeul sponsored in 1443 "to facilitate literacy"; "initially denounced and disparaged by the Korean educated class"; "by 1912 literacy in Korea was only 20 percent. It rose to over 60 percent by 1940." These figures are the **reviewer's own**, unsourced and not marked as the book's (the reference list does not support them); do not describe them as what Cha and Pacheco Pardo report. I have not checked them against a census. +- **ivanova2026process** — Ivanova, Y. (2026). A process of implementing the Korean alphabet on the Korean peninsula. *Ezikov Svyat, 24*(1), DOI 10.37708/ezs.swu.bg.v24i1.9. Read in full (card: library/ivanova2026process.md; the card adds statements about spacing in the body that are new to the project, and dates the Kabo reforms to 1894 and the first joint hangul-and-hanja newspaper to 1886, p. 138). Abstract: "for centuries the Korean alphabet was marginalized and associated with women, children, and the lower social classes." (DOI not Crossref-checked.) +- **becker2024conquering** — Becker, S. O., et al. (2024). Conquering Korea for Jesus. *Labour Economics, 91*, 102618, DOI 10.1016/j.labeco.2024.102618. Read in full (card: library/becker2024conquering.md). Printed abstract: Protestant church density predicts 1930 literacy, instrumented on missionary location, and Presbyterians did better than Methodists. The quotation earlier given here ("Protestantism played a crucial role in the accumulation of human capital") does not appear in the PDF (abstract or text) and is withdrawn. Flag: Becker and Won date the start of government efforts to teach Hangul to "the Gabo Reform in 1886" (p. 6), while Ivanova (p. 138) dates the Kabo reforms to 1894 and the joint newspaper to 1886; one has the date wrong or they mean different things. +- **hong2018colonization** — Hong, J., et al. (2018). Colonization and education. *Economic History Review, 71*(3), 938-964, DOI 10.1111/ehr.12538. Abstract only: early-colonial literacy "was strongly correlated with the historical presence of civil exam passers" and with private schools. +- *Bearing:* **Refutes the interface-spreads-competence model on the best test case in the set.** A script built for ease of learning existed from 1443-1446 and sat largely unused by the educated for centuries. Literacy rose in the twentieth century with mission and colonial schools, nationalism, and post-1945 campaigns. Credit goes to schools and institutions in every source I read. + +**China.** I found no causal estimate for simplification or pinyin and did not read any full text. Context only (abstracts): Weng (2023, *Harvard Journal of Asiatic Studies*, DOI 10.1353/jas.2023.a938222) on the 1935-36 simplification being withdrawn for printing-cost reasons; Kuzuoglu (2022, *Modern Chinese Literature and Culture*, DOI 10.3366/mclc.2022.0003) on the 1920s-30s mass literacy movement and "basic Chinese"; Li (2004, *Written Communication*, DOI 10.1177/0741088303262843) on the "protoliteracy" view. None isolates the script. Gap. + +--- + +## (c) Synthesis + +### 1. Drivers the historians name + +| Work | Named drivers | Weight given to page conventions | +|---|---|---| +| Clanchy 1993 | royal government, law, landholders' need for documents, a "literate mentality" | index entry "word separation, 130" under "format and layout", and a sentence crediting Irish and Anglo-Saxon monks with developing word separation (sentence and page: NEEDS LIBRARY CHECK); he lists no effect on who could read | +| Stock 1983 | heterodox and reform groups organised around texts | none (0 hits for every term) | +| McKitterick 1989 | Carolingian legislation, schools, church, lay pragmatic literacy | none (0 hits for legibility, punctuation, word division) | +| Graff 1987 | continuity and contradiction; schooling, church, state; the "literacy" and "printing" myths | punctuation and silent reading mentioned as supporting developments | +| Houston 2002 | schooling, church, social and economic need, print as secondary | none | +| Cressy 1980 | "a response to need", schools, religion | none | +| Cipolla 1969 | economic surplus, rulers, Reformation, print | none | +| Eisenstein 1979 | the printing press | punctuation and layout are print-era aids | +| Goody 1986 | writing's institutional effects | none | +| Street 1984 | social practice, not the script ("ideological" model) | Vai example only | +| Scribner and Cole 1981 | schooling vs. practice; Vai script | unspaced, learned informally | +| Havelock 1963 | "supply of readers" by primary schooling | none | +| Harris 1989 | preconditions: cities, schools, cheap copying, state, belief | none reported by reviews | + +Silence is the finding on the second question: of ten standard histories, I found word separation indexed or discussed in one (Clanchy, an index entry and one sentence crediting Irish and Anglo-Saxon monks) and silent reading or punctuation mentioned in two others (Graff, Eisenstein). Four of the ten (Stock, McKitterick, Cipolla, Goody) return zero on every spacing or silent-reading term. The OCR caveat applies. + +### 2. Timeline against numbers + +- **Spacing** (dates for facet A to settle; these are Saenger's, as reported in Part 2 facet E and in the lending scan's contents): Insular scribes seventh and eighth centuries; Continental scribes did not adopt separation "until the mid-tenth century" (Saenger p. 100), and separation was "standard and ubiquitous" by the mid-eleventh century (p. 176 [page to check in a library copy]); reference-reading skills reached the Continent "beginning in the late tenth century" (p. 4); chapter 14, "The Twelfth and Thirteenth Centuries" (p. 243), with the section "The Spread of Canonical Separation" (p. 244). +- **Literacy.** Antiquity, unspaced: Attica 5-10 per cent, Hellenistic high points 20-30 per cent, Roman empire at most 15 per cent (Harris, via Zanker and Sherratt reviews; disputed; the 20-30 per cent for Rhodes is Zanker's report of Harris and is unverified in Harris: see AUDIT_01). England: male illiteracy "around 80%" and female "close to 95%" at Elizabeth's accession in 1558 (Cressy, "a reasonable guess"). Cressy's text also reports "nearly 90%" adult male literacy by 1800 for Scotland (not England; the audit confirmed the neighbouring text is on Scotland). +- **The gap.** Spaced writing was standard on the Continent by the mid-eleventh century (Saenger p. 176 [page to check in a library copy]) and in England well before the Conquest, yet English male literacy was a fifth by 1558. Taking c. 1050 as "the mid-eleventh century", 1558 − 1050 = 508, about five centuries during which the interface was fixed and the rate stayed low. (The earlier "about the year 1000" and "five and a half centuries", 1558 − 1000 = 558, rested on a date that is not Saenger's.) The historians' account of what filled the gap is in the table above: documents for government and landholding (Clanchy), lay schooling and Carolingian reform (McKitterick), Reformation and Counter-Reformation schooling (Houston), economic need (Cressy, Cipolla), print (Cipolla, Eisenstein, qualified by Graff and Houston), and the state (Turkey, Korea). +- **Another gap.** Zanker reports that Harris puts unspaced Rhodes at what England reached in 1580-1700 [unverified: see AUDIT_01; in Harris's own text 20-30% is the male-literacy benchmark for England 1580-1700 and antiquity falls "well below" it]. The inference that an unspaced script is compatible with a literacy rate of 20-30 per cent in a city with schools is withdrawn until Harris is checked. + +### 3. Evidence on teaching + +- Saenger describes a "syllabary method of presenting text to novice readers" before separation, and says beginners' texts got extra marks; he never shows separation reduced the length of training or the number of people trained. His chapter in Olson and Torrance says separation simplified the reading **process**. +- Cressy documents the Tudor and Stuart pedagogy of basic literacy as a teaching problem solved by schoolmasters. +- Havelock: "reading trauma" imposed at the primary level. +- Scribner and Cole: Vai (unspaced) learned informally in adulthood without long study. +- Seymour et al.: among European orthographies in school, the authors attribute the slower English rate (about 2.5:1 in the body, "more than twice as slow" in the abstract, against the shallow orthographies) to syllabic complexity and orthographic depth; word spacing is not tested. +- **Nobody I found argues that separation lowered the training needed to teach a population to read.** The nearest is Saenger's process claim, which is about speed and ease for readers who already read. + +### 4. Cross-script test of "interface change spreads a competence" + +The population-scale tests (Turkey, Korea) do not isolate the script from the campaign. In both, the ease of the new script is plausible and cited, and in both the variation in outcomes tracks schooling, state capacity, missions and campaigns. The within-school test (Seymour) shows that orthographic design changes the rate of acquisition, so H-cost has support there; Ziegler et al. (2010) report predictors of reading in Grade 2 in alphabetic scripts, not rates of acquisition, and add nothing on this point. Spacing is a different property from orthographic depth, and none of these studies tests it. + +### 5. Verdict on H-cause against H-cost + +- **H-cost: supported where the sources speak.** The physiology and efficiency claim (Saenger 1991, 1997) is the only place a source argues a mechanism, and the within-school orthography studies show the principle. Part 2 facet E and the planned facet D hold the experimental reading side. +- **H-cause: not supported by any historian of literacy I could read.** The drivers named are institutions, need, schooling, religion, state and print. Saenger, the strongest candidate, disclaims a literacy history ("not a general history of literacy"), and on p. 11 he lists "the greater diffusion of literacy throughout all social strata of the population" among the "advantages that modern readers perceive as accruing from ease of reading", which the ancients "seldom viewed as advantages"; he attaches this to ease of reading in general, presents it as a modern value, and offers no evidence that spacing raised the number of readers. Nobody credible quoted here argues "ignition". +- **"Necessary-ish enabling condition, not ignition"?** Even "necessary" is hard to support. Unspaced Vai spread informally, and unspaced Greek cities reached 15-30 per cent [unverified: see AUDIT_01]; spaced Hebrew from antiquity did not produce mass literacy. What remains defensible is a narrower enabling claim: spacing plausibly made **fast silent reading of difficult texts by non-native readers** easier, which mattered for scholastic and devotional reading from the twelfth century. That is a claim about a kind of reading, not about who could read. +- **The causal arrow may run backwards.** Part 2 facet E reports Parkes (1992, p. 1) that where Latin was "a foreign language, there was a need for conventions which made it easier to read" (snippet-level). On that account, new learners (the Irish, the English) produced spacing. The facet does not settle it, but a demand-side reading is in the literature, and Saenger's own explanation of ancient absence is demand-side too. + +### 6. The most defensible sentence + +*Agent-drafted suggestion for the author to accept, revise or discard:* + +> Separating words moved some of the work of parsing from the reader to the page, which made fluent silent reading of difficult Latin easier (Saenger 1991); the historians of literacy do not credit it with the spread of reading, which they attribute to schools, religion, administration and commerce, and, later, print. + +A tighter fallback: "Word separation lowered a cost of reading; it did not by itself raise the number of readers." + +### 7. Claims in the author's framing the sources contradict + +1. **"Ignited literacy."** No history of literacy read here credits word separation. Clanchy, Cressy, Houston and Cipolla name administrative, economic, religious and schooling causes; Saenger says his book is "not a general history of literacy". +2. **"Word separation did not train the populace to read *scriptio continua*; it changed the interface so that reading no longer required the training."** (From `draft.md:263`, quoted in the research plan.) Qualified at three points: Vai script is an unspaced syllabary learned informally in adulthood (Scribner and Cole), which does not show whether spacing lowers the cost for learners; where training was heavy, the sources locate it in schooling (Havelock's "reading trauma", Houston, Cressy); spaced scripts (Hebrew, per Saenger's editors) coexisted with restricted literacy. +3. **Timing.** Continental spacing standard by the mid-eleventh century (Saenger p. 176) against English male literacy of about 20 per cent in 1558 (Cressy); mass literacy came in the eighteenth and nineteenth centuries (Houston, Graff). +4. **The interface-change model at population scale.** Hangul (from 1443) and the Turkish alphabet (1928) both show the script alone is not what the sources credit; schools and campaigns are. +5. **The framing "one move".** Not tested here (facet A and the research plan, problem 4), but Graff's and Eisenstein's "printing" and "literacy" myths are exactly the move from one convention to a general cause. +6. **Possibly: "reading as a skill created by the page."** Zieman and Thomas argue that decoding without comprehension was a legitimate kind of literacy in its setting. + +--- + +## (d) Gaps and unverified leads + +- **No book was read in full.** Every history of literacy above was queried by search-inside; none was borrowed. A library user should open: Clanchy 1993, ch. on "Writing" around p. 130 and the index entry; Saenger 1997 preface and the passage around "greater diffusion of literacy throughout all social strata"; Saenger's conclusion; McKitterick ch. 1 and the script chapter; Houston ch. 1-2 and the printing chapter; Graff 1987, the Middle Ages chapter near the "silent reading" and "Latin punctuation" hits; Scribner and Cole ch. 3 and the schooling chapters. +- **Printed page numbers** are missing for all snippet quotations. The archive.org full-text API gives scan indices only. +- **Harris, *Ancient Literacy***: no accessible copy; read only through two reviews. Whether Harris discusses interpuncts or word division is unchecked (copy needed from a library or a lending copy). +- **Saenger 1982 (*Viator*) and 1989 (Chartier)**: not accessible. The 1989 chapter's pagination and publisher are unverified. Brepols is bot-walled; the Chartier volume is a library book. +- **Ziegler and Goswami 2005** (paywalled, abstract not retrieved); **Seymour et al. 2003** (now read in full, card). +- **Stock 1983 and McKitterick 1989**: the zero-hit results rely on OCR; the McKitterick scan is 314 pages and Stock's 626, which suggests the OCR covers the book, but I did not verify page counts against the printed books. +- **Eisenstein and Johns**: no full text; relied on search-inside and repo cards that are themselves secondary. +- **Goody, Havelock, Ong**: no new reading beyond snippets and the repo cards; Goody 1986 is TOC-level in the repo and 0-hit here. +- **Zieman 2008**: the criticism of "phonetic literacy" needs the full page (Zieman's introduction) before it carries weight. +- **Turkey and Korea numbers**: the "8 per cent to 20 per cent" (Turkey) and "20 per cent in 1912, 60 per cent in 1940" (Korea, via a review) figures are not verified against census sources. A source for Turkish census literacy (1927, 1935, 1945) and for Korean colonial-period literacy is needed. +- **China**: nothing causal on simplification or pinyin; Rawski (1979) unchecked because OpenAlex rate-limited me. +- **Reviews of Saenger** (Aronoff, Richter, Mulchahey, Hanna, Walker, Baeuml; named in `REUSE.md`): not retrieved; facet B covers them. +- **Anyone arguing ignition?** I found no one credible. Candidates a human could check: later popularisations that cite Saenger; Olson's *The World on Paper* (1994) and Stock's later work are not examined here. A claim in a source I could not open (Saenger 1982) could be stronger than his 1997 disclaimer. +- **Scholar Gateway**: `ACCESS_DENIED`. **OpenAlex**: free daily budget exhausted. + +## Changes after full-text cards (2026-10-13) + +- `becker2024conquering`: the quotation "Protestantism played a crucial role in the accumulation of human capital" withdrawn (not in the PDF); replaced with the printed abstract's claim; added the Gabo 1886 versus Ivanova 1894 date flag; read status "read in full (card)". +- `clemens2025korea`: the literacy figures are the reviewer's own, unsourced, not the book's; read status "read in full (card)". +- `ivanova2026process`: read status "read in full (card)"; noted that the body's statements on spacing are new to the project and that the card dates the 1894 Kabo reforms and 1886 newspaper. +- `seymour2003foundation`: added the body's "about 2.5:1" ratio, that letter-sound knowledge does not differ, and that the gap sits in word and nonword reading; read status "read in full (card)"; Seymour paragraph and gaps list updated. +- `ziegler2010orthographic`: "same bearing as Seymour" replaced; the paper measures predictors of Grade 2 reading, not acquisition rates, and restricts its claim to alphabetic scripts (p. 557). + +Synthesis verdicts changed: one. Old: "The within-school tests (Seymour, Ziegler) show that orthographic design changes the rate of acquisition, so H-cost has support there." New: "The within-school test (Seymour) shows ... so H-cost has support there; Ziegler et al. (2010) report predictors of reading ... not rates of acquisition, and add nothing on this point." diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_D_reading_science.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_D_reading_science.bib new file mode 100644 index 00000000..946120fb --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_D_reading_science.bib @@ -0,0 +1,669 @@ +% Facet D — reading science. Agent-drafted 2026-10-12. Read status is in each note. +% Keys follow surnameYEARfirstword; rayner1998unspaced and mirault2019reading reuse the keys in +% decision_experience_design/part2_ethics/literature/facets/E_reading_history_meaning.bib. + +@article{rayner1998unspaced, + author = {Rayner, Keith and Fischer, Martin H. and Pollatsek, Alexander}, + title = {Unspaced text interferes with both word identification and eye movement control}, + journal = {Vision Research}, + year = {1998}, volume = {38}, number = {8}, pages = {1129--1144}, + doi = {10.1016/S0042-6989(97)00274-5}, + note = {Abstract only (publisher PDF bot-walled). Abstract: reading rate decreased by approx. 50 percent without spaces; larger word-frequency effect; initial landing positions changed. PMID 9666972.} +} + +@article{pollatsek1982eye, + author = {Pollatsek, Alexander and Rayner, Keith}, + title = {Eye movement control in reading: The role of word boundaries}, + journal = {Journal of Experimental Psychology: Human Perception and Performance}, + year = {1982}, volume = {8}, number = {6}, pages = {817--833}, + doi = {10.1037/0096-1523.8.6.817}, + note = {Not accessible; no abstract found. Known only through citations in Bai et al. 2008 and Winskel et al. 2009.} +} + +@article{morris1990parafoveal, + author = {Morris, Robin K. and Rayner, Keith and Pollatsek, Alexander}, + title = {Eye movement guidance in reading: The role of parafoveal letter and space information}, + journal = {Journal of Experimental Psychology: Human Perception and Performance}, + year = {1990}, volume = {16}, number = {2}, pages = {268--281}, + doi = {10.1037/0096-1523.16.2.268}, + note = {Abstract only (PubMed PMID 2142198).} +} + +@article{epelboim1994reading, + author = {Epelboim, Julie and Booth, James R. and Steinman, Robert M.}, + title = {Reading unspaced text: Implications for theories of reading eye movements}, + journal = {Vision Research}, + year = {1994}, volume = {34}, number = {13}, pages = {1735--1766}, + doi = {10.1016/0042-6989(94)90130-9}, + note = {Abstract only (PubMed PMID 7941379). Subjects read aloud and silently; modest cost reported.} +} + +@article{raynerpollatsek1996reading, + author = {Rayner, Keith and Pollatsek, Alexander}, + title = {Reading unspaced text is not easy: Comments on the implications of Epelboim et al.'s (1994) study for models of eye movement control in reading}, + journal = {Vision Research}, + year = {1996}, volume = {36}, number = {3}, pages = {461--465}, + doi = {10.1016/0042-6989(95)00132-8}, + note = {Abstract only (PubMed PMID 8746235); publisher link redirected to a cookie page.} +} + +@article{epelboim1997fillers, + author = {Epelboim, Julie and Booth, James R. and Ashkenazy, Rebecca and Taleghani, Arash and Steinman, Robert M.}, + title = {Fillers and spaces in text: The importance of word recognition during reading}, + journal = {Vision Research}, + year = {1997}, volume = {37}, number = {20}, pages = {2899--2914}, + doi = {10.1016/S0042-6989(97)00095-3}, + note = {Abstract only (PubMed PMID 9415369).} +} + +@article{perea2009space, + author = {Perea, Manuel and Acha, Joana}, + title = {Space information is important for reading}, + journal = {Vision Research}, + year = {2009}, volume = {49}, number = {15}, pages = {1994--2000}, + doi = {10.1016/j.visres.2009.05.009}, + note = {Abstract only (PubMed PMID 19463847).} +} + +@article{sheridan2013unsegmented, + author = {Sheridan, Heather and Rayner, Keith and Reingold, Eyal M.}, + title = {Unsegmented text delays word identification: Evidence from a survival analysis of fixation durations}, + journal = {Visual Cognition}, + year = {2013}, volume = {21}, number = {1}, pages = {38--60}, + doi = {10.1080/13506285.2013.767296}, + note = {Abstract only (OpenAlex). Frequency effect begins 112 ms (spaced) versus 152 ms (digits for spaces).} +} + +@article{mirault2019reading, + author = {Mirault, Jonathan and Snell, Joshua and Grainger, Jonathan}, + title = {Reading without spaces revisited: The role of word identification and sentence-level constraints}, + journal = {Acta Psychologica}, + year = {2019}, volume = {195}, pages = {22--29}, + doi = {10.1016/j.actpsy.2019.03.001}, + note = {Read in full (card: library/mirault2019reading.md). The earlier arm (Part 2 facet E) read the HAL copy and records p. 22 quotations. The 40-70% figure is unverified (AUDIT_02).} +} + +@article{mirault2019precise, + author = {Mirault, Jonathan and Snell, Joshua and Grainger, Jonathan}, + title = {Reading without spaces: The role of precise letter order}, + journal = {Attention, Perception, \& Psychophysics}, + year = {2019}, volume = {81}, number = {3}, pages = {846--860}, + doi = {10.3758/s13414-018-01648-6}, + note = {Read in full (card: library/mirault2019precise.md). Open copy: PMC6407912.} +} + +@article{veldre2017spelling, + author = {Veldre, Aaron and Drieghe, Denis and Andrews, Sally}, + title = {Spelling ability selectively predicts the magnitude of disruption in unspaced text reading}, + journal = {Journal of Experimental Psychology: Human Perception and Performance}, + year = {2017}, volume = {43}, number = {9}, pages = {1612--1628}, + doi = {10.1037/xhp0000425}, + note = {Abstract only (Consensus).} +} + +@article{drieghe2017parafoveal, + author = {Drieghe, Denis and Fitzsimmons, Gemma and Liversedge, Simon P.}, + title = {Parafoveal preview effects in reading unspaced text}, + journal = {Journal of Experimental Psychology: Human Perception and Performance}, + year = {2017}, volume = {43}, number = {10}, pages = {1701--1716}, + doi = {10.1037/xhp0000441}, + note = {Abstract only (Consensus). Southampton eprints PDF returned a bot page.} +} + +@article{christofalos2024impact, + author = {Christofalos, Andriana L. and Arco, Nicole M. and Laks, Madison and Sheridan, Heather}, + title = {The impact of interword spacing on inference processing during text reading: Evidence from eye movements}, + journal = {Discourse Processes}, + year = {2024}, volume = {62}, pages = {1--15}, + doi = {10.1080/0163853X.2024.2413811}, + note = {Abstract only (Consensus).} +} + +@article{degno2019coregistration, + author = {Degno, Federica and Loberg, Otto and Zang, Chuanli and Zhang, Manman and Donnelly, Nick and Liversedge, Simon P.}, + title = {A co-registration investigation of inter-word spacing and parafoveal preview: Eye movements and fixation-related potentials}, + journal = {PLOS ONE}, + year = {2019}, volume = {14}, number = {12}, pages = {e0225819}, + doi = {10.1371/journal.pone.0225819}, + note = {Read in part (card: library/degno2019coregistration.md). EEG, not fMRI.} +} + +@article{ozkan2026wordinitial, + author = {Özkan, Zeynep G. and Perea, Manuel}, + title = {Word-initial capitalization preserves lexical processing in unspaced text}, + journal = {Scientific Reports}, + year = {2026}, + doi = {10.1038/s41598-026-67473-2}, + note = {Read in full (card: library/ozkan2026wordinitial.md).} +} + +@article{bai2008reading, + author = {Bai, Xuejun and Yan, Guoli and Liversedge, Simon P. and Zang, Chuanli and Rayner, Keith}, + title = {Reading spaced and unspaced Chinese text: Evidence from eye movements}, + journal = {Journal of Experimental Psychology: Human Perception and Performance}, + year = {2008}, volume = {34}, number = {5}, pages = {1277--1287}, + doi = {10.1037/0096-1523.34.5.1277}, + note = {Read in full (card: library/bai2008reading.md). 1 method and results, general discussion, Tables 1--2.} +} + +@article{zang2013interword, + author = {Zang, Chuanli and Liang, Feifei and Bai, Xuejun and Yan, Guoli and Liversedge, Simon P.}, + title = {Interword spacing and landing position effects during Chinese reading in children and adults}, + journal = {Journal of Experimental Psychology: Human Perception and Performance}, + year = {2013}, volume = {39}, number = {3}, pages = {720--734}, + doi = {10.1037/a0030097}, + note = {Abstract only (PubMed PMID 23067120).} +} + +@article{shen2012eye, + author = {Shen, Deli and Liversedge, Simon P. and Tian, Jin and Zang, Chuanli and Cui, Lei and Bai, Xuejun and Yan, Guoli and Rayner, Keith}, + title = {Eye movements of second language learners when reading spaced and unspaced Chinese text}, + journal = {Journal of Experimental Psychology: Applied}, + year = {2012}, volume = {18}, number = {2}, pages = {192--202}, + doi = {10.1037/a0027485}, + note = {Abstract only (PubMed PMID 22545927).} +} + +@article{blythe2012inserting, + author = {Blythe, Hazel I. and Liang, Feifei and Zang, Chuanli and Wang, Jingxin and Yan, Guoli and Bai, Xuejun and Liversedge, Simon P.}, + title = {Inserting spaces into {Chinese} text helps readers to learn new words: An eye movement study}, + journal = {Journal of Memory and Language}, + year = {2012}, volume = {67}, number = {2}, pages = {241--254}, + doi = {10.1016/j.jml.2012.05.004}, + note = {Abstract only (Consensus). Adults and children aged 7--10.} +} + +@article{bai2013interword, + author = {Bai, Xuejun and Liang, Feifei and Blythe, Hazel I. and Zang, Chuanli and Yan, Guoli and Liversedge, Simon P.}, + title = {Interword spacing effects on the acquisition of new vocabulary for readers of {Chinese} as a second language}, + journal = {Journal of Research in Reading}, + year = {2013}, volume = {36}, number = {S1}, + doi = {10.1111/j.1467-9817.2013.01554.x}, + note = {Abstract only (OpenAlex). No page range in Crossref.} +} + +@article{cui2023eye, + author = {Cui, Yaqiong}, + title = {Eye movements of second language learners when reading spaced and unspaced {Chinese} texts}, + journal = {Frontiers in Psychology}, + year = {2023}, volume = {14}, pages = {783960}, + doi = {10.3389/fpsyg.2023.783960}, + note = {Read in full (card: library/cui2023eye.md).} +} + +@article{li2021effects, + author = {Li, Sainan and Wang, Yongsheng and Lan, Zebo and Yuan, Xiaoyuan and Zhang, Li and Yan, Guoli}, + title = {Effects of word spacing on children's reading: Evidence from eye movements}, + journal = {Reading and Writing}, + year = {2022}, volume = {35}, pages = {1019--1033}, + doi = {10.1007/s11145-021-10215-9}, + note = {Abstract only (Consensus). First online 2021.} +} + +@article{bassetti2012effects, + author = {Bassetti, Benedetta and Masterson, Jackie}, + title = {Effects of removing morphemic information and adding interword spacing on reading in {Chinese} experienced and inexperienced readers}, + journal = {Reading and Writing}, + year = {2012}, volume = {25}, number = {9}, pages = {2291--2314}, + doi = {10.1007/s11145-012-9358-7}, + note = {Read in full (card: library/bassetti2012effects.md).} +} + +@article{chen2021tradeoff, + author = {Chen, Mingjing and Wang, Yongsheng and Zhao, Bingjie and Li, Xin and Bai, Xuejun}, + title = {The trade-off between format familiarity and word-segmentation facilitation in {Chinese} reading}, + journal = {Frontiers in Psychology}, + year = {2021}, volume = {12}, pages = {602931}, + doi = {10.3389/fpsyg.2021.602931}, + note = {Read in full (card: library/chen2021tradeoff.md). A Corrigendum exists: 10.3389/fpsyg.2021.665606 (Frontiers in Psychology 12:665606);} +} + +@article{lin2017beautiful, + author = {Lin, Yu-Cheng and Lin, Pei-Ying}, + title = {The beautiful and the ugly: Reading ability modulates word spacing effects in {Chinese} children}, + journal = {The Journal of Experimental Education}, + year = {2017}, volume = {85}, number = {3}, pages = {501--512}, + doi = {10.1080/00220973.2016.1252999}, + note = {Abstract only (Consensus). Online 2016.} +} + +@inproceedings{kohsom1997adding, + author = {Kohsom, Chananda and Gobet, Fernand}, + title = {Adding spaces to {Thai} and {English}: Effects on reading}, + booktitle = {Proceedings of the 19th Annual Meeting of the Cognitive Science Society}, + year = {1997}, pages = {388--393}, publisher = {Erlbaum}, + url = {https://bura.brunel.ac.uk/bitstream/2438/2122/1/Cogsci%2797-thai.pdf}, + note = {Read in full (card: library/kohsom1997adding.md). No DOI. N = 8 Thai native speakers.} +} + +@article{winskel2009eye, + author = {Winskel, Heather and Radach, Ralph and Luksaneeyanawin, Sudaporn}, + title = {Eye movements when reading spaced and unspaced {Thai} and {English}: A comparison of {Thai}--{English} bilinguals and {English} monolinguals}, + journal = {Journal of Memory and Language}, + year = {2009}, volume = {61}, number = {3}, pages = {339--351}, + doi = {10.1016/j.jml.2009.07.002}, + note = {Read in full (card: library/winskel2009eye.md).} +} + +@article{kasisopa2013eye, + author = {Kasisopa, Benjawan and G. Reilly, Ronan and Luksaneeyanawin, Sudaporn and Burnham, Denis}, + title = {Eye movements while reading an unspaced writing system: The case of {Thai}}, + journal = {Vision Research}, + year = {2013}, volume = {86}, pages = {71--80}, + doi = {10.1016/j.visres.2013.04.007}, + note = {Abstract only (PubMed PMID 23608059).} +} + +@article{chantavarin2025effects, + author = {Chantavarin, Suphasiree and Leung, Tommi Tsz-Cheung}, + title = {The effects of interword spacing and morphological complexity in reading {Thai}: An eye-tracking study}, + journal = {Language and Cognition}, + year = {2025}, volume = {17}, + doi = {10.1017/langcog.2025.10028}, + note = {Read in full (card: library/chantavarin2025effects.md).} +} + +@article{sainio2007role, + author = {Sainio, Miia and Hyönä, Jukka and Bingushi, Kazuo and Bertram, Raymond}, + title = {The role of interword spacing in reading {Japanese}: An eye movement study}, + journal = {Vision Research}, + year = {2007}, volume = {47}, number = {20}, pages = {2575--2584}, + doi = {10.1016/j.visres.2007.05.017}, + note = {Abstract only (PubMed PMID 17697693).} +} + +@article{wang2025spaces, + author = {Wang, Xiaoyun and Halisiana, Hermiyanti Tri and Perea, Manuel}, + title = {Spaces or colors? The role of marking word boundaries on reading aloud in {Javanese} script}, + journal = {Journal of Cultural Cognitive Science}, + year = {2025}, volume = {9}, pages = {89--103}, + doi = {10.1007/s41809-024-00159-1}, + note = {Abstract only (Consensus).} +} + +@article{wang2024tibetan, + author = {Wang, Danhui and Niu, Dingyi and Li, Tianzhi and Gao, Xiaolei}, + title = {The effect of visual word segmentation cues in {Tibetan} reading}, + journal = {Brain Sciences}, + year = {2024}, volume = {14}, number = {10}, pages = {964}, + doi = {10.3390/brainsci14100964}, + note = {Read in full (card: library/wang2024tibetan.md). Two further 2026 Tibetan papers are cited in the md from Consensus abstracts only: 10.3390/bs16071119 and 10.3389/fpsyg.2026.1922806.} +} + +@article{baek2022korean, + author = {Baek, Hyunah and Choi, Wonil and Gordon, Peter C}, + title = {Reading spaced and unspaced {Korean} text: Evidence from eye-tracking during reading}, + journal = {Quarterly Journal of Experimental Psychology}, + year = {2022}, volume = {76}, number = {5}, pages = {1072--1085}, + doi = {10.1177/17470218221104736}, + note = {Abstract only (Consensus).} +} + +@article{bouma1970interaction, + author = {BOUMA, H.}, + title = {Interaction effects in parafoveal letter recognition}, + journal = {Nature}, + year = {1970}, volume = {226}, number = {5241}, pages = {177--178}, + doi = {10.1038/226177a0}, + note = {Not accessible. Content known only through Pelli and Tillman 2008 (critical spacing roughly half the eccentricity).} +} + +@article{pelli2008uncrowded, + author = {Pelli, Denis G and Tillman, Katharine A}, + title = {The uncrowded window of object recognition}, + journal = {Nature Neuroscience}, + year = {2008}, volume = {11}, number = {10}, pages = {1129--1135}, + doi = {10.1038/nn.2187}, + note = {Read in part (PMC2772078, NIH author manuscript): abstract, Bouma law, reading and dyslexia sections. An Erratum exists: 10.1038/nn1208-1463b (Nature Neuroscience 11:1463); not seen, so check it before quoting any figure.} +} + +@article{zorzi2012extra, + author = {Zorzi, Marco and Barbiero, Chiara and Facoetti, Andrea and Lonciari, Isabella and Carrozzi, Marco and Montico, Marcella and Bravar, Laura and George, Florence and Pech-Georgel, Catherine and Ziegler, Johannes C.}, + title = {Extra-large letter spacing improves reading in dyslexia}, + journal = {Proceedings of the National Academy of Sciences}, + year = {2012}, volume = {109}, number = {28}, pages = {11455--11459}, + doi = {10.1073/pnas.1205566109}, + note = {Read in part via a WebFetch extraction of the PMC3396504 HTML; the downloaded PMC XML held only the abstract. Numbers are the extractor's output; 74 children, factor of two and about 11% are unverified (AUDIT_02).} +} + +@article{zorzi2012reply, + author = {Zorzi, Marco and Barbiero, Chiara and Facoetti, Andrea and Ziegler, Johannes C.}, + title = {Reply to {Skottun} and {Skoyles}: Statistical and practical significance of extra-wide letter spacing for dyslexic children}, + journal = {Proceedings of the National Academy of Sciences}, + year = {2012}, volume = {109}, number = {44}, + doi = {10.1073/pnas.1213265109}, + note = {Not read (PMC3497756 returned front matter only).} +} + +@article{skottun2012interletter, + author = {Skottun, Bernt C. and Skoyles, John R.}, + title = {Interletter spacing and dyslexia}, + journal = {Proceedings of the National Academy of Sciences}, + year = {2012}, volume = {109}, + doi = {10.1073/pnas.1212877109}, + note = {Not read (OpenAlex returned journal boilerplate).} +} + +@article{hakvoort2017improvements, + author = {Hakvoort, Britt and van den Boer, Madelon and Leenaars, Tineke and Bos, Petra and Tijms, Jurgen}, + title = {Improvements in reading accuracy as a result of increased interletter spacing are not specific to children with dyslexia}, + journal = {Journal of Experimental Child Psychology}, + year = {2017}, volume = {164}, pages = {101--116}, + doi = {10.1016/j.jecp.2017.07.010}, + note = {Read in full (card: library/hakvoort2017improvements.md).} +} + +@article{luniewska2022effect, + author = {Łuniewska, Magdalena and Wójcik, Marta and Jednoróg, Katarzyna}, + title = {The effect of inter-letter spacing on reading performance and eye movements in typically reading and dyslexic children}, + journal = {Learning and Instruction}, + year = {2022}, volume = {80}, pages = {101576}, + doi = {10.1016/j.learninstruc.2021.101576}, + note = {Abstract only (Consensus). First online 2021.} +} + +@article{vandenboer2015default, + author = {van den Boer, Madelon and Hakvoort, Britt E.}, + title = {Default spacing is the optimal spacing for word reading}, + journal = {Quarterly Journal of Experimental Psychology}, + year = {2015}, volume = {68}, number = {4}, pages = {697--709}, + doi = {10.1080/17470218.2014.964272}, + note = {Abstract only (Consensus).} +} + +@article{perea2012effects, + author = {Perea, Manuel and Panadero, Victoria and Moret-Tatay, Carmen and Gómez, Pablo}, + title = {The effects of inter-letter spacing in visual-word recognition: Evidence with young normal readers and developmental dyslexics}, + journal = {Learning and Instruction}, + year = {2012}, volume = {22}, number = {6}, pages = {420--430}, + doi = {10.1016/j.learninstruc.2012.04.001}, + note = {Abstract only (Consensus).} +} + +@article{perea2012subtle, + author = {Perea, Manuel and Gomez, Pablo}, + title = {Subtle increases in interletter spacing facilitate the encoding of words during normal reading}, + journal = {PLoS ONE}, + year = {2012}, volume = {7}, number = {10}, pages = {e47568}, + doi = {10.1371/journal.pone.0047568}, + note = {Read in full (card: library/perea2012subtle.md). Open access; full text} +} + +@article{slattery2016interword, + author = {Slattery, Timothy J. and Yates, Mark and Angele, Bernhard}, + title = {Interword and interletter spacing effects during reading revisited: Interactions with word and font characteristics}, + journal = {Journal of Experimental Psychology: Applied}, + year = {2016}, volume = {22}, number = {4}, pages = {406--422}, + doi = {10.1037/xap0000104}, + note = {Abstract only (Consensus).} +} + +@article{galliussi2020inter, + author = {Galliussi, Jessica and Perondi, Luciano and Chia, Giuseppe and Gerbino, Walter and Bernardis, Paolo}, + title = {Inter-letter spacing, inter-word spacing, and font with dyslexia-friendly features: Testing text readability in people with and without dyslexia}, + journal = {Annals of Dyslexia}, + year = {2020}, volume = {70}, pages = {141--152}, + doi = {10.1007/s11881-020-00194-x}, + note = {Read in full (card: library/galliussi2020inter.md).} +} + +@article{duranovic2018influence, + author = {Duranovic, Mirela and Senka, Smajlagic and Babic-Gavric, Branka}, + title = {Influence of increased letter spacing and font type on the reading ability of dyslexic children}, + journal = {Annals of Dyslexia}, + year = {2018}, volume = {68}, number = {3}, pages = {218--228}, + doi = {10.1007/s11881-018-0164-z}, + note = {Abstract only (PubMed PMID 30094714).} +} + +@article{stagg2021room, + author = {Stagg, Steven D. and Kiss, Nathan}, + title = {Room to read: The effect of extra-large letter spacing and coloured overlays on reading speed and accuracy in adolescents with dyslexia}, + journal = {Research in Developmental Disabilities}, + year = {2021}, volume = {119}, pages = {104065}, + doi = {10.1016/j.ridd.2021.104065}, + note = {Abstract only (PubMed PMID 34600780).} +} + +@article{dotan2018mind, + author = {Dotan, Shahar and Katzir, Tami}, + title = {Mind the gap: Increased inter-letter spacing as a means of improving reading performance}, + journal = {Journal of Experimental Child Psychology}, + year = {2018}, volume = {174}, pages = {13--28}, + doi = {10.1016/j.jecp.2018.04.010}, + note = {Abstract only (Consensus).} +} + +@article{rayner1986eye, + author = {Rayner, Keith}, + title = {Eye movements and the perceptual span in beginning and skilled readers}, + journal = {Journal of Experimental Child Psychology}, + year = {1986}, volume = {41}, number = {2}, pages = {211--236}, + doi = {10.1016/0022-0965(86)90037-8}, + note = {Abstract only (PubMed PMID 3701249): 14--15 characters right of fixation for skilled readers, about 11 for beginners.} +} + +@article{rayner1998eye, + author = {Rayner, Keith}, + title = {Eye movements in reading and information processing: 20 years of research}, + journal = {Psychological Bulletin}, + year = {1998}, volume = {124}, number = {3}, pages = {372--422}, + doi = {10.1037/0033-2909.124.3.372}, + note = {Abstract only (PubMed PMID 9849112). Primary source for the standard eye-movement figures; not opened.} +} + +@article{rayner2009thirtyfifth, + author = {Rayner, Keith}, + title = {The 35th {Sir Frederick Bartlett Lecture}: Eye movements and attention in reading, scene perception, and visual search}, + journal = {Quarterly Journal of Experimental Psychology}, + year = {2009}, volume = {62}, number = {8}, pages = {1457--1506}, + doi = {10.1080/17470210902816461}, + note = {Abstract only (PubMed PMID 19449261).} +} + +@article{frey2026correlational, + author = {Frey, Aline and Meary, David and Loichot, Murielle and Mirault, Jonathan and Bosse, Marie-Line}, + title = {Correlational analysis of distinct contributions and overlaps between visual, visual attention, and perceptual spans}, + journal = {Scientific Reports}, + year = {2026}, volume = {16}, + doi = {10.1038/s41598-026-38243-x}, + note = {Read in full (card: library/frey2026correlational.md). Secondhand for Rayner.} +} + +@article{wertli2023influence, + author = {Wertli, Jason and Schötzau, Andreas and Palmowski-Wolfe, Anja}, + title = {The influence of age on eye movements during reading in early elementary school children}, + journal = {Klinische Monatsbl{\"a}tter f{\"u}r Augenheilkunde}, + year = {2023}, volume = {240}, pages = {591--598}, + doi = {10.1055/a-2045-7271}, + note = {Read in full (card: library/wertli2023influence.md). Secondhand.} +} + +@article{yu2023altered, + author = {Yu, Haojue and Kwon, MiYoung}, + title = {Altered eye movements during reading with simulated central and peripheral visual field defects}, + journal = {Investigative Ophthalmology \& Visual Science}, + year = {2023}, volume = {64}, number = {13}, pages = {21}, + doi = {10.1167/iovs.64.13.21}, + note = {Read in full (card: library/yu2023altered.md).} +} + +@article{cattell1886time, + author = {CATTELL, JAMES MCKEEN}, + title = {The time it takes to see and name objects}, + journal = {Mind}, + year = {1886}, volume = {os-XI}, number = {41}, pages = {63--65}, + doi = {10.1093/mind/os-XI.41.63}, + note = {Not accessible (OUP paywall). Described by Jordan et al. 2024 and Houpt et al. 2014.} +} + +@article{reicher1969perceptual, + author = {Reicher, Gerald M.}, + title = {Perceptual recognition as a function of meaningfulness of stimulus material}, + journal = {Journal of Experimental Psychology}, + year = {1969}, volume = {81}, number = {2}, pages = {275--280}, + doi = {10.1037/h0027768}, + note = {Abstract only (OpenAlex).} +} + +@article{wheeler1970processes, + author = {Wheeler, Daniel D}, + title = {Processes in word recognition}, + journal = {Cognitive Psychology}, + year = {1970}, volume = {1}, number = {1}, pages = {59--85}, + doi = {10.1016/0010-0285(70)90005-8}, + note = {Not accessible; no abstract found.} +} + +@article{jordan2024reicher, + author = {Jordan, Timothy R. and Akkaya, Aziz M. and Göçmüş, Fatma Zehra and Kalan, Aleynanur and Morgul, Ebru and Önalan, Kübra and Sheen, Mercedes Kier}, + title = {The {Reicher-Wheeler} paradigm in word recognition research: a cautionary note on its actual contributions and published misconceptions}, + journal = {Frontiers in Psychology}, + year = {2024}, volume = {15}, pages = {1399237}, + doi = {10.3389/fpsyg.2024.1399237}, + note = {Read in full (card: library/jordan2024reicher.md).} +} + +@article{houpt2014new, + author = {Houpt, Joseph W. and Townsend, James T. and Donkin, Christopher}, + title = {A new perspective on visual word processing efficiency}, + journal = {Acta Psychologica}, + year = {2014}, volume = {145}, pages = {118--127}, + doi = {10.1016/j.actpsy.2013.10.013}, + note = {Read in full (card: library/houpt2014new.md).} +} + +@article{arditi2007letter, + author = {Arditi, Aries and Cho, Jianna}, + title = {Letter case and text legibility in normal and low vision}, + journal = {Vision Research}, + year = {2007}, volume = {47}, number = {19}, pages = {2499--2505}, + doi = {10.1016/j.visres.2007.06.010}, + note = {Abstract only (PMC2016788; the PMC XML held the abstract only).} +} + +@article{dehaene2007cultural, + author = {Dehaene, Stanislas and Cohen, Laurent}, + title = {Cultural recycling of cortical maps}, + journal = {Neuron}, + year = {2007}, volume = {56}, number = {2}, pages = {384--398}, + doi = {10.1016/j.neuron.2007.10.004}, + note = {Abstract only (PubMed PMID 17964253).} +} + +@article{dehaene2010learning, + author = {Dehaene, Stanislas and Pegado, Felipe and Braga, Lucia W. and Ventura, Paulo and Filho, Gilberto Nunes and Jobert, Antoinette and Dehaene-Lambertz, Ghislaine and Kolinsky, Régine and Morais, José and Cohen, Laurent}, + title = {How learning to read changes the cortical networks for vision and language}, + journal = {Science}, + year = {2010}, volume = {330}, number = {6009}, pages = {1359--1364}, + doi = {10.1126/science.1194140}, + note = {Abstract only (PubMed PMID 21071632).} +} + +@article{dehaene2015illiterate, + author = {Dehaene, Stanislas and Cohen, Laurent and Morais, José and Kolinsky, Régine}, + title = {Illiterate to literate: behavioural and cerebral changes induced by reading acquisition}, + journal = {Nature Reviews Neuroscience}, + year = {2015}, volume = {16}, number = {4}, pages = {234--244}, + doi = {10.1038/nrn3924}, + note = {Abstract only (PubMed PMID 25783611). Repo key dehaene-etal-2015-illiterate-to-literate in dial_response_algorithmacy cards.} +} + +@article{price2003myth, + author = {Price, Cathy J and Devlin, Joseph T}, + title = {The myth of the visual word form area}, + journal = {NeuroImage}, + year = {2003}, volume = {19}, number = {3}, pages = {473--481}, + doi = {10.1016/S1053-8119(03)00084-3}, + note = {Abstract only (PubMed PMID 12880781).} +} + +@article{price2011interactive, + author = {Price, Cathy J. and Devlin, Joseph T.}, + title = {The interactive account of ventral occipitotemporal contributions to reading}, + journal = {Trends in Cognitive Sciences}, + year = {2011}, volume = {15}, number = {6}, pages = {246--253}, + doi = {10.1016/j.tics.2011.04.001}, + note = {Read in full (card: library/price2011interactive.md).} +} + +@article{hervais2019learning, + author = {Hervais-Adelman, Alexis and Kumar, Uttam and Mishra, Ramesh K. and Tripathi, Viveka N. and Guleria, Anupam and Singh, Jay P. and Eisner, Frank and Huettig, Falk}, + title = {Learning to read recycles visual cortical networks without destruction}, + journal = {Science Advances}, + year = {2019}, volume = {5}, number = {9}, pages = {eaax0262}, + doi = {10.1126/sciadv.aax0262}, + note = {Read in full (card: library/hervais2019learning.md). Repo key hervais-adelman-etal-2019-... in dial_response_algorithmacy cards.} +} + +@article{kubota2024rethinking, + author = {Kubota, Emily and Grill-Spector, Kalanit and Nordt, Marisa}, + title = {Rethinking cortical recycling in ventral temporal cortex}, + journal = {Trends in Cognitive Sciences}, + year = {2024}, volume = {28}, number = {1}, pages = {8--17}, + doi = {10.1016/j.tics.2023.09.006}, + note = {Read in full (card: library/kubota2024rethinking.md). Repo key kubota-etal-2024-rethinking-cortical-recycling.} +} + +@article{cohen2008reading, + author = {Cohen, Laurent and Dehaene, Stanislas and Vinckier, Fabien and Jobert, Antoinette and Montavont, Alexandra}, + title = {Reading normal and degraded words: Contribution of the dorsal and ventral visual pathways}, + journal = {NeuroImage}, + year = {2008}, volume = {40}, number = {1}, pages = {353--366}, + doi = {10.1016/j.neuroimage.2007.11.036}, + note = {Abstract only (PubMed PMID 18182174).} +} + +@article{zhou2023fixation, + author = {Zhou, Wei and Wang, Sile and Yan, Ming}, + title = {Fixation-related {fMRI} analysis reveals the neural basis of natural reading of unspaced and spaced {Chinese} sentences}, + journal = {Cerebral Cortex}, + year = {2023}, volume = {33}, number = {20}, pages = {10401--10410}, + doi = {10.1093/cercor/bhad290}, + note = {Abstract only (Consensus); OUP PDF returned a login page. Issue number not checked.} +} + +@article{kutas2011thirty, + author = {Kutas, Marta and Federmeier, Kara D.}, + title = {Thirty years and counting: Finding meaning in the {N400} component of the event-related brain potential ({ERP})}, + journal = {Annual Review of Psychology}, + year = {2011}, volume = {62}, pages = {621--647}, + doi = {10.1146/annurev.psych.093008.131123}, + note = {Read in full (card: library/kutas2011thirty.md).} +} + +@article{vatri2012physiology, + author = {VATRI, ALESSANDRO}, + title = {The physiology of ancient {Greek} reading}, + journal = {The Classical Quarterly}, + year = {2012}, volume = {62}, number = {2}, pages = {633--647}, + doi = {10.1017/S0009838812000213}, + note = {Not accessible (Cambridge paywall; no abstract in OpenAlex). A proxy-hosted copy exists on a third-party university proxy domain; not used. Part 2 facet E read the opening paragraphs.} +} + +@article{wickens1995proximity, + author = {Wickens, Christopher D. and Carswell, C. Melody}, + title = {The proximity compatibility principle: Its psychological foundation and relevance to display design}, + journal = {Human Factors}, + year = {1995}, volume = {37}, number = {3}, pages = {473--494}, + doi = {10.1518/001872095779049408}, + note = {Abstract only (OpenAlex).} +} + +@article{wagemans2012century, + author = {Wagemans, Johan and Elder, James H. and Kubovy, Michael and Palmer, Stephen E. and Peterson, Mary A. and Singh, Manish and von der Heydt, Rüdiger}, + title = {A century of {Gestalt} psychology in visual perception: {I}. Perceptual grouping and figure--ground organization}, + journal = {Psychological Bulletin}, + year = {2012}, volume = {138}, number = {6}, pages = {1172--1217}, + doi = {10.1037/a0029333}, + note = {Read in part (card: library/wagemans2012century.md).} +} + +@article{li2020integrated, + author = {Li, Xingshan and Pollatsek, Alexander}, + title = {An integrated model of word processing and eye-movement control during {Chinese} reading}, + journal = {Psychological Review}, + year = {2020}, volume = {127}, number = {6}, pages = {1139--1162}, + doi = {10.1037/rev0000248}, + note = {Abstract only (Consensus). Used only for the phrase "word segmentation problem in Chinese reading".} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_D_reading_science.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_D_reading_science.md new file mode 100644 index 00000000..68dddd5b --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_D_reading_science.md @@ -0,0 +1,487 @@ +# D — Reading science: what the space between words does for a reader + +*Corrected on 2026-10-12 against `../audit/AUDIT_02.md`: 12 factual fixes.* + +*Facet D of the foundation paper for Decision Experience Design. Agent-drafted on 2026-10-12. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** Experimental reading research on word spacing: (1) what happens when spaces are removed from a spaced script (English and other alphabetic scripts); (2) what happens when spaces are added to an unspaced script (Chinese, Thai, Japanese, Tibetan, Javanese), for fluent readers, learners and children; (3) spacing and crowding in beginning and dyslexic readers; (4) the perceptual mechanism and the standard figures (perceptual span, saccade length, fixation duration), with sources; (5) the word superiority effect; (6) literacy and the visual word form area (VWFA), under the standing ruling; (7) any experiment that simulates *scriptio continua* in Latin or Greek; (8) a named human-factors construct for a display change that removes a segmentation task from the operator. A final block checks the figures asserted in the two Gemini essays. + +**The question the synthesis answers.** Does experimental reading research support H-cost (spacing lowers the cost of reading, most for novices) and H-cause (spacing caused the spread of silent reading and literacy)? What effect sizes can the author quote, and which claims in the Gemini essays survive? + +**Standing ruling that governs this facet.** The paper does not claim that a design change restructures cognition. This facet therefore reports what a spacing change does to measured performance (reading time, fixations, accuracy) and treats the neuroscience as background only. + +## Legend for read status + +- *Read in full*: I read the whole text this session. +- *Read in part (sections named)*: I had the full text (open-access XML/PDF or author manuscript) and read the sections named. Where the text came from an HTML extraction by a summarising tool (WebFetch), the entry says "extractor" and the quotations are the extractor's output, not my own reading of the page. +- *Abstract only*: abstract retrieved from PubMed, Crossref, OpenAlex or Consensus; the paper itself was not opened. Paraphrase only. +- *Secondhand*: a figure that I found stated in an open paper that cites the primary source; the primary source itself was not opened. +- *Not accessible*: no text or abstract reached. + +Full texts sit in `~/iit-playground/dxd_foundation_local/pdfs/D/` (PMC XML via NCBI efetch; three PDFs). + +--- + +## (a) Search log + +All runs on 2026-10-12. "Hits" is the count the source reported, or "found / not found" for a lookup of a known work. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | repo read | `RESEARCH_PLAN.md`, `literature/REUSE.md` | — | Rules, format, leads. | +| 2 | repo read | `dial_response_algorithmacy/sources/ux_evolution_scriptio_continua.md` and `ecological_ergonomics_algorithmacy_uxd.md`, grep for figures | — | Lists of asserted figures in block (c-5). | +| 3 | repo read | `decision_experience_design/part2_ethics/literature/facets/E_reading_history_meaning.md`, grep Mirault/Rayner/Epelboim | — | Earlier arm records Mirault et al. (2019, p. 22): unspaced reading "slower by about 40% to 70%"; I could not reopen the HAL copy to check. | +| 4 | Crossref `query.bibliographic` | Rayner/Fischer/Pollatsek 1998; Pollatsek & Rayner 1982; Morris et al. 1990; Epelboim 1994, 1997; Perea & Acha 2009; Sheridan et al.; Mirault et al. | found 8 of 8 | Epelboim 1997 is "Fillers and spaces in text" (the plan's "1994 and 1997"). Sheridan et al. 2013 is *Visual Cognition*, "Unsegmented text delays word identification". | +| 5 | Crossref `query.bibliographic` | Bai et al. 2008; Zang et al. 2013; Shen et al. 2012; Blythe et al. 2012; Kohsom & Gobet 1997; Winskel et al. 2009; Sainio et al. 2007; Bouma 1970; Pelli & Tillman 2008; Zorzi et al. 2012 | found 9 of 10 | Kohsom & Gobet 1997 has no DOI (conference paper); found by WebSearch (row 17). Zang et al. 2013 JEP:HPP is the children-and-adults study. | +| 6 | Crossref `query.bibliographic` | Rayner 1998 review; Rayner 2009; Perea 2012 inter-letter; Hakvoort; Marinus; Skottun & Skoyles; Reicher 1969; Wheeler 1970; Cattell 1886; Dehaene & Cohen 2007; Dehaene 2010, 2015; Price & Devlin 2003, 2011; Kubota 2024; Hervais-Adelman 2019; Wickens & Carswell 1995; Kutas & Hillyard; Kutas & Federmeier; Vatri 2012; Pelli et al. 2003 | found 20 of 22 | Hakvoort 2017 not found by this query (it surfaced in Consensus, row 13); no Marinus et al. paper found. | +| 7 | Unpaywall `v2/{doi}` | 30 DOIs | OA copy for Bai 2008 (PMC2662925), Pelli & Tillman (PMC2772078), Zorzi 2012 (PMC3396504), Hervais-Adelman 2019, Kubota 2024, Price & Devlin 2011, Kutas & Federmeier 2011, Degno 2019, Bassetti 2012 (White Rose) | Publisher PDFs for Rayner 1998, Epelboim 1997, Perea & Acha, Mirault 2019 (HAL) all returned as bot or redirect pages (row 8). | +| 8 | direct `curl` | ScienceDirect PDF links (Rayner 1998, Epelboim 1997, Perea & Acha 2009); `doi.org` for Rayner & Pollatsek 1996 and Sainio 2007; HAL `hal-02138873/document` on `amu.hal.science` and `hal.science`; Springer PDF for Schotter et al. 2012; SAGE PDF for Rayner et al. 2016; PMC and Europe PMC PDF render for Zorzi; PNAS PDF; Cambridge page for Vatri 2012 | 0 of 14 | All returned a Cloudflare, Anubis, redirect or login page, not the paper. I did not try to get around them. | +| 9 | NCBI efetch (PMC XML) | `db=pmc&id=` for 17 PMC IDs | 17 of 17 | Works for author manuscripts and OA articles. PMC3396504 (Zorzi) returned front matter and abstract only, not the body. | +| 10 | PubMed esearch/efetch by DOI | abstracts for 25 DOIs | 20 abstracts | Rate limit (HTTP 429) on a few; retried. Pollatsek & Rayner 1982, Sheridan 2013, Blythe 2012, Cattell 1886 have no PubMed abstract. | +| 11 | Scholar Gateway | `semanticSearch`: "How much does removing spaces between words slow reading, and is the slowdown due to word identification or eye movement control?" | ACCESS_DENIED | Failed, as in earlier runs. | +| 12 | Consensus | "experimental eye-tracking study of reading Latin or ancient Greek text without word spaces (scriptio continua)" | 20 | No Latin or Greek experiment. Surfaced Mirault 2019, Rayner 1998, Perea & Acha 2009, Epelboim 1997, Drieghe 2017, Christofalos 2024, Arabic, Thai. | +| 13 | Consensus | "effect of interword spacing on reading by children learning to read or beginning readers" | 19 | Surfaced Li et al. 2021, Zang 2013, Blythe 2012, Bassetti & Masterson 2012, Hakvoort 2017, Łuniewska 2022, Dotan 2018, Perea 2012, Slattery 2016. | +| 14 | Consensus | "word spacing visual word form area fMRI unspaced text" | 20 | Surfaced Zhou et al. 2023 (fixation-related fMRI, unspaced Chinese), Degno 2019, Cohen 2008. No fMRI study of unspaced alphabetic text. | +| 15 | Consensus | "practice or training reduces cost of reading text without spaces; adaptation to unspaced text over time" | 20 | Surfaced Chen et al. 2021 (ten-day training), Lin & Lin 2016/17, Veldre et al. 2017, Baik 2022, Baek et al. 2022. | +| 16 | Consensus | "word segmentation reading Khmer Lao Burmese Tibetan Javanese interword spaces eye movements" | 20 | Tibetan (four studies), Javanese (one). **Nothing on Lao, Khmer or Burmese.** | +| 17 | WebSearch | Rayner et al. 1998 pdf; Rayner 1998 review pdf; Perea & Acha pdf; Mirault pdf; Blythe 2012; Winskel 2009; Sainio 2007; Vatri 2012; Cattell 1886 | — | Found Kohsom & Gobet (Brunel BURA PDF) and Winskel et al. (Wuppertal author PDF). Others: abstracts only. A proxy-hosted copy of Vatri 2012 exists on a university proxy domain; I did not use it. | +| 18 | direct `curl` | Brunel BURA `Cogsci'97-thai.pdf`; Wuppertal `Winskel_Radach_Thai_English_JML_2009.pdf`; White Rose `BassettiMasterson4pure.pdf` | 3 of 3 | Read (Kohsom, Winskel in part; Bassetti in part). | +| 19 | Europe PMC `search` (full text) | `"perceptual span" AND "14-15 characters" AND OPEN_ACCESS:y`; `... "3-4 characters" ...`; `"saccade" AND "7-9 characters" ...` | 4 each | Found three open papers that restate the figures: Frey et al. 2026, Yu & Kwon 2023, Wertli et al. 2023 (secondhand for Rayner). | +| 20 | WebFetch | PMC3396504 (Zorzi 2012) | extractor output | Numbers in the Zorzi entry come from this extraction. | +| 21 | Europe PMC `search` | `TITLE:"Letter case and text legibility"` | 1 | Arditi & Cho 2007 (abstract only). | +| 22 | Archive.org advanced search | Cattell 1886 *Mind*; *Mind* 1886 volume | 0 | Not on archive.org. OUP page is PDF-only behind a paywall (USD 44 for 24 h). | +| 23 | OpenAlex `works?search=` | `scriptio continua reading experiment`; `ancient Greek reading unspaced text eye movements experiment` | rate-limited (no `results`) | Retried through Consensus (row 12). | + +--- + +## (b) Entries + +Entries are grouped by the eight questions. Each gives the full reference, the bib key, how I verified it exists, read status, what it argues, quotations with locators, and its bearing on H-cost, H-cause and the four problems in `RESEARCH_PLAN.md` §4. + +### 1. Removing spaces from spaced scripts + +**rayner1998unspaced** — Rayner, K., Fischer, M. H., & Pollatsek, A. (1998). Unspaced text interferes with both word identification and eye movement control. *Vision Research, 38*(8), 1129–1144. https://doi.org/10.1016/S0042-6989(97)00274-5 +- *Verified:* Crossref title match; PubMed PMID 9666972. +- *Read status:* **abstract only** (publisher PDF bot-walled; row 8). +- *What it argues:* Two experiments with normal text, text with spaces removed (filled with x) and text with spaces kept but words flanked by x. Removing space information slows reading, enlarges the word-frequency effect on target words (word identification harmed) and moves initial landing positions (eye-movement control harmed). +- *Quotation (abstract):* "reading rate decreased by approx. 50% when space information was not available". Also: "The removal of space information increased the effect of word frequency on the fixation times for selected target words, indicating that word identification was interfered with by the lack of spaces." +- *Secondhand detail:* Bai et al. 2008 (Introduction) report that without spaces readers "fixated proportionally longer on low-frequency words than on high-frequency words" and "fixated much earlier in the word (as their average saccade lengths were much shorter when the spaces were removed)". Treat as Bai et al.'s account. +- *Bearing:* **Supports H-cost** for readers of spaced scripts: roughly a halving of reading rate when the artefact stops doing the segmenting. It says nothing about where the cost falls in a reader who never had spaces (see entries in section 2). **Complicates** any claim that the cost is only oculomotor. + +**pollatsek1982eye** — Pollatsek, A., & Rayner, K. (1982). Eye movement control in reading: The role of word boundaries. *Journal of Experimental Psychology: Human Perception and Performance, 8*(6), 817–833. https://doi.org/10.1037/0096-1523.8.6.817 +- *Verified:* Crossref title match. +- *Read status:* **not accessible** (APA; no PubMed abstract; OpenAlex abstract empty). Cited only through Winskel et al. 2009 and Bai et al. 2008, which list it among the studies behind "30–50%" and "up to 50%". +- *Bearing:* none directly. Listed so a human with library access can open it (section d). + +**morris1990parafoveal** — Morris, R. K., Rayner, K., & Pollatsek, A. (1990). Eye movement guidance in reading: The role of parafoveal letter and space information. *Journal of Experimental Psychology: Human Perception and Performance, 16*(2), 268–281. https://doi.org/10.1037/0096-1523.16.2.268 +- *Verified:* Crossref title match; PubMed PMID 2142198. +- *Read status:* **abstract only.** +- *What it argues:* Gaze-contingent manipulation of what the reader can see to the right of fixation (spaces only; spaces then letters; both at once), with the onset delayed 0–250 ms. "Although both letter and space information influenced saccade length and initial landing positions within words, only letter information had an effect on fixation duration." Two processes run in tandem: when to move (letters) and where (letters and spaces). +- *Bearing:* **Supports** a two-part mechanism: spaces serve targeting; word identification serves timing. Names the split that Rayner et al. 1998 and later work test. + +**epelboim1994reading** — Epelboim, J., Booth, J. R., & Steinman, R. M. (1994). Reading unspaced text: Implications for theories of reading eye movements. *Vision Research, 34*(13), 1735–1766. https://doi.org/10.1016/0042-6989(94)90130-9 +- *Verified:* Crossref; PubMed PMID 7941379. +- *Read status:* **abstract only.** +- *What it argues:* High-precision eye-tracking while subjects read spaced and unspaced passages silently and aloud. "Modest increases in fixation durations and decreases in overall reading speed were observed when unspaced texts were read. However, subjects read unspaced texts with the same level of comprehension and percentage of regressions as they read spaced texts." Progressive saccades were shorter in unspaced text, in proportion to the text's higher density; the same oculomotor strategy held "for both silent reading and reading aloud". Conclusion: "Words, not spaces, may serve as the perceptual units that guide the line of sight through the text." +- *Bearing:* **Complicates** H-cost: the authors report a modest cost, not a collapse (the abstract says "subjects"; the 1997 abstract says "skilled readers"). This is the one study I found that tested unspaced text read aloud against read silently (entry 7). The abstract gives no percentage; I did not read the paper. + +**raynerpollatsek1996reading** — Rayner, K., & Pollatsek, A. (1996). Reading unspaced text is not easy: Comments on the implications of Epelboim et al.'s (1994) study for models of eye movement control in reading. *Vision Research, 36*(3), 461–465. https://doi.org/10.1016/0042-6989(95)00132-8 +- *Verified:* Crossref; PubMed PMID 8746235. +- *Read status:* **abstract only.** The publisher link redirected to a cookie page. +- *Quotation (abstract):* Epelboim et al. argued that "unspaced text is relatively easy to read"; Rayner and Pollatsek "argue that (1) reading unspaced text is not easy for most readers and (2) there are more diagnostic ways to examine the role of spacing". +- *Bearing:* Records the dispute the plan asks for. The reply says "most readers", which points to variation by reader (see Veldre et al. 2017 below). + +**epelboim1997fillers** — Epelboim, J., Booth, J. R., Ashkenazy, R., Taleghani, A., & Steinman, R. M. (1997). Fillers and spaces in text: The importance of word recognition during reading. *Vision Research, 37*(20), 2899–2914. https://doi.org/10.1016/S0042-6989(97)00095-3 +- *Verified:* Crossref; PubMed PMID 9415369. +- *Read status:* **abstract only.** +- *Quotations (abstract):* "Surrounding each word with randomly chosen letters, while preserving spaces, slowed reading by 44-75%--as much as, or more than, removing spaces from normal text. Removing spaces from text with Latin-letter fillers slowed reading by only 10-20% more." "fillers in text disrupt reading by affecting word recognition directly, without necessarily affecting the eye movement pattern." +- *Bearing:* **Complicates** the picture of spaces as a pure targeting aid. In this design the cost of junk characters, not the absence of spaces, does most of the damage. Skilled readers read aloud and silently. + +**perea2009space** — Perea, M., & Acha, J. (2009). Space information is important for reading. *Vision Research, 49*(15), 1994–2000. https://doi.org/10.1016/j.visres.2009.05.009 +- *Verified:* Crossref; PubMed PMID 19463847. +- *Read status:* **abstract only.** +- *Quotation (abstract):* "Although there was a reading cost in the unspaced sentences relative to the normally written sentences, this cost was much smaller in alternatingbold unspaced sentences than in regular unspaced sentences." +- *Bearing:* **Supports** the segmentation reading of the cost. A different cue (alternating bold) recovers much of what the space did. The cost belongs to missing word boundaries, not to missing whitespace as such. + +**sheridan2013unsegmented** — Sheridan, H., Rayner, K., & Reingold, E. M. (2013). Unsegmented text delays word identification: Evidence from a survival analysis of fixation durations. *Visual Cognition, 21*(1), 38–60. https://doi.org/10.1080/13506285.2013.767296 +- *Verified:* Crossref; OpenAlex abstract. +- *Read status:* **abstract only.** (Crossref record: vol. 21, issue 1, pp. 38–60.) +- *Quotation (abstract):* "the earliest discernible influence of word frequency was 112 ms from the start of fixation in the normal text condition, and 152 ms in the unsegmented text condition." Random digits stood in for the spaces. +- *Bearing:* **Supports** the word-identification half of the mechanism with a time figure: a 40 ms delay (my subtraction) in the earliest frequency effect. This is the only paper I found that is about "alternating-shade" or unsegmented text from the Sheridan group; I found no Sheridan paper on alternating shades. + +**mirault2019reading** — Mirault, J., Snell, J., & Grainger, J. (2019). Reading without spaces revisited: The role of word identification and sentence-level constraints. *Acta Psychologica, 195*, 22–29. https://doi.org/10.1016/j.actpsy.2019.03.001 +- *Verified:* Crossref; PubMed PMID 30861421. +- *Read status:* **read in full (card: library/mirault2019reading.md)**; first read as abstract only (HAL blocked). The card confirms the abstract quotations. The "40–70%" range appears on p. 22, where the authors attribute it to Rayner & Pollatsek (1996) and Rayner et al. (1998), not to their own data. +- *Quotation (abstract):* "the relative ease with which skilled readers can read unspaced text is due to a combination of an increased use of bottom-up word identification in guiding the timing and targeting of eye movements, plus an increased interactivity between word identification and sentence-level processing." And: "the length of the currently fixated word determined the amplitude of forward saccades leaving that word during the reading of unspaced text." +- *Bearing:* **Complicates** the oculomotor story: unspaced text still lets skilled readers target by word length once the word is identified. **Supports** the point that skilled readers adapt (relevant to "reader skill"). + +**mirault2019precise** — Mirault, J., Snell, J., & Grainger, J. (2019). Reading without spaces: The role of precise letter order. *Attention, Perception, & Psychophysics, 81*(3), 846–860. https://doi.org/10.3758/s13414-018-01648-6 +- *Verified:* Crossref. *Read status:* **read in full (card: library/mirault2019precise.md)**; first listed as not read. The full text confirms the secondary summary that letter transpositions hurt unspaced reading more, and adds that the interaction was absent in first-fixation and gaze duration and was weaker when boundary letters were frequent. + +**veldre2017spelling** — Veldre, A., Drieghe, D., & Andrews, S. (2017). Spelling ability selectively predicts the magnitude of disruption in unspaced text reading. *Journal of Experimental Psychology: Human Perception and Performance, 43*(9), 1612–1628. https://doi.org/10.1037/xhp0000425 +- *Verified:* Crossref; Consensus abstract. *Read status:* **abstract only.** +- *Quotation (abstract):* "Higher reading ability was associated with greater overall reading speed and fluency in all conditions. In contrast, spelling ability selectively modulated the effect of interword spacing with poorer spelling ability predicting greater difficulty across the majority of sentence- and word-level measures." +- *Bearing:* **Supports** "for whom": among skilled adult readers, the cost of removing spaces varies with the quality of lexical representations. Better spellers are "inoculated". + +**drieghe2017parafoveal** — Drieghe, D., Fitzsimmons, G., & Liversedge, S. P. (2017). Parafoveal preview effects in reading unspaced text. *JEP: HPP, 43*(10), 1701–1716. https://doi.org/10.1037/xhp0000441 +- *Verified:* Crossref. *Read status:* **abstract only.** Replicates "increased reading times during unspaced reading" and finds "rather limited evidence for more distributed processing". + +**christofalos2024impact** — Christofalos, A. L., Arco, L., Laks, K., et al. (2024). The impact of interword spacing on inference processing during text reading: Evidence from eye movements. *Discourse Processes, 62*, 1–15. https://doi.org/10.1080/0163853X.2024.2413811 +- *Verified:* Crossref (first three authors only). *Read status:* **abstract only.** Unspaced passages gave "longer reading times, longer fixation durations, reduced skipping, and shorter saccades", and the effect of inferential constraint seen in spaced passages was "absent for unspaced passages". Bearing: the cost reaches above the word level in the study; one study, abstract only. + +**degno2019coregistration** — Degno, F., Loberg, O., Zang, C., Zhang, M., Donnelly, N., & Liversedge, S. P. (2019). A co-registration investigation of inter-word spacing and parafoveal preview: Eye movements and fixation-related potentials. *PLOS ONE, 14*(12), e0225819. https://doi.org/10.1371/journal.pone.0225819 +- *Verified:* Crossref; PMC6919607. *Read status:* **read in part (card: library/degno2019coregistration.md)** (abstract, introduction, participants, discussion paragraphs on the 300–500 ms window). The card adds that "unspaced" means unspaced in the parafoveal preview, so the study does not show how an entire unspaced page reads. +- *What it argues:* 39 Southampton students read English sentences with spaces replaced by random letters, with and without a valid preview of the next word, while eye movements and EEG were recorded. +- *Quotation (Results/Discussion):* in the 300–500 ms window, "none of the differences between unspaced and spaced preview conditions as observed in the FRP topographies reached significance in the permutation analysis." +- *Bearing:* Brain evidence about unspaced text exists only as scalp potentials, not fMRI. The window that corresponds to the N400 showed no reliable spacing effect here. **Complicates** any claim that unspaced text visibly changes meaning-level brain response in alphabetic readers (see section 6). + +**ozkan2026wordinitial** — Özkan, Z., & Perea, M. (2026). Word-initial capitalization preserves lexical processing in unspaced text. *Scientific Reports.* https://doi.org/10.1038/s41598-026-67473-2 +- *Verified:* Crossref. *Read status:* **read in full (card: library/ozkan2026wordinitial.md)**; first read as abstract only (via Consensus). Initial capitals leave the unspaced cost "additive" with word frequency, but "interword spaces provide additional visual and oculomotor cues that capitalization alone does not fully restore." Correction from the card: the design has no lowercase spaced baseline and no unspaced lowercase condition, so it cannot show what capitals restore relative to ordinary text. It shows only that the frequency effect is the same size with and without spaces once capitals are present. The authors say the design "does not test any additional cost or benefit that word-initial capitals themselves may impose relative to ordinary lowercase spaced text" (Discussion, PDF p. 11). The global reading-time cost also remained (about 7%, reviewer arithmetic from Table 2). The earlier "capitals restored lexical processing" reading is withdrawn; the Perea & Acha substitution point stands on its own. + +### 2. Adding spaces to unspaced scripts + +**bai2008reading** — Bai, X., Yan, G., Liversedge, S. P., Zang, C., & Rayner, K. (2008). Reading spaced and unspaced Chinese text: Evidence from eye movements. *JEP: HPP, 34*(5), 1277–1287. https://doi.org/10.1037/0096-1523.34.5.1277 +- *Verified:* Crossref; PMC2662925 (NIH author manuscript, no page numbers; I cite section names). +- *Read status:* **read in full (card: library/bai2008reading.md)**; first read in part (abstract, Introduction, Experiment 1 method, global and local results, General Discussion, Tables 1–2). The card adds no contradiction. +- *What it argues:* Sixteen undergraduate native readers of Chinese (Tianjin Normal University), four spacing conditions: normal unspaced, word-spaced, nonword-spaced, single-character-spaced. Experiment 2 used highlighting for the same four conditions. +- *Quotations:* "sentences with unfamiliar word spaced format were as easy to read as visually familiar unspaced text" (abstract). "inserting spaces between words (or highlighting word boundaries) did not facilitate reading Chinese, at least beyond the level observed for normal unspaced text. On the other hand, and perhaps more importantly, demarcating word boundaries either through the use of spaces or highlighting did not interfere with reading" (General Discussion). "text presented with word spacing was as easy to read as normal unspaced text" (Exp. 1, total sentence reading times). +- *Numbers (Table 1, Exp. 1, means):* total sentence reading time 5,221 ms unspaced vs 5,239 ms word-spaced vs 5,867 ms nonword-spaced vs 5,686 ms single-character-spaced; reading speed about 239 characters/min unspaced and word-spaced; mean fixation duration 246 ms unspaced vs 227 ms word-spaced; mean saccade length 2.7 vs 3.6 character spaces; fixations per sentence 17.7 vs 19.3; regressive saccades 3.5 vs 4.0. +- *Bearing:* **Complicates H-cost for fluent readers of an unspaced script**: adding the aid changed fixations and saccade length but not total time. **Supports** the claim that the word is a unit for these readers, and that a *wrong* segmentation cost them about 12% (5,867 vs 5,221 ms; my arithmetic). Also gives the secondhand figure "reading speed typically decreases by up to 50%" for spaced scripts (Introduction). From the card: the null result is on N = 16 (Exp. 1) and N = 24 (Exp. 2) skilled adults only; the "words are the unit" claim rests on the character-spacing and nonword-spacing costs; the source says nothing about learners, so it cannot carry the learner half of P4. Bai found no gain over unspaced text (5,239 ms word-spaced against 5,221 ms unspaced), which agrees with Bassetti & Masterson on no gain. + +**zang2013interword** — Zang, C., Liang, F., Bai, X., Yan, G., & Liversedge, S. P. (2013). Interword spacing and landing position effects during Chinese reading in children and adults. *JEP: HPP, 39*(3), 720–734. https://doi.org/10.1037/a0030097 +- *Verified:* Crossref; PubMed PMID 23067120. *Read status:* **abstract only.** +- *Quotation (abstract):* "interword spacing reduced children and adults' first pass reading times and refixation probabilities indicating spaces between words facilitated word identification. Word spacing effects occurred to a similar degree for both children and adults". +- *Bearing:* Spaces helped adults too, in contrast to Bai et al. 2008; children did not benefit more than adults here. **Complicates** "most for novices". The abstract gives no effect sizes and no ages; I did not read the paper. + +**shen2012eye** — Shen, D., Liversedge, S. P., Tian, J., Zang, C., Cui, L., Bai, X., Yan, G., & Rayner, K. (2012). Eye movements of second language learners when reading spaced and unspaced Chinese text. *Journal of Experimental Psychology: Applied, 18*(2), 192–202. https://doi.org/10.1037/a0027485 +- *Verified:* Crossref; PubMed PMID 22545927. *Read status:* **abstract only.** +- *Quotation (abstract):* "Eye movement measures showed least disruption to reading for word-spaced text and longer reading times for unspaced than character-spaced text, with nonword-spaced text yielding the most disruption. These effects were uninfluenced by native language". Four L2 groups (American, Korean, Japanese, Thai). +- *Bearing:* **Supports H-cost for novices**: for second-language learners word spacing was the least disruptive format. Note the native-language null: an L1 with spaces or without did not change it. + +**blythe2012inserting** — Blythe, H. I., Liang, F., Zang, C., Wang, J., Yan, G., Bai, X., & Liversedge, S. P. (2012). Inserting spaces into Chinese text helps readers to learn new words: An eye movement study. *Journal of Memory and Language, 67*(2), 241–254. https://doi.org/10.1016/j.jml.2012.05.004 +- *Verified:* Crossref; Consensus abstract. *Read status:* **abstract only.** +- *Quotation (abstract):* "We recorded adults' and 7- to 10-year-old children's eye movements as they read new 2-character words... In the learning session, participants in the spaced group read the new words more quickly than participants in the unspaced group. Further, children in the spaced group maintained this benefit in the test session (unspaced text)." +- *Bearing:* **Supports** the effect on the cost of *becoming* a reader: spaces helped children learn words and the gain persisted in unspaced text. This is the closest experiment to the novice-cost claim, but it is about vocabulary, not about learning to read. The author's note that this study is about "children learning new words" is correct. + +**bai2013interword** — Bai, X., Liang, F., Blythe, H. I., Zang, C., Yan, G., & Liversedge, S. P. (2013). Interword spacing effects on the acquisition of new vocabulary for readers of Chinese as a second language. *Journal of Research in Reading, 36*(S1). https://doi.org/10.1111/j.1467-9817.2013.01554.x +- *Verified:* Crossref; OpenAlex abstract. *Read status:* **abstract only.** "Participants in the spaced learning group read the target words more quickly than those in the unspaced learning group. This benefit was maintained at test". (Crossref gives volume and issue but no page range; the pages shown in the heading line are therefore omitted below.) + +**cui2023eye** — Cui, Y. (2023). Eye movements of second language learners when reading spaced and unspaced Chinese texts. *Frontiers in Psychology, 14*, 783960. https://doi.org/10.3389/fpsyg.2023.783960 +- *Verified:* Consensus abstract and DOI. *Read status:* **read in full (card: library/cui2023eye.md)**; first read as abstract only. Twenty-four intermediate CSL learners and 20 native controls; the abstract says learners "generally spent less time reading Chinese texts with spaces between words". Correction from the card: the word-spaced against unspaced contrast was not significant (ps > .11; about 9.5 s in 160 s, Table 4, pp. 6–7); the significant contrasts were against the non-word and pinyin versions, the regression-path contrasts missed the author's alpha of .007, and comprehension was at ceiling. The native-reader null replicates Bai et al. 2008. + +**li2021effects** — Li, S., Wang, Y., Lan, Z., Yuan, X., Zhang, L., & Yan, G. (2021/2022). Effects of word spacing on children's reading: Evidence from eye movements. *Reading and Writing, 35*, 1019–1033. https://doi.org/10.1007/s11145-021-10215-9 +- *Verified:* Crossref (volume and pages from the Crossref record; year of online publication 2021). *Read status:* **abstract only.** +- *Quotation (abstract):* "global analyses from both Experiments consistently showed that word spacing facilitated Grade 1 reading, with much reduced facilitation for higher grade readers... we obtained greater facilitatory effects of word spacing for younger compared with elder readers, which provides strong evidence for the trade-off hypothesis." +- *Bearing:* **Supports H-cost, most for novices**, directly: Chinese Grades 1–3, spaces help the youngest most. The abstract gives no effect sizes. This is the strongest single piece of evidence in the facet for "the aid matters more the earlier the reader". Its caution: the "trade-off hypothesis" is that unfamiliarity with the spaced format offsets the segmentation benefit. + +**bassetti2012effects** — Bassetti, B., & Masterson, J. (2012). Effects of removing morphemic information and adding interword spacing on reading in Chinese experienced and inexperienced readers. *Reading and Writing, 25*(9), 2291–2314. https://doi.org/10.1007/s11145-012-9358-7 +- *Verified:* Crossref; White Rose author PDF (`bassetti2012.pdf`). *Read status:* **read in full (card: library/bassetti2012effects.md)**; first read in part. +- *Quotation (abstract):* "Adding interword spacing had no facilitative effects, and had some negative effects, especially with children." +- *Also (Introduction):* "Previous research found no facilitative effects in children reading hanzi materials (Shen et al., 2010)". That is Bassetti's report of a study I did not open. +- *Bearing:* **Refutes a simple "most for novices"** in one task: primary-school children were slowed by interword spacing in a sentence-picture verification task. It conflicts with Li et al. 2021, Blythe et al. 2012 and Zang et al. 2013. The designs differ (task, materials, hanzi and pinyin); I cannot say which result a replication would favour. From the card: the child effect is negative, not null (children were slower, F(1, 21) = 11.80, p = .002, and less accurate, 95.67% against 90.91%, in hanzi with word spacing; Exp. 2, Table 2), and the adult hanzi slowdown depends on block order (0.77 sps for readers who met hanzi after pinyin, 0.03 for those who read hanzi first). Bassetti and Masterson say Bai et al. 2008 "found faster reading with interword spacing"; that holds only against character spacing. Bai found no gain over unspaced text (Table 1: 5,239 ms word-spaced, 5,686 ms character-spaced, 5,221 ms unspaced). The "twice the width" explanation is Bassetti and Masterson's, not a finding of Bai et al. + +**chen2021tradeoff** — Chen, M., Wang, Y., Zhao, B., Li, X., & Bai, X. (2021). The trade-off between format familiarity and word-segmentation facilitation in Chinese reading. *Frontiers in Psychology, 12*, 602931. https://doi.org/10.3389/fpsyg.2021.602931 +- *Verified:* Crossref. *Read status:* **read in full (card: library/chen2021tradeoff.md)**; first read as abstract only. The card confirms the Experiment 2 null on total time and speed. +- *Quotation (abstract):* In right-to-left Chinese, "faster reading speed and shorter total reading time for the inter-word spaced format"; after ten days of reading training "There was no significant difference between the total reading time and reading speed in the inter-word spaced format and unspaced format". +- *Bearing:* **Supports** that the cost of a format shrinks with practice, and that spacing's gain shrinks with it. One experiment (40 + 40 native undergraduates). The authors read the null difference as a smaller facilitation effect, not an erased one; a corrigendum is reported (10.3389/fpsyg.2021.665606) but the open copy read for the card carries none, so that point is unconfirmed. It is the nearest thing to a training result on the cost. + +**lin2017beautiful** — Lin, Y.-C., & Lin, P.-Y. (2017). The beautiful and the ugly: Reading ability modulates word spacing effects in Chinese children. *Journal of Experimental Education, 85*(3), 501–512. https://doi.org/10.1080/00220973.2016.1252999 +- *Verified:* Crossref (online 2016). *Read status:* **abstract only.** "spaced texts slow down reading speeds in children without RD. In contrast, spaced texts improved reading speeds in children with reading difficulties after some training took place." Supports the idea that the benefit goes to weaker readers; conflicts with a benefit to all children. + +**kohsom1997adding** — Kohsom, C., & Gobet, F. (1997). Adding spaces to Thai and English: Effects on reading. In *Proceedings of the 19th Annual Meeting of the Cognitive Science Society* (pp. 388–393). Erlbaum. Open copy: https://bura.brunel.ac.uk/bitstream/2438/2122/1/Cogsci%2797-thai.pdf +- *Verified:* WebSearch result with the Brunel repository record; PDF opened (6 pages). No DOI. +- *Read status:* **read in full (card: library/kohsom1997adding.md)**; first read in part. The card finds nothing in the paper that matches "gain nothing". +- *Design:* eight Thai native speakers (five women; mean age 26; students at Pittsburgh), reading aloud paragraphs in Thai and English, spaced or unspaced, coherent or scrambled. +- *Quotations (Results):* "With Thai texts, subjects were slower to read meaningless texts [F (1, 7) = 38.68, p < .001], while the addition of spaces made reading faster (the effect is marginally significant: F (1, 7) = 4.21, p < .08)." "the effect of spacing is stronger with English (on average, a difference of 33.72 sec between the spaced and unspaced conditions) than with Thai (on average, a difference of 1.37 sec) texts [t = 6.76, p < .001]." "Thai/Coherent/Unspaced... took 0.47 sec per word on average... or 46.2 sec per paragraph." "Subjects' retrospective reports indicate that, if anything, they thought that spaces were making reading harder." +- *Calculation (mine):* 1.37 s on a 46.2 s paragraph is about 3%, from a marginal effect in 8 people. +- *Transfer:* "We found that there was no such transfer": Thai readers' unspaced English coherent text took 55% longer than spaced, against 30% to 48% for English natives (the authors' citation of Rayner & Pollatsek 1996). +- *Bearing:* The abstract ("subjects were faster in reading and made less errors when spaces were added") overstates the Thai result. **Complicates** the oft-cited claim that spaces speed Thai reading; the effect is small, marginal and n = 8. **Supports** that unspaced skill does not transfer to a spaced script. From the card: the paper's own claim is a small gain for natives of an unspaced script (time, marginal; six of eight faster on coherent text; fewer errors, though only "word added" errors reliably) and a large gain for the second-language English text, where readers have no practice without spaces (33.72 s against 1.37 s per paragraph). That fits P3 and the learners clause of P4, not "gain nothing". + +**winskel2009eye** — Winskel, H., Radach, R., & Luksaneeyanawin, S. (2009). Eye movements when reading spaced and unspaced Thai and English: A comparison of Thai–English bilinguals and English monolinguals. *Journal of Memory and Language, 61*(3), 339–351. https://doi.org/10.1016/j.jml.2009.07.002 +- *Verified:* Crossref; Wuppertal author PDF (24 pages, `winskel2009.pdf`). +- *Read status:* **read in full (card: library/winskel2009eye.md)**; first read in part. The card confirms the cost of removing spaces and notes the Thai result is the more informative one for the unspaced-script debate: a 5% slowdown at sentence level with a word-level gain (shorter gaze and total fixation times), and large differences between readers (15 of 36 slower, 6 faster, 15 unaffected, p. 345). +- *Design:* 36 Thai–English bilinguals (Chulalongkorn University) and 36 English monolinguals (Western Sydney); high- and low-frequency target words. +- *Quotations:* "Removing spaces in English typically slows reading by 30–50%, disrupting both the way the eyes move through the text and the word identification process" (Introduction; Winskel's summary of Morris et al. 1990, Pollatsek & Rayner 1982, Rayner et al. 1998, Spragins et al. 1976). Thai: "There was a 5% decrement in reading rate when spaces were introduced into the text. Reading times may have been longer in the spaced condition because the sentences were 12.8% longer than in the unspaced condition." English: "There was a 45% decrement in reading rate in the unspaced condition" (bilinguals); "only 33% decrement in reading rate in the monolinguals". "Some participants' global reading measures were deleteriously affected by the unusual spaced format (42%), yet the reading of other participants was either facilitated (16%) or not affected (42%)." "Interword spaces have a selective facilitatory effect on reading in Thai, as word processing is facilitated, but eye guidance (word targeting and lexical segmentation) is not facilitated". Transfer: "the prediction that there would be a positive transference effect when reading English due to their life time experience of reading Thai without spaces was not supported. It appears that language ability in the specific language being read affects performance more than prior experience with an unrelated orthography." +- *Bearing:* The best single dataset for H-cost's two sides. **Supports** a cost of 33% (monolingual) and 45% (L2-proficiency-limited) when spaces are removed from English; **complicates** the Thai side: Thai readers took about 5% longer on spaced sentences (p = .05 by participants, minF' p = .13; the spaced sentences were 12.8% longer), with word recognition helped and eye guidance not. By participant, 42% were slowed, 16% sped up, 42% unaffected. + +**kasisopa2013eye** — Kasisopa, B., Reilly, R. G., Luksaneeyanawin, S., & Burnham, D. (2013). Eye movements while reading an unspaced writing system: The case of Thai. *Vision Research, 86*, 71–80. https://doi.org/10.1016/j.visres.2013.04.007 +- *Verified:* Crossref; PubMed PMID 23608059. *Read status:* **abstract only.** "Thai readers tended to land their first fixation at or near the centre of words, just as readers of spaced texts do", guided by "word boundary characters". Bearing: skilled readers of an unspaced alphabet target word centres without spaces, using orthographic regularities (this is the mechanism behind "skilled readers do not need spaces"). + +**chantavarin2025effects** — Chantavarin, S., & Leung, T. T.-C. (2025). The effects of interword spacing and morphological complexity in reading Thai: An eye-tracking study. *Language and Cognition.* https://doi.org/10.1017/langcog.2025.10028 +- *Verified:* Consensus abstract and DOI only; not resolved against Crossref. *Read status:* **read in full (card: library/chantavarin2025effects.md)**; first read as abstract only. "interword spacing has a facilitative effect on word identification... shorter first fixation duration, gaze duration and total fixation time", larger for compounds. Bearing: Thai spacing helps word identification, as Winskel et al. found. The abstract did not report sentence-level rates; the body does: sentence reading time was longer with spaces (3,804 ms against 3,676 ms for compounds; 3,675 against 3,546 for simple nouns; spacing estimate 132 ms, SE 35; Tables 2–3, pp. 11–12), while first fixation on the target fell from 330 to 297 ms for compounds (Table 2, p. 11). The authors say the facilitative effect "is not due to familiarity with spaced text" (p. 17). So adult native Thai readers gain at the word level and lose at the sentence level; "gain nothing" (P4) holds only for whole-sentence speed. + +**sainio2007role** — Sainio, M., Hyönä, J., Bingushi, K., & Bertram, R. (2007). The role of interword spacing in reading Japanese: An eye movement study. *Vision Research, 47*(20), 2575–2584. https://doi.org/10.1016/j.visres.2007.05.017 +- *Verified:* Crossref; PubMed PMID 17697693. *Read status:* **abstract only.** +- *Quotation (abstract):* "Interword spacing facilitated both word identification and eye guidance when reading syllabic script, but not when the script contained ideographic characters. We conclude that in reading Hiragana interword spacing serves as an effective segmentation cue. In contrast, spacing information in mixed Kanji-Hiragana text is redundant, since the visually salient Kanji characters serve as effective segmentation cues by themselves." +- *Bearing:* **Supports** a clean rule: spacing helps where the script gives no other boundary cue, and adds nothing where it does. This is the most general statement of "what it is not needed for". + +**wang2025spaces** — Wang, X., Halisiana, F., & Perea, M. (2025). Spaces or colors? The role of marking word boundaries on reading aloud in Javanese script. *Journal of Cultural Cognitive Science, 9*, 89–103. https://doi.org/10.1007/s41809-024-00159-1 +- *Verified:* Crossref. *Read status:* **abstract only.** "developing readers benefited from both types of word segmentation cues. For adult readers, only the insertion of interword spaces facilitated reading aloud." Reading aloud, an unspaced alphasyllabary, developing and adult readers. + +**wang2024tibetan** — Wang, D., Niu, D., Li, et al. (2024). The effect of visual word segmentation cues in Tibetan reading. *Brain Sciences, 14*(10), 964. https://doi.org/10.3390/brainsci14100964 +- *Verified:* Crossref (first authors). *Read status:* **read in full (card: library/wang2024tibetan.md)**; first read as abstract only. "interword spaces facilitate reading and lexical recognition and aid in saccade target selection during reading." From the card: word spacing shortened sentence reading time, 5,566 ms against 6,019 ms for normal text (Exp. 1, Table 1), with fewer fixations and shorter gaze and total fixation durations on target words; the authors had predicted no global effect. The reading confirms spacing helps; it does not support "no gain for fluent readers". The sample is university students, not learners, and the authors tie the gain to first-language attrition and reliance on the tsheg. Two further Tibetan studies by the same group (not opened): Liang et al. 2026, *Behavioral Sciences* 16, 1119, https://doi.org/10.3390/bs16071119: "inter-word spaces significantly facilitated Tibetan reading" under low format familiarity, the effect "disappeared as format familiarity increased"; Zhang et al. 2026, *Frontiers in Psychology* 17, https://doi.org/10.3389/fpsyg.2026.1922806: "full-width spaces provided a stronger facilitative effect on lexical recognition for low-proficiency readers". Bearing: Tibetan, like Chinese, shows the familiarity trade-off and a gain for low-proficiency readers. These are 2024–2026 papers from one group; none replicated elsewhere. + +**baek2022korean** — Baek, H., Choi, W., Gordon, P. C., et al. (2022). Reading spaced and unspaced Korean text. *Quarterly Journal of Experimental Psychology, 76*(5), 1072–1085. https://doi.org/10.1177/17470218221104736 +- *Verified:* Crossref (first authors). *Read status:* **abstract only.** Removing spaces from Korean text (which has spaces between phrasal units) "hindered reading, but this effect was not as disruptive as previously found in English". Bearing: the cost of removing spaces varies by script; it was smaller where particles supply a segmentation cue. + +**Lao, Khmer, Burmese:** I found no eye-movement or reading experiment on spaced versus unspaced text in any of the three (rows 16 and 5). Tibetan and Javanese are the only additional scripts reached. + +### 3. Beginning and struggling readers: spacing and crowding + +**bouma1970interaction** — Bouma, H. (1970). Interaction effects in parafoveal letter recognition. *Nature, 226*, 177–178. https://doi.org/10.1038/226177a0 +- *Verified:* Crossref. *Read status:* **not accessible** (Nature; no abstract). I cite its content only as Pelli and Tillman state it. +- *Via Pelli & Tillman (section "The Bouma law"):* "a generalization of an observation that Herman Bouma reported in 1970, that the critical spacing for identification of small letters is roughly half the eccentricity". + +**pelli2008uncrowded** — Pelli, D. G., & Tillman, K. A. (2008). The uncrowded window of object recognition. *Nature Neuroscience, 11*(10), 1129–1135. https://doi.org/10.1038/nn.2187 +- *Verified:* Crossref; PMC2772078 (NIH author manuscript). An erratum exists (10.1038/nn1208-1463b); I could not see what it corrects (AUDIT_02), so check it before quoting any figure. +- *Read status:* **read in part** (abstract, "The Bouma law", the reading and dyslexia sections, figure captions; not the object-recognition sections in full). +- *Quotations:* "the Bouma law: for an object that can be identified in isolation, our ability to identify it among similar objects depends solely on the ratio of the object spacing to the observer's critical spacing at that location. The object is crowded whenever the ratio is less than one." "It has long been known that reading consists of a series of eye fixations, 4 per second". "Reading speed depends on letter spacing and not size." "Incidentally, note that letters at the ends of words are much less crowded and have a larger uncrowded window." On development and dyslexia: "For all the normal readers, including both children and adults, reading speed was fairly well predicted (with no degrees of freedom) by the product of span and the standard 4-Hz rate of fixations" and dyslexics "read much more slowly than is predicted by their span: They were all well below the normal line, reading at less than half of their span-predicted speed." +- *Bearing:* **Supports** the mechanism for why a space helps: letters at word ends are less crowded (my inference, not Pelli and Tillman's: a space beside a word end leaves that side free), and the uncrowded span (the figure caption demonstration puts individual spans at 3 to 9 letters) bounds reading speed. **Complicates** crowding as the cause of dyslexia: the review finds dyslexic readers below the crowding-predicted line, so "something else" accounts for their deficit. + +**zorzi2012extra** — Zorzi, M., Barbiero, C., Facoetti, A., Lonciari, I., Carrozzi, M., Montico, M., Bravar, L., George, F., Pech-Georgel, C., & Ziegler, J. C. (2012). Extra-large letter spacing improves reading in dyslexia. *PNAS, 109*(28), 11455–11459. https://doi.org/10.1073/pnas.1205566109 +- *Verified:* Crossref; PMC3396504. +- *Read status:* **read in part via extractor** (WebFetch of the PMC HTML; the PMC XML I downloaded holds only the abstract). The numbers below are as the extractor returned them; "74", "factor of two" and "about 11%" are [unverified: see AUDIT_02]. +- *Numbers:* 74 dyslexic children (34 Italian, 40 French; ages 8 to 14, mean 10.4). Standard interletter spacing was increased by 2.5 pt; words were "separated by three space characters" and line spacing doubled in the spaced text. "The manipulation of letter spacing improved accuracy in text reading by a factor of two." Speed gain "about 0.3 syllables/s". Against reading-level controls: dyslexics 1.75 vs 1.94 syllables/s (normal vs spaced, P = 0.001); controls 1.87 vs 1.93 (P > 0.3). A second experiment (20 Italian dyslexic children) had 6 vs 11.8 errors, and 2 sessions later 4.95 vs 13.5. Letter-identification (crowding) correlation r = −0.60, P = 0.005 in 20 French children. +- *Calculation (mine):* 1.94 vs 1.75 syllables/s is about +11% speed [unverified: see AUDIT_02; the audit notes this does not match the 0.3 syllables/s above]. +- *Bearing:* Large accuracy effect, modest speed effect, for readers who are weak at the task. **The widening included word spacing**, so the design cannot separate letter from word spacing (the extractor said as much). **Supports** "spacing helps the struggling reader most", but see the replications. + +**zorzi2012reply, skottun2012interletter** — Zorzi et al. (2012) Reply to Skottun and Skoyles: Statistical and practical significance of extra-wide letter spacing for dyslexic children. *PNAS, 109*(44). https://doi.org/10.1073/pnas.1213265109; Skottun, B. C., & Skoyles, J. R. (2012). Interletter spacing and dyslexia. *PNAS, 109*. https://doi.org/10.1073/pnas.1212877109. *Read status:* **not read** (PMC3497756 returned front matter; Skottun and Skoyles returned journal boilerplate). The existence of the exchange is confirmed by Crossref; its content is not. + +**Replications and failures to replicate.** All abstract only. +- **hakvoort2017improvements** — Hakvoort, B., van den Boer, M., Leenaars, T., Bos, P., & Tijms, J. (2017). Improvements in reading accuracy as a result of increased interletter spacing are not specific to children with dyslexia. *J. Experimental Child Psychology, 164*, 101–116. https://doi.org/10.1016/j.jecp.2017.07.010. "Increased LS resulted in fewer errors, not faster reading rates"; "no indications that the effect of LS was related to reading ability"; "the effect of LS seems to occur at the interword level" in the abstract; the body (General Discussion, p. 112) says "sentence level", so cite the body (30 dyslexic and 30 control children; 189 unselected readers). [read in full: card library/hakvoort2017improvements.md] +- **luniewska2022effect** — Łuniewska, M., Wójcik, M., & Jednoróg, K. (2022). The effect of inter-letter spacing on reading performance and eye movements in typically reading and dyslexic children. *Learning and Instruction, 80*, 101576 (2022; first online 2021). https://doi.org/10.1016/j.learninstruc.2021.101576. 38 dyslexic and 32 typical readers, ages 10–14: "Inter-letter spacing did not affect reading accuracy, comprehension, or speed." A failure to replicate. +- **vandenboer2015default** — van den Boer, M., & Hakvoort, B. (2015). Default spacing is the optimal spacing for word reading. *QJEP, 68*(4), 697–709. https://doi.org/10.1080/17470218.2014.964272. "neither beginning nor advanced readers benefited from an increase in interletter spacing"; poorer fluency under tighter-than-default spacing. +- **perea2012effects** — Perea, M., Panadero, V., Moret-Tatay, C., & Gómez, P. (2012). The effects of inter-letter spacing in visual-word recognition: Evidence with young normal readers and developmental dyslexics. *Learning and Instruction, 22*(6), 420–430. https://doi.org/10.1016/j.learninstruc.2012.04.001. Increased spacing shortened word-identification times in adults, Grade 2 and 4 children, "and, even to a larger degree, with readers with dyslexia". +- **perea2012subtle** — Perea, M., & Gómez, P. (2012). Subtle increases in interletter spacing facilitate the encoding of words during normal reading. *PLoS ONE, 7*(10), e47568. https://doi.org/10.1371/journal.pone.0047568. Fixations briefest at +1.5 spacing in skilled readers. [read in full: card library/perea2012subtle.md; total reading time did not follow the briefest fixations] +- **slattery2016interword** — Slattery, T. J., Yates, M., & Angele, B. (2016). Interword and interletter spacing effects during reading revisited: Interactions with word and font characteristics. *J. Experimental Psychology: Applied, 22*(4), 406–422. https://doi.org/10.1037/xap0000104. "the direction of spacing effects depend on the size of the default spacing chosen by font developers"; "interletter spacing interacts with interword spacing". +- **galliussi2020inter** — Galliussi, J., Perondi, L., Chia, G., Gerbino, W., & Bernardis, P. (2020). Inter-letter spacing, inter-word spacing, and font with dyslexia-friendly features. *Annals of Dyslexia, 70*, 141–152. https://doi.org/10.1007/s11881-020-00194-x. Two groups of 64 children: "reading speed is impaired by an increase in inter-letter spacing not combined with an adequate increase in inter-word spacing." [read in full: card library/galliussi2020inter.md] The authors call the effect practically negligible (under 0.20 syll/s, p. 150). The letterform had no main effect, but there was a letterform × inter-word-spacing interaction in the dyslexic group (p. 147). +- **duranovic2018influence** — Duranovic, M., Smajlagic, S., & Babic-Gavric, B. (2018). Influence of increased letter spacing and font type on the reading ability of dyslexic children. *Annals of Dyslexia, 68*, 218–228. https://doi.org/10.1007/s11881-018-0164-z. "only spacing plays a role in enhancing dyslexic individuals' reading performance". +- **stagg2021room** — Stagg, S. D., & Kiss, N. (2021). Room to read. *Research in Developmental Disabilities, 119*, 104065. https://doi.org/10.1016/j.ridd.2021.104065. 32 dyslexic and 27 control adolescents (mean age 13): extra-large spacing "significantly improved reading speed more substantially for the dyslexia group". +- **dotan2018mind** — Dotan, S., & Katzir, T. (2018). Mind the gap. *J. Experimental Child Psychology, 174*, 13–28. https://doi.org/10.1016/j.jecp.2018.04.010. 132 Hebrew first and third graders: "increased spacing contributed to reading accuracy without affecting reading rate"; "all first graders benefited"; in third grade only low achievers. +- **bellocchi2018effects** (Bellocchi et al., 2018, *Child Neuropsychology*; DOI from Consensus did not resolve at Crossref, so no bib entry). Extra spacing disrupted saccade landing in beginning readers and dyslexics; the authors say the benefit "can be better explained in terms of decreased visual crowding than improved saccade targeting." +- **Inter-letter versus inter-word.** In the studies above, the gain from letter spacing is mostly accuracy, sometimes speed, and depends on default spacing and on the word spacing that goes with it (Slattery; Galliussi; for Hakvoort, Experiment 1 widened word spaces too, and Experiment 2, letter spacing only, gave the same result, p. 113, so the abstract's "interword level" is better read as the body's "sentence level"). This is the point at which the facet's topic (word spacing) and the dyslexia literature (letter spacing) meet. + +**rayner1986eye** — Rayner, K. (1986). Eye movements and the perceptual span in beginning and skilled readers. *J. Experimental Child Psychology, 41*(2), 211–236. https://doi.org/10.1016/0022-0965(86)90037-8 +- *Verified:* Crossref; PubMed PMID 3701249. *Read status:* **abstract only.** +- *Quotation (abstract):* "the perceptual span of beginning readers extends about 11 character spaces to the right of fixation; for skilled readers, the span extends 14-15 spaces to the right of fixation." "It is concluded that the size of the perceptual span does not cause beginning readers' slow reading rates." (Second-, fourth- and sixth-grade children and adults.) +- *Bearing:* Sizes the beginner gap in the mechanism: 11 against 14–15 characters. The span is not the cause of slow reading; this complicates a story in which spaces enlarge the span for novices. + +### 4. The perceptual mechanism and the standard figures + +**rayner1998eye** — Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. *Psychological Bulletin, 124*(3), 372–422. https://doi.org/10.1037/0033-2909.124.3.372 +**rayner2009thirtyfifth** — Rayner, K. (2009). The 35th Sir Frederick Bartlett Lecture: Eye movements and attention in reading, scene perception, and visual search. *QJEP, 62*(8), 1457–1506. https://doi.org/10.1080/17470210902816461 +- *Verified:* Crossref; PubMed PMIDs 9849112 and 19449261. *Read status:* **abstract only** for both (APA and T&F paywalls). The abstracts list topics (perceptual span, preview benefit, eye-movement control) and give no figures. + +**Standard figures, with the best source I reached for each.** No primary Rayner text was open; the figures below come from open papers that restate them. Each is labelled secondhand where the primary was not opened. + +| Figure | Source I opened | Locator and status | +|---|---|---| +| Perceptual span: 14–15 characters right of fixation (skilled readers) | Rayner 1986 (abstract only); restated by Yu & Kwon 2023 (PMC10584020) and Frey et al. 2026 (PMC12972110) | Rayner 1986 abstract (primary, abstract). | +| Perceptual span: 3–4 characters left of fixation | Frey et al. 2026, *Scientific Reports* 16 (Introduction): "the PS size is estimated to extend about 14–15 characters to the right of fixation and 3–4 characters to the left of fixation, in English"; Yu & Kwon 2023 (Discussion section; locator approximate) gives "14 to 15 (skilled readers) or 11 (less-skilled readers) characters to the right of fixation and 3-4 charact[ers to the left]" (text cut in my extraction) | **Secondhand.** The primary studies (McConkie and Rayner; Rayner and colleagues) were not opened. The Frey quotation is accurate (card: library/frey2026correlational.md), but the article's own measured perceptual span was 8.87 letters, well below the classical figure; the authors attribute the gap to wider letters (0.54° against 0.46° in earlier work) and suggest the span is more stable in degrees than in letters (PDF p. 8). Yu & Kwon (card: library/yu2023altered.md) measured a span of about 11 to 14 letters in their task, which they call "slightly" smaller than the 14 to 19 letters they cite. | +| Beginning readers' span: about 11 characters right | Rayner 1986 abstract | Primary, abstract. | +| Typical forward saccade: 7–9 characters | Wertli et al. 2023 (Introduction): "The saccades are typically forward movements of 7 – 9 characters" | **Secondhand** (card library/wertli2023influence.md confirms the Introduction figures are taken from other work). | +| Fixation rate: about 4 per second | Pelli & Tillman 2008 (section on reading): "reading consists of a series of eye fixations, 4 per second" | Secondhand, open paper. | +| Fixation duration "200–250 ms" | No source opened states it. Measured values I did open: mean fixation 214 ms and 217 ms under intact viewing (Yu & Kwon 2023, Results, fixation-duration paragraphs); 246 ms for skilled readers of unspaced Chinese (Bai et al. 2008, Table 1) | The range is plausible; I did not find the sentence in an opened source. | +| Regressions "10–15%" of eye movements | **Not found.** Bai et al. 2008 Table 1 gives 3.5 regressive saccades against 17.7 fixations per sentence for unspaced Chinese (about 20%, my ratio) | Not confirmed. | +| Word-length cues drive saccade targeting | Mirault et al. 2019 (abstract): in unspaced text "the length of the currently fixated word determined the amplitude of forward saccades"; Morris et al. 1990 (abstract); Winskel et al. 2009 (Introduction): saccades aim at the "optimal viewing position" near the word centre, with landing positions "between the word beginning and word center" | Abstract and part. | +| Bouma law: critical spacing about half the eccentricity | Pelli & Tillman 2008, quoted above | Read in part. | +| Reading rate falls only if letter spacing, not size, passes a threshold | Pelli & Tillman 2008 | Read in part. | + +**Bearing on the unsourced notes.** Every number in Gemini essay 2 (3–4 and 14–15; 7–9; 200–250 ms; 10–15% regressions) now has one of three statuses: confirmed secondhand (the span, the saccade length), measured but not as a range (fixation duration), or not found (regressions). + +### 5. The word superiority effect + +**cattell1886time** — Cattell, J. M. (1886). The time it takes to see and name objects. *Mind, os-XI*(41), 63–65. https://doi.org/10.1093/mind/os-XI.41.63 +- *Verified:* Crossref; OUP page returns "available only as a PDF" behind a paywall. *Read status:* **not accessible.** Described by Jordan et al. 2024 and Houpt et al. 2014, who cite it. +- *Via Jordan et al. 2024 (Introduction):* "reports of this phenomenon date from some of the earliest experiments in the history of psychology (Cattell, 1886; Erdmann and Dodge, 1898; see also Huey, 1908)": words were easier to identify than nonwords and even single letters under brief display. *Via Houpt et al. 2014 (Introduction):* "letters were displayed sequentially to participants at faster and faster rates until they could no longer correctly identify the letters. They found that participants maintained accuracy with shorter durations when the letters were presented as part of a word compared with random letter sequences (Cattell, 1886)." + +**reicher1969perceptual** — Reicher, G. M. (1969). Perceptual recognition as a function of meaningfulness of stimulus material. *J. Experimental Psychology, 81*(2), 275–280. https://doi.org/10.1037/h0027768. *Verified:* Crossref; OpenAlex abstract. *Read status:* **abstract only.** "Performance on single words was clearly better than performance on single letters. The data suggest that the first stages of information processing are done in parallel, but scanning of the resultant highly processed information is done serially." + +**wheeler1970processes** — Wheeler, D. D. (1970). Processes in word recognition. *Cognitive Psychology, 1*(1), 59–85. https://doi.org/10.1016/0010-0285(70)90005-8. *Verified:* Crossref. *Read status:* **not accessible** (no abstract). Jordan et al. 2024 and the Wikipedia-level summary place it as the second source of the Reicher–Wheeler paradigm. + +**jordan2024reicher** — Jordan, T. R., Akkaya, A. M., Göçmüş, F. Z., Kalan, A., Morgul, E., Önalan, K., & Sheen, M. K. (2024). The Reicher-Wheeler paradigm in word recognition research: a cautionary note on its actual contributions and published misconceptions. *Frontiers in Psychology, 15*, 1399237. https://doi.org/10.3389/fpsyg.2024.1399237 + +- *Verified:* Crossref title and DOI match; PMC11532128 (open). *Read status:* **read in full (card: library/jordan2024reicher.md)**; first read in part. The card confirms the quotations and corrects the author initials (as now given). +- *Quotations:* "words are easier to identify than nonwords, and even single letters, and reports of this phenomenon date from some of the earliest experiments in the history of psychology (Cattell, 1886; Erdmann and Dodge, 1898; see also Huey, 1908)." The early findings were "confounded by the possibility that at least part of the advantage for words... reflected the ability of participants to use sophisticated, non-perceptual guesswork". "this became known as the Reicher-Wheeler Paradigm (hereafter, R-WP) after it was first reported by Reicher (1969) and Wheeler (1970)": a brief display followed by two alternative letters that both fit the rest of the string, so guessing cannot help. +- *Bearing:* **Corrects the author's note.** Cattell (1886) reported the observation; Reicher (1969) and Wheeler (1970) turned it into a controlled test. Neither the observation nor the paradigm is about lowercase versus capitals. + +**houpt2014new** — Houpt, J. W., Townsend, J. T., & Donkin, C. (2014). A new perspective on visual word processing efficiency. *Acta Psychologica, 145*, 118–127. https://doi.org/10.1016/j.actpsy.2013.10.013. *Verified:* Crossref; PMC3877166. *Read status:* **read in part (card: library/houpt2014new.md)** (one Introduction sentence about Cattell; Experiment 2 describes capital-letter stimuli). Cited only for its account of Cattell. The card's caution: the experiments are about single four-letter words in isolation, so do not cite the paper for word-level processing in running text. It supports one narrow point for D6: word-level units are processed more efficiently than letter-by-letter, which the authors tie to perceptual learning, not to any property of spacing. + +**arditi2007letter** — Arditi, A., & Cho, J. (2007). Letter case and text legibility in normal and low vision. *Vision Research, 47*(19), 2499–2505. https://doi.org/10.1016/j.visres.2007.06.010 (DOI from Europe PMC; Crossref confirms vol. 47, no. 19, pp. 2499–2505). *Verified:* PMC2016788. *Read status:* **abstract only.** +- *Quotation (abstract):* "It is thought by cognitive scientists and typographers alike, that lower-case text is more legible than upper-case. Yet lower-case letters are, on average, smaller in height and width than upper-case characters, which suggests an upper-case advantage." Results: "Reading speeds were higher for upper- than for mixed-case text at sizes twice acuity size; at larger sizes, the upper-case advantage disappeared." +- *Bearing:* The lowercase-over-capitals claim is a belief with a counter-result under particular conditions; it is not what the word superiority effect shows. + +### 6. Literacy and the visual word form area (under the ruling) + +**dehaene2007cultural** — Dehaene, S., & Cohen, L. (2007). Cultural recycling of cortical maps. *Neuron, 56*(2), 384–398. https://doi.org/10.1016/j.neuron.2007.10.004. *Verified:* Crossref; PubMed PMID 17964253. *Read status:* **abstract only.** "we propose a neuronal recycling hypothesis, according to which cultural inventions invade evolutionarily older brain circuits and inherit many of their structural constraints." + +**dehaene2010learning** — Dehaene, S., Pegado, F., Braga, L. W., et al. (2010). How learning to read changes the cortical networks for vision and language. *Science, 330*(6009), 1359–1364. https://doi.org/10.1126/science.1194140. *Verified:* Crossref; PubMed PMID 21071632. *Read status:* **abstract only.** 10 illiterate, 22 adult-literate, 31 childhood-literate adults: literacy "induced a small competition with faces at this location, but also broadly enhanced visual responses in fusiform and occipital cortex, extending to area V1." "Most changes occurred even when literacy was acquired in adulthood." + +**dehaene2015illiterate** — Dehaene, S., Cohen, L., Morais, J., & Kolinsky, R. (2015). Illiterate to literate: behavioural and cerebral changes induced by reading acquisition. *Nature Reviews Neuroscience, 16*(4), 234–244. https://doi.org/10.1038/nrn3924. *Verified:* Crossref; PubMed PMID 25783611. *Read status:* **abstract only.** "responses to written characters are increased in the left occipito-temporal sulcus, whereas responses to faces shift towards the right hemisphere." + +**price2003myth** — Price, C. J., & Devlin, J. T. (2003). The myth of the visual word form area. *NeuroImage, 19*(3), 473–481. https://doi.org/10.1016/S1053-8119(03)00084-3. *Verified:* Crossref; PubMed PMID 12880781. *Read status:* **abstract only.** "there is no evidence that visual word form representations are subtended by a single patch of neuronal cortex and it is misleading to label the left midfusiform region as the visual word form area." + +**price2011interactive** — Price, C. J., & Devlin, J. T. (2011). The interactive account of ventral occipitotemporal contributions to reading. *Trends in Cognitive Sciences, 15*(6), 246–253. https://doi.org/10.1016/j.tics.2011.04.001. *Verified:* Crossref; PMC3223525. *Read status:* **read in full (card: library/price2011interactive.md)**; first read in part. The card adds that the account is a framework argued from existing studies, not a new experiment; it sets itself against Dehaene's recycling and bigram-detector proposals; and it makes the point that task and attention reverse selectivity, which weakens any claim that vOT activation shows "word-specific" tuning. "specialization for orthography emerges from regional interactions without assuming that vOT is selectively tuned to orthographic features." + +**hervais2019learning** — Hervais-Adelman, A., Kumar, U., Mishra, R. K., Tripathi, V. N., Guleria, A., Singh, J. P., Eisner, F., & Huettig, F. (2019). Learning to read recycles visual cortical networks without destruction. *Science Advances, 5*(9), eaax0262. https://doi.org/10.1126/sciadv.aax0262. *Verified:* Crossref; PMC6750915. *Read status:* **read in full (card: library/hervais2019learning.md)**; first read in part. The card confirms both quotations. Addition from the card: no decrease in response was seen for faces, houses, tools, false fonts or checkerboards (p. 4 of 10), but a similar post-training change was not seen for nonwords, so the sentence effect might reflect words learned in training. +- *Quotation (abstract):* "We found that literacy enhances responses to other visual input in early visual areas and enhances representational similarity between text and faces, without reducing the extent of response to nonorthographic input. Thus, acquisition of literacy in childhood recycles existing object representation mechanisms but without destructive competition." Results: "There is no obvious decrease in the magnitude of the peak response to any category of stimulus as a function of literacy". + +**kubota2024rethinking** — Kubota, E., Grill-Spector, K., & Nordt, M. (2024). Rethinking cortical recycling in ventral temporal cortex. *Trends in Cognitive Sciences, 28*(1), 8–17. https://doi.org/10.1016/j.tics.2023.09.006. *Verified:* Crossref; PMC10841108. *Read status:* **read in full (card: library/kubota2024rethinking.md)**; first read in part. The card confirms both quotations and adds three cautions: the before-and-after study had a sample of 10; body-selective cortex stayed stable while limb-selective cortex was recycled; and the authors admit the evidence for lateralization of faces in preschoolers is inconsistent. +- *Quotation (abstract):* "Prevailing theories suggest that face-selective regions become right lateralized due to competition with word-selective regions in the left hemisphere. However, recent longitudinal studies examining face and word-selective responses in childhood do not provide support for this theory. Instead, there is evidence that word representations recycle cortex previously involved in processing other stimuli such as limbs." + +**cohen2008reading** — Cohen, L., Dehaene, S., Vinckier, F., Jobert, A., & Montavont, A. (2008). Reading normal and degraded words: Contribution of the dorsal and ventral visual pathways. *NeuroImage, 40*(1), 353–366. https://doi.org/10.1016/j.neuroimage.2007.11.036. *Verified:* Crossref; PubMed PMID 18182174. *Read status:* **abstract only.** Letter spacing was one of three degradations: "words displayed in formats outside this field of expertise should be read serially, under supervision of dorsal parietal attention systems"; a word-length effect on latency appeared above degradation thresholds. This is the closest fMRI result to the Gemini claim about a shift to serial dorsal processing, but the manipulation was widened letter spacing in single words, not removal of word spaces from sentences. + +**zhou2023fixation** — Zhou, W., Wang, S., & Yan, M. (2023). Fixation-related fMRI analysis reveals the neural basis of natural reading of unspaced and spaced Chinese sentences. *Cerebral Cortex, 33*(20), 10401–10410. https://doi.org/10.1093/cercor/bhad290. *Verified:* Crossref. *Read status:* **abstract only** (OUP link was a login page). 33 adults; "activity pattern in the left middle temporal gyrus could significantly predict... unspaced vs. spaced conditions"; "bidirectional brain connections between the left middle temporal gyrus and the left inferior occipital cortex in the unspaced Chinese sentence reading but not in the spaced reading." This is the only fMRI study of spaced against unspaced text I found. It is Chinese, abstract only, and does not report a VWFA effect in the abstract. + +**kutas2011thirty** — Kutas, M., & Federmeier, K. D. (2011). Thirty years and counting: Finding meaning in the N400 component of the event-related brain potential (ERP). *Annual Review of Psychology, 62*, 621–647. https://doi.org/10.1146/annurev.psych.093008.131123. *Verified:* Crossref; PMC4052444. *Read status:* **read in full (card: library/kutas2011thirty.md)**; first read in part (abstract, introduction). +- *Quotations:* "an event-related brain potential response linked to meaning processing." "The N400 was labeled as such because it was a relative negativity peaking around 400 ms. More precisely, it is negative-going at particular scalp locations relative to a specific reference derivation". +- *Bearing:* **Confirms the flag**: the N400 is a scalp-recorded electrical potential (EEG/ERP), not an fMRI measure. From the card: the quotation "The N400 was labeled as such..." sits in "Names and definitions", not the introduction; fMRI appears once in the body, and the review does not mention word spacing, boundaries, segmentation or degraded text. + +**VWFA and unspaced text.** I found no study of the VWFA with unspaced alphabetic text. The nearest items are Zhou et al. 2023 (Chinese, abstract only), Degno et al. 2019 (EEG; no reliable spacing effect in the N400 window), and Cohen et al. 2008 (letter spacing as degradation). Dehaene's work covers literacy, not spacing; Price and Devlin dispute the single-region framing; Kubota et al. 2024 and Hervais-Adelman et al. 2019 dispute the "destructive competition" version. Under the ruling, the paper should treat all of this as background it does not rely on. + +### 7. Experiments that simulate scriptio continua in Latin or Greek + +**vatri2012physiology** — Vatri, A. (2012). The physiology of ancient Greek reading. *Classical Quarterly, 62*(2), 633–647. https://doi.org/10.1017/S0009838812000213. *Verified:* Crossref. *Read status:* **not accessible** (Cambridge; no abstract in OpenAlex). A search snippet described its introduction as a response to Saenger's claim that unspaced script forced vocal rehearsal. I did not open it, so I cannot say what experimental literature it uses. A human with library access should check whether it cites Epelboim et al. or Rayner et al. + +I found **no experiment** that has trained readers of Latin or Greek read unspaced classical text under eye-tracking or timed conditions (rows 12, 17, 23). The nearest experimental work is: Epelboim et al. 1994 and 1997 (skilled English readers, unspaced text read aloud and silently; the 1994 abstract reports the same oculomotor strategy in both); Kohsom & Gobet 1997 (Thai readers read aloud); Wang et al. 2025 (Javanese read aloud, where adults gained from spaces). None compares aloud against silent reading with a spacing manipulation as its main variable. Facet B is the place to weigh Vatri's argument. + +### 8. A named construct for "a display change that removes a segmentation task from the operator" + +I found no single established term that means exactly this. The nearest named items: + +**wickens1995proximity** — Wickens, C. D., & Carswell, C. M. (1995). The proximity compatibility principle: Its psychological foundation and relevance to display design. *Human Factors, 37*(3), 473–494. https://doi.org/10.1518/001872095779049408. *Verified:* Crossref; OpenAlex abstract. *Read status:* **abstract only.** "Displays relevant to a common task or mental operation (close task or mental proximity) should be rendered close together in perceptual space (close display proximity)." The principle matches display grouping to task grouping. It speaks to how a display should cluster related items; it does not say that a display change removes a segmentation job. + +**wagemans2012century** — Wagemans, J., Elder, J. H., Kubovy, M., Palmer, S. E., Peterson, M. A., Singh, M., & von der Heydt, R. (2012). A century of Gestalt psychology in visual perception: I. Perceptual grouping and figure–ground organization. *Psychological Bulletin, 138*(6), 1172–1217. https://doi.org/10.1037/a0029333. *Verified:* Europe PMC (PMC3482144; DOI from the record); *Read status:* **read in part (card: library/wagemans2012century.md)** (the paragraphs introducing grouping and proximity). The card finds the quotation matches §3.1 and that the review does not mention word spacing. +- *Quotation:* "Max Wertheimer first posed the problem of perceptual grouping in his ground-breaking 1923 paper by asking what stimulus factors influence the perceived grouping of discrete elements... when he altered the spacing between adjacent dots so that some dots were closer than others, the closer ones grouped together strongly into pairs. This factor of relative distance, which Wertheimer called proximity, was the first of his famous laws or (more accurately) principles of grouping." +- *Bearing:* **Names the perceptual mechanism** (the facet's inference, not the review's: the review does not mention word spaces): word spaces would work by proximity grouping, which the review presents as the first of Wertheimer's principles. That is the standard term for the mechanism. "Word segmentation" is the term in reading science itself (Li et al. 2020, *Psychological Review*, https://doi.org/10.1037/rev0000248, abstract only via Consensus: "the word segmentation problem in Chinese reading"). + +I did not open sources for chunking (Miller 1956; Cowan 2001, both in the repo's DXD bib) or pre-attentive segmentation (Treisman & Gelade 1980, in the Part 4 bib). The Gemini essay's "pre-attentive" and "Guided Search" language belongs to facet E. + +--- + +## (c) Synthesis + +### c-1. What the space does, how big it is, and where it acts + +1. **Removing spaces from a spaced script costs roughly a third to a half of reading speed.** Rayner et al. 1998 report "approx. 50%" (abstract). Winskel et al. 2009 measured 33% (36 English monolinguals) and 45% (36 Thai–English bilinguals reading English). The 30–50% range quoted in the Gemini essay is Winskel et al.'s own summary of earlier work (Introduction). Mirault et al. (2019, p. 22) give "about 40% to 70%" slower than normally spaced text, and attribute that range to Rayner & Pollatsek (1996) and Rayner et al. (1998), not to their own data. Epelboim et al. 1994 found only "modest" costs (the abstract says "subjects", not "skilled readers"); Rayner and Pollatsek 1996 replied that unspaced text "is not easy for most readers". The figures differ with the filler used, the readers and the measure; a paper should quote a range with its source, not one number. +2. **The cost has two parts.** Word identification: the frequency effect is larger without spaces (Rayner et al. 1998) and arrives later (112 ms against 152 ms; Sheridan et al. 2013, abstract). Eye guidance: initial landing positions move toward the word's start and saccades shorten (Rayner et al. 1998; Winskel et al. 2009 for English). Morris et al. 1990 separate the roles: letters govern fixation duration; spaces and letters both govern saccade length and landing site. Epelboim et al. 1997 and Winskel et al. 2009 (Thai) argue that word recognition carries most of the load. +3. **Other cues can replace the space.** Alternating bold cut the cost "much" (Perea & Acha 2009, abstract). With initial capitals the word-frequency effect is the same size with and without spaces, and a global reading-time cost of about 7% remains; the design has no lowercase baseline, so it does not show what capitals restore relative to ordinary text (Özkan & Perea 2026, Discussion, PDF p. 11). Kanji does the work of spaces in mixed Japanese (Sainio et al. 2007, abstract). Korean particles do too (Baek et al. 2022, abstract). The space is one way of doing a job, segmentation, and the job can be done other ways. +4. **A mechanism for why a space helps at all.** Letters at word ends are much less crowded: "letters at the ends of words are much less crowded and have a larger uncrowded window" (Pelli & Tillman 2008; an aside in their paper, and the paper has an erratum). I infer, without a test of word spacing, that a space beside a word end is what leaves that side free. Pelli and Tillman's account of reading speed as span times fixation rate (4 per second) puts the space inside a general limit on the span of recognisable letters. + +### c-2. For whom: fluent readers of unspaced scripts versus novices + +| Reader | Evidence | Direction | +|---|---|---| +| Fluent adult, unspaced script (Chinese) | Bai et al. 2008: 5,221 ms unspaced vs 5,239 ms word-spaced, 16 readers; "as easy to read" | No gain, no cost. Fixations and saccade length changed. | +| Fluent adult, Thai | Chantavarin & Leung 2025; Winskel et al. 2009: Thai readers took 5% longer on spaced sentences (p = .05 by participants, minF' p = .13; spaced sentences were 12.8% longer); by participant, 42% were slowed, 16% sped up, 42% unaffected. Kohsom & Gobet 1997: about 1.4 s per paragraph faster with spaces, on a coherent-unspaced baseline of 46 s, p < .08, n = 8 | Small and mixed: a word-level gain with a sentence-level cost in Thai (Winskel; Chantavarin & Leung 2025, read in full: sentence time 132 ms longer, word measures shorter); a small, marginal global gain in Kohsom & Gobet. The "faster with spaces" abstract of 1997 overstates a marginal effect. | +| Fluent adult, Japanese | Sainio et al. 2007: helps in Hiragana; redundant with Kanji | Depends on whether the script has its own cues. | +| Fluent adult, Tibetan | Wang et al. 2024 (read in full, card: 5,566 ms word-spaced against 6,019 ms normal) and two 2026 papers (abstracts) | Spaces help on global and word measures, with the gain falling as format familiarity rises and rising for low-proficiency readers. | +| Second-language learner | Shen et al. 2012: least disruption for word-spaced text, not modulated by native language; Bai et al. 2013 (new words read faster, benefit kept at test); Cui 2023: word-spaced against unspaced not significant (the significant contrasts are against non-word and pinyin text) | Helps (Shen; Bai 2013); Cui 2023 is directional and mostly non-significant against unspaced text. | +| Children, Chinese | Li et al. 2021: helps Grade 1 most, "much reduced" later; Zang et al. 2013: helps children and adults "to a similar degree"; Blythe et al. 2012: helps new-word learning; kept in unspaced test; Lin & Lin 2016: slows children without reading difficulty, helps those with difficulty after training; Bassetti & Masterson 2012: children slower and less accurate with word spacing (negative, not null) | **Mixed; the benefit for Grade 1 (Li et al.) is not replicated for Grades 2-3 (Liao et al. 2024; Bassetti & Masterson 2012).** | +| Developing and adult readers, Javanese, reading aloud | Wang et al. 2025 (abstract) | Developing readers gained from spaces and colour; adults gained from spaces only. | +| Skilled English readers with weaker spelling | Veldre et al. 2017 (abstract) | The cost of removing spaces is bigger for weaker spellers. | +| Practice | Chen et al. 2021 (abstract): ten days of reading training left no significant difference between spaced and unspaced right-to-left text (the authors read this as a smaller facilitation effect; a corrigendum is reported but not confirmed in the open copy). Kohsom & Gobet 1997 and Winskel et al. 2009: a lifetime's skill with unspaced Thai gave no help with unspaced English | Skill with a script lowers the cost of that script's format; it does not transfer. | + +**What this means for the paper's wording.** The data support this statement: *for readers who lack the habit of a script, word spacing speeds identification; for habituated readers it adds little at the level of the whole sentence, and can slow them while the format is unfamiliar, though it can still speed word identification (Thai, Tibetan).* That fits the paper's claim that spacing "changed the interface so that reading no longer required the training" (the author's draft) only in part. The evidence says the aid matters most before the training has happened. It does not say the training was ever unobtainable: fluent Thai, Chinese and Japanese readers show little or no cost from the absence of spaces in their own script. + +### c-3. Beginning and struggling readers, crowding and letter spacing + +The dyslexia literature is a different question (letter spacing, not word spacing) and its results are unsettled. Zorzi et al. 2012 report that extra-large letter spacing substantially improved text reading in dyslexic children (abstract); "74 children", "factor of two" and "about 11%" [unverified: see AUDIT_02], with the widening including word spaces. Of the replications I reached, two find accuracy gains with no speed gain (Hakvoort et al. 2017; Dotan & Katzir 2018), two find faster reading or word identification (Perea et al. 2012; Stagg & Kiss 2021), and one reports a general gain without separating the two (Duranovic et al. 2018), one finds nothing (Łuniewska et al. 2022: no effect on accuracy, comprehension or speed) and one finds none for beginning readers (van den Boer & Hakvoort 2015). One study (Hakvoort et al. 2017) shows the effect is not specific to dyslexia, and others show it depends on word spacing and default spacing (Slattery et al. 2016; Galliussi et al. 2020). The safest reading is that wider spacing helps weaker readers' accuracy modestly, speed less reliably, and that word spacing is part of the effect. Pelli & Tillman 2008 find that crowding does not account for the whole dyslexic deficit. + +### c-4. H-cost and H-cause + +**H-cost** (spacing moved parsing work from the reader to the page, lowering the cost of reading and widening the group who could read without long training). *The reading science supports the cost half and gives partial, mixed support to the "most for novices" half.* +- For: a cost of one-third to one-half of speed when spaces are taken from a spaced script; faster word identification (about 40 ms earlier frequency effect); benefits for second-language learners and, in several studies, for the youngest Chinese readers and for low-proficiency Tibetan readers; a gain that shrinks as the format becomes familiar. +- Against or complicating: fluent readers of unspaced Chinese show no net gain from spaces, and fluent Thai readers lose sentence-level speed when spaces appear, though they gain at the word level (Winskel; Chantavarin & Leung 2025), and Tibetan readers gain on global measures (Wang et al. 2024); two studies of Chinese children report no benefit or a cost (Bassetti & Masterson 2012; Lin & Lin 2016 for children without difficulty); Epelboim's group reports only modest costs for skilled English readers; skilled readers adapt (Mirault et al. 2019; Veldre et al. 2017). +- Gaps in the claim: no study I found compares the *time to acquire* reading in spaced against unspaced instruction. Blythe et al. 2012 (new-word learning) and Chen et al. 2021 (format training) are the closest. The effect sizes are seconds per paragraph or tens of percent of reading speed per reader, not differences in who can read. + +**H-cause** (word separation caused the spread of silent reading and literacy). *The reading science is silent.* Every study measures speed, fixations or accuracy in a lab in readers who already read. None measures who becomes literate, or whether reading is silent. Fluent unspaced reading in Thai, Chinese, Japanese and Tibetan That literate populations read Thai, Chinese and Japanese without spaces is known independently; these experiments show the cost to such readers of lacking spaces is small or absent. This facet therefore neither supports nor refutes H-cause. It does refute a strong version: that spacing was necessary for fast or private reading. Facet C holds the historical test. + +**Problem 2 (spacing as a legibility remedy).** The reading science gives the author a defensible distinction: a space acts on the page, before the reader's effort (proximity grouping, a facet inference; Wagemans et al. 2012 do not discuss word spaces), whereas the aids that need the reader to do more are not in this literature. Nothing here tests the observability side, which belongs to facets E and G. + +### c-5. Claims in the Gemini essays + +Files: `ux_evolution_scriptio_continua.md` (essay 1) and `ecological_ergonomics_algorithmacy_uxd.md` (essay 2), under `dial_response_algorithmacy/sources/`. + +| # | Claim (essay) | Verdict | Source and note | +|---|---|---|---| +| 1 | Cattell discovered the word superiority effect in 1886, showing "humans process whole words faster than individual letters" (essay 1) | **Correct in part.** Cattell (1886) made the early observation, but the controlled effect is Reicher (1969) and Wheeler (1970); the early work was confounded by guessing. "Faster" is not the measure in the sources I read: the effect is accuracy for letters under brief display. | Jordan et al. 2024 (Introduction; R-WP section); Houpt et al. 2014 (Introduction). Cattell's paper itself not accessible. | +| 2 | The word superiority effect shows "lowercase text is read faster than all-caps text" (essay 1) | **Not found; contradicted in part.** The word superiority literature I read is about words against nonwords and letters, not case. A study of letter case found upper-case read faster than mixed case at small sizes, and the advantage vanished at larger sizes; the abstract calls lowercase's advantage something "thought" by scientists and typographers. | Jordan et al. 2024; Arditi & Cho 2007 (abstract only). I found no source for the lowercase claim. | +| 3 | The "Bouma shape" model, named for Herman Bouma, says readers recognise words by outline (essay 1) | **Not confirmed.** The Bouma source I could reach describes Bouma's 1970 observation as a crowding result: critical spacing about half the eccentricity. I did not find the word-outline model attributed to Bouma in an opened source. | Pelli & Tillman 2008 ("The Bouma law"). Bouma 1970 itself not accessible. | +| 4 | Perceptual span "3 to 4 character spaces to the left... 14 to 15... to the right" (essay 2) | **Confirmed (secondhand).** | Frey et al. 2026; Yu & Kwon 2023; Rayner 1986 abstract (14–15 right). | +| 5 | "ballistic saccades averaging 7 to 9 letter spaces" (essay 2) | **Confirmed (secondhand).** | Wertli et al. 2023. | +| 6 | "fixations lasting 200 to 250 milliseconds" (essay 2) | **Plausible, no source found for the range.** Measured means: 214 and 217 ms (adults, intact viewing); 246 ms (Chinese). | Yu & Kwon 2023; Bai et al. 2008 Table 1. | +| 7 | "regressive saccades—normally comprising 10% to 15% of eye movements" (essay 2) | **Not found.** | Bai et al. 2008 Table 1 gives roughly 20% for unspaced Chinese (my ratio of 3.5 to 17.7). | +| 8 | Removing spaces makes "reading velocity plummet by 30% to 50%" (essay 2) | **Confirmed as a range, with caveats.** Winskel et al. (2009) put it at "30–50%"; Rayner et al. 1998: "approx. 50%"; Winskel's own data 33% and 45%. Mirault et al. (2019, p. 22) quote 40–70%, which they attribute to Rayner & Pollatsek (1996) and Rayner et al. (1998), not to their own data. | See c-1. The range depends on the readers and the filler. | +| 9 | In unspaced text saccades are "erratic" and fixations "swell" (essay 2) | **Partly confirmed.** Landing positions move toward word starts and saccades shorten; the 1998 abstract does not use "erratic". Epelboim et al. 1994 found "modest increases in fixation durations". | Rayner et al. 1998 (abstract); Epelboim et al. 1994 (abstract). | +| 10 | VWFA recognises letters and words "within 150 to 200 milliseconds of fixation" (essay 2) | **Not found.** | — | +| 11 | "Without visual spacing, the ventral stream fails to isolate orthographic entities, kicking processing... into high-load, serial parieto-frontal attentional networks and mandatory auditory loops" (essay 2) | **Not found for word spacing.** Cohen et al. 2008 (abstract) show a shift toward serial dorsal parietal processing for degraded words, with widened letter spacing among the degradations. No fMRI study of unspaced alphabetic text found. The "mandatory auditory loops" are not in any source I read. | Cohen et al. 2008; Zhou et al. 2023 (Chinese). | +| 12 | "fMRI studies show that degraded text... triggering increased N400 electrophysiological responses" (essay 1) | **Contradicted.** N400 is an ERP component recorded with EEG. | Kutas & Federmeier 2011 (abstract and opening). | +| 13 | Reading acquisition makes visual words "invade" cortex "weakly specialized for tools" and "subtly displaces face recognition", pushing it right (essay 1) | **Contested.** Dehaene et al. 2010 report "a small competition with faces"; Dehaene et al. 2015 report face responses "shift towards the right hemisphere". Hervais-Adelman et al. 2019 find "without destructive competition"; Kubota et al. 2024 say longitudinal studies "do not provide support" and point to recycling of cortex that processed "limbs". | Abstracts for Dehaene; Hervais-Adelman and Kubota read in full (cards). | +| 14 | Ancient readers "had to subvocalize... to identify word breaks" and silent reading was rare "because the interface imposed a punishing working-memory bottleneck" (essay 2; essay 1 similar) | **Not tested by the experiments I found.** Unspaced English was read both aloud and silently with "the same oculomotor strategy" in one study (Epelboim et al. 1994, abstract); the evidence on Greek and Latin readers is historical (facet B). | Epelboim et al. 1994 (abstract only). | +| 15 | "Dehaene's... parallel recognition engine requires discrete visual boundaries" (essay 2) | **Not found.** Skilled readers of Thai, Chinese and Japanese read unspaced text fluently. | Bai et al. 2008; Winskel et al. 2009; Sainio et al. 2007. | + +Claims the sources contradict, in one list: (i) N400 as an fMRI result; (ii) lowercase advantage as part of the word superiority effect; (iii) "faster" as the word superiority measure; (iv) the "30–50%" as one study's finding; (v) the face-displacement story as settled; (vi) the abstract of Kohsom & Gobet (1997) read as a robust Thai speed-up. + +--- + +## (d) Gaps and unverified leads + +**Sources I could not open, and where a human with library access would find them.** +- Pollatsek & Rayner (1982), *JEP:HPP* 8(6), 817–833, DOI 10.1037/0096-1523.8.6.817 (APA PsycNet). No abstract found. +- Morris, Rayner & Pollatsek (1990), full text (APA PsycNet). +- Rayner, Fischer & Pollatsek (1998), *Vision Research* 38, 1129–1144 (ScienceDirect; open archive but bot-walled). This is the key paper for the 50% figure and the two-part mechanism; read it for the word-length and landing-position effects, the number of subjects, and the reading-rate table. +- Epelboim et al. (1994, 1997) and Rayner & Pollatsek (1996), *Vision Research* (ScienceDirect). The size of the 1994 "modest" cost is not in the abstract. +- Perea & Acha (2009), *Vision Research* 49, 1994–2000 (ScienceDirect). +- Mirault, Snell & Grainger (2019), *Acta Psychologica* 195, 22–29: HAL `hal-02138873` (Anubis wall today). The page-22 quotations and the 40–70% figure are now checked against the full text (card): the range is Rayner & Pollatsek's (1996) and Rayner et al.'s (1998). +- Sheridan, Rayner & Reingold (2013), *Visual Cognition* (T&F); volume and page range unchecked. +- Shen et al. (2012), Zang et al. (2013), Blythe et al. (2012), Bai et al. (2013), Li et al. (2021), Lin & Lin (2016/17): effect sizes, ages and Ns beyond the abstracts. Blythe's adult and child group sizes are the first thing to read. +- Sainio et al. (2007), the reading-time numbers. +- Rayner (1998, *Psych Bull*) and Rayner (2009, *QJEP*): the primary statements of perceptual span, saccade length, fixation duration and regression rates. Until read, the 3–4 left figure, the 7–9 saccade length, the 200–250 ms range and the regression rate stay secondhand or unfound. Clifton et al. (2016), "Eye movements in reading and information processing: Keith Rayner's 40 year legacy", *J. Memory and Language*, may be a shorter route. +- Cattell (1886), *Mind* os-XI(41), 63–65 (OUP, paywalled). Check what Cattell measured and whether he said anything about case. +- Wheeler (1970), *Cognitive Psychology* 1, 59–85. +- Bouma (1970), *Nature* 226, 177–178; and any source for the word-outline ("Bouma shape") model. +- Vatri (2012), *Classical Quarterly* 62, 633–647. Facet B territory. +- Zorzi et al. (2012) full text (own reading, not extractor); Skottun & Skoyles and the reply in PNAS 109(38). +- Price & Devlin (2003); Dehaene et al. (2010, 2015); Dehaene & Cohen (2007): abstracts only. Cohen et al. (2008), Zhou et al. (2023): abstracts only. +- Mirault et al. (2019, *Attention, Perception, & Psychophysics* 81, 846–860), PMC6407912 (open; I did not read it). + +**Gaps in the evidence.** +- No Lao, Khmer or Burmese study found. +- No experiment with Latin or Greek text and trained readers. +- No study of time to acquire reading under spaced against unspaced instruction. +- No study of practice effects on unspaced alphabetic reading beyond Chen et al. 2021 (Chinese, right-to-left) and the Thai-to-English transfer null. +- The VWFA literature contains nothing on word spacing; the single fMRI study of spaced against unspaced text is in Chinese and I read only its abstract. +- Primary sources for the standard figures were not opened (see the table in section 4). +- Where I give "et al." or three authors in an entry, I did not compare the full author list with Crossref. + +**Bibliographic checks.** Volume, issue and pages for the journal articles were taken from the Crossref record on 2026-10-12. Exceptions: Bai et al. (2013), Cui (2023), Hervais-Adelman et al. (2019), Frey et al. (2026), Özkan & Perea (2026), Chantavarin et al. (2025) carry no page range in Crossref. Chantavarin et al. were listed by Crossref with two authors; I show "et al." because the full list was not checked. + +## Changes after full-text cards (2026-10-13) + +- `mirault2019reading`: status "abstract only" to "read in full (card)"; the "40–70%" range is Rayner & Pollatsek (1996) and Rayner et al. (1998), quoted on p. 22, not Mirault's data (entry, c-1 item 1, c-5 row 8, section d). +- `mirault2019precise`: "not read" to "read in full (card)"; transposition summary confirmed; interaction absent in first-fixation and gaze duration, weaker with frequent boundary letters. +- `ozkan2026wordinitial`: "abstract only" to "read in full (card)"; "capitals restored lexical processing" withdrawn (no lowercase baseline; about 7% global cost remains; Discussion PDF p. 11) in the entry and in c-1 item 3. +- `hakvoort2017improvements`: abstract's "interword level" annotated; body says "sentence level" (p. 112); the c-3-adjacent "interword level" list item rewritten (Experiment 2 result, p. 113). +- `hervais2019learning`: author list now eight (Eisner added); status updated; nonword caveat added. +- `jordan2024reicher`: author initials corrected; status updated. +- `cui2023eye`: learner contrast against unspaced text not significant (ps > .11); abstract's "generally spent less time" qualified; c-2 learner row revised. +- `bassetti2012effects`: child effect negative (slower and less accurate), adult slowdown depends on block order; Bai found no gain over unspaced text (agrees with Bassetti); the width explanation is Bassetti and Masterson's; c-2 children row revised. +- `bai2008reading`: status updated; N = 16 and 24, character/nonword-spacing basis for "words are the unit", silent on learners. +- `chantavarin2025effects`: "et al." replaced with the two authors; status updated; sentence-level slowing and word-level gain added from the body. +- `kohsom1997adding`: status updated; paper's claim is a small gain for Thai natives and a large one for English text; nothing matches "gain nothing". +- `winskel2009eye`: status updated; Thai word-level gain with 5% sentence-level cost and between-reader variation stressed. +- `wang2024tibetan`: status updated; global gains for fluent Tibetan readers (5,566 against 6,019 ms); c-2 Tibetan row revised. +- `galliussi2020inter`: effect practically negligible (under 0.20 syll/s); letterform × inter-word spacing interaction for dyslexic readers (p. 147). +- `kutas2011thirty`, `wagemans2012century`: statuses updated; neither mentions word spacing; the "proximity grouping, the oldest grouping principle" sentence is the facet's inference (entry and c-4 Problem 2). +- `chen2021tradeoff`, `degno2019coregistration`, `frey2026correlational`, `houpt2014new`, `kubota2024rethinking`, `perea2012subtle`, `price2011interactive`, `wertli2023influence`, `yu2023altered`: statuses updated and the card's cautions added (Frey's own span 8.87 letters; Yu and Kwon 11 to 14 letters; Houpt not for running text; Chen corrigendum unconfirmed). +- c-2 and c-4: Thai row, H-cost "against" list and the closing wording sentence revised for word-level gains by fluent readers of unspaced scripts. + +Synthesis verdicts changed: +- c-1 item 3. Old: "Initial capitals restored lexical processing but not all of the oculomotor cost (Özkan & Perea 2026, abstract)." New: with capitals the frequency effect is the same size with and without spaces and a global cost of about 7% remains; the design has no lowercase baseline. +- c-2 table, second-language learner row. Old: "Helps." New: "Helps (Shen; Bai 2013); Cui 2023 is directional and mostly non-significant against unspaced text." +- c-2 table, Thai row. Old: "Small and mixed." New: word-level gain with sentence-level cost (Winskel; Chantavarin & Leung), small marginal global gain (Kohsom & Gobet). +- c-2 table, Tibetan row. Old: gain falling with familiarity (abstracts). New: gains on global and word measures (Wang 2024, read in full). +- c-2 wording sentence. Old: "for habituated readers it adds little and can slow them while the format is unfamiliar." New: adds little at the level of the whole sentence, can slow them, but can still speed word identification (Thai, Tibetan). +- c-4 H-cost "against" list. Old: fluent readers of unspaced scripts "show no cost to being without spaces and sometimes lose speed when spaces appear". New: no net gain in Chinese, sentence-level loss with word-level gain in Thai, global gain in Tibetan. +- c-4 Problem 2. Old: "proximity grouping; Wagemans et al. 2012". New: marked as the facet's inference. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_E_human_factors_frame.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_E_human_factors_frame.md new file mode 100644 index 00000000..8c5121ec --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_E_human_factors_frame.md @@ -0,0 +1,360 @@ +# E — The human-factors frame: spacing as reallocation of work from reader to artefact + +*Corrected on 2026-10-12 against `../audit/AUDIT_02.md`: 14 factual fixes.* + +*Facet E of the foundation paper for Decision Experience Design. 2026-10-12. Agent-drafted. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** The human-factors and cognitive-science vocabulary for a move in which a task leaves the human and enters the artefact: the representational effect (Larkin and Simon; Zhang and Norman), cognitive artefacts and external cognition (Norman, Hutchins, Scaife and Rogers, Kirsh), cognitive offloading (Risko and Gilbert), function allocation and its critics (Fitts, Jordan, Dekker and Woods, Parasuraman, Sheridan and Wickens), ecological interface design and display principles (Vicente and Rasmussen, Wickens and Carswell, Woods), design versus training (hierarchy of controls, expertise reversal, desirable difficulties), the contrast case of monitoring (Bainbridge, only as far as the line needs), HCI and document-design uses of writing history, and representations that coordinate people through a system (Erickson and Kellogg, Heath and Luff, Gutwin and Greenberg). + +**The question the synthesis answers.** (i) What is the best-supported name for the scribes' move, with the definition quoted and located? (ii) Is "the artefact absorbs the task" versus "the human is asked to read or monitor more" a principled line, and which source is strongest for and against? (iii) Whom does such a move help, novice or expert? (iv) Where does the analogy from page to coordination break? (v) Which claims in the author's framing do the sources contradict? This is problem 2 in `RESEARCH_PLAN.md`; the answer to (iv) also feeds problem 4. + +**Legend for read status.** *Read in full*: I read the whole text this session, or the entry reuses a repo card built from full text. *Read in part*: I obtained the text and read the passages named (pages given). *Abstract only*: a database abstract or publisher record. *Not accessible*: no text and no abstract. Where a page locator is "ms." it refers to an author manuscript, not the version of record. Quotations from OCR text are normalised for spacing only. Every entry's DOI was resolved against Crossref on 2026-10-12 unless the entry says otherwise. + +--- + +## (a) Search log + +All searches ran on 2026-10-12. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | repo read | `RESEARCH_PLAN.md`, `literature/REUSE.md`, `SOURCE.md`; lab F3, F4; library `INDEX.md` 127-184; cards bainbridge1983, parasuraman2000, norman1990, hutchins1995; `phd1750_topic_and_rationale.md` | — | The lab holds none of Zhang and Norman, Larkin and Simon, Norman 1991, Kirsh, Risko and Gilbert (REUSE.md confirms). F4 holds de Winter and Dodou (read in full there) and names Dekker and Woods 2002 by DOI only. | +| 2 | Crossref (DOI resolution) | 38 DOIs, among them 10.1111/j.1551-6708.1987.tb00863.x; 10.1207/s15516709cog1801_3; 10.1207/s15516709cog1903_1; 10.1207/s15516709cog1804_1; 10.1016/0004-3702(94)00017-U; 10.1006/ijhc.1996.0048; 10.1016/j.tics.2016.07.002; 10.1007/s101110200022; 10.1109/3468.844354 | all resolved to the expected title | The DOI 10.1016/0364-0213(94)90021-3 that a WebSearch summary gave for Zhang and Norman 1994 does not resolve; Crossref returns 10.1207/s15516709cog1801_3 for the title. I use the latter. Crossref gives Larkin and Simon pp. 65-100; the PDF header reads 65-99. | +| 3 | Unpaywall | same DOIs, `?email=` as in `oa_acquire.py` | OA copy located for Hutchins 1995 (publisher bronze, blocked at fetch), Scaife and Rogers (green), Vicente and Rasmussen 1992 (DTU), Rasmussen and Vicente 1989 (DTU landing page only), Risko and Gilbert (UCL), Kalyuga et al. (UNSW, metadata-only) | Publisher "bronze" links returned HTML bot pages for Hutchins, Kirsh 1995 and Kasisopa 2013. | +| 4 | WebSearch | `Zhang Norman 1994 "Representations in distributed cognitive tasks" pdf` | 10 | Author-course copy at pages.ucsd.edu/~scoulson/203/zhang.pdf retrieved, 22 pp. | +| 5 | WebSearch | `Larkin Simon 1987 "Why a diagram is (sometimes) worth ten thousand words" pdf` | 10 | UCSD course copy retrieved (OCR text, noisy). | +| 6 | WebSearch | `Norman "Cognitive artifacts" 1991 Designing Interaction pdf "system view" "personal view"` | 10 | No free copy of the chapter. Only a glossary summary; I do not quote it. Norman's two views are quoted from Zhang and Norman 1994, p. 119. | +| 7 | WebSearch | `Dekker Woods "MABA-MABA or abracadabra" substitution myth pdf` and a second variant | 9 / 9 | Publisher closed (Unpaywall: no OA location; Springer redirects to login). A blog's reading notes carry eleven short excerpts; I cannot verify them and do not quote them as the paper. | +| 8 | WebSearch | `Risko Gilbert "Cognitive offloading" Trends in Cognitive Sciences 2016 pdf` | 10 | UCL Discovery author manuscript retrieved. | +| 9 | WebSearch | `Kirsh Maglio "On distinguishing epistemic from pragmatic action" pdf` | 9 | adrenaline.ucsd.edu copy retrieved. | +| 10 | WebSearch | `Kirsh "The intelligent use of space" Artificial Intelligence 1995 pdf` | 10 | The archive link returned 404; the revised author version at interactivity.ucsd.edu/articles/Space/AIJ1.html (HTML) retrieved. | +| 11 | WebSearch | `Scaife Rogers "External cognition: how do graphical representations work" computational offloading pdf` | 9 | Dourish course copy with journal pagination retrieved. | +| 12 | WebSearch | `Hutchins "How a cockpit remembers its speeds" pdf` | 9 | Berkeley course copy with journal pagination retrieved. | +| 13 | WebSearch | `Parasuraman Sheridan Wickens 2000 ... pdf` | 10 | No clean copy; I reuse the repo card (full text, page-verified). | +| 14 | WebSearch | `Kalyuga Ayres Chandler Sweller "The expertise reversal effect" 2003 pdf` | 10 | Wollongong repository lists a file; download returned empty. Abstract only (OpenAlex). | +| 15 | WebSearch | `Soderstrom Bjork "Learning versus performance" pdf` | 10 | hwb.gov.wales copy of the SAGE article retrieved. | +| 16 | WebSearch | `NIOSH hierarchy of controls elimination substitution engineering controls ...` | 10 | cdc.gov blocked `curl` (Access Denied); read through WebFetch, which returns a model-written extraction, so the wording is not verified verbatim. | +| 17 | WebSearch | `Erickson Kellogg "Social translucence ..." pdf` | 10 | Northwestern copy with journal pagination retrieved. | +| 18 | WebSearch | `word spacing scriptio continua as interface design OR "human factors" ... Saenger history of reading HCI` | 9 | One HCI paper found (Termuende et al. 2025, arXiv); no HCI or human-factors work that uses the history of writing as a design case. | +| 19 | WebSearch | `Waller "Text as diagram" typography ...` | 9 | Bibliographic record only (Jonassen, ed., 1982, pp. 137-166). Not read. | +| 20 | WebSearch | `Heath Luff 1992 "Collaboration and control" ... pdf` | 9 | Repository record with abstract; no full text. | +| 21 | WebFetch | dl.eusset.eu record for Heath and Luff | 1 | Abstract only. | +| 22 | OpenAlex | `search=` for Pea 1985; Gutwin and Greenberg 2002; Kohsom and Gobet 1997; Shen et al. 2012; Kasisopa et al. 2013; Clark and Brennan 1991; Dourish and Bellotti 1992; Salomon et al. 1991; Endsley 1995 | 3 each | Found all but Dourish and Bellotti 1992 and Salomon, Perkins and Globerson 1991 (no record returned; not pursued). Pea abstract and Kohsom and Gobet PDF (Brunel) retrieved. | +| 23 | OpenAlex abstracts | DOIs for Jordan, Kalyuga et al., Vicente 2002, Wickens and Carswell, Dillon 1992 | 5 | Abstracts only. Woods 1984 returned no abstract. | +| 24 | archive.org | advanced search for Norman 1993, Norman *Design of Everyday Things*, Hutchins 1995 | 1 / 6 / 1 | *Things That Make Us Smart* and *Cognition in the Wild* are borrow-only. Two uploads of the Norman design book exist but are user uploads of a copyrighted book; I did not use them. | +| 25 | arXiv | 2508.16914 (Termuende, Larson, Nacenta) | 1 | Full text read in part at the time; since read in full (card). | +| 26 | Scholar Gateway | one natural-language query on design versus training in function allocation | error | `ACCESS_DENIED: Access denied`. Not retried. | +| 27 | Consensus | `designing out a task versus training the operator: representation design reduces task difficulty human factors` | 20 | Surfaced Casner and Larkin 1989, Kirsh 1996, Moritz et al. 2020, Eisma et al. 2020 and Kalyuga et al. 2001 (abstracts only). One call only. | +| 28 | Philpapers | Kirsh 1996 PDF | blocked | Returned an HTML bot page. Abstract only. | + +--- + +## (b) Entries + +Order: representation and external cognition; function allocation; display design; skill, training and design as substitutes; the contrast case; HCI and writing; coordination through a system. + +### 1. Representation and external cognition + +**larkin1987why** — Larkin, J. H., & Simon, H. A. (1987). Why a diagram is (sometimes) worth ten thousand words. *Cognitive Science, 11*(1), 65-100. https://doi.org/10.1111/j.1551-6708.1987.tb00863.x +- *Verification:* Crossref title and volume match; UCSD course copy carries journal pagination (pp. 65-99 in the header). +- *Read status:* **read in full (card: library/larkin1987why.md).** OCR text is noisy. +- *What it argues:* Two representations can hold the same information and still differ in how cheaply a given processor can use it, because they differ in search, recognition and inference. +- *Quotations:* + - p. 67: "Two representations are informationally equivalent if all of the information in the one is also inferable from the other, and vice versa." And: "Two representations are computationally equivalent if they are informationally equivalent and, in addition, any inference that can be drawn easily and quickly from the information given explicitly in the one can also be drawn easily and quickly from the information given explicitly in the other, and vice versa." + - p. 65 (abstract): "Sentential representations are sequential, like the propositions in a text. Diagrammatic representations are indexed by location in a plane." + - pp. 69-70, on recognition: "Human abilities to recognize information are highly sensitive to the exact form (representation) in which the information is presented to the senses (or to memory)." + - p. 91, on a geometry problem: "In the sentential representation, the perceptual work of recognition is explicit and extensive; in the diagrammatic representation, it is automatic and easy." +- *Bearing:* **Supports H-cost and problem 2.** Spaced and unspaced text fit the definition of informationally equivalent strings with unequal computational cost (the paper's definition of computational equivalence includes the informational condition and turns on "easily and quickly", which the authors call "not precise terms", p. 67). The sentence at p. 91 states the scribes' gain in the authors' own terms: the perceptual work moves from the reader's recognition process to the drawn form. One caution is mine, not the authors': where an unspaced string has two lawful parses, the spaced text commits to one and is no longer informationally equivalent to its source. The authors' "diagram" is a spatial index over a plane, not a line of text, so the match to spacing is by function, not by kind. + +**zhang1994representations** — Zhang, J., & Norman, D. A. (1994). Representations in distributed cognitive tasks. *Cognitive Science, 18*(1), 87-122. https://doi.org/10.1207/s15516709cog1801_3 +- *Verification:* Crossref; UCSD course copy with journal pagination. +- *Read status:* **read in full (card: library/zhang1994representations.md).** +- *What it argues:* The representational system of a task is the set of its internal and external representations; isomorphic representations of one formal structure produce different behaviour, because they activate different processes. +- *Quotations:* + - p. 88: "The representational effect refers to the phenomenon that different isomorphic representations of a common formal structure can cause dramatically different cognitive behaviors." The authors' example is Arabic against Roman numerals for multiplication (73 x 27 against LXXIII x XXVII), also p. 88. + - p. 118: "different processes are activated by different representations, but not vice versa. For example, perceptual processes are activated by external representations, while cognitive processes are usually activated by internal representations." + - p. 118: "External objects are not just peripheral aids to cognition, they provide a different form of representation." + - p. 118, property 2: external representations "can provide information that can be directly perceived and used without being interpreted and formulated explicitly"; in the physically constrained version of the puzzle, "When the subjects were asked to explicitly formulate the rules after the games, few could do it." + - p. 119, property 4, citing Norman (1991): "From the system's view (internal + external representations), external representations can make a task easier; from the person's view (internal representations only), external representations change the nature of the task." + - p. 110: "the more rules externalized, the easier the problem. In addition, external rules seem to be error proof." +- *Bearing:* **The core source for the vocabulary (problem 2).** Spacing is a change of external representation that leaves the formal task intact and moves word-boundary work from a cognitive process in the reader to a perceptual process on the page. Two features cut against the author's wording. First, the page-119 sentence says the task changes for the person; it does not say it disappears. Second, the p. 118 finding that subjects followed rules they could not state is a cost the paper's framing should face: knowledge embedded in the artefact is used without being held, so it is also not inspectable or contestable by the user. The experiments are four laboratory studies on Tower of Hanoi isomorphs (N = 24, 24, 18, 36 undergraduates); the paper is silent on text, reading and writing (card). + +**zhang1995representational** — Zhang, J., & Norman, D. A. (1995). A representational analysis of numeration systems. *Cognition, 57*(3), 271-295. https://doi.org/10.1016/0010-0277(95)00674-3 +- *Verification:* Crossref title match. *Read status:* **not accessible** (publisher closed). It is the numeral-system case that Zhang and Norman 1994 (p. 88) and Scaife and Rogers (p. 189) cite. Do not describe its findings from this facet. A human with library access should read it for facet F, where place-value numerals are the parallel case. + +**norman1991cognitive** — Norman, D. A. (1991). Cognitive artifacts. In J. M. Carroll (Ed.), *Designing interaction: Psychology at the human-computer interface* (pp. 17-38). Cambridge University Press. +- *Verification:* Cambridge catalogue record seen through search results; no DOI. *Read status:* **not accessible.** I quote Norman's two views only as Zhang and Norman 1994 (p. 119) report them, above. The sentence "artifacts don't make us smarter ... they change the task" that circulates in glossaries is not quoted here because I could not check it against the chapter. A library copy is needed before the paper uses it. + +**hutchins1995cockpit** — Hutchins, E. (1995). How a cockpit remembers its speeds. *Cognitive Science, 19*(3), 265-288. https://doi.org/10.1207/s15516709cog1903_1 +- *Verification:* Crossref; Berkeley course copy with journal pagination. +- *Read status:* **read in full (card: library/hutchins1995cockpit.md).** +- *What it argues:* The unit that remembers landing speeds is the cockpit, not the pilot. Speed cards and airspeed-indicator "bugs" re-represent the same speeds so that a different, cheaper set of processes can do the job. +- *Quotations:* + - p. 282: "Adding speed bugs to the system does nothing to alter the memory of the pilots, but it does permit a different set of processes to be assembled into a functional system that achieves the same results as the system without speed bugs." + - p. 282: "Individual pilot memory has not been enhanced; rather, the memory function has now become a property of a larger system in which the individual engages in a different sort of cognitive behavior." And, p. 283: "To call speed bugs a 'memory aide' for the pilots is to mistake the cognitive properties of the reorganized functional system for the cognitive properties of one of its human components." + - p. 283, on the final approach: "a memory and scale reading task is transformed into a judgment of spatial adjacency." + - p. 280: "Setting any single speed bug to a particular value requires the pilot to hold the target speed in memory, read speed scale, locate the target speed on the speed scale ... and then, manually, move the speed bug to the scale position." + - p. 277: "The external speed bug settings capture a regularity in the environment that is of a shorter time scale than the weight/speed correspondences that are represented in the speed card booklet. The speed bug settings are a memory that is malleable, and that fits a particular period of time." +- *Bearing:* **Supports H-cost and the vocabulary; sharpest source on what "moves."** Hutchins's phrase "reorganized functional system" is the nearest human-factors name for the scribes' move, and his "transformed into a judgment of spatial adjacency" is the located statement that a task becomes a perceptual one. Two further points. The work is relocated in time and person: someone must set the bugs with working-memory cost (p. 280), as someone must space the text (facet A). And Hutchins says the speed bugs "capture a regularity in the environment that is of a shorter time scale" than the speed card, a memory "that is malleable" (p. 277); the page-to-coordination analogy fails exactly where the regularity drifts (see (c)(iv)). The bugs do not take the monitoring away: the flying pilot must still monitor airspeed, glide path and runway alignment together (p. 283), and the pilot not flying watches airspeed and calls deviations (p. 285), so both pilots keep reading the instruments (card). All Hutchins quotations here come from the 1995 *Cognitive Science* article, pp. 277-283, and are cited to it. The repo's card for the 1995 book *Cognition in the Wild* is marked unverified and I did not reopen it; no quotation here is cited to the book. + +**kirshmaglio1994distinguishing** — Kirsh, D., & Maglio, P. (1994). On distinguishing epistemic from pragmatic action. *Cognitive Science, 18*(4), 513-549. https://doi.org/10.1207/s15516709cog1804_1 +- *Verification:* Crossref; UCSD author copy. *Read status:* **read in full (card: library/kirshmaglio1994distinguishing.md).** +- *What it argues:* Tetris players rotate and translate pieces to change the world so that a perceptual or computational problem gets easier, which the authors call epistemic action. +- *Quotation:* pp. 513 and 515 (the epistemic definition is on p. 513, "own computational state" on p. 514, and the pragmatic-action definition, "closer to his or her physical goal", on p. 515): "we distinguish pragmatic actions—actions performed to bring one physically closer to a goal—from epistemic actions—actions performed to uncover information that is hidden or hard to compute mentally." +- *Bearing:* **Complicates the vocabulary.** Epistemic action is something the agent does to its own task. Unspaced readers have such actions available (reading aloud and pointing are plausible cases; that link is my inference, not Kirsh and Maglio's, and the evidence is in facet A). Spacing by a scribe is a change made by a third person before the reader acts. The term fits the reader's workarounds, not the scribes' move. + +**kirsh1995intelligent** — Kirsh, D. (1995). The intelligent use of space. *Artificial Intelligence, 73*(1-2), 31-68. https://doi.org/10.1016/0004-3702(94)00017-U +- *Verification:* Crossref; revised author version (HTML) at interactivity.ucsd.edu/articles/Space/AIJ1.html. *Read status:* **read in full (card: library/kirsh1995intelligent.md).** No page numbers in the HTML; locators are section names. +- *Quotation (section "Spatial Arrangements That Simplify Perception"):* "Perhaps the most obvious way of simplifying perception is to arrange objects in space so that they form equivalence classes, or partitions, that reflect relevant preconditions, or properties that are useful to track, notice, or exploit." And: "By segregating the two groups we highlight their differences." +- *Bearing:* **Supports (vocabulary, problem 2).** Spatial partitioning to ease categorisation is the closest cognitive-science description of what a space between words does. Kirsh's focus is the expert agent arranging its own workspace, but his text is wider than that: he says arrangements cue "cognitive events or processes in himself or herself, or another agent", his example is supermarket managers who bias "consumer choice" with displays, and his fourth assumption has work environments pre-structured by others (section "How to Think About Everyday Activity"). An arrangement made by one party for another is therefore inside his frame; the supermarket case is one where the arranger's interest differs from the shopper's, which Kirsh does not draw out (card). + +**kirsh1996adapting** — Kirsh, D. (1996). Adapting the environment instead of oneself. *Adaptive Behavior, 4*(3-4), 415-452. https://doi.org/10.1177/105971239600400307 +- *Verification:* Crossref (via Consensus record, title and DOI); *Read status:* **abstract only** (the Philpapers PDF was behind a bot wall). +- *What it argues (abstract):* "creatures with the capacity to redesign their environments have an adaptive advantage over those who can adapt only passively," and one major method is to "change its cognitive congeniality," changes that "leave the state space formally intact but reduce the number and cost of mental operations needed for task success." +- *Bearing:* **Supports (vocabulary).** The title is the cleanest statement I found of the contrast between redesigning the environment and adapting the human, but it describes an agent redesigning its own environment. Read the full text before relying on it; a human with library access should start here. + +**scaife1996external** — Scaife, M., & Rogers, Y. (1996). External cognition: How do graphical representations work? *International Journal of Human-Computer Studies, 45*(2), 185-213. https://doi.org/10.1006/ijhc.1996.0048 +- *Verification:* Crossref; Dourish course copy with journal pagination. *Read status:* **read in full (card: library/scaife1996external.md).** +- *Quotations:* + - p. 188: "Computational offloading. This refers to the extent to which differential external representations reduce the amount of cognitive effort required to solve informationally equivalent problems." + - p. 189: "Re-representation. This refers to how different external representations, that have the same abstract structure, make problem-solving easier or more difficult." + - p. 210: "Many of the presumed benefits of good-old fashioned graphical representations (i.e. static diagrams) were considered to be due to years of practice of perceptual processing of visual stimuli and the learning of graphical conventions." And: "Similarly we have even less understanding of how (and if) computational offloading works in such dynamic environments." +- *Bearing:* **Supports the vocabulary and states the limit asked for in item 8.** "Computational offloading" and "re-representation" are defined for representations, not for acts, so they fit the scribes' move better than Risko and Gilbert's term. The authors also say the benefit of static representations rests on practice with conventions, which weakens "no training needed," and they say the offloading story is not understood for dynamic settings. + +**risko2016cognitive** — Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. *Trends in Cognitive Sciences, 20*(9), 676-688. https://doi.org/10.1016/j.tics.2016.07.002 +- *Verification:* Crossref; UCL Discovery author manuscript. *Read status:* **read in full (card: library/risko2016cognitive.md).** +- *Quotation:* ms. pp. 2-3: "cognitive offloading: the use of physical action to alter the information processing requirements of a task in order to reduce cognitive demand." Box 3: "In many cases the decision to offload is not made by the individual. Rather, offloading is 'built-in' to the task environment by design." Costs named there: spatial-memory impairment with GPS, impaired memory for photographed objects, automation complacency and bias, and "skill decay." +- *Bearing:* **Complicates.** The definition says physical action, which points at the person's own act, but Box 3 counts offloading that a designer builds into the task environment as an instance (ms. p. 33), so the review also covers designer-built cases and the definition does not by itself exclude a scribe's spacing; the text never mentions writing, reading or conventions, so applying it to spacing is the paper's own step. Box 3 ties the built-in case to costs that appear when a task is absorbed by design. Use "offloading" for the reader's own moves; use "re-representation" for the scribe's. + +**clark1998extended** — Clark, A., & Chalmers, D. (1998). The extended mind. *Analysis, 58*(1), 7-19. https://doi.org/10.1093/analys/58.1.7. *Read status:* **not accessible**; pointer only, as instructed. I make no claim from it. + +### 2. Function allocation + +**dewinter2014fitts** (reused) — de Winter, J. C. F., & Dodou, D. (2014). Why the Fitts list has persisted throughout the history of function allocation. *Cognition, Technology & Work, 16*(1), 1-11. https://doi.org/10.1007/s10111-011-0188-1 +- *Read status:* **read in full by the F4 run and in the card library (card: library/dewinter2014fitts.md)**; I did not reopen it (the publisher PDF returned an HTML page and the Delft resolver failed). Per F4, it reproduces the Fitts (1951) list and treats Jordan (1963) and Dekker and Woods (2002) as critiques it survives. I do not add any claim about it. + +**jordan1963allocation** — Jordan, N. (1963). Allocation of functions between man and machines in automated systems. *Journal of Applied Psychology, 47*(3), 161-165. https://doi.org/10.1037/h0043729 +- *Verification:* Crossref (abstract supplied). *Read status:* **abstract only.** +- *Abstract (Crossref):* "men and machines are not comparable, but complementary, and the term 'allocation of tasks to men and machine' becomes meaningless. Rather a task must be done by men and machines." +- *Bearing:* **Complicates.** The oldest statement in the function-allocation literature that rejects the either-or framing, on grounds of complementarity. It can be cited for the claim that "moving the task" is a poor description even in 1963. Abstract only; read the article before quoting more. + +**dekker2002maba** — Dekker, S. W. A., & Woods, D. D. (2002). MABA-MABA or abracadabra? Progress on human-automation co-ordination. *Cognition, Technology & Work, 4*(4), 240-244. https://doi.org/10.1007/s101110200022 +- *Verification:* Crossref; OpenAlex (328 citations). *Read status:* **not accessible** (publisher closed; no OA copy; no abstract in Crossref or OpenAlex). +- *What I can say:* Search summaries and a third-party blog's notes agree that the paper argues against substitution-based function allocation: that automating a function transforms the work and does not replace it, and that "who does what" is the wrong question. I did not verify any wording, and I do not quote the blog's excerpts as the paper's. The repo's `parasuraman2000` card also names the paper as an attack on MABA-MABA allocation, without detail. +- *Bearing:* **Likely complicates, unverified.** If the paper says what the summaries say, it warns that "the task disappears" misdescribes any reallocation. The fix is a human-factors library copy before this source carries weight; it is the single most important unread paper in this facet. + +**parasuraman2000types** (reused) — Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. *IEEE Transactions on Systems, Man, and Cybernetics, Part A, 30*(3), 286-297. https://doi.org/10.1109/3468.844354 +- *Read status:* **read in full** per repo card `algorithmacy_design_ethics/literature/library/parasuraman2000.md` (page-verified in an S2 pass); not reopened. +- *Per the card:* automation "is not all or none" but applies at a level to each of four functions: information acquisition, information analysis, decision and action selection, and action implementation (p. 287); "high levels should be implemented only for low-risk situations" for decision and action (p. 293); adaptive automation is set aside (p. 289). +- *Bearing:* **Complicates.** The model is built for automation, which acts. A static page does not act, so "moving parsing from reader to page" is not automation in this model's sense. It is nearest to the information-acquisition and analysis stages, delivered without an agent. The same model would class a log or audit trail as automation of acquisition that leaves analysis with the human, a point for (c)(ii). + +**hollnagel2005joint / woods2006joint** (pointers) — Hollnagel & Woods, *Joint cognitive systems* (2005, 2006); named in F3 entry 35 as unread. I did not open them. Facet G or a library visit should read the joint-cognitive-system framing. + +### 3. Display design principles + +**vicente1992ecological** (key reused from the repo bib) — Vicente, K. J., & Rasmussen, J. (1992). Ecological interface design: Theoretical foundations. *IEEE Transactions on Systems, Man, and Cybernetics, 22*(4), 589-606. https://doi.org/10.1109/21.156574 +- *Verification:* Crossref; DTU accepted manuscript (no journal pagination; locators are section headings). *Read status:* **read in full (card: library/vicente1992ecological.md).** +- *Quotations:* + - IV.A: "In general, perceptual processing is fast, effortless, and proceeds in parallel, whereas analytical problem solving is slow, laborious, and proceeds in a serial fashion." And: "It is important to note, however, that the lower levels can only be activated in familiar situations because they require that the operator be attuned to the perceptual features of the environment." + - IV.A, why perceptual design suits process control: "operators of such systems are highly skilled and have extensive experience in controlling the system" and "interface design for complex systems consists of specifying an interface for a single, specific application." + - V.B: "the approach will only succeed to the extent that designers understand the system they are building." And on generality: "the generalizability of the EID framework remains an open question." + - Abstract (as retrieved): "The basic goals of EID are not to force processing to a higher level than the demands of the task require ... an EID interface should not contribute to the difficulty of the task." +- *Bearing:* **Supports in principle; sets three limits that matter for item 8.** Spacing fits the EID aim of making a task perceptual instead of analytical, and a space between words is plausibly skill-based support by perceptual means; that application is mine, since the paper is about process control and never treats text. But EID assumes (1) trained, repeated-use operators who attune to the display, (2) one known application, and (3) a designer who knows the constraints. Each fails for a lay user coordinating through an adaptive mediator. + +**rasmussen1989coping** — Rasmussen, J., & Vicente, K. J. (1989). Coping with human errors through system design: Implications for ecological interface design. *International Journal of Man-Machine Studies, 31*(5), 517-534. https://doi.org/10.1016/0020-7373(89)90014-X. *Verification:* Crossref. *Read status:* **not accessible** (DTU record returned HTML only). Cited as the earlier statement of EID; no claim made. + +**vicente2002ecological** — Vicente, K. J. (2002). Ecological interface design: Progress and challenges. *Human Factors, 44*(1), 62-78. https://doi.org/10.1518/0018720024494829 +- *Verification:* Crossref; OpenAlex abstract. *Read status:* **abstract only.** +- *Abstract:* "This literature review shows that in situations requiring problem solving, EID improves performance when compared with current design approaches in industry." Also: "An experimental evaluation of EID using a full-fidelity simulator with professional workers has yet to be conducted." +- *Bearing:* **Supports with a limit.** The evidence base for "make constraints visible so perception does the work" is real but is studies with simulated or student operators in industrial domains, not with lay coordinators. + +**wickens1995proximity** — Wickens, C. D., & Carswell, C. M. (1995). The proximity compatibility principle: Its psychological foundation and relevance to display design. *Human Factors, 37*(3), 473-494. https://doi.org/10.1518/001872095779049408 +- *Read status:* **abstract only.** *Abstract:* "Displays relevant to a common task or mental operation (close task or mental proximity) should be rendered close together in perceptual space (close display proximity)." +- *Bearing:* **Supports (vocabulary, tentative).** Read on its face, letters belonging to one lexical unit sit close and lexical units sit apart. That mapping is my inference. The paper is about displays, and I did not read it. + +**woods1984visual** — Woods, D. D. (1984). Visual momentum: A concept to improve the cognitive coupling of person and computer. *International Journal of Man-Machine Studies, 21*(3), 229-244. https://doi.org/10.1016/S0020-7373(84)80043-7. *Verification:* Crossref. *Read status:* **not accessible** (no abstract in OpenAlex). Not used. + +**bennett-flach** — Bennett & Flach (2011), *Display and interface design*; and the 1992 *Human Factors* paper on configural displays. **Not located or opened.** Gap. + +### 4. Skill, training and design as substitutes + +**niosh-hierarchy** — National Institute for Occupational Safety and Health. (n.d.). *Hierarchy of controls.* U.S. Centers for Disease Control and Prevention. https://www.cdc.gov/niosh/hierarchy-of-controls/about/index.html (accessed 2026-10-12). +- *Read status:* **read in part** through a WebFetch extraction (the page is bot-walled for `curl`). The wording below is the extraction's quotation and needs one check against the live page. +- *Extraction:* "Elimination removes the hazard at the source." Substitution, engineering controls, administrative controls and PPE follow in that order. "Elimination, substitution, and engineering controls are more effective" because they "control exposures without significant human interaction." Training is listed under administrative controls ("Work process training"). +- *Bearing:* **Supports problem 2 as a recognised distinction, from safety engineering and not from cognitive science.** The ranking puts "remove it from the work" above "train the worker," for the reason that the control does not depend on the person. This is the nearest field-wide statement of the author's line. It is a hierarchy of effectiveness against hazards, not a finding about cognition, and nothing in it concerns reading. + +**erickson2000social** — Erickson, T., & Kellogg, W. A. (2000). Social translucence: An approach to designing systems that support social processes. *ACM Transactions on Computer-Human Interaction, 7*(1), 59-83. https://doi.org/10.1145/344949.345004 +- *Verification:* Crossref; Northwestern copy with journal pagination. *Read status:* **read in full (card: library/erickson2000social.md).** +- *Quotations:* p. 61: a stairwell door slammed into people, a sign reading "Please Open Slowly" failed, and "putting a glass window in the door. The glass window approach means that the sign is no longer required." Three reasons (pp. 61-62): the window "makes socially significant information visible," "supports awareness," and gives accountability: "I know that you know that I know you are there." The abstract: "digital systems [can] support coherent behavior by making participants and their activities visible to one another." +- *Bearing:* **Supports problem 2 as a visibility example, not as design over training (AUDIT_02).** The sign asks the human to do something; the window changes what the human perceives so the sign is unnecessary. It also shows that "making visible" is a respectable artefact-side move in human factors, so the author's rejection of "legibility" cannot rest on visibility itself. (c)(ii) develops this. It is a design argument illustrated with several vignettes (restaurant, reception, wine shop, book-ordering workshop, limerick game, as well as the stairwell door) and a chat system (Babble), not an experiment on the door. The accountability argument assumes the parties see each other with symmetric information and that the system is a neutral medium; the authors note that "assurances of communicative symmetry" are "not automatically present in the digital realm" (p. 64) but treat the cure as more visibility and never consider a system owner with interests of its own (card). + +**sweller1988cognitive / sweller1998cognitive** — Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. *Cognitive Science, 12*(2), 257-285. https://doi.org/10.1207/s15516709cog1202_4. Sweller, J., van Merriënboer, J. J. G., & Paas, F. G. W. C. (1998). Cognitive architecture and instructional design. *Educational Psychology Review, 10*(3), 251-296. https://doi.org/10.1023/A:1022193728205. *Verification:* Crossref. *Read status:* **not accessible.** The repo holds `sweller2019cognitive`. Extraneous load is the construct that spacing would reduce; I make no claim from these two papers. + +**kalyuga2003expertise** — Kalyuga, S., Ayres, P., Chandler, P., & Sweller, J. (2003). The expertise reversal effect. *Educational Psychologist, 38*(1), 23-31. https://doi.org/10.1207/S15326985EP3801_4 +- *Verification:* Crossref; OpenAlex abstract. *Read status:* **abstract only.** +- *Abstract:* "Instructional techniques that are highly effective with inexperienced learners can lose their effectiveness and even have negative consequences when used with more experienced learners. We call this phenomenon the expertise reversal effect." +- *Bearing:* **Complicates (who benefits).** It predicts that an aid that helps novices can hinder experts. It is a result about instructional techniques; applying it to reading aids is an extension. Section (c)(iii) tests the extension against the spacing data. + +**kalyuga2001learner** — Kalyuga, S., Chandler, P., & Sweller, J. (2001). Learner experience and efficiency of instructional guidance. *Educational Psychology.* https://doi.org/10.1080/01443410124681 (journal, year and DOI from the Consensus record; volume and pages not verified; DOI not separately resolved against Crossref). *Read status:* **abstract only** (Consensus). Apprentices given worked examples did better on complex tasks, and "the advantage of a worked examples procedure disappeared" as experience rose. + +**soderstrom2015learning** — Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. *Perspectives on Psychological Science, 10*(2), 176-199. https://doi.org/10.1177/1745691615569000 +- *Verification:* Crossref; hwb.gov.wales copy of the article. *Read status:* **read in part** (pp. 192-193 and the concluding comments). +- *Quotations:* p. 192: "the general idea that what can hurt performance can help learning is captured in the desirable difficulties framework." p. 193: "when the difficulties cannot be overcome by the learner ... they become undesirable." Concluding comments: "conditions that appear to degrade acquisition performance are often the very conditions that yield the most durable and flexible learning." +- *Bearing:* **Complicates.** A task removed from the reader is a difficulty removed. Bjork's framework says that removal raises performance now and may lower learning. The framework is about learning outcomes; if the goal is to read, not to become a better parser, the difficulty is extraneous. If the goal is competence in a coordination, as in the author's definition of algorithmacy, the difficulty may be the learning. Everything depends on which goal the paper sets. + +**moritz2019control** — Moritz, J., et al. (2020; Crossref issue date 2020, the facet's earlier "2019" was the PsycINFO year). Control over spatial representation format enhances information extraction but prevents long-term learning. *Journal of Educational Psychology.* https://doi.org/10.1037/edu0000364 *(Consensus returned the first author, title, year, journal and DOI only; volume, pages and co-authors not verified. The DOI was not separately resolved against Crossref.)* +- *Read status:* **abstract only** (Consensus). *Abstract:* participants who offloaded a spatial transformation did better at once, but "in subsequent problem solving without representation control (up to 24 hr later), participants who had been using representation control previously fell back to the level of an untrained baseline condition." +- *Bearing:* **Complicates.** A direct test that offloading by representation does not build the offloaded skill. The effect did not generalise to a near-transfer task. Verify the reference details before the paper cites it. + +**casner1989cognitive** — Casner, S., & Larkin, J. (1989). *Cognitive efficiency considerations for good graphic design* (technical report). https://doi.org/10.21236/ada218976. *Read status:* **abstract only** (Consensus). *Abstract:* designing "external data structures that (a) allow users to substitute less effortful visual operators for more effortful logical operators, and (b) reduce search"; display variants differed by a factor of two in response time (7 subjects). *Bearing:* **Supports (vocabulary).** "Substituting visual operators for logical operators" is a precise name for the perceptual-for-cognitive swap. DTIC report; full author list not checked. + +### 5. The contrast case: monitoring + +**bainbridge1983ironies** (reused; repo bib key) — Bainbridge, L. (1983). Ironies of automation. *Automatica, 19*(6), 775-779. https://doi.org/10.1016/0005-1098(83)90046-8 +- *Read status:* **read in full** per repo card `bainbridge1983.md` (all twenty quotations page-verified); I did not reopen it. Facet G covers it in depth. +- *Per the card:* p. 776: "there is therefore no way in which the human operator can check in real-time that the computer is following its rules correctly ... The human monitor has been given an impossible task"; p. 777: the computer should decide "using methods and criteria, and at a rate, which the operator can follow." +- *Bearing:* **Supplies the contrast.** What separates the cases is not that one is automation and the other is not. A spaced page asks the reader to do a task that remains within real-time human capability, whereas a monitoring task grows past it. Bainbridge's p. 777 remedy is itself an artefact-side constraint, so the HF literature offers support for bounding the mediator and not only for displaying it. + +**eisma2020augmented** — Eisma, Y. B., et al. (2020). Augmented visual feedback: Cure or distraction? *Human Factors.* https://doi.org/10.1177/0018720820924602 *Read status:* **abstract only** (Consensus). Novices with a solution space diagram for air-traffic conflicts found the task subjectively easier and detected conflicts better, but showed "visual tunneling and misunderstanding," and the authors advise "Caution ... when human operators are expected to reproduce task solutions that are provided by augmented visual feedback." *Bearing:* **Complicates.** An air-traffic case in which an artefact absorbs a judgment task and the absorbed answer then has to be trusted. Verify the author list and pages before citing (n = 13 and 11). + +**norman1990problem / sarter-woods / parasuraman1997humans** — reused lab cards (`norman1990.md` etc.); not reopened. Facet G owns them. + +### 6. HCI, document design and the history of writing + +**termuende2025impact** — Termuende, M., Larson, K., & Nacenta, M. (2025). The impact of visual segmentation on lexical word recognition. arXiv:2508.16914 [cs.HC]. https://arxiv.org/abs/2508.16914 +- *Verification:* arXiv listing (23 Aug 2025); preprint, not peer reviewed. *Read status:* **read in full (card: library/termuende2025impact.md)**, preprint v1. +- *Quotations:* Introduction: "According to Saenger [58], the addition of space between words in written text was a pivotal graphical innovation that led to faster, silent reading." Abstract: "even with a moderate amount of training, these visual interventions always slow down word identification," by "between 32.7ms—color technique—and 70.7ms—connection technique" (192 participants). The two figures appear only in the abstract; the body gives 25-84 ms (Exp. 1), about 17-63 ms (Exp. 3) and "at least 17ms" (section 10). "192" counts collected datasets (48 + 48 + 96), including Exp. 2, whose results are withheld. +- *Bearing:* **The one HCI use of the scriptio continua history that I found, and it complicates H-cost.** It treats word spacing as the model case of "visual segmentation" and then shows that further segmentation cues (colour, outlines, underlines, extra spacing within words) slowed lexical decision for typical adult readers, even after training. Two cautions. It tests sub-word units (syllables, phonemes), not word spacing. And its introduction states Saenger's account as settled, but section 9.1 hedges the same account and notes opposing evidence, though the paper never names Saenger's critics (it cites Saenger once). Its lesson for the paper is that "put the segmentation in the artefact" is not self-justifying; the segmentation has to match a unit the reader already uses. + +**waller1982text** — Waller, R. H. W. (1982). Text as diagram: Using typography to improve access and understanding. In D. H. Jonassen (Ed.), *The technology of text* (pp. 137-166). Educational Technology Publications. *Verification:* bibliographic record via a reading list (Tam) and a related 1980 chapter, "Graphic aspects of complex texts: Typography as macropunctuation" (Kolers, Wrolstad & Bouma, Eds., *Processing of Visible Language*). *Read status:* **not accessible.** The titles show that a document-design tradition already treats typography as a structuring device ("macropunctuation"), which is relevant to Parkes's "grammar of legibility." I cannot say what Waller argues. + +**dillon1992reading** — Dillon, A. (1992). Reading from paper versus screens: A critical review of the empirical literature. *Ergonomics, 35*(10), 1297-1326. https://doi.org/10.1080/00140139208967394. *Read status:* **abstract only** ("single variable explanations are insufficient"). It shows the human-factors tradition of studying reading as a task, not text as an interface genre with a history. + +**Not found.** I found no human-factors or HCI source that treats the history of writing or page layout as a case of interface design apart from Termuende et al. (2025), whose use is incidental. I did not open Tufte, Schriver, or Card, Mackinlay and Shneiderman (1999). The closest accepted precedent in the HF canon for "notation as interface" is numerals: Zhang and Norman (1994, p. 88) use Arabic against Roman numerals, and Norman (1993) discusses the same; that is facet F's ground. + +### 7. Who gains: the spacing data + +**kohsom1997adding** — Kohsom, C., & Gobet, F. (1997). Adding spaces to Thai and English: Effects on reading. *Proceedings of the Cognitive Science Society* (1997 conference; volume and pages not verified). Brunel repository: http://bura.brunel.ac.uk/bitstream/2438/2122/1/Cogsci%2797-thai.pdf +- *Verification:* OpenAlex record; PDF retrieved (6 pp.). *Read status:* **read in part.** +- *Abstract:* "subjects were faster in reading and made less errors when spaces were added ... ability to read unspaced texts in Thai does not transfer to English." Eight native-Thai graduate students at one US university (mean age 26); in the coherent-text Thai condition 6 of 8 were faster with spaces. +- *Bearing:* **Complicates the expertise-reversal prediction.** Skilled native readers of an unspaced script still read faster with spaces. Tiny sample, conference paper, text on a screen. Facet D should place it. + +**shen2012eye** — Shen, D., Liversedge, S. P., Tian, J., Zang, C., Cui, L., et al. (2012). Eye movements of second language learners when reading spaced and unspaced Chinese text. *Journal of Experimental Psychology: Applied, 18*(2), 192-202. https://doi.org/10.1037/a0027485 (OpenAlex lists the first five authors; the rest not checked). *Read status:* **abstract only.** *Abstract:* "Eye movement measures showed least disruption to reading for word-spaced text," and "reading times differed between groups as a result of Chinese reading experience." *Bearing:* **Supports** the claim that spacing helps learners; it also shows experience still matters, so spacing reduces and does not remove the need for training. + +**kasisopa2013eye** — Kasisopa, B., Reilly, R. B., Luksaneeyanawin, S., & Burnham, D. (2013). Eye movements while reading an unspaced writing system: The case of Thai. *Vision Research, 86*, 71-80. https://doi.org/10.1016/j.visres.2013.04.007. *Read status:* **not accessible** (publisher PDF returned HTML; no abstract from OpenAlex). Listed as a lead for facet D. + +**pea1985beyond** — Pea, R. D. (1985). Beyond amplification: Using the computer to reorganize mental functioning. *Educational Psychologist, 20*(4), 167-182. https://doi.org/10.1207/s15326985ep2004_2. *Read status:* **read in full (card: library/pea1985beyond.md).** *Abstract:* "Computers are classically viewed as amplifiers of cognition. An alternative conceptualization is offered of computer as reorganizer of mental functioning." *Bearing:* **Supports (vocabulary).** An independent, earlier statement of the amplifier-versus-reorganizer contrast that Norman's "personal" and "system" views and Hutchins's "reorganized functional system" make; the full text confirms this and shows the contrast is built on Cole and Griffin (1980) (card). + +### 8. Coordination through a system + +**heath1992collaboration** — Heath, C., & Luff, P. (1992). Collaboration and control: Crisis management and multimedia technology in London Underground Line Control Rooms. *Computer Supported Cooperative Work, 1*(1-2), 69-94. https://doi.org/10.1007/BF00752451. *Read status:* **abstract only** (repository record, via WebFetch). The abstract says collaborative technology fails when it ignores how people work, and that staff pass information and coordinate through tacit practices. *Bearing:* **Complicates, unread.** A control-room study of coordination among several people through a display system; a human with library access should read it first for item 8. + +**gutwin2002descriptive** — Gutwin, C., & Greenberg, S. (2002). A descriptive framework of workspace awareness for real-time groupware. *Computer Supported Cooperative Work, 11*(3-4), 411-446. https://doi.org/10.1023/A:1021271517844. *Read status:* **not accessible** (closed). Existence and title verified (Crossref, OpenAlex, 1,185 citations). Not used beyond pointing to it. + +**clark1991grounding** — Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, & S. D. Teasley (Eds.), *Perspectives on socially shared cognition* (pp. 127-149). APA. https://doi.org/10.1037/10096-006. *Read status:* **not accessible.** Verified by Crossref. The common-ground source for item 8; unread. + +**CPDLC party-line studies** (reused) — Midkiff & Hansman (1993), Pritchett & Hansman (1994), Hansman, Pritchett & Midkiff (1995); repo cards `midkiff1993.md`, `hansman1995.md`. I did not reopen them. Per the library index, they are "a real, citable field case of suppressing the direct channel," and the index records that no classic HF automation-trust source models a second human party the automation also serves (Cluster C(i) finding). That absence is the main finding for item 8. + +--- + +## (c) Synthesis + +### (i) The best-supported vocabulary for the scribes' move + +Ranked by how well the located definition fits. + +1. **Re-representation, with the representational effect as its finding.** Zhang and Norman (1994, p. 88): "different isomorphic representations of a common formal structure can cause dramatically different cognitive behaviors." Scaife and Rogers (1996, p. 189): re-representation is "how different external representations, that have the same abstract structure, make problem-solving easier or more difficult." Larkin and Simon (1987, p. 67) give the cleaner formal pair: informationally equivalent, computationally unequal (my gloss; their computational equivalence includes the informational condition and rests on "easily and quickly", which they call "not precise terms"). This family names an artefact property, which suits a page. +2. **Perceptual for cognitive substitution; computational offloading.** Larkin and Simon (p. 91): the recognition work is "explicit and extensive" in one form and "automatic and easy" in the other. Zhang and Norman (p. 118): "perceptual processes are activated by external representations, while cognitive processes are usually activated by internal representations." Casner and Larkin (1989, a DTIC technical report) call it substituting "less effortful visual operators for more effortful logical operators" (abstract only). Scaife and Rogers (p. 188) call the saving computational offloading. +3. **Reorganised functional system; task transformation.** Hutchins (1995, "How a cockpit remembers its speeds", *Cognitive Science* 19, p. 283): "a memory and scale reading task is transformed into a judgment of spatial adjacency," and "Individual pilot memory has not been enhanced" (p. 282). This is the best phrase for the claim that the author wants and the one that protects it from the "training" reading. Norman's personal and system views (via Zhang and Norman, p. 119) and Pea's amplifier-versus-reorganizer contrast (abstract only) say the same. +4. **Design versus training as a ranked choice: elimination or engineering control above administrative control.** NIOSH (read through a WebFetch extraction) ranks controls that work "without significant human interaction" above work-process training. Erickson and Kellogg (2000, pp. 61-62) give a worked case of an artefact change that makes a posted instruction unnecessary: a glass window replaces the sign "Please Open Slowly", by making people visible to one another. Their account is of visibility and accountability, not of design versus training. The NIOSH ranking is the recognised form of the distinction the author draws, and it comes from safety engineering. +5. **Environmental redesign.** Kirsh (1996), abstract only: adapt the environment "instead of oneself," and change its "cognitive congeniality." Strong title, unread text. + +Terms that do not fit, with the reason. *Cognitive offloading* in Risko and Gilbert's definition is the use of physical action (ms. pp. 2-3), which points at the person's own act; their Box 3 also counts designer-built offloading (ms. p. 33), so the term is narrower than the scribe's move but not strictly excluded. *Epistemic action* (Kirsh and Maglio 1994, pp. 513-514) likewise describes an agent changing its own task; both fit the unspaced reader's workarounds, not the scribe's move. *Function allocation* presumes two agents; PSW's four stages and levels describe automation that acts, and a page does not. *Affordance* is used loosely in Zhang and Norman (p. 118, citing Gibson) and I would avoid it. *Ecological interface design* fits only as a family resemblance: EID is built for process-control operators, and the sentence "a space is skill-based support by perceptual means" is my application, not a claim from Vicente and Rasmussen. + +Does "moving parsing from reader to page" count as function allocation? Not in the Fitts-to-PSW sense, because the page is not an agent. It does count as a change of who bears a task. The cost went to the scribe (who must now decide boundaries, as the setter of a speed bug must hold and place a value, Hutchins p. 280) and the benefit went to every later reader. That distribution of work across people and time, with an artefact as the carrier, is what Hutchins's cockpit shows. The substitution myth (Dekker and Woods) is the likely warning here, with the stated caveat that I could not read the paper: the work changes hands and changes shape, and new errors and new dependencies appear. For spacing, the new dependency is that the reader inherits one parse chosen by someone else, and may not be able to recover another. That is my inference, not a sourced claim, and it links to facet A. + +### (ii) Is "artefact absorbs the task" versus "human reads more" a principled line? + +**Yes, in three separate literatures, and none of them states it in the author's words.** NIOSH's ranking is the field-wide backing for the line: change what the person meets so the instruction is unnecessary, and do not ask the person to do more (that wording is not any source's). Erickson and Kellogg (2000, pp. 61-62) give a worked case of an artefact change that makes a posted instruction unnecessary, by making people visible to one another; their account is of visibility and accountability, not of design versus training. Hutchins (1995, "How a cockpit remembers its speeds", *Cognitive Science* 19, p. 283; the article, not *Cognition in the Wild*) gives the cognitive account of why that works. The strongest source for the line from the other side, the monitoring case, is Bainbridge (1983, p. 776): when the machine decides on more criteria than a human can follow in real time, "The human monitor has been given an impossible task." That is an argument that "observe more" does not scale, which supports the author's contrast with Nadella's seven principles. + +**The line breaks in three places, and each is a source against the wording "legibility is rejected."** + +1. *Making things visible is mainstream artefact-side design.* Ecological interface design is a programme of making constraints visible so that perception does the reasoning (Vicente and Rasmussen 1992; Vicente 2002), and social translucence is a programme of making participants visible to one another (Erickson and Kellogg 2000). If the paper rejects "legibility" as a category, the human-factors literature will read it as rejecting its own best examples. The principled cut is different: an artefact change counts on the author's side when it turns the user's task into a perceptual one at a load the user can carry in real time, and counts on Nadella's side when it leaves the user's task analytical and raises its volume. By that cut, "observe everything" and "human readable evidence" fall on the monitoring side, while a display that lets a coordinating user see at a glance what the mediator is doing could fall on the author's side. Parkes's "grammar of legibility" and the ruling's "legibility" are two senses of one word: graphic legibility of a script, and interpretability of an opaque system. The first is a perceptual aid; the second is a demand to read. +2. *Containment is on the artefact side.* In the hierarchy of controls, engineering controls rank above administrative ones because they do not depend on the person. Several of Nadella's principles that I take from the repo's summary (controls outside the model, containment, independent controls) are of that kind, whatever his seventh "observe everything." Facet G should test this; it bears on problem 3. +3. *Absorbing the task has costs the sources name.* Risko and Gilbert (Box 3, ms. pp. 33-34): offloading "built-in" by design can bring complacency, automation bias and "skill decay" (the source says "could"). Moritz et al. (2020, abstract only): users who controlled a transformation fell back to the level of an untrained baseline in Experiments 1 and 2, but not in a near-transfer task. Eisma et al. (2020, abstract only): an air-traffic display that handed over the answer made the task subjectively easier and conflict detection better, with side effects of visual tunnelling and misunderstanding. Dekker and Woods (unread) warn that reallocation changes the work. The author's sentence that reading "no longer required the training" is the claim these sources push against. + +**Strongest sources for:** NIOSH; Erickson and Kellogg (as a visibility case, not design versus training); Hutchins; Zhang and Norman; NIOSH. **Strongest against:** Bainbridge supports the contrast but Risko and Gilbert, Moritz, Eisma, and Dekker and Woods qualify the claim that the task "disappears." + +### (iii) Who benefits, novice or expert + +The expertise-reversal literature predicts that aids tuned to novices can hurt experts (Kalyuga et al. 2003, abstract). The spacing data point the other way for word spacing itself, and the same data show that adding further cues hurts. + +- Thai adults read marginally faster with spaces in one small study (Kohsom and Gobet 1997; n = 8, p < .08, conference paper); Winskel et al. (2009) found them slightly slower. Second-language learners of Chinese read best with word spaces, and their reading times differed between groups with Chinese reading experience (Shen et al. 2012, abstract only). The second result shows that experience still matters. +- For adult readers of ordinary spaced text, extra segmentation cues slowed word identification even after training (Termuende et al. 2025, preprint, sub-word units). The aid that works is the one already integrated into the reader's skill. +- Evidence on the cost of removing spaces from habituated readers sits in facet D and in the repo's Part 2 facet E (Rayner et al. 1998; Mirault et al. 2019) and I did not reopen it. + +The prediction for the paper: a stable, conventional aid that every reader meets from the start can serve novices and experts, and a bolt-on aid layered over expert practice is the case the expertise-reversal effect warns about. Vicente and Rasmussen's EID cuts the same way: it presupposes trained operators who attune to a fixed display (IV.A), so it shows how experts gain, not novices. Bjork's framework adds that an aid that raises performance may fail to produce learning (Soderstrom and Bjork 2015, pp. 192-193), and Moritz et al. (2020, abstract only) found no carry-over: participants returned to the untrained baseline. + +### (iv) Where the analogy from page to coordination breaks + +The sources support the author's caution ("Traditional media hold still; an algorithmic intermediary does not") and give four located reasons. + +1. *Regularity and time scale.* Hutchins's bugs "capture a regularity in the environment that is of a shorter time scale" than the card, a memory "that is malleable" (p. 277); the argument that static media hold still needs a different source. Scaife and Rogers (p. 210) say the offloading story is not understood for "dynamic environments." Vicente and Rasmussen (V.B): the approach "will only succeed to the extent that designers understand the system they are building." An adaptive mediator has no stable regularity for a representation to freeze, and its designer may not know its constraints. +2. *The user and the application.* EID assumes skilled operators on one known application (IV.A). A lay party using many platforms is the reverse. Vicente's 2002 review also reports that no full-fidelity test of EID with professional workers had yet been done (abstract). +3. *A third party with its own objective.* Every study above puts one user and one artefact (or several humans who share a goal) in the frame. Erickson and Kellogg make participants visible to each other, and their accountability argument works because all parties see each other and the system is assumed to be a neutral medium (card). A mediator that reads both parties and commits determinations neither controls is outside what these papers model. The repo's literature finds the same absence: no classic HF automation-trust source models a second human party the automation also serves, and the CPDLC party-line studies are the one field case with two humans and a mediator. +4. *Where the hidden thing sits.* Word boundaries are a property of the signal that the reader could recover with effort. In a coordination, what is hidden is the mediator's state and aims, which the user cannot recover by effort. Heath and Luff (abstract only) describe control rooms that coordinate by tacit practice and peripheral awareness; nothing I read shows that a static display can replace that when the mediator adapts. + +### (v) Claims in the author's framing the sources contradict or qualify + +- **"it changed the interface so that reading no longer required the training"** (`draft.md:263`). Hutchins (pp. 282-283), Zhang and Norman (p. 119) and Scaife and Rogers (p. 210) speak of a task transformed, not removed, and of benefits resting on practice with conventions. Shen et al. (2012) found that groups differed in reading time with Chinese experience, whatever the spacing. Soften to "lowered the training the task required" or keep the sentence and flag it. Per the standing rule, I flag and do not rewrite. +- **"Algorithmacy is not legibility"** (`README.md:38-43`). The human-factors literature treats making constraints visible as a core artefact-side move (Vicente and Rasmussen; Erickson and Kellogg). The ruling needs a narrower statement, such as "not more reading load placed on the user," to survive a human-factors reader. +- **Course statement: "I reject that novices should get abstracted heuristics while only experts engage structural variation"** (`phd1750_topic_and_rationale.md`). The expertise-reversal literature predicts that guidance should differ by expertise (Kalyuga et al. 2003, 2001, abstracts only), and the spacing data show the exception for a stable convention. The paper can argue that its aid is of the conventional kind, but a human-factors reader will expect the argument. +- **"The history is an analogy, not a mechanism"** (Part 2). Consistent with the sources: none of them shows that word spacing caused anything beyond a reduction in reading cost, and the one HCI paper that cites Saenger on causation (Termuende et al.) states the causal claim in its introduction without testing it, and hedges it in section 9.1. This facet gives H-cause no support. +- **"Traditional media hold still"** (Part 1, line 118). Supported by Hutchins, Scaife and Rogers, and Vicente and Rasmussen, as in (iv). + +### The course headings + +| Course heading | Carries | Sources | +|---|---|---| +| **Perception** | Spacing and its relatives as perceptual re-representation; the cost of recognition and search | Larkin and Simon 1987 (pp. 67, 69-70, 91); Zhang and Norman 1994 (p. 118); Kirsh 1995 (clustering to categorize); Wickens and Carswell 1995 (abstract); Vicente and Rasmussen 1992 (IV.A); Kohsom and Gobet 1997; Shen et al. 2012 (abstract); Termuende et al. 2025. Folk theories (Eslami; already in the course statement) belong here, from the DIAL cards. | +| **Cognition** | Task transformation, offloading, the learning-versus-performance cost, expertise effects | Hutchins 1995 (pp. 277-283); Norman 1991 via Zhang and Norman (p. 119); Pea 1985 (abstract); Scaife and Rogers 1996; Risko and Gilbert 2016; Kirsh and Maglio 1994; Kalyuga et al. 2003 (abstract); Soderstrom and Bjork 2015; Moritz et al. 2020 (abstract); Sweller (not read). | +| **Human factors and design** | The design-versus-train line; function allocation and its critics; monitoring; coordination through a system | NIOSH hierarchy; Erickson and Kellogg 2000; Vicente 2002 (abstract); Jordan 1963 (abstract); Dekker and Woods 2002 (unread); Parasuraman, Sheridan and Wickens 2000 (card); Bainbridge 1983 (card); Eisma et al. 2020 (abstract); Heath and Luff 1992 (abstract); the CPDLC cards. | + +--- + +## (d) Gaps and unverified leads + +Sources I could not open, with where a human would find them. + +- **Dekker and Woods (2002)**, *Cognition, Technology & Work* 4, 240-244. The key unread paper. Springer link above; a university library with Springer access, or Dekker's own site. +- **Norman (1991)**, "Cognitive artifacts," in Carroll (ed.), pp. 17-38; **Norman (1993)**, *Things That Make Us Smart* (archive.org borrow, `thingsthatmakeus00norm_0`); **Norman (2013)**, *The Design of Everyday Things* (for "knowledge in the world and in the head"; the repo has `norman2013design` as a key but I did not read it). Library or borrow. +- **Zhang and Norman (1995)**, *Cognition* 57, 271-295: library. +- **Kirsh (1996)**, *Adaptive Behavior* 4, 415-452: Philpapers archive link blocked; library or Kirsh's site. +- **Wickens and Carswell (1995)**, **Woods (1984)**, **Vicente (2002)**, **Rasmussen and Vicente (1989)**, **Kalyuga et al. (2003)**: abstracts or nothing; library. +- **Bennett and Flach** (2011 book; 1992 *Human Factors*): not located. +- **Hutchins (1995)**, *Cognition in the Wild*, and **Hutchins (1990)**, "The technology of team navigation": the navigation case is the best-known account of coordination through artefacts and I did not read either (the repo's card is marked unverified). Library or archive.org borrow (`cognitioninwild0000hutc`). +- **Heath and Luff (1992)**, **Gutwin and Greenberg (2002)**, **Clark and Brennan (1991)**, **Dourish and Bellotti (1992)** (not found in OpenAlex), **Endsley (1995)** (repo key `endsley1995toward`, not reopened), and the **team situation-awareness** literature (Stanton et al. on distributed SA; Salas et al.): item 8 needs at least Heath and Luff and a distributed-SA paper read in full. +- **Sheridan** on supervisory control (*Telerobotics, Automation and Human Supervisory Control*, 1992; Sheridan and Verplank 1978): not opened. Facet G may cover it. +- **Jordan (1963)** beyond the abstract; **Fitts (1951)** itself (via de Winter and Dodou only). +- **Typography and document design:** Waller (1980, 1982), Schriver (1997), Tufte, Card, Mackinlay and Shneiderman (1999) on external cognition, and the legibility tradition (Tinker; Paterson and Tinker). I did not reach them. The legibility tradition matters for the equivocation in (c)(ii): it is the human-factors home of "legibility" in the graphic sense. +- **Facet D material I did not verify here:** Kasisopa et al. (2013); the Rayner removal costs; Winskel et al. on Thai. +- **Moritz et al. (2020), Eisma et al. (2020), Casner and Larkin (1989), Kalyuga et al. (2001):** bibliographic details (authors, pages) rest on the Consensus record and my memory and should be checked against Crossref or the journal before any citation. Where this file marks authors as unverified, they are. +- **NIOSH page:** one verbatim check against the live page. +- **Scholar Gateway:** access denied again (row 26). + +**Search claims, stated as findings and not as novelty claims.** I found no human-factors or HCI work, in the sources searched (rows 18-19 and 27), that uses the history of word separation as a case of interface design, apart from the incidental citation in Termuende et al. (2025). I found no source that defines the contrast "artefact absorbs the task versus human reads more" under that name; the nearest are Hutchins's "reorganized functional system," Pea's "reorganizer," Erickson and Kellogg's sign-versus-window example, and the NIOSH hierarchy. I found no human-factors study of representations for coordination among people through an adaptive, opaque mediator. + +## Changes after full-text cards (2026-10-13) + +- `zhang1994representations`: representational-effect definition and the Arabic/Roman example moved from p. 89 to p. 88 (here and in the vocabulary list and the "Not found" paragraph); the "8-to-20-subject designs" replaced by four experiments, N = 24, 24, 18, 36; paper silent on text; read status "read in full (card)". +- `hutchins1995cockpit`: read status "read in full (card)"; added that both pilots keep monitoring the instruments (pp. 283, 285). +- `kirshmaglio1994distinguishing`: pragmatic-action definition located at p. 515 (epistemic p. 513, "own computational state" p. 514); "Terms that do not fit" page changed to pp. 513-514; read status "read in full (card)". The 800-1200 ms figure and its three-player pilot basis (p. 530) are not in this facet (it is in facet I), so nothing to change here. +- `kirsh1995intelligent`: replaced "Kirsh's agent arranges its own workspace; the scribe arranges another person's" with the card's reading (arrangements cue "himself or herself, or another agent"; supermarket displays; others pre-structure the environment; arranger's interest may differ); read status "read in full (card)". +- `risko2016cognitive`: definition located at ms. pp. 2-3; Box 3 counts designer-built offloading, so "the definition requires the person's own physical action, so it does not cover a scribe's spacing" was revised (also in "Terms that do not fit"); read status "read in full (card)". +- `scaife1996external`: both quotations ("presumed benefits", "even less understanding") relocated from pp. 209-210 to p. 210; read status "read in full (card)". +- `erickson2000social`: "one anecdote and a chat system" replaced by several vignettes plus Babble; added that the accountability argument assumes symmetric visibility and a neutral medium, with p. 64 on communicative symmetry (also in the "third party" limit); read status "read in full (card)". +- `termuende2025impact`: "states Saenger's account as settled ... do not know whether the paper discusses his critics" revised (section 9.1 hedges it and notes opposing evidence, critics not named); abstract-only 32.7/70.7 ms figures flagged against body ranges (25-84, about 17-63, at least 17 ms); 192 = datasets 48 + 48 + 96 including withheld Exp. 2; "repeats the causal claim without testing it" in the claims list qualified; read status "read in full (card)". +- `larkin1987why`: "informationally equivalent, computationally unequal" marked as a gloss (definition includes the informational condition; "easily and quickly" called "not precise terms", p. 67); read status "read in full (card)". +- `pea1985beyond`, `vicente1992ecological`, `dewinter2014fitts`: read status updated to "read in full (card)"; Pea's contrast noted as built on Cole and Griffin (1980). + +Synthesis verdicts changed: none to the headline verdicts. Two terminology verdicts were qualified: "cognitive offloading ... is the person's own physical action" now says Box 3 also counts designer-built offloading, so the term is narrower than the scribe's move but not strictly excluded; and Kirsh's "agent arranges its own workspace; the scribe arranges another's" no longer marks a gap in fit. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_F_parallel_moves.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_F_parallel_moves.bib new file mode 100644 index 00000000..5eb310b9 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_F_parallel_moves.bib @@ -0,0 +1,749 @@ +% F_parallel_moves.bib -- Facet F (other single conventions credited with opening a competence) +% Agent-drafted 2026-10-12. Not reviewed by the author. +% Every DOI below was resolved against its Crossref record on 2026-10-12 (title, authors, container, volume, pages compared), +% except where the note says otherwise. Keys hutchins1985direct, larkin1987why, peters2006numeracy, shneiderman1983direct and +% zhang1995representational already exist in other repo bibs and are reused. Read status is in each note; see F_parallel_moves.md. + +% ---- Case 1: Hindu-Arabic numerals ---- +@article{zhang1995representational, + author = {Zhang, Jiajie and Norman, Donald A.}, + title = {A representational analysis of numeration systems}, + journal = {Cognition}, + volume = {57}, + number = {3}, + pages = {271--295}, + year = {1995}, + doi = {10.1016/0010-0277(95)00674-3}, + note = {Not accessible (no open copy; paywalled). Known only through Schlimm and Neth 2008, Buijsman 2020 and Chrisomalis 2020 (p. 22 snippet). Crossref gives pages 271--295.} +} + +@inproceedings{schlimm2008modeling, + author = {Schlimm, Dirk and Neth, Hansj{\"o}rg}, + title = {Modeling ancient and modern arithmetic practices: Addition and multiplication with Arabic and Roman numerals}, + booktitle = {Proceedings of the 30th Annual Meeting of the Cognitive Science Society}, + editor = {Sloutsky, V. and Love, B. and McRae, K.}, + publisher = {Cognitive Science Society}, + address = {Austin, TX}, + pages = {2097--2102}, + year = {2008}, + url = {https://www.psych.uni-goettingen.de/de/cognition/publikationen-dateien-neth-3/2008_schlimmneth_modeling-ancient-and-modern-arithmetic-practices.pdf/@@download/file/2008_SchlimmNeth_Modeling%20ancient%20and%20modern%20arithmetic%20practices.pdf}, + note = {Read in full (author copy; card: library/schlimm2008modeling.md). Not in Crossref; verified from the paper's own header. Simulation study, no human participants.} +} + +@article{buijsman2020numerals, + author = {Buijsman, Stefan}, + title = {How numerals support new cognitive capacities}, + journal = {Synthese}, + volume = {197}, + pages = {3779--3796}, + year = {2020}, + doi = {10.1007/s11229-018-01989-7}, + note = {Read in full (card: library/buijsman2020numerals.md). Open access. Reports Krajcsi and Szabo (2012) on sign-value versus place-value learning; Krajcsi and Szabo not read.} +} + +@book{chrisomalis2010numerical, + author = {Chrisomalis, Stephen}, + title = {Numerical Notation: A Comparative History}, + publisher = {Cambridge University Press}, + year = {2010}, + doi = {10.1017/CBO9780511676062}, + note = {Read in part, by Google Books search-inside snippets only (id ux--OWgWvBQC; pp. 21, 28--32, 116--117, 123, 125 and others). Every quotation is a fragment; check sense in the book.} +} + +@book{chrisomalis2020reckonings, + author = {Chrisomalis, Stephen}, + title = {Reckonings: Numerals, Cognition, and History}, + publisher = {MIT Press}, + year = {2020}, + doi = {10.7551/mitpress/13381.001.0001}, + note = {Read in part, by Google Books search-inside snippets only (id j14JEAAAQBAJ; pp. xii, xiii, 21--24, 75, 79, 108). MIT preview PDF bot-walled. Fragments only.} +} + +@article{overmann2022review, + author = {Overmann, Karenleigh A.}, + title = {Review of {S}tephen {C}hrisomalis, {R}eckonings: {N}umerals, {C}ognition, and {H}istory}, + journal = {Journal of Anthropological Research}, + volume = {78}, + number = {1}, + pages = {152--153}, + year = {2022}, + doi = {10.1086/717832}, + note = {Not accessible (record only; OSF preprint 10.31219/osf.io/3zwp4 did not render).} +} + +@article{nothaft2020satanic, + author = {Nothaft, C. Philipp E.}, + title = {Medieval {E}urope's satanic ciphers: On the genesis of a modern myth}, + journal = {British Journal for the History of Mathematics}, + volume = {35}, + number = {2}, + pages = {107--136}, + year = {2020}, + doi = {10.1080/26375451.2020.1726050}, + note = {Read in full (card: library/nothaft2020satanic.md): accepted manuscript from Oxford ORA (56 pages; reference list from p. 44 skimmed). Page numbers in F_parallel_moves.md are those of the manuscript.} +} + +@techreport{danna2022numerical, + author = {Danna, Raffaele and Iori, Martina and Mina, Andrea}, + title = {A Numerical Revolution: The Diffusion of Practical Mathematics and the Growth of Pre-Modern {E}uropean Economies}, + institution = {Laboratory of Economics and Management, Scuola Superiore Sant'Anna}, + type = {LEM Working Paper}, + number = {2022/18}, + year = {2022}, + url = {http://hdl.handle.net/10419/266088}, + note = {Read in part (card: library/danna2022numerical.md): open working paper, 50 pages; main text read in full, appendix tables and references skimmed. Not peer reviewed in this form. Regression identification not checked.} +} + +@article{danna2021figuring, + author = {Danna, Raffaele}, + title = {Figuring out: The spread of {H}indu-{A}rabic numerals in the {E}uropean tradition of practical mathematics (13th--16th centuries)}, + journal = {Nuncius}, + volume = {36}, + number = {1}, + pages = {5--48}, + year = {2021}, + doi = {10.1163/18253911-bja10004}, + note = {Abstract only (Brill PDF returned 403). Title as in the Cambridge working-paper version; Crossref gives the title as "Figuring Out".} +} + +@article{biggs2020decoding, + author = {Biggs, Norman}, + title = {Decoding chancery records from the 1240s}, + journal = {British Journal for the History of Mathematics}, + volume = {36}, + number = {1}, + pages = {1--12}, + year = {2020}, + doi = {10.1080/26375451.2020.1834796}, + note = {Read in full (card: library/biggs2020decoding.md; LSE accepted manuscript). Crossref lists the container as British Journal for the History of Mathematics; the accepted manuscript names BSHM Bulletin.} +} + +@incollection{burnett2024learning, + author = {Burnett, Charles}, + title = {Learning {I}ndian arithmetic in the early thirteenth century}, + booktitle = {Numerals and Arithmetic in the Middle Ages}, + chapter = {IX}, + publisher = {Routledge}, + year = {2024}, + doi = {10.4324/9781003554448-9}, + note = {Abstract only (Consensus). DOI resolved in Crossref; pages given there as IX-15 to IX-26.} +} + +@article{james2024otis, + author = {James, Kathryn}, + title = {Review of {J}essica {M}arie {O}tis, {B}y the {N}umbers: {N}umeracy, {R}eligion, and the {Q}uantitative {T}ransformation of {E}arly {M}odern {E}ngland}, + journal = {Journal of Religious History}, + volume = {49}, + number = {1}, + pages = {94--95}, + year = {2024}, + doi = {10.1111/1467-9809.13111}, + note = {Review text read in full via Consensus. Otis (OUP, 2024) not read; quotations from Otis (pp. 64, 65) are the reviewer's.} +} + +@book{cajori1928notations1, + author = {Cajori, Florian}, + title = {A History of Mathematical Notations, Vol. 1: Notations in Elementary Mathematics}, + publisher = {Open Court}, + address = {Chicago and London}, + year = {1928}, + url = {https://archive.org/details/historyofmathema031756mbp}, + note = {Read in part: archive.org OCR text searched by keyword (section 54 on Hindu-Arabic numerals and the Peruzzi bank; Vieta sections). Public domain.} +} + +@book{cajori1929notations2, + author = {Cajori, Florian}, + title = {A History of Mathematical Notations, Vol. 2: Notations Mainly in Higher Mathematics}, + publisher = {Open Court}, + address = {Chicago and London}, + year = {1929}, + url = {https://archive.org/details/historyofmathema027671mbp}, + note = {Read in part: archive.org OCR text searched by keyword (fluxion and differential sections, sections 581--588). Public domain.} +} + +% ---- Case 2: Staff notation ---- +@article{haines2008origins, + author = {Haines, John}, + title = {The origins of the musical staff}, + journal = {The Musical Quarterly}, + volume = {91}, + number = {3--4}, + pages = {327--378}, + year = {2008}, + doi = {10.1093/musqtl/gdp002}, + note = {Read in full (card: library/haines2008origins.md; copy at cartusiana.org; Advance Access 2009). The 'untenable' sentence is on p. 331 and 'some seventy manuscript sources' on p. 328 (AUDIT_02).} +} + +@incollection{strunk1998guido, + author = {Guido of Arezzo}, + title = {Prologue to his {A}ntiphoner; {E}pistola de ignoto cantu (excerpts)}, + booktitle = {Source Readings in Music History}, + editor = {Strunk, Oliver and Treitler, Leo}, + publisher = {W. W. Norton}, + address = {New York}, + year = {1998}, + url = {https://www.armusicanthology.com/media/PDFs/Workbook-Medieval-Part1.pdf}, + note = {Read in full (card: library/strunk1998guido.md), as reproduced in the A-R Music Anthology workbook (Medieval Period, Part 1), Supplemental Reading 1; translation not checked against Strunk and Treitler or Pesce 1999. Quoted briefly only.} +} + +@book{pesce1999guido, + author = {Pesce, Dolores}, + title = {Guido d'{A}rezzo's {R}egule rithmice, {P}rologus in antiphonarium, and {E}pistola ad {M}ichahelem: A Critical Text and Translation, with an Introduction, Annotations, Indices, and New Manuscript Inventories}, + series = {Musicological Studies}, + number = {73}, + publisher = {Institute of Mediaeval Music}, + address = {Ottawa}, + year = {1999}, + note = {Not accessible. No DOI found. Needed to settle what Guido claims about learning time. Cited in Haines 2008 (nn. 8, 18, 28).} +} + +@book{mengozzi2010renaissance, + author = {Mengozzi, Stefano}, + title = {The Renaissance Reform of Medieval Music Theory: Guido of Arezzo between Myth and History}, + publisher = {Cambridge University Press}, + year = {2010}, + note = {Not accessible (borrowable at archive.org, inside-search blocked). Table of contents seen via Consensus. No DOI found.} +} + +@article{mengozzi2007si, + author = {Mengozzi, Stefano}, + title = {`Si quis manus non habeat': Charting non-hexachordal musical practices in the age of solmisation}, + journal = {Early Music History}, + volume = {26}, + pages = {181--218}, + year = {2007}, + doi = {10.1017/s0261127907000228}, + note = {Abstract only.} +} + +@article{mews2024guidonian, + author = {Mews, Constant J.}, + title = {Implementing the {G}uidonian revolution: {J}ohn of {A}ffligem and the cause of reform 1030--1100}, + journal = {Viator}, + volume = {55}, + number = {1}, + pages = {311--341}, + year = {2024}, + doi = {10.1484/j.viator.5.144515}, + note = {Abstract only.} +} + +% ---- Case 3: Finding devices ---- +@incollection{rouse1982statim, + author = {Rouse, Richard H. and Rouse, Mary A.}, + title = {{Statim invenire}: Schools, preachers, and new attitudes to the page}, + booktitle = {Renaissance and Renewal in the Twelfth Century}, + editor = {Benson, Robert L. and Constable, Giles}, + publisher = {Harvard University Press}, + address = {Cambridge, MA}, + year = {1982}, + note = {Not accessible. Bibliographic details from memory and the brief; no database record or full text seen this session. Verify before use.} +} + +@book{blair2010too, + author = {Blair, Ann M.}, + title = {Too Much to Know: Managing Scholarly Information before the Modern Age}, + publisher = {Yale University Press}, + year = {2010}, + note = {Not accessible. Details from the brief; not verified this session.} +} + +@incollection{conraulewis2021index, + author = {Conrau-Lewis, Kyle}, + title = {The index in the premodern and modern world}, + booktitle = {Oxford Research Encyclopedia of Literature}, + publisher = {Oxford University Press}, + year = {2021}, + doi = {10.1093/acrefore/9780190201098.013.1224}, + note = {Abstract only (OUP page redirected; not fetched).} +} + +@article{abbamonte2023indexing, + author = {Abbamonte, Giancarlo and Kallendorf, Craig}, + title = {On indexing: The birth and early development of an idea}, + journal = {Journal of the History of Ideas}, + volume = {84}, + number = {3}, + pages = {465--486}, + year = {2023}, + doi = {10.1353/jhi.2023.a901490}, + note = {Abstract only.} +} + +@article{flanders2023place, + author = {Flanders, Judith S.}, + title = {A place for everything, and everything that came before: The development of alphabetical order}, + journal = {The Indexer}, + volume = {41}, + number = {3}, + pages = {223--234}, + year = {2023}, + doi = {10.3828/index.2023.29}, + note = {Abstract only.} +} + +@article{thomann2020serial, + author = {Thomann, Johannes}, + title = {From serial access to random access: Tables of contents, chapter headings and hierarchical text structures in fourth/tenth-century scientific books}, + journal = {Journal of Abbasid Studies}, + volume = {7}, + number = {2}, + pages = {207--228}, + year = {2020}, + doi = {10.1163/22142371-12340056}, + note = {Abstract only.} +} + +@article{dacosta2018tabulae, + author = {da Costa, Alexandra}, + title = {``That ye mowe redely fynde \ldots{} what ye desyre'': Printed tables of contents and indices, 1476--1550}, + journal = {Huntington Library Quarterly}, + volume = {81}, + number = {3}, + pages = {291--313}, + year = {2018}, + doi = {10.1353/hlq.2018.0013}, + note = {Read in full (card: library/dacosta2018tabulae.md; Cambridge repository accepted manuscript, which lacks the abstract). Crossref author field reads "Costa".} +} + +% ---- Case 4: Double entry ---- +@article{sangster2024emergence, + author = {Sangster, Alan}, + title = {The emergence of double entry bookkeeping}, + journal = {The Economic History Review}, + volume = {78}, + number = {2}, + pages = {499--528}, + year = {2024}, + doi = {10.1111/ehr.13358}, + note = {Read in full (card: library/sangster2024emergence.md; Aberdeen repository copy). Online 2024-05-23; print date is May 2025 (vol. 78(2)), so cite as Sangster (2025 [online 2024]).} +} + +@article{sangster2016genesis, + author = {Sangster, Alan}, + title = {The genesis of double entry bookkeeping}, + journal = {The Accounting Review}, + volume = {91}, + number = {1}, + pages = {299--315}, + year = {2016}, + doi = {10.2308/accr-51115}, + note = {Record only; not read. Crossref issued date is 2015 (online first); volume year 2016.} +} + +@article{sangster2025reboot, + author = {Sangster, Alan}, + title = {Time to reboot accounting history -- evidence from scholarship on double entry}, + journal = {Accounting History}, + volume = {30}, + number = {3}, + pages = {368--391}, + year = {2025}, + doi = {10.1177/10323732251357606}, + note = {Read in full (card: library/sangster2025reboot.md).} +} + +@article{yamey1949scientific, + author = {Yamey, B. S.}, + title = {Scientific bookkeeping and the rise of capitalism}, + journal = {The Economic History Review}, + volume = {1}, + number = {2/3}, + pages = {99}, + year = {1949}, + doi = {10.2307/2589824}, + note = {Record only; not read. Crossref lists first page 99 only; end page not found.} +} + +@article{yamey1964accounting, + author = {Yamey, Basil S.}, + title = {Accounting and the rise of capitalism: Further notes on a theme by {S}ombart}, + journal = {Journal of Accounting Research}, + volume = {2}, + number = {2}, + pages = {117}, + year = {1964}, + doi = {10.2307/2489995}, + note = {Abstract only. Crossref lists first page 117 only; end page not found.} +} + +@article{yamey2005historical, + author = {Yamey, Basil S.}, + title = {The historical significance of double-entry bookkeeping: Some non-{S}ombartian claims}, + journal = {Accounting, Business \& Financial History}, + volume = {15}, + number = {1}, + pages = {77--88}, + year = {2005}, + doi = {10.1080/09585200500033089}, + note = {Abstract only. Consensus names the journal Accounting History Review; Crossref gives Accounting, Business & Financial History.} +} + +@article{ryan2014historical, + author = {Ryan, Joseph}, + title = {Historical note: Did double-entry bookkeeping contribute to economic development, specifically the introduction of capitalism?}, + journal = {Australasian Accounting, Business and Finance Journal}, + volume = {8}, + number = {3}, + pages = {85--97}, + year = {2014}, + doi = {10.14453/aabfj.v8i3.8}, + note = {Read in full (card: library/ryan2014historical.md); earlier PDF download returned an empty file.} +} + +@article{carruthers1991accounting, + author = {Carruthers, Bruce G. and Espeland, Wendy Nelson}, + title = {Accounting for rationality: Double-entry bookkeeping and the rhetoric of economic rationality}, + journal = {American Journal of Sociology}, + volume = {97}, + number = {1}, + pages = {31--69}, + year = {1991}, + doi = {10.1086/229739}, + note = {Abstract only.} +} + +% ---- Case 7: Computing ---- +@article{shneiderman1983direct, + author = {Shneiderman, Ben}, + title = {Direct manipulation: A step beyond programming languages}, + journal = {Computer}, + volume = {16}, + number = {8}, + pages = {57--69}, + year = {1983}, + doi = {10.1109/mc.1983.1654471}, + note = {Read in full (card: library/shneiderman1983direct.md). Reused key. This paper reports no experiment: the N = 30 experiment is Margono and Shneiderman 1987 (key margono1987study).} +} + +@article{hutchins1985direct, + author = {Hutchins, Edwin L. and Hollan, James D. and Norman, Donald A.}, + title = {Direct manipulation interfaces}, + journal = {Human--Computer Interaction}, + volume = {1}, + number = {4}, + pages = {311--338}, + year = {1985}, + doi = {10.1207/s15327051hci0104_2}, + note = {Not accessible (no open copy). Reused key.} +} + +@techreport{margono1987study, + author = {Margono, Sepeedeh and Shneiderman, Ben}, + title = {A study of file manipulation by novices using commands vs. direct manipulation}, + institution = {University of Maryland, Department of Computer Science}, + type = {Technical Report}, + number = {CAR-TR-264}, + year = {1987}, + url = {https://www.cs.umd.edu/users/ben/papers/Margono1987Study.pdf}, + note = {Read in full (card: library/margono1987study.md). Report number and June 1987 date from a search-result record, not from the PDF text.} +} + +@article{power2000history, + author = {Power, Daniel J.}, + title = {A history of microcomputer spreadsheets}, + journal = {Communications of the Association for Information Systems}, + volume = {4}, + number = {9}, + year = {2000}, + doi = {10.17705/1cais.00409}, + note = {Not accessible (403). Article number 9 from the AIS URL (vol4/iss1/9).} +} + +% ---- Case 8: Writing systems ---- +@article{walker1993early, + author = {Walker, Willard and Sarbaugh, James}, + title = {The early history of the {C}herokee syllabary}, + journal = {Ethnohistory}, + volume = {40}, + number = {1}, + pages = {70}, + year = {1993}, + doi = {10.2307/482159}, + note = {Abstract only (Consensus excerpt). Crossref lists first page 70 only; the full range is not confirmed. The abstract says the syllabary was disseminated without the aid of Euroamerican institutions and that most missionary groups adopted it by 1826 (AUDIT_02).} +} + +@article{cushman2011syllabary, + author = {Cushman, Ellen}, + title = {The {C}herokee syllabary: A writing system in its own right}, + journal = {Written Communication}, + volume = {28}, + number = {3}, + pages = {255--281}, + year = {2011}, + doi = {10.1177/0741088311410172}, + note = {Abstract only.} +} + +@article{montgomeryanderson2014review, + author = {Montgomery-Anderson, Brad}, + title = {Review of {C}ushman, {T}he {C}herokee syllabary: {W}riting the people's perseverance}, + journal = {Language}, + volume = {90}, + number = {2}, + pages = {537--540}, + year = {2014}, + doi = {10.1353/lan.2014.0030}, + note = {Review text read via Consensus (excerpt up to the first chapters). Cushman's book itself not read.} +} + +@book{parins2013literacy, + author = {Parins, James W.}, + title = {Literacy and Intellectual Life in the {C}herokee {N}ation, 1820--1906}, + publisher = {University of Oklahoma Press}, + year = {2013}, + note = {Not accessible. Quoted figures come from B. B. Schweiger's review in the Arkansas Historical Quarterly (2014), read via Consensus; the Consensus record carries a mis-assigned Choice DOI (10.5860/choice.51-5784), so no DOI is given for the review.} +} + +@article{unseth2016international, + author = {Unseth, Peter}, + title = {The international impact of {S}equoyah's {C}herokee syllabary}, + journal = {Written Language \& Literacy}, + volume = {19}, + number = {1}, + pages = {75--93}, + year = {2016}, + doi = {10.1075/wll.19.1.03uns}, + note = {Abstract only.} +} + +@book{bender2002signs, + author = {Bender, Margaret}, + title = {Signs of {C}herokee Culture: {S}equoyah's Syllabary in {E}astern {C}herokee Life}, + publisher = {University of North Carolina Press}, + year = {2002}, + note = {Not accessible. Publisher blurb read via a Consensus record (Choice review DOI 10.5860/choice.40-3482, not a DOI for the book). Publication year 2002 taken from Cushman and Montgomery-Anderson's citation of Bender (2002a,b); not confirmed against a catalogue.} +} + +@article{white1962revival, + author = {White, John}, + title = {On the revival of printing in the {C}herokee language}, + journal = {Current Anthropology}, + volume = {3}, + number = {5}, + pages = {511--514}, + year = {1962}, + doi = {10.1086/200321}, + note = {Opening paragraphs read via Consensus. Cited only as an example of an uncited popular claim ("Overnight the whole nation became a school").} +} + +@misc{handel2016learn, + author = {Handel, Zev}, + title = {Learn to read {K}orean: An introduction to the {H}angul alphabet}, + year = {2016}, + note = {Lecture abstract seen via Consensus; abstract only. No DOI (the one attached in Consensus belongs to an unrelated conference). Source and wording of the "morning ... ten days" quotation unverified.} +} + +@misc{hangeulmuseum, + author = {{National Hangeul Museum}}, + title = {Read society through its {H}angeul letters}, + howpublished = {Exhibition web page}, + url = {https://www.hangeul.go.kr/exhibition/485}, + note = {Read in full (short page; card: library/hangeulmuseum.md). Accessed 2026-10-12.} +} + +@article{baek2024spread, + author = {Baek, Nak-cheon}, + title = {On the use and spread of {H}angeul in the {J}oseon dynasty}, + journal = {Liberal Arts Innovation Center}, + volume = {14}, + pages = {449--476}, + year = {2024}, + doi = {10.54698/kl.2024.14.449}, + note = {Abstract only. Crossref lists the first author given name as missing; surname Baek.} +} + +@article{ivanova2026process, + author = {Ivanova, Yanitsa}, + title = {A process of implementing the {K}orean alphabet on the {K}orean peninsula}, + journal = {Ezikov Svyat (Orbis Linguarum)}, + volume = {24}, + number = {1}, + pages = {136--151}, + year = {2026}, + doi = {10.37708/ezs.swu.bg.v24i1.9}, + note = {Read in full (card: library/ivanova2026process.md); Bulgarian body.} +} + +@misc{park2025dokkyo, + author = {Park, Jong-Bae}, + title = {Diversification of literacy patterns by social class in late {J}oseon (forum abstract)}, + howpublished = {D{\=o}kky{\=o} University International Forum web page}, + url = {https://www.dokkyo.ac.jp/international/36-10.html}, + note = {Forum abstract, confirmed by the card (library/park2025dokkyo.md) to contain no counts or rates. Abstract read via WebFetch (Japanese); the year (2025) is inferred from a navigation link and is unconfirmed.} +} + +% ---- Case 9: Counter-cases ---- +@article{liebowitz1990fable, + author = {Liebowitz, S. J. and Margolis, Stephen E.}, + title = {The fable of the keys}, + journal = {The Journal of Law and Economics}, + volume = {33}, + number = {1}, + pages = {1--25}, + year = {1990}, + doi = {10.1086/467198}, + note = {Abstract only (SSRN copy bot-walled).} +} + +@article{norman1982alphabetic, + author = {Norman, Donald A. and Fisher, Dennis}, + title = {Why alphabetic keyboards are not easy to use: Keyboard layout doesn't much matter}, + journal = {Human Factors}, + volume = {24}, + number = {5}, + pages = {509--519}, + year = {1982}, + doi = {10.1177/001872088202400502}, + note = {Abstract only.} +} + +@article{kay2013rerun, + author = {Kay, Neil M.}, + title = {Rerun the tape of history and {QWERTY} always wins}, + journal = {Research Policy}, + volume = {42}, + number = {6--7}, + pages = {1175--1185}, + year = {2013}, + doi = {10.1016/j.respol.2013.03.007}, + note = {Abstract only.} +} + +@article{hossain2009quest, + author = {Hossain, Tanjim and Morgan, John}, + title = {The quest for {QWERTY}}, + journal = {American Economic Review}, + volume = {99}, + number = {2}, + pages = {435--440}, + year = {2009}, + doi = {10.1257/aer.99.2.435}, + note = {Abstract only. Cited for its summary of David 1985 and Liebowitz and Margolis 1990.} +} + +@article{vera2009decimal, + author = {Vera, H{\'e}ctor}, + title = {Decimal time: Misadventures of a revolutionary idea, 1793--2008}, + journal = {KronoScope}, + volume = {9}, + number = {1--2}, + pages = {29--48}, + year = {2009}, + doi = {10.1163/156771509x12638154745382}, + note = {Abstract only.} +} + +@article{zerubavel1977republican, + author = {Zerubavel, Eviatar}, + title = {The {F}rench {R}epublican calendar: A case study in the sociology of time}, + journal = {American Sociological Review}, + volume = {42}, + number = {6}, + pages = {868}, + year = {1977}, + doi = {10.2307/2094573}, + note = {Abstract only. Crossref lists first page 868 only; end page not found.} +} + +@article{buncic2017factors, + author = {Bun{\v c}i{\'c}, Daniel}, + title = {Factors influencing the success and failure of writing reforms}, + journal = {Studi Slavistici}, + pages = {21--46}, + year = {2017}, + doi = {10.13128/studi_slavis-21937}, + note = {Read in full (card: library/buncic2017factors.md).} +} + +% ---- Numeracy-format set ---- +@article{peters2006numeracy, + author = {Peters, Ellen and V{\"a}stfj{\"a}ll, Daniel and Slovic, Paul and Mertz, C. K. and Mazzocco, Ketti and Dickert, Stephan}, + title = {Numeracy and decision making}, + journal = {Psychological Science}, + volume = {17}, + number = {5}, + pages = {407--413}, + year = {2006}, + doi = {10.1111/j.1467-9280.2006.01720.x}, + note = {Abstract only (Europe PMC abstract). Reused key.} +} + +@article{peters2007less, + author = {Peters, Ellen and Dieckmann, Nathan and Dixon, Anna and Hibbard, Judith H. and Mertz, C. K.}, + title = {Less is more in presenting quality information to consumers}, + journal = {Medical Care Research and Review}, + volume = {64}, + number = {2}, + pages = {169--190}, + year = {2007}, + doi = {10.1177/10775587070640020301}, + note = {Abstract only (Europe PMC).} +} + +@article{hibbard2003supporting, + author = {Hibbard, Judith H. and Peters, Ellen}, + title = {Supporting informed consumer health care decisions: Data presentation approaches that facilitate the use of information in choice}, + journal = {Annual Review of Public Health}, + volume = {24}, + pages = {413--433}, + year = {2003}, + doi = {10.1146/annurev.publhealth.24.100901.141005}, + note = {Abstract only (Europe PMC).} +} + +@article{peters2007numeracy, + author = {Peters, Ellen and Hibbard, Judith and Slovic, Paul and Dieckmann, Nathan}, + title = {Numeracy skill and the communication, comprehension, and use of risk-benefit information}, + journal = {Health Affairs}, + volume = {26}, + number = {3}, + pages = {741--748}, + year = {2007}, + doi = {10.1377/hlthaff.26.3.741}, + note = {Abstract only (Europe PMC). Crossref lists four authors in the order shown.} +} + +@article{gigerenzer1995improve, + author = {Gigerenzer, Gerd and Hoffrage, Ulrich}, + title = {How to improve {B}ayesian reasoning without instruction: Frequency formats}, + journal = {Psychological Review}, + volume = {102}, + number = {4}, + pages = {684--704}, + year = {1995}, + doi = {10.1037/0033-295x.102.4.684}, + note = {Record only; abstract not retrieved (MPG copy returned 403).} +} + +@article{garciaretamero2013visual, + author = {Garcia-Retamero, Rocio and Cokely, Edward T.}, + title = {Communicating health risks with visual aids}, + journal = {Current Directions in Psychological Science}, + volume = {22}, + number = {5}, + pages = {392--399}, + year = {2013}, + doi = {10.1177/0963721413491570}, + note = {Record only; not read, no abstract retrieved.} +} + +@article{larkin1987why, + author = {Larkin, Jill H. and Simon, Herbert A.}, + title = {Why a diagram is (sometimes) worth ten thousand words}, + journal = {Cognitive Science}, + volume = {11}, + number = {1}, + pages = {65--100}, + year = {1987}, + doi = {10.1111/j.1551-6708.1987.tb00863.x}, + note = {Not read; cited via Schlimm and Neth 2008 (p. 2098). Reused key.} +} + +@article{david1985clio, + author = {David, Paul A.}, + title = {Clio and the economics of {QWERTY}}, + journal = {American Economic Review}, + year = {1985}, + note = {Not accessible; not in Crossref under the DOI guessed (10.1257/aer.p); volume, issue and pages deliberately omitted (not verified); the title and year come from the Hossain and Morgan abstract and the brief. No DOI found.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_F_parallel_moves.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_F_parallel_moves.md new file mode 100644 index 00000000..afa9fab2 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_F_parallel_moves.md @@ -0,0 +1,376 @@ +# F — Parallel moves: other single conventions credited with opening a competence + +*Corrected on 2026-10-12 against `../audit/AUDIT_02.md`: 22 factual fixes.* + +*Facet F of the foundation paper's literature run. Agent-drafted on 2026-10-12. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** Cases outside the history of reading where one representational or interface convention is credited with widening who can do something: Hindu-Arabic place value (with the 1299 Florence "ban" and the Roman-numeral dispute), the musical staff and Guido of Arezzo, finding devices in the book (alphabetical order, index, chapter division), double-entry bookkeeping, algebraic symbolism and calculus notation, direct manipulation and the spreadsheet, the Cherokee syllabary and Hangul, and counter-cases that were clear improvements and did not spread (Dvorak, spelling reform, decimal time). I add one set the brief did not name: format interventions inside the numeracy literature itself (Peters, Hibbard, Gigerenzer), because the author's standing model for algorithmacy is numeracy. + +**The question the synthesis answers.** Is "one move on the page widened a competence" a recurring, well-evidenced pattern, or a retrospective selection? Which parallels help the paper, which embarrass it, and what do they imply for H-cost, H-cause and problem 4 ("one similar move may be the wrong shape")? + +**Legend for read status.** +- *Read in full*: I read the whole text this session. +- *Read in part*: I read the text but not all of it, or I read it by keyword search and say where. For Chrisomalis (2010, 2020) the only access was Google Books search-inside snippets of about 40 words, so every Chrisomalis quotation is a fragment and its sense needs checking in context. +- *Abstract only*: database abstract or record, no body text. +- *Not accessible*: no text and no abstract. + +Quotations carry a locator. "OCR §" means a section number in an archive.org OCR text. Locators for Guido are the part of the text (Prologue or *Epistola* excerpt), because the workbook's page breaks split sentences and I could not map them to Strunk and Treitler's pagination. + +--- + +## (a) Search log + +All searches ran on 2026-10-12. Downloaded files sit under `~/iit-playground/dxd_foundation_local/pdfs/F/`. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | repo | Read `RESEARCH_PLAN.md`, `literature/REUSE.md`; grep existing `.bib` keys | n/a | Reused keys: `hutchins1985direct`, `larkin1987why`, `peters2006numeracy`, `shneiderman1983direct`, `zhang1995representational`. | +| 2 | Crossref `query.bibliographic` | About 45 title/author queries (Zhang and Norman; Schlimm and Neth; Yamey; Carruthers and Espeland; Hutchins et al.; Shneiderman; Liebowitz and Margolis; Walker and Sarbaugh; Peters et al.; Chrisomalis; Haines; Gigerenzer and Hoffrage; Hibbard and Peters; others) | target found for most | Schlimm and Neth 2008 is **not** in Crossref (CogSci proceedings). Rouse and Rouse, David 1985 (AER 75), Pesce 1999, Strunk and Treitler not found. | +| 3 | Crossref `/works/{doi}` | 60 DOIs checked for title, authors, container, volume, pages | 60/60 resolved | Zhang and Norman pages are 271–295 (a web search returned "217–295", which is wrong). Walker and Sarbaugh: Crossref lists only first page 70. The DOI attached to Handel's lecture abstract in Consensus (`10.18429/jacow-ipac2016-thea01`) belongs to an accelerator conference; I did not use it. | +| 4 | Unpaywall | About 45 DOIs | OA copy for 10 | Closed or no URL: Zhang and Norman, Chrisomalis chapter, Carruthers and Espeland, Hutchins et al., Shneiderman, Walker and Sarbaugh, Nothaft (publisher copy), Yamey 2005, Cushman. | +| 5 | Scholar Gateway `semanticSearch` | `Hindu-Arabic numerals diffusion medieval Europe abacus schools resistance` | 0 | `ACCESS_DENIED`. Not retried. | +| 6 | Consensus | Nine queries: Florence 1299 ban; Cherokee syllabary spread; Hangul adoption among commoners; Guido staff and training; finding devices; double entry and Sombart/Yamey; decimal time; Dvorak/QWERTY; spelling reform and Shavian | 20 each (19 for double entry) | Source of most abstracts below. Outputs for Guido and finding devices were saved to the tool-results folder and read as titles and abstracts only. | +| 7 | WebSearch | Zhang and Norman PDF; Schlimm and Neth PDF; Guido "ten years"; Guido Prologue wording (two queries); 1299 statute (two); Chrisomalis efficiency (two); Cherokee literacy figures; Hunmin jeongeum postface wording; Hangul uptake/1945 literacy; Margono and Shneiderman; Liebowitz and Margolis/Navy study | 8–10 each | Found the Göttingen copy of Schlimm and Neth and the Cartusiana copy of Haines. **Failed** to surface Guido's "ten years" statement, any primary text of the 1299 statute, or Ledyard's Hangul postface translation. Treat web-search summaries as leads only; I used none as evidence except where a fetched page confirmed them. | +| 8 | curl downloads (OA) | Schlimm and Neth (Göttingen); Nothaft accepted manuscript (Oxford ORA); Danna, Iori and Mina LEM working paper (EconStor); Haines (cartusiana.org); Margono and Shneiderman (UMD); Shneiderman 1983 (UMD); Biggs 2020 (LSE); Buijsman 2020 (Springer OA); Cajori vols 1–2 (archive.org djvu text); A-R Music Anthology workbook PDF | 10 obtained | Read as described in the entries. | +| 9 | curl downloads (blocked or empty) | Liebowitz and Margolis (SSRN); Chrisomalis *Reckonings* preview (MIT); Danna 2021 (Brill); Sangster 2024 (Wiley); Ryan 2014 (UOW); Power 2000 (AIS, 403 by WebFetch); Hutchins et al. (two author-site guesses); Gigerenzer and Hoffrage (MPG, 403); Danna 2023 chapter and Sebregondi chapter (OAPEN, HTML returned); Oxford Research Encyclopedia article (OUP redirect); utdallas.edu Liebowitz page (dead, redirects to a portal notice) | 0 text | Cloudflare, 403 or HTML in place of a PDF. I did not try to get round any block. | +| 10 | archive.org | Advanced search for Guido editions; borrowable *Hucbald, Guido, and John on Music* and Mengozzi (`hucbaldguidojohn0000unse`, `renaissancerefor0000meng`) and full-text-inside requests; Gerbert's *Scriptores* (two items) | items found | Both borrowable books are access-restricted ("Item not available" to the inside-search endpoint). The two Gerbert items I found are volumes that do not contain Guido's *Epistola* (no hits for *Michaheli*, *ignoto cantu*, *antiphonarii*). | +| 11 | Google Books search-inside (`jscmd=SearchWithinVolume2`) | Chrisomalis 2010 (`ux--OWgWvBQC`): efficient, forgery, incomplete, mousetrap, Arte del Cambio, others. Chrisomalis 2020 (`j14JEAAAQBAJ`): numeracy, going hand in hand, cognitive technology, efficiency, others. Rouse and Rouse (`uK6JVxh92QUC`): alphabetical, statim invenire | 9 + 7 + 10 snippets for the first two; 0 for the third | `uK6JVxh92QUC` turned out to be a different book (Hufton on the French Revolution), so Rouse and Rouse remained unreached. Google's REST API returned a daily-quota error. Hyphenated or multi-word queries often returned zero; a zero is not evidence of absence. | +| 12 | Europe PMC REST | Abstracts by DOI for Peters et al. 2006, Hibbard and Peters 2003, Peters et al. 2007 (two), Garcia-Retamero and Cokely, Gigerenzer and Hoffrage | 4 of 6 had abstracts | Gigerenzer and Hoffrage and Garcia-Retamero and Cokely: no abstract retrieved. | +| 13 | WebFetch | Hangeul Museum page; Dokkyo University abstract page; OSF page for Overmann's review | 2 of 3 usable | OSF page is JavaScript-rendered and returned only "OSF". | +| 14 | grep over downloaded OCR/text | Nothaft, Danna et al., Cajori, Haines, Margono and Shneiderman, Shneiderman, workbook | n/a | Passages located by keyword, then read in context. | + +--- + +## (b) Entries + +Order follows the brief: numerals (1), staff (2), finding devices (3), double entry (4), algebra and calculus (5), computing (7), writing systems (8), counter-cases (9), then the numeracy-format set. Cases 5 (maps, coordinate grids) and 6 (logarithm table, slide rule) were not researched beyond the notation entries below; see section (d). + +### Case 1. Hindu-Arabic numerals, place value and zero + +**zhang1995representational** — Zhang, J. and Norman, D. A. (1995). A representational analysis of numeration systems. *Cognition*, 57(3), 271–295. https://doi.org/10.1016/0010-0277(95)00674-3 +- *Verification:* Crossref record resolved (title, authors, volume, pages match). +- *Read status:* **not accessible.** No open copy found; the article is not in Unpaywall. I know it only through three readers: Schlimm and Neth (2008), Buijsman (2020) and Chrisomalis (2020), all below. +- *What it is cited for:* a framework in which numeration systems are representations whose structure affects how easily a task is performed, so the same arithmetic costs different amounts in different notations. Schlimm and Neth (p. 2097) place it with Larkin and Simon (1987) and Norman (1993). Chrisomalis (2020, p. 22 snippet) criticises it: "numeracy (e.g., Zhang and Norman 1995) under the guise of the assumption that it is reasonable to compare numerical notations on the basis of their efficiency for written computations." +- *Bearing:* supplies the strongest theory for H-cost (work moves to the representation). I could not check its own claims, so the paper should not cite specific findings from it until a human with library access reads it. + +**schlimm2008modeling** — Schlimm, D. and Neth, H. (2008). Modeling ancient and modern arithmetic practices: Addition and multiplication with Arabic and Roman numerals. In V. Sloutsky, B. Love and K. McRae (Eds.), *Proceedings of the 30th Annual Meeting of the Cognitive Science Society* (pp. 2097–2102). Cognitive Science Society. https://www.psych.uni-goettingen.de/de/cognition/publikationen-dateien-neth-3/2008_schlimmneth_modeling-ancient-and-modern-arithmetic-practices.pdf/@@download/file/2008_SchlimmNeth_Modeling%20ancient%20and%20modern%20arithmetic%20practices.pdf +- *Verification:* author copy downloaded; header carries the citation; not in Crossref. +- *Read status:* **read in full** (6 pages; card: library/schlimm2008modeling.md, no contradiction). +- *What it argues:* computational models (four agents, not human participants) show that addition and multiplication are possible with Roman numerals. Abstract: Roman numerals "are not only informationally equivalent to Arabic ones but also computationally similar—a claim that is widely disputed" (p. 2097). Conclusion (p. 2102): "the Achilles heel of Roman numerals is not that they are (qualitatively) unsuitable for algorithmic usage, but their greater cost in terms of the number of necessary elementary information processes, especially those using external resources." Roman addition needs no memorised addition facts (p. 2101) but many more writing and perceptual-motor steps; the Arabic agent's advantage depends on the problem being laid out in right-justified columns ("the precise format of external inputs", p. 2101). +- *Limit:* simulation of 1,000,000 additions and 10,000 multiplications of numbers up to 100, no human data (p. 2098 explains why). The authors say the models would change with an abacus. +- *Bearing:* complicates H-cost. "Cheaper" depends on which resource you count (memory, motor steps, columnar layout). It also supports the point that the place-value advantage is partly an advantage of the page layout, which is a design move of the right kind for the paper. + +**buijsman2020numerals** — Buijsman, S. (2020). How numerals support new cognitive capacities. *Synthese*, 197, 3779–3796. https://doi.org/10.1007/s11229-018-01989-7 +- *Verification:* Crossref; open access (PDF downloaded). +- *Read status:* read in full (card: library/buijsman2020numerals.md). +- *What it says:* reports Krajcsi and Szabó (2012): when taught a completely new sign-value system and a new place-value system, "performance on simple addition and comparison tasks is better for the sign-value system", so "sign-value systems are easier to learn and work with for those kinds of tasks" (pp. 3787–3788). I did not read Krajcsi and Szabó. +- *Bearing:* complicates H-cost. For novices the simpler, older notation can be easier. Buijsman's text states neither a practice effect nor a column-layout effect for place value; the column point comes only from Menary's multiplication layouts (pp. 3786–3787), without a cost figure (card). The Krajcsi and Szabó sentence runs from the foot of p. 3787 to the top of p. 3788. + +**chrisomalis2010numerical** — Chrisomalis, S. (2010). *Numerical Notation: A Comparative History*. Cambridge University Press. https://doi.org/10.1017/CBO9780511676062 +- *Verification:* Crossref record. +- *Read status:* **read in part, by Google Books snippets only** (pp. 21, 28–32, 116–117, 123, 125, 131, 173, 194–195, 272, 277, 350). Not the body text. +- *What the fragments show:* a section titled "Technology, Function, and Efficiency" (p. 29) begins from the historian's habit of ranking notations. P. 32 (fragments): efficiency "for computation is a Western-centered and historically inaccurate benchmark for comparing"; a notation needs "not only … a better mousetrap, but also that an entire social group learn it before its usefulness will be evident. A system with many users is functional for that reason alone". P. 116: "efficiency of the Roman numerals is a common subject in the history of mathematics. The consensus of these …". P. 123: "Yet the debate was not only about efficiency. In 1299, the Arte del Cambio or moneychangers' guild of Florence prohibited the use of Western numerals in its …". P. 125: the view that Roman numerals were worse for computation "is correct but incomplete. The Roman numerals were used alongside the Greek alphabetic numerals for well over a thousand years". +- *Bearing:* refutes the cleanest version of "better notation wins" and supplies the social-adoption alternative (network effects). Chrisomalis (2010, p. 123 snippet) still treats the 1299 prohibition as a fact; Nothaft (2020) below shows what it covered. + +**chrisomalis2020reckonings** — Chrisomalis, S. (2020). *Reckonings: Numerals, Cognition, and History*. MIT Press. https://doi.org/10.7551/mitpress/13381.001.0001 +- *Verification:* Crossref record. +- *Read status:* **read in part, snippets only** (pp. xii, xiii, 21–24, 75, 79, 108). MIT preview PDF was bot-walled. Overmann's review (see `overmann2022review`) was not read. +- *What the fragments show:* p. xii: "numeracy is not preordained". P. 22: the criticism of Zhang and Norman quoted above. P. 24: "cognitive effects from written number notations is not to say that there are no such effects" (fragment; sense incomplete). P. 75: "the East Asian evidence rebuts the idea that an integrated computational-representational system should presumptively be considered superior to a division of cognitive labor." P. 79 on abacists and algorists: "there were proponents of each position, with texts emphasizing one over the other. But few if any of these texts cast aspersions on the other method." P. 108: literacy and numeracy "going hand in hand, linked to formal schooling … but this is a modern idea, a product of a particular historical moment." +- *Bearing:* the sceptic's position in the author's own field. Two consequences for the paper: do not assert that numeracy and literacy travel together, and do not rank notations by task efficiency. See synthesis (iv). + +**nothaft2020satanic** — Nothaft, C. P. E. (2020). Medieval Europe's satanic ciphers: On the genesis of a modern myth. *British Journal for the History of Mathematics*, 35(2), 107–136. https://doi.org/10.1080/26375451.2020.1726050 +- *Verification:* Crossref; accepted manuscript from Oxford ORA (CC BY-NC-ND). Page numbers below are those of the 56-page accepted manuscript, not the journal. +- *Read status:* read in full (card: library/nothaft2020satanic.md; the reference list from p. 44 skimmed). +- *What it argues:* the story that zero or Hindu-Arabic numerals were banned as diabolical is "false or unsubstantiated at nearly every level of analysis" (abstract). On Florence: the provision is Article 102 of the statutes of the Arte del Cambio, edited by Camerani Marri (1955, 72–73), headed *Quod nullus de arte scribat in suo libro per abbacum*. It bars members from writing, or letting others write, credits and debits in their account books "by means of or in the letter of the abacus, but let him write it openly and in full by way of letters" (*aliquod quod per modum vel licteram abbachi intelligatur, sed aperte et extense scribat per licteram*; pp. 14–15, translation from Struik 1968). "The reasons behind this prohibition are a matter of speculation." Nothaft lists fraud-prevention as a "popular" explanation but does not endorse it, and "no such comment appears in the source text" (p. 15, n. 22, on Gazalé). The same article reappears in the revisions of 1300, 1313 and 1316 (n. 21). Other provisions: Padua statutes of 1331 (not 1348; the substance of the "1348" rule is found in 1331) required booksellers to display prices in clear letters rather than ciphers, a rule about display to buyers, not about the booksellers' own accounts (pp. 16–17); Trieste 1350 barred money changers and merchants from keeping books *per abachum*; Frankfurt 1494 told clerks to stop using the numerals in public books (p. 17). "Neither of these examples documents a blanket 'ban' of the sort imagined by some modern authors" (p. 17). The popular "non per cifras sed per literas claras" wording appears in Padua's 1465 statutes, and "the year 1348 … may have been a mistake on Ball's part" (p. 16). +- *Bearing:* **contradicts** the often-repeated 1299 story as evidence of church or civic resistance to a better notation. What survives is a narrow guild rule about account books, with motives unknown. Note also what the rule shows: by 1299 the numerals were common enough among Florentine money changers for the guild to regulate them. The paper should drop the ban or cite it with Nothaft's wording. + +**danna2022numerical** — Danna, R., Iori, M. and Mina, A. (2022). *A numerical revolution: The diffusion of practical mathematics and the growth of pre-modern European economies* (LEM Working Paper 2022/18). Scuola Superiore Sant'Anna. http://hdl.handle.net/10419/266088 +- *Verification:* EconStor record; PDF downloaded. Working paper, not peer reviewed in this form. +- *Read status:* read in part (card: library/danna2022numerical.md; main text read in full, appendix tables A1–A9 and reference list skimmed). +- *What it argues:* abstract: "This numerical revolution laid the foundations for the commercial revolution of the 13th century, and the diffusion of knowledge through organised learning had positive and significant effects on the growth of pre-modern European economies." Evidence: a dataset of "over 1050 manuals" (abstract; the body gives over 1,290 in the full dataset, p. 9, and 1,054 in the 815-city growth sample, p. 13) of practical arithmetic and city-level growth regressions. Fibonacci's *Liber abaci* (first completed 1202) is "not … a text exclusively addressed to merchants" (p. 5). "While there is no direct evidence of the first adoption of Hindu-Arabic numerals for economic purposes, they surely had been adopted by the end of the 13th century" (p. 5). The carrier was the abacus school: "lay and vernacular schools where pupils between the age of ten and thirteen (the curriculum lasted on average two years) were trained to use Hindu-Arabic numerals" (p. 6). Fees were "low enough to make it possible also for children of small shopkeepers to afford" it (p. 6). The manuals "only present calculations with Hindu-Arabic numerals" (n. 3). +- *Bearing:* the best available evidence that the numerals mattered, and in the same paper the evidence that they travelled with a schooling system, a vernacular text genre and a commercial demand. The measured unit is the manual (and so the school), not the numerals alone, which is why I read it as supporting "bundle" (Danna et al. do not use the word; they say "organised learning"). Their identification (distance-based instrument) is theirs; I have not checked it. Main limitation (card): city growth proxies income and manuals proxy schools and skills, the instrument's exclusion restriction is argued but untestable, and nothing here concerns a notation's effect on how fast an individual learns. + +**danna2021figuring** — Danna, R. (2021). Figuring out: The spread of Hindu-Arabic numerals in the European tradition of practical mathematics (13th–16th centuries). *Nuncius*, 36(1), 5–48. https://doi.org/10.1163/18253911-bja10004 +- *Read status:* abstract only (Brill PDF blocked). Abstract: "It can be argued that this is the tradition which drove the adoption of Hindu-Arabic numerals in Europe." +- *Bearing:* same as above, from the sole author; the 2023 book chapter version (OAPEN) was also not retrievable. + +**biggs2020decoding** — Biggs, N. (2020). Decoding chancery records from the 1240s. *British Journal for the History of Mathematics*, 36(1), 1–12. https://doi.org/10.1080/26375451.2020.1834796 +- *Read status:* read in full (card: library/biggs2020decoding.md; LSE accepted manuscript, which names *BSHM Bulletin* as the journal). +- *What it says:* "The use of Roman numerals in official records continued long after the underlying calculations were being done with the Hindu-Arabic numerals and algorithms. The people who did the calculations tended to keep their methods secret, and their expertise was not recognised publicly at the time" (p. 3). The paper describes Fibonacci's book as "the book that sparked the adoption of Hindu-Arabic algorithms throughout Western Europe". +- *Bearing:* complicates diffusion, but the lag is Biggs's premise (introduction, p. 3), not a finding of the records: in §4 he says no firm answer exists on when the officers learned the methods (p. 9), the multiplications he decodes are within reach of an abacus (p. 8), and the methods were known to few and probably passed on by word of mouth (p. 8). A historian's credit of the "spark" to one book appears here as a flat statement. + +**cajori1928notations1** — Cajori, F. (1928). *A History of Mathematical Notations, Vol. 1: Notations in Elementary Mathematics*. Open Court. Archive.org item `historyofmathema031756mbp` (public domain). +- *Read status:* read in part, by keyword search of the OCR text (§54 and the Viète sections). +- *Quotations:* "While the Hindu-Arabic numerals became generally known in Europe about 1275, the Roman numerals continued to hold a commanding place" (OCR §54). The Peruzzi bank in Florence "did not use Arabic numerals in their account-books" (§54, citing Nagl). Viète "was the first to do this systematically" (using letters for general magnitudes; §176, pp. 182–183 in the OCR page headers). +- *Bearing:* an old historian gives a date (about 1275) for being "generally known" and a continuing Roman-numeral habit, which is consistent with a slow, uneven spread. + +**burnett2024learning** — Burnett, C. (2024). Learning Indian arithmetic in the early thirteenth century. In *Numerals and Arithmetic in the Middle Ages* (Ch. IX). Routledge. https://doi.org/10.4324/9781003554448-9 +- *Read status:* abstract only (Consensus). +- *What it says:* the shift to Indian numerals is "the most momentous development in the history of pre-modern mathematics"; Latin adaptations began "from the early twelfth century"; the *Liber abbaci* (1202, 1220) "ensured" acceptance alongside the popular manuals of Alexander de Villa Dei and Sacrobosco; and "such was the novelty of Indian calculation that it took some time for it to be understood and accepted." +- *Bearing:* the Fibonacci-as-introducer story is itself a simplification: the numerals and algorism were in Latin circulation nearly a century earlier. + +**james2024otis** — James, K. (2024). Review of J. M. Otis, *By the Numbers: Numeracy, Religion, and the Quantitative Transformation of Early Modern England* (OUP, 2024). *Journal of Religious History*, 49(1), 94–95. https://doi.org/10.1111/1467-9809.13111 +- *Read status:* read in full as a review text via Consensus. I did not read Otis. +- *What it reports:* in early modern England participants in numeracy "could be 'simultaneously numerate and illiterate'" (reviewer quoting Otis, p. 64), because counters and tally sticks needed no writing, whereas "the use of Arabic numerals presumed the ability to write as a prerequisite (p. 65)". The review also notes that written computations were seen as open to tampering, while counters were trusted. +- *Bearing:* **complicates "widened".** A notation that widened numeracy among the writing could narrow it, or move it, among those who counted with objects. Any claim about "who could do arithmetic" needs a baseline of what non-literate people already did. + +**overmann2022review** — Overmann, K. A. (2022). Review of S. Chrisomalis, *Reckonings*. *Journal of Anthropological Research*, 78(1), 152–153. https://doi.org/10.1086/717832 (preprint https://doi.org/10.31219/osf.io/3zwp4). *Not accessible* (OSF page did not render). Listed for the lead to open. + +### Case 2. Musical staff notation and Guido of Arezzo + +**haines2008origins** — Haines, J. (2008). The origins of the musical staff. *The Musical Quarterly*, 91(3–4), 327–378. https://doi.org/10.1093/musqtl/gdp002 +- *Verification:* Crossref; PDF from cartusiana.org. +- *Read status:* read in full (card: library/haines2008origins.md; endnotes skimmed). +- *What it argues:* historians credit Guido with inventing the staff (Apel, Palisca, Lang, the *Oxford History of Western Music*; pp. 327–328). "In sum, the claim implied by Guido's status as 'the one who invented the staff' is untenable. The Arezzo master's role in the staff's development lies less in any one of his presumed contributions than in a clever pedagogy of synthesis backed up by a fair bit of self-promotion" (p. 331 by the PDF's page headers). The ruled staff with letters at the start of each line is in the *Musica enchiriadis* tradition a century and a half before Guido; "This popularization, then, rather than an invention of the staff, was Guido's historical contribution" (p. 344). Guido's own humility topos, quoted: when "those who come after learn with the greatest ease the chants that I and all before me learned with the greatest difficulty" (about p. 331, from Pesce 1999, 445). +- *Bearing:* a documented case where a single-inventor credit for a page convention was taken apart by specialists. Guido's own contribution is dated only through manuscripts of the late 11th or early 12th century, since no copy of the prologue survives from his time (p. 345), so any claim about what Guido proposed rests on later copies. It also places the claimed widening in "pedagogy" and in moving the layout into ordinary liturgical books, not in the notation as such. + +**strunk1998guido** — Guido d'Arezzo, Prologue to his Antiphoner and *Epistola de ignoto cantu* (excerpts), trans. in O. Strunk and L. Treitler (Eds.), *Source Readings in Music History* (Norton, 1998), reproduced in the A-R Music Anthology *Workbook: The Medieval Period, Part 1*, Supplemental Reading 1 (Prologue and *Epistola* excerpts). https://www.armusicanthology.com/media/PDFs/Workbook-Medieval-Part1.pdf +- *Verification:* the workbook states its source; I could not check the translation against Strunk and Treitler or against Pesce (1999). Marked "Do Not Copy" by its publisher; I quote only short passages. +- *Read status:* read in full as the workbook reproduces the excerpts (card: library/strunk1998guido.md; page breaks split one sentence). The Prologue excerpt does describe the ruled lines, letters and colours (workbook p. 3); the "Epistola ... not about the ruled staff" point holds for the *Epistola*. +- *What Guido says:* "the wretched singers and their pupils, though they sing every day for a hundred years, will never sing by themselves without a master a single antiphon, not even a short one" (Prologue, first paragraph). He writes the antiphoner so that "any intelligent and studious person may learn the chant by means of it; after he has thoroughly learned a part of it through a master, he will unhesitatingly understand the rest of it by himself without one" (Prologue, close of the paragraph on his plan for the antiphoner). Boys "who until now have been beaten for their gross ignorance of the Psalms and vulgar letters" sing the antiphon "correctly by themselves without a master" (Prologue). The notation alone is not enough: "however perfect the formation of the neumes might be, it is altogether meaningless and worthless without the addition of letters or colors" (Prologue). The *Epistola* excerpt (about his memorising method, with a teacher present, not about the ruled staff): "After I began teaching this procedure to boys, some of them were able [page break] before the third day to sing an unknown melody with ease, which by other methods would not have been possible in many weeks." The *Epistola* method presupposes "someone present to teach the pupil, not merely from a written textbook, but rather by our practice of informal discussion", and, per the Prologue, learning "by heart a fair number of melodies" so that the pupil recognises the sounds. +- **The "ten years to one or two" claim is not in these excerpts.** The figures I found in Guido's own words are rhetorical ("a hundred years") and "before the third day … in many weeks" (the latter is about his memorising method, not the staff). I found "ten years" attributed to Guido only in popular web pages that gave no citation. Mark the claim **unverified**; the full Prologue and *Epistola* in Pesce (1999), pp. 437–531 for the *Epistola* (page range from a catalogue record), would settle what Guido actually claims. +- *Bearing:* Guido's claim is about a bundle (lines, letters, colours, the hymn mnemonic, memorised melodies, a teacher) that makes learning faster for boys, and he states it as such. It is evidence for H-cost in its most careful form and against "one move". + +**mengozzi2010renaissance** — Mengozzi, S. (2010). *The Renaissance Reform of Medieval Music Theory: Guido of Arezzo between Myth and History*. Cambridge University Press. *Not accessible* (borrowable book not reachable). Table of contents seen: "Guido's musical syllables: conflicting views from modern historiography", "Hands off! Singing without syllables in the Middle Ages", "Discarding the Guidonian image of early music". **mengozzi2007si** — Mengozzi, S. (2007). 'Si quis manus non habeat': charting non-hexachordal musical practices in the age of solmisation. *Early Music History*, 26, 181–218. https://doi.org/10.1017/s0261127907000228. Abstract only. It says our picture of medieval musical training "is frustratingly generic and fragmentary" and that the usual claim that musicians were "well versed" in the Guidonian hand needs testing. **mews2024guidonian** — Mews, C. (2024). Implementing the Guidonian revolution. *Viator*, 55(1), 311–341. https://doi.org/10.1484/j.viator.5.144515. Abstract only: it treats uptake of Guido's pedagogy as a project carried by particular reformers and abbeys (John of Affligem, Canterbury, Bec), with German theorists "more cautious". +- *Bearing:* uptake came through named teachers and reform networks, within decades, and was contested. + +### Case 3. Finding devices in the book + +*None of the primary scholarship on this case was reached.* Rouse and Rouse, "*Statim invenire*" (in Benson and Constable (Eds.), *Renaissance and Renewal in the Twelfth Century*, Harvard UP, 1982), Blair, *Too Much to Know* (Yale UP, 2010) and Duncan's history of the index are **not accessible**; the entries below are abstracts of later work and give a picture, not a verdict. + +**conraulewis2021index** — Conrau-Lewis, K. (2021). The index in the premodern and modern world. *Oxford Research Encyclopedia of Literature*. https://doi.org/10.1093/acrefore/9780190201098.013.1224. Abstract only. "The earliest use of indexes is found in legal and then scholastic and patristic texts in continental Europe; they were particularly useful for university students and preachers"; "the manuscript evidence shows a great degree of experimentation, combining alphabetic, vocalic, and systematic orders of arrangement"; "The use of indexes in Greek, Hebrew, and Islamic book culture only began in earnest in the early modern period." + +**abbamonte2023indexing** — Abbamonte, G. and Kallendorf, C. (2023). On indexing: The birth and early development of an idea. *Journal of the History of Ideas*, 84(3), 465–486. https://doi.org/10.1353/jhi.2023.a901490. Abstract only. It "traces the effective origin of indexing to the early printed editions of" Valla's *Elegantie* and Perotti's *Cornu copiae*, and ties indexing to "a new way of consuming books, which is labeled 'transverse reading'". This dates the index proper later than the thirteenth-century account I know from Rouse and Rouse; the abstract states no century, so my earlier "three centuries later" is [unverified: see AUDIT_02] (the audit puts the early printed editions in the 15th century, about two centuries later). The two sources use "index" differently; the dispute is one for a specialist to resolve. + +**flanders2023place** — Flanders, J. S. (2023). A place for everything, and everything that came before: The development of alphabetical order. *The Indexer*, 41(3), 223–234. https://doi.org/10.3828/index.2023.29. Abstract only. Alphabetisation "took control in the fourteenth century"; "hierarchy, geography, subject – even caste and class" were long-standing orders, some thought superior as late as the nineteenth century; printing and encyclopedias then promoted alphabetical order. + +**thomann2020serial** — Thomann, J. (2020). From serial access to random access: Tables of contents, chapter headings and hierarchical text structures in fourth/tenth-century scientific books. *Journal of Abbasid Studies*, 7(2), 207–228. https://doi.org/10.1163/22142371-12340056. Abstract only. Chapter headings, tables of contents and page numbering become standard in tenth-century Arabic scientific books; "These innovations point to the use of books as a source of knowledge outside traditional learning contexts." The author thinks they may be borrowings from India and Central Asia. + +**dacosta2018tabulae** — da Costa, A. (2018). "That ye mowe redely fynde … what ye desyre": Printed tables of contents and indices, 1476–1550. *Huntington Library Quarterly*, 81(3), 291–313. https://doi.org/10.1353/hlq.2018.0013. Read in full (card: library/dacosta2018tabulae.md; accepted manuscript, which lacks the abstract, so the phrase "marketed these features as time-saving innovations" is not checkable there; the body shows printers selling tables as quick finding aids, p. 17, and explaining reference schemes to readers at length, pp. 10–12). Abstract: printers "marketed these features as time-saving innovations" and, in early Reformation works, used them "as means to shaping and constraining how readers read". + +- *Bearing for case 3:* the finding devices show a multi-century, multi-site accumulation with no single inventor, different orders (alphabetic, systematic) in competition, and uses that served specialist readers (lawyers, preachers, students) first. Nothing I read credits them with widening the base of readers. The abstract-level evidence points to bundle and to expert users, and the claim "ignited" has no support here. The da Costa abstract also warns that the same device can narrow how people read. + +### Case 4. Double-entry bookkeeping + +**sangster2024emergence** — Sangster, A. (2024). The emergence of double entry bookkeeping. *The Economic History Review*, 78(2), 499–528. https://doi.org/10.1111/ehr.13358. Read in full (card: library/sangster2024emergence.md; Aberdeen repository version of record). The method appears in a Florentine moneychanger-bankers' ledger of 1211: "the origin of this bookkeeping method is attributed to northern Italian moneychanger-bankers in the twelfth century", and it "addressed the evidential demands of multiple legal systems relating to use of credit". The abstract reads "The reason for the emergence of such a detailed bookkeeping method is unknown." **Under Sangster's input definition (accounts with duality of entry, p. 501) the 1211 ledger has double entries (pp. 504–507), 283 years before Pacioli's 1494; under the output definition earlier historians used, the earliest case is 1340 (p. 501). Contra entries almost never survive, and double entry is confirmed mainly from cross-references and offset (p. 506). He dates the origin to the twelfth century, earlier than 1211 (p. 520); the ledger is evidence of a method already running. The text never discusses Pacioli's *Summa* (the name appears only in a cited title, n. 48): Sangster targets the Yamey–Besta–Sombart definitional tradition, and the Pacioli comparison is this facet's. He does not rebut "double entry produced capitalism": he reports Yamey's view and cites Yamey 1949, p. 103, for credit dealings leading to systematic accounting (n. 154, p. 523).** Print date 2025 (online 2024-05-23). + +**sangster2025reboot** — Sangster, A. (2025). Time to reboot accounting history – evidence from scholarship on double entry. *Accounting History*, 30(3), 368–391. https://doi.org/10.1177/10323732251357606. Read in full (card: library/sangster2025reboot.md). The 1211 ledger and the entity-wide double entry details cited from this paper in earlier drafts belong to Sangster 2024; here 1211 is the earliest paragraph-format ledger (p. 378) and the confirmed double-entry dates are 1296 and 1299. Sangster does not overturn Yamey's finding that few identified profit before 1800; he argues that Yamey's definitions and synonyms invite readers to take it as a finding about bookkeeping (pp. 373–374). Abstract: "To economic historians, double entry bookkeeping served two primary purposes before 1800: control over debt and distant agents, factors, and partners. Few used it to identify overall profit or wealth, doing so only when accountable to others." The traditional accounting-history literature, it says, contains "unsupportable assertions … accepted as axiomatic fact". + +**yamey1964accounting** — Yamey, B. S. (1964). Accounting and the rise of capitalism: Further notes on a theme by Sombart. *Journal of Accounting Research*, 2(2), 117 ff. https://doi.org/10.2307/2489995 (Crossref lists only the first page). Abstract only. Yamey sets out Sombart's thesis: double entry "clarified the aims of acquisitive business", "provided the rational basis on which the capitalist could choose" where to employ capital, and "made possible the separation of the business firm from its owners". Further entries: **yamey1949scientific** — Yamey (1949), Scientific bookkeeping and the rise of capitalism, *Economic History Review*, 1(2/3), https://doi.org/10.2307/2589824, record only; **yamey2005historical** — Yamey (2005), The historical significance of double-entry bookkeeping: Some non-Sombartian claims, *Accounting History Review* (Crossref: *Accounting, Business & Financial History*), 15(1), 77–88, https://doi.org/10.1080/09585200500033089, abstract only; **ryan2014historical** — Ryan, J. (2014), Historical note: Did double-entry bookkeeping contribute to economic development, specifically the introduction of capitalism?, *Australasian Accounting, Business and Finance Journal*, 8(3), 85–97, https://doi.org/10.14453/aabfj.v8i3.8, read in full (card: library/ryan2014historical.md). The abstract says only that it reviews "Yamey's criticisms of Sombart and other historians regarding the significance they attributed to the introduction of double-entry bookkeeping"; the body defends the causal claim at the macro level (pp. 92–96) and says nothing about control of debt or agents. Cite it for the summary of Yamey and for Pacioli describing a practice about 200 years old (p. 86, n. 2), not for deflating Sombart. + +**carruthers1991accounting** — Carruthers, B. G. and Espeland, W. N. (1991). Accounting for rationality: Double-entry bookkeeping and the rhetoric of economic rationality. *American Journal of Sociology*, 97(1), 31–69. https://doi.org/10.1086/229739. Abstract only. Weber, Schumpeter and Sombart held that accounting enhanced rationality and capitalist production. The authors conclude: "As a practical method, double-entry bookkeeping appears to have increased 'rationality,' but the rhetorical side of double entry is also critical", and its significance "can be appreciated only if its rhetorical and technical aspects are considered." + +**sangster2016genesis** — Sangster, A. (2016). The genesis of double entry bookkeeping. *The Accounting Review*, 91(1), 299–315. https://doi.org/10.2308/accr-51115. Record only. Listed for the lead. + +- *Bearing for case 4:* a cautionary case on three counts. (a) The "invention" credited to Pacioli (1494) was a description of an older practice (Ryan 2014, p. 86, n. 2: about 200 years). (b) The big causal claim (Sombart, with Weber and Schumpeter in the abstract of Carruthers and Espeland) was contested for decades by Yamey (who found little evidence of profit calculation) and is defended at the macro level by Ryan (2014); the restriction of double entry to control of debt and agents, not profit measurement, is the view of economic historians as Sangster (2025) reports it (abstract), and Sangster argues that Yamey's definitions invite readers to take Yamey's result as a finding about bookkeeping (pp. 373–374). (c) The sociologists' correction keeps a role for the notation, but as persuasion and as method together. Funnell (2001, abstract via Consensus; not entered) argues that Sombart's account of double entry was shaped by anti-Semitism; I did not read it. I did not read Sombart, Weber, Yamey 1949 or Pacioli. + +### Case 5. Algebraic symbolism and calculus notation + +**cajori1929notations2** — Cajori, F. (1929). *A History of Mathematical Notations, Vol. 2: Notations Mainly in Higher Mathematics*. Open Court. Archive.org item `historyofmathema027671mbp` (public domain). +- *Read status:* read in part, by keyword search of the OCR text (the fluxion and differential sections around §§581–588). +- *Quotations:* "The adoption of the Leibnizian notation in England, advocated in 1802 by Woodhouse, was finally secured through the efforts of J. F. W. Herschel, Peacock, and Babbage at Cambridge" (OCR §583). "The calculus notation used in the United States before 1824 was almost exclusively Newtonian, from 1824 to 1841 almost entirely Leibnizian" (OCR §588). "Nor had Lagrange any prejudice against the Newtonian dots"; he used both notations, which Thomson and Tait judged "combined … with admirable skill and taste" (OCR, immediately before §582). +- *Bearing:* Cajori is the historian most associated with notation. In his account, which notation a country used was settled by schools, textbooks and a Cambridge reform group, over decades. He does not credit Leibniz's notation with widening who could do the calculus; I found no such claim in the passages I read. I did not read the sections on Descartes or on the general principles. + +**shneiderman1983direct** (see case 7) carries as its epigraph a Leibniz line on signs that "picture" a thing so that "the labor of thought is [wonderfully] diminished" (p. 57, OCR-damaged in the text I read). That is the same design intuition, stated about calculus notation, which the paper's author might cite as a lineage; I did not trace the source (it is quoted from F. Kreiling in *Scientific American*, 1968). + +- *Case 5 status:* weak evidence. Viète and Descartes were read only through Cajori's snippets of Viète. No source I read argues that symbolic algebra widened the population who could do mathematics. + +### Case 7. Computing: direct manipulation, spreadsheet, web, touch + +**shneiderman1983direct** — Shneiderman, B. (1983). Direct manipulation: A step beyond programming languages. *Computer*, 16(8), 57–69. https://doi.org/10.1109/mc.1983.1654471 +- *Verification:* Crossref; PDF from Shneiderman's UMD page. +- *Read status:* read in full (card: library/shneiderman1983direct.md; the earlier read omitted the "Problems with direct manipulation" section, p. 64, and the cited experiments). +- *What it says:* the author defines the style after talking to enthusiastic users: "The central ideas seemed to be visibility of the object of interest; rapid, reversible, incremental actions; and replacement of complex command language syntax by direct manipulation of the object of interest" (p. 57). He lists several properties, so the definition is a bundle. On the spreadsheet: "Since Visicalc simulates an accountant's worksheet, it is easy for novices to comprehend", and "the command language for setting up the worksheet can be tricky for novices to learn" (p. 59). He cites Roberts for display editors needing half the time of line editors (p. 57) and "office automation evaluations" favouring them; I did not check either. +- *Bearing:* the framing text for the computing parallel. Its evidence is the author's observation and secondary reports, not a controlled study. + +**hutchins1985direct** — Hutchins, E. L., Hollan, J. D. and Norman, D. A. (1985). Direct manipulation interfaces. *Human–Computer Interaction*, 1(4), 311–338. https://doi.org/10.1207/s15327051hci0104_2. *Not accessible* (no open copy; two author-site guesses returned HTML). Bib key reused. The paper is the usual theoretical account of why direct manipulation lowers cognitive cost; I cannot summarise it. + +**margono1987study** — Margono, S. and Shneiderman, B. (1987). A study of file manipulation by novices using commands vs. direct manipulation (Technical Report CAR-TR-264). University of Maryland. https://www.cs.umd.edu/users/ben/papers/Margono1987Study.pdf +- *Verification:* PDF from the UMD page; the report is also reprinted in Shneiderman (Ed.), *Sparks of Innovation in HCI* (1993) per a search result. +- *Read status:* read in full (card: library/margono1987study.md). +- *What it reports:* 30 novices (18 psychology students with no computer experience, 12 recruited personally) compared file manipulation on a Macintosh with MS-DOS on an IBM PC after "a brief training period". Abstract: novices "accomplished file manipulation tasks more rapidly, with fewer errors and greater satisfaction with the Apple Macintosh." Results: test time "5.77 minutes and 4.80 minutes respectively, statistically significant by t-test at the p < 0.10 level"; errors 2.03 against 0.80 (p < 0.01); satisfaction 3.80 against 5.37 of 7 (p < 0.01). "Out of thirty subjects, eighteen of them preferred the Macintosh, ten preferred the IBM PC" (§3). "Five subjects did have a very difficult time with the mouse the entire time" (§4). +- *Limits:* the comparison is of whole machines (hardware, mouse, software, tasks), so it cannot isolate the interface style; the time difference is marginal; a third of the participants preferred the command system; the satisfaction figure rests on an ad hoc composite, and the test time is under five minutes in both conditions (card). +- *Bearing:* the only controlled evidence I found for the claim that direct manipulation lowers cost for novices. It is small, from the advocate's own group, and supports H-cost at best weakly. + +**power2000history** — Power, D. J. (2000). A history of microcomputer spreadsheets. *Communications of the AIS*, 4, Article 9. https://doi.org/10.17705/1cais.00409. *Not accessible* (403). Listed for the lead. I found no source this session that quantifies the spreadsheet's effect on the user population; Shneiderman's article is the earliest I read, and he credits VisiCalc to a Harvard MBA student's frustration (p. 59). + +- *Web browser, hyperlink, touch interface:* **not researched to the evidence level.** I ran no search that produced a usable source for adoption by non-specialists, so I make no claim about them beyond the gap note in (d). +- *Bearing for case 7:* the computing cases are the only ones where the claimed widening is recent, measured, and attributed to a design movement with a manifesto. The manifesto itself names a bundle. The evidence for population widening (as opposed to task performance) is thin in what I could read. + +### Case 8. Writing systems designed for learnability + +**walker1993early** — Walker, W. and Sarbaugh, J. (1993). The early history of the Cherokee syllabary. *Ethnohistory*, 40(1), 70 ff. https://doi.org/10.2307/482159 +- *Read status:* abstract only (JSTOR; Consensus excerpt). +- *What it says:* the printed syllabary has been virtually unchanged since 1828; documents by Sequoyah and contemporaries "all suggest that Cherokees alone developed the syllabary". Unlike European-derived writing systems, "the inventor was a monolingual speaker of Cherokee without links to Christianity or to Euroamerican schooling. The syllabary was disseminated without the aid of Euroamerican institutions, but seems to have had an almost immediate impact on all Cherokee-speaking communities", and it "was enthusiastically embraced" in both east and west. "Most missionary groups concerned with Cherokees adopted the syllabary by 1826." + +**cushman2011syllabary** — Cushman, E. (2011). The Cherokee syllabary: A writing system in its own right. *Written Communication*, 28(3), 255–281. https://doi.org/10.1177/0741088311410172. Abstract only. The 86-character system "was learned in manuscript form and became widely used by the Cherokee within the span of a few years"; Worcester's 1827 rearrangement fitted it to Latin-alphabet order, and scholars who use that arrangement "have misunderstood the syllabary by viewing it through an alphabetic lens". **montgomeryanderson2014review** — Montgomery-Anderson, B. (2014). Review of Cushman, *The Cherokee Syllabary: Writing the People's Perseverance*. *Language*, 90(2), 537–540. https://doi.org/10.1353/lan.2014.0030. Read as a review text via Consensus: Cushman claims Sequoyah was fluent and literate in English, which departs from the "traditional portrait"; the syllabary was "a powerful technology for recording traditional knowledge" that also "helped them to resist the assimilative pressures" of alphabetic literacy. + +**parins2013literacy** — Parins, J. W. (2013). *Literacy and Intellectual Life in the Cherokee Nation, 1820–1906*. University of Oklahoma Press. *Not accessible.* I read the review by B. B. Schweiger in the *Arkansas Historical Quarterly* (2014; the Consensus record carries a mis-assigned Choice DOI, so no DOI is given here), as a text via Consensus. The review reports: "Within seven years of the syllabary's introduction around 1820, it was reported that almost all young and middle-aged men could read it"; "the syllabary was so well formed that most could do so in three days"; "By the mid-1830s, as many as two-thirds of Cherokees were literate in Cherokee and/or English"; and that "less than a quarter of mixed-blood Cherokees were literate in English" beforehand. The review adds a causal question the book leaves open: "Questions remain as to why Sequoyah's handiwork found such a warm reception", and cites Barry O'Connell on missionaries' plan to write the language. + +**unseth2016international** — Unseth, P. (2016). The international impact of Sequoyah's Cherokee syllabary. *Written Language & Literacy*, 19(1), 75–93. https://doi.org/10.1075/wll.19.1.03uns. Abstract only: the script's influence led to "21 scripts for over 65 languages". + +**bender2002signs** — Bender, M. (2002). *Signs of Cherokee Culture: Sequoyah's Syllabary in Eastern Cherokee Life*. University of North Carolina Press. Record only (Choice review record, DOI 10.5860/choice.40-3482, from Consensus). Publisher blurb: the syllabary "quickly replaced alternative writing systems for Cherokee and was reportedly in widespread use by the mid-nineteenth century. After that, literacy in Cherokee declined, except in specialized religious contexts." + +- *Unreliable anecdotes I did not accept:* "the entire Cherokee Nation was literate within months" (attributed in a web-search summary to a University of Cincinnati anthropologist; no source found); "70–90 percent by 1830" and "almost 100% within a quarter-century" (Wikipedia and popular sites, no primary source). A 1962 *Current Anthropology* note (White, https://doi.org/10.1086/200321) says "Overnight the whole nation became a school" without citation. All are **unverified**. +- *Bearing:* the best-documented fast case, with caveats the author should keep. The syllabary was a whole new script, with a one-to-one syllable mapping, not an added convention on an existing page; it was adopted by a community with political reasons (a national council and, from 1828, a printing press; missionary groups adopted the syllabary by 1826, after it had spread, per Walker and Sarbaugh) and in a language where a syllabic script suits the phonology; literacy in Cherokee later declined. It supports a version of H-cause for a *new script*; it does not transfer to a *spacing* convention in a script people already read. + +**handel2016learn** — Handel, Z. (2016). Learn to read Korean: An introduction to the Hangul alphabet (lecture abstract). Record only; the DOI given in Consensus belongs to another conference, so none is given. Abstract: the 1446 text "that introduced Hangul" says it was so simple "that 'a wise man can master it in a morning, and even a stupid person can learn it in ten days'". *Abstract only.* I could not trace the wording to Ledyard's 1966 translation or to Jeong Inji's postface. **Unverified locator.** + +**hangeulmuseum** — National Hangeul Museum. "Read Society Through Its Hangeul Letters" (exhibition page). https://www.hangeul.go.kr/exhibition/485. Read in full (a short page). "The illiteracy rate of Korean people was 78% in 1945." "It went down to 4.1% in 1958 as the government-led measures to eradicate illiteracy had been implemented." "However, it's not easy to figure out the rate accurately because proper surveys had not been taken place until 1990." A news opinion piece gave "more than 90 percent" for 1945; that conflict is another reason to treat any single figure as soft. The page itself gives no source for the 78% or 4.1% figures (card). + +**baek2024spread** — Baek, N. (2024). On the use and spread of Hangeul in the Joseon dynasty. *Liberal Arts Innovation Center*, 14, 449–476. https://doi.org/10.54698/kl.2024.14.449. Abstract only: Hangeul "was especially welcomed by women", "spread very quickly in Seoul, but it took much longer to reach local areas", driven by letters, novels and the Hangeul Bible. **ivanova2026process** — Ivanova, Y. (2026). A process of implementing the Korean alphabet on the Korean peninsula. *Ezikov Svyat*, 24(1), 136–151. https://doi.org/10.37708/ezs.swu.bg.v24i1.9. Read in full (card: library/ivanova2026process.md; Bulgarian body, translations the card writer's). On word separation: the script was to be written "without division between words" (p. 137), and the Unified Orthography of 1933 set as its third principle that each word be written separately (p. 144), with a chapter on spaces (p. 145); she does not say when Koreans first spaced Hangul or link the 1933 rule to readership. Abstract: the script "faced strong resistance from the Confucian elite" and for centuries "was marginalized and associated with women, children, and the lower social classes"; the late-nineteenth-century turn came from modernisation, missions and nationalism. **park2025dokkyo** — Park, J.-B. (2025?). Diversification of literacy patterns by social class in late Joseon (forum abstract, Dokkyo University). https://www.dokkyo.ac.jp/international/36-10.html. The date is unconfirmed. Abstract (Japanese; my translation): most yangban and chungin in late Joseon were literate in both Hangul and Chinese characters, and commoners' literate life centred on Hangul "relatively". Card (library/park2025dokkyo.md): the page is a forum abstract with no counts or rates, so it cannot be cited for literacy levels. These three are second- or third-tier sources; Ledyard (1966), Kim-Renaud (1997) and Sampson (1985) are the standard works and I did not reach them. + +- *Bearing for Hangul:* the design intention (to let ordinary people read) is documented; the learnability claim comes from the 1446 text itself, and rapid learning is a property of the alphabet, not evidence of widespread use. Uptake was socially patterned (women, letters, novels; Baek says it spread very quickly in Seoul and slowly to local areas) and mass literacy came with schooling and state campaigns in the twentieth. This is the cleanest case that a very learnable script does not by itself spread literacy. It supports H-cost (low learning cost) and counts against H-cause. + +### Case 9. Counter-cases + +**liebowitz1990fable** — Liebowitz, S. J. and Margolis, S. E. (1990). The fable of the keys. *Journal of Law and Economics*, 33(1), 1–25. https://doi.org/10.1086/467198. Abstract only (SSRN copy bot-walled). "Contrary to the claims made by Paul David and Brian Arthur, we find virtually no evidence to support a view that QWERTY is inferior to DVORAK." **norman1982alphabetic** — Norman, D. A. and Fisher, D. (1982). Why alphabetic keyboards are not easy to use: Keyboard layout doesn't much matter. *Human Factors*, 24(5), 509–519. https://doi.org/10.1177/001872088202400502. Abstract only: "Performance with the Dvorak keyboard is only improved by about 5% over performance on the Sholes keyboard. The conclusion is that it is not worthwhile … to change to the Dvorak layout for experts." **kay2013rerun** — Kay, N. M. (2013). Rerun the tape of history and QWERTY always wins. *Research Policy*, 42(6–7), 1175–1185. https://doi.org/10.1016/j.respol.2013.03.007. Abstract only. **hossain2009quest** — Hossain, T. and Morgan, J. (2009). The quest for QWERTY. *American Economic Review*, 99(2), 435–440. https://doi.org/10.1257/aer.99.2.435. Abstract only. The case is contested: David (1985, AER; not in Crossref, not read) argued lock-in on an inferior layout; the sources above dispute the inferiority; one working paper (Torres, 2013, seen in Consensus; not entered) computes Dvorak at over 10 per cent more efficient. I did not trace the 1944 US Navy study. **Dvorak is therefore not a clean "clear improvement that failed".** + +**vera2009decimal** — Vera, H. (2009). Decimal time: Misadventures of a revolutionary idea, 1793–2008. *KronoScope*, 9(1–2), 29–48. https://doi.org/10.1163/156771509x12638154745382. Abstract only: decimal time was "the product of a broader rationalizing project that successfully reformed weights and measures … and money … but failed in bringing the chronological and calendrical reform into fruition". **zerubavel1977republican** — Zerubavel, E. (1977). The French Republican calendar: A case study in the sociology of time. *American Sociological Review*, 42(6), 868 ff. https://doi.org/10.2307/2094573. Abstract only; three factors for failure: "totality of the obliteration of the traditional sociotemporal order, overemphasis on secularization, and advocation of nationalistic particularism at the expense of a practical and cognitive disruption of temporal coordination on a global level." The decimal metre and franc succeeded; decimal time did not, in the same revolutionary bundle. + +**buncic2017factors** — Bunčić, D. (2017). Factors influencing the success and failure of writing reforms. *Studi Slavistici*, 21–46. https://doi.org/10.13128/studi_slavis-21937. Read in full (card: library/buncic2017factors.md); the body makes the same two points as the abstract (pp. 26, 39–41), and installed base is treated at p. 36. After comparing Slavic and post-Soviet reforms: "linguistic factors turn out to play a very minor role compared to extralinguistic factors. In fact, the most decisive factor seems to be timing." Successes and failures named: the Russian reform of 1917–1918 (success), the 1930 Latinisation, the 1977 Chinese second simplification scheme (withdrawn 1986), the French 1990 reform (ignored), Uzbek Latinisation 1993–1995. + +- *Shavian and English spelling reform:* I found only abstracts and Unicode background (Consensus: a proposal that Shaw's will funded a competition won by Kingsley Read; a 2018 review of a Victorian reform that obtained a knighthood for Pitman and "the complete failure of Victorian-era attempts at spelling reform" in A. R. Hoffman's words). No source with an adoption measure. Not entered in the bib. +- *Bearing for case 9:* clear design improvements on paper (decimal time, phonetic spelling) fail when they break the existing coordination of users, and adoption turns on timing, authority and installed base, not on the merit of the form. This is a strong warning for the author's "a similar move" and bears on the choice of mechanism (a move that needs everyone to switch is a different kind of proposal from one that a mediator can adopt unilaterally). + +### Numeracy-format set (not in the brief) + +**peters2006numeracy** — Peters, E., Västfjäll, D., Slovic, P., Mertz, C. K., Mazzocco, K. and Dickert, S. (2006). Numeracy and decision making. *Psychological Science*, 17(5), 407–413. https://doi.org/10.1111/j.1467-9280.2006.01720.x +- *Verification:* Crossref (authors shown: Peters, Västfjäll, Slovic, Mertz, …); abstract from Europe PMC. +- *Read status:* **abstract only** (as in the Part 1 bib). +- *What it argues:* "highly numerate individuals [are] more likely to retrieve and use appropriate numerical principles, thus making themselves less susceptible to framing effects", draw "different (generally stronger or more precise) affective meaning from numbers", and the effect "was not due to general intelligence." +- *Bearing:* the author's standing model measures *individual differences in a competence*. It contains no claim that a notation produced the competence. For the paper's history, that means "numeracy" is borrowed as a competence-and-decision-quality link, and the parallel with notation history has to be argued separately. This entry matters because the sources below connect the same research programme to format design. + +**peters2007less** — Peters, E., Dieckmann, N., Dixon, A., Hibbard, J. H. and Mertz, C. K. (2007). Less is more in presenting quality information to consumers. *Medical Care Research and Review*, 64(2), 169–190. https://doi.org/10.1177/10775587070640020301. Abstract only. "Results were particularly strong for those lower in numeracy, who had higher comprehension and made better choices when the information-presentation format was designed to ease the cognitive burden and highlight the meaning of important information." **hibbard2003supporting** — Hibbard, J. H. and Peters, E. (2003). Supporting informed consumer health care decisions: Data presentation approaches that facilitate the use of information in choice. *Annual Review of Public Health*, 24, 413–433. https://doi.org/10.1146/annurev.publhealth.24.100901.141005. Abstract only: "The challenge is not merely to communicate accurate information to consumers, but to understand how to present and target that information so that it is actually used in decision-making." **peters2007numeracy** — Peters, E., Hibbard, J., Slovic, P. and Dieckmann, N. (2007). Numeracy skill and the communication, comprehension, and use of risk-benefit information. *Health Affairs*, 26(3), 741–748. https://doi.org/10.1377/hlthaff.26.3.741. Abstract only. **gigerenzer1995improve** — Gigerenzer, G. and Hoffrage, U. (1995). How to improve Bayesian reasoning without instruction: Frequency formats. *Psychological Review*, 102(4), 684–704. https://doi.org/10.1037/0033-295x.102.4.684. Record only; abstract not retrieved. **garciaretamero2013visual** — Garcia-Retamero, R. and Cokely, E. T. (2013). Communicating health risks with visual aids. *Current Directions in Psychological Science*, 22(5), 392–399. https://doi.org/10.1177/0963721413491570. Record only. **larkin1987why** — Larkin, J. H. and Simon, H. A. (1987). Why a diagram is (sometimes) worth ten thousand words. *Cognitive Science*, 11(1), 65–100. https://doi.org/10.1111/j.1551-6708.1987.tb00863.x. Not read; cited by Schlimm and Neth as the source of the informational/computational equivalence distinction (p. 2098). +- *Bearing:* these are experiments on a format change that helps lower-numeracy people decide better. They are the nearest thing to H-cost with human participants in the numeracy tradition, and they concern what the paper's DXD ethic proposes (move work from the person to the artefact). I read abstracts only; the effect sizes, the populations and the kind of format are unknown to me. + +--- + +## (c) Synthesis + +### (i) Case table + +"Evidence" grades the evidence for the *widening* claim, not for the history of the convention. "One move or bundle" is what the scholars I read say, not what I infer. + +| Case | Convention and credited agent | Claimed effect | Evidence quality | Diffusion time | One move or bundle | +|---|---|---|---|---|---| +| 1. Hindu-Arabic numerals | Place value and zero; Fibonacci 1202 | Opened commercial arithmetic; "numerical revolution" | Historical plus one working-paper regression (Danna et al.); lab and simulation work (Schlimm and Neth; Krajcsi and Szabó via Buijsman) cuts both ways | Generally known in Europe by about 1275 (Cajori); in use by merchants by end of 13th c. (Danna et al.; their inference, "no direct evidence of the first adoption"); printed manuals to the 16th c. [unverified: see AUDIT_02 for "spread to England and the Atlantic", which I removed]; Roman numerals persisted in records | Danna et al. measure manuals and abacus schooling, not the numerals alone; reading that as a bundle (numerals, vernacular manuals, abacus schools with 2-year curricula, commercial demand) is my inference; Chrisomalis stresses user networks | +| 2. Staff and Guido | Ruled staff plus letters/colours; Guido c. 1030 | Faster learning of chant by boys | Guido's testimony (Epistola) concerns his memorising method with a teacher present, not the ruled staff: "before the third day … many weeks"; his Prologue makes the staff-and-letters claim separately; Haines on origins; no outcome data | Some seventy manuscript sources, 11th–16th c. (Haines, p. 328); reform networks (Mews) | Bundle by Guido's own account; the inventor credit is "untenable" (Haines) | +| 3. Finding devices | Alphabetical order, index, chapter division, tables | Faster retrieval for specialist readers | Abstract-level only; primary scholarship not reached | Centuries; alphabetisation "took control in the fourteenth century" (Flanders); Abbamonte and Kallendorf trace the "effective origin of indexing" to the early printed editions of Valla's *Elegantie* and Perotti's *Cornu copiae* (abstract; the century is not stated there) | Accumulation across sites; no single move; benefit to specialists first | +| 4. Double entry | Pacioli 1494; Sombart, Weber | Rationalised business; helped create capitalism | Contested since at least Yamey (1949); defended at the macro level by Ryan (2014); Sangster (2025) faults Yamey's definitions | Under Sangster's input definition a Florentine moneychanger-bankers' ledger of 1211 already has double entries (the output definition gives 1340 at the earliest), and ledgers of 1296 and 1299 confirm entity-wide double entry by the end of the 13th century (Sangster 2024 [online 2024], pp. 501, 507) | Method plus rhetoric (Carruthers and Espeland); the big causal claim contested, not rebutted | +| 5. Algebraic and calculus notation | Viète; Leibniz | Better mathematics | Historian's narrative (Cajori); no widening claim found | Decades per country; England switched to Leibniz after the Cambridge reformers (1802 to about 1820s) | Social and educational adoption, not a single move | +| 7. Direct manipulation, spreadsheet | Shneiderman 1983; Xerox/Apple; VisiCalc | Novices can compute without commands | One small experiment (Margono and Shneiderman 1987: N = 30 novices; Macintosh vs MS-DOS machine, so not isolating interface style; time 4.80 vs 5.77 min, p < 0.10; errors p < 0.01); Shneiderman 1983 reports no experiment; advocate's observation | [unverified: see AUDIT_02: the "decade from Star to Mac to mass use" has no source in this facet]; population effect not measured in what I read | Defined as a bundle by its author (visibility, reversible incremental actions, no syntax) | +| 8a. Cherokee syllabary | Sequoyah, 1821 | National literacy within years | Reports in secondary literature; Parins as reviewed by Schweiger (reported, via a review; not checked against Parins): almost all young and middle-aged men within seven years, about two-thirds literate in Cherokee or English by mid-1830s | Years; Bender's publisher summary says Cherokee literacy declined after the mid-nineteenth century (Eastern Cherokee focus; book not read) | A whole new script plus a community with a political reason, a national council and, from 1828, a press; missionary groups adopted the syllabary by 1826 after it had spread (Walker and Sarbaugh) | +| 8b. Hangul | Sejong 1443–46 | Commoners learn in days | Design claim in the 1446 text (via Handel, unverified locator); uptake sources are second tier | Marginalised in elite and official life for centuries, yet spread through letters and novels (Baek 2024; Ivanova 2026); written without word division until the 1933 orthography required separate words (Ivanova 2026, p. 144); illiteracy reported at 78% in 1945 (Hangeul Museum; no proper surveys before 1990) | The script alone did not widen use; schooling and state campaigns did | +| 9. Counter-cases | Dvorak; spelling reform; decimal time | None realised | Contested (Dvorak); no adoption data (spelling); historical analyses (decimal time, Bunčić) | n/a | Failure traced to timing and installed base (Bunčić; "the revolutionary bundle" is not Bunčić's phrase and he does not treat reforms as bundles) | +| N. Numeracy formats | Frequency formats, layout and "less is more" displays | Lower-numeracy people decide better | Peters et al. 2007 (Medical Care Research and Review; three studies, abstract only) found that format changes helped lower-numeracy participants; Hibbard and Peters 2003 and Peters et al. 2007 (Health Affairs) are reviews | n/a | A single format change, tested as such, in the author's own model family | + +### (ii) Conditions under which a notational or interface change widened a competence + +What the sources support, no more: + +1. **The cost saving is task- and user-specific.** Place value cuts perceptual-motor steps when the problem is laid out in columns, but loads memory (100 addition facts); Roman numerals cut memory and raise perceptual-motor steps (Schlimm and Neth 2008, p. 2101); sign-value systems were easier for novices to learn (Krajcsi and Szabó via Buijsman); Dvorak's gain was about 5% in simulation (Norman and Fisher). A claim that a notation "lowered the cost" needs "of what, for whom, after how much practice". +2. **In most cases a teaching or carrying institution is present; Cherokee is the exception its own source stresses.** Abacus schools and vernacular manuals (numerals), a teacher plus memorised melodies (Guido), community networks and a national council (Cherokee; Walker and Sarbaugh stress dissemination "without the aid of Euroamerican institutions"), schooling and state campaigns (Hangul in the 20th century), the Cambridge group (Leibniz in England). +3. **A demand or incentive, usually commercial or political, is present.** Merchants (numerals), moneychanger-bankers and lawyers (double entry, indexes), political self-determination (Cherokee). +4. **A widening for one group can narrow it for another.** Arabic numerals presumed writing; counters did not (James on Otis). +5. **Users must adopt together.** Chrisomalis: a notation's usefulness is "evident" only when "an entire social group" learns it, and "a system with many users is functional for that reason alone". Bunčić finds that, in mainly Slavic and post-Soviet reforms, linguistic factors play "a very minor role" and "the most decisive factor seems to be timing". +6. **Specialist users adopt first and the benefit to a wider population is inferred, not measured.** This holds for indexes and double entry; for the spreadsheet it is my inference [unverified: see AUDIT_02]. + +### (iii) Which cases serve the paper, and which embarrass it + +Best parallels, in order: +1. **Format interventions in the numeracy programme (Peters, Hibbard, Gigerenzer).** Peters et al. 2007 (Medical Care Research and Review; three studies) report that a format change helped the less numerate decide better; Hibbard and Peters 2003 and Peters et al. 2007 (Health Affairs) are reviews. It is a design-precedent argument for H-cost with human data, and it sits where the author wants it (numeracy). *Caveat:* abstracts only; effect sizes unknown. +2. **Hindu-Arabic numerals with the abacus schools.** The best-documented historical case with a numeracy label, a quantitative study of the carrier, and an active scholarly argument against both the efficiency story and the ban story. It has to be told as a bundle. +3. **Direct manipulation and the spreadsheet.** The closest interface precedent to the paper's own subject, with a design manifesto by the person who named it. Use it for the shape of the argument (visibility, reversibility, no syntax), not for a population-level effect, which I could not find. + +A fourth, with care: **Cherokee**, as the single documented case of a writing system followed by fast literacy, with the caveats in section (b). + +Cases that would embarrass the paper if used as written: +- **The 1299 Florence "ban".** It is not a ban on a better notation, and its motive is unknown. Using it as resistance to a superior system repeats a modern myth documented by Nothaft. +- **Guido "invented the staff"** and **"cut training from ten years to one or two"**. Both fail (Haines; no source found for the figure). +- **Pacioli and double entry as the cause of capitalism.** The causal claim is contested by Yamey and defended at the macro level by Ryan (2014); Sangster does not rebut it. Under Sangster's input definition double entries are 283 years older than Pacioli (the 1211 ledger; 1340 under the output definition), with entity-wide double entry roughly two centuries older (Sangster 2024 [online 2024], pp. 501, 507). +- **Hangul "built so common people learn it in days".** True of the design intention, false as a story of use. It shows spacing's problem (low cost does not spread a script) and so tells against H-cause. +- **Dvorak** as a "clear improvement that did not spread". The dispute is live. + +### (iv) What Chrisomalis and the other sceptics imply + +I read Chrisomalis only in fragments, so what follows is bounded by those fragments (pp. 22, 32, 79, 108, 123, 125) and by Nothaft and Schlimm and Neth. + +- **Efficiency is not the explanation.** Chrisomalis calls efficiency for computation "a Western-centered and historically inaccurate benchmark" and says adoption needs "an entire social group" to learn the system. For the paper this removes the sentence "spaces were better, so they won". It leaves the more modest statement that spaces lowered a cost. +- **The numeracy–literacy link is historically contingent.** Chrisomalis (2020, p. 108 fragment) calls the pairing of literacy and numeracy as going "hand in hand" "a modern idea". The author's pairing of literacy and numeracy as the analogy structure needs a caveat; the model is borrowed from decision research, not from history. +- **No single origin.** Haines on Guido, Nothaft on the zero myth, Sangster on the definitional tradition behind the Pacioli story (the Pacioli comparison is this facet's), and Cajori on Leibniz's notation in England each take apart a one-agent, one-device story. The reading-history claim about spaces and "ignition" is of the same genre, and facets A to C need to test it with the same rigour. This is a pattern in how credits are made, not evidence about spaces. +- **Cognitive effects of notation are real but narrower than the tradition claims.** Chrisomalis (2020, p. 24 fragment) allows that there are effects; Schlimm and Neth show them as trade-offs; Zhang and Norman (unread) argue for representational effects. The author's "similar move" is defensible as design only if the claim stays at this scale. +- **I did not find a sceptic who denies that a convention can lower a cost.** The sceptics deny the *explanatory* weight of the convention in adoption and the *ranking* of notations. + +### (v) Bearing on H-cost and H-cause + +**H-cost.** Supported in a qualified form by the parallels. Each case where I could read evidence shows that a representation can move work from the person to the page (columnar place value, a staff that holds pitch, a display that holds state) and that some users then do the task more easily. The evidence that this is true of *spacing* is facet D's task, not mine. What the parallels add are two conditions: the saving depends on task and user (Schlimm and Neth; Krajcsi and Szabó; Norman and Fisher), and a convention that helps one group can burden another (James on Otis). The brief's "widened the group who could read without long training" survives only if the paper names which group and which training. + +**H-cause.** Not supported by any parallel as a claim about a convention alone. The cases that came nearest are Cherokee (a new script, a motivated community, institutions improvised in a decade) and Guido (boys, a teacher, memorised melodies). The cases with the most literal "design for the common people" intention, Hangul, show the opposite. In numerals, the one quantitative study measures schools and manuals, not numerals. If H-cause is kept, it should read: "a convention was one necessary component of a bundle that widened literacy", which is H-cost plus a weak necessity claim. The parallels do not let me say more. + +### (vi) Problem 4: is "one similar move" the right shape? + +**No, in the shape the author uses.** Every case with usable evidence is a bundle, and the scholars who set the standard (Danna et al. on numerals; Haines on the staff; Carruthers and Espeland on double entry; Shneiderman on his own definition) say something close to it in their own words (Danna et al.: "organised learning"; Haines: "popularization", "pedagogy of synthesis"); only Guido and Shneiderman state a bundle themselves. The pattern the author reaches for is real as a genre of historical credit and as an observation that some conventions measurably lower a task's cost; the "single move ignites" form is the part that comes from retrospective hero-credit. Two readings keep the paper's idea intact without the weak form: +- Treat the space between words as *one component with a measured effect on cost* (H-cost) and say the paper wants the equivalent component for algorithmacy, accompanied by the equivalents of schools, manuals and incentives. DXD's own claim that the expertise is structural intervention (Part 1, lines 34, 142, 150) fits a bundle better than a move. +- The adoption cases (decimal time, spelling reform, Dvorak) imply the "move" must be adoptable by the party that controls the mediator, since users cannot switch together. This fits Part 1 line 118 (a coordination, not a medium). + +### Claims in the author's framing, or in the wider story, that the sources contradict + +1. "Florence banned Arabic numerals in 1299 as a rejection of the new system." Nothaft (2020, pp. 14–17): a guild rule on account books, motive "a matter of speculation"; the fraud rationale is not in the text. +2. "Roman numerals could not do arithmetic." Schlimm and Neth (2008): they could; the cost is in steps, not feasibility. +3. "Place-value notation is simply the more efficient system, which is why it spread." Chrisomalis (2010, p. 32; 2020, p. 22 fragments); Danna et al. (spread through schools); James on Otis (counters served people who could not write). +4. "Guido invented the staff." Haines (2008, p. 331): "untenable". +5. "The staff cut singers' training from ten years to one or two." Not found in Guido's text; the figures I found are "a hundred years" and "before the third day … in many weeks" (his memorising method, not the staff). Unverified; see the Pesce gap. +6. "Pacioli introduced double entry" and "double entry produced capitalism". Sangster (2024) on the origin date; Yamey and Sangster (2025) on the evidence for the causal claim; Ryan (2014) defends the causal claim; Carruthers and Espeland. +7. "The Cherokee nation was literate within months." No scholarly source supports "months"; Parins (as reviewed) gives seven years for almost all young and middle-aged men and about two-thirds by the mid-1830s. +8. "Hangul was built so that common people learned it in days, and it was therefore adopted by them." The design claim is in the 1446 text; use was marginal in elite and official life for centuries (Ivanova; Baek; Hangeul Museum for 1945). +9. "Dvorak is clearly better and lost only to habit." Contested (Liebowitz and Margolis; Norman and Fisher; Kay). +10. "Literacy and numeracy go together." Chrisomalis (2020, p. 108 fragment): a modern idea; in England counters served the unlettered (James on Otis). + +--- + +## (d) Gaps and unverified leads + +**Not accessible; a human with library access should open these.** +- Zhang and Norman (1995), *Cognition* 57(3), 271–295: ScienceDirect or a library. Needed before the paper relies on the representational effect. +- Chrisomalis (2010) and (2020): read the chapters on the Roman numerals, the Western numerals' spread and the "efficiency" section (2010, pp. 29–32, 116–125; 2020, chs. on cognition and on the abacus/algorism debate). MIT Press has an open preview that was bot-walled to me. Overmann's review (*J. Anthropological Research* 78(1), 152–153; OSF preprint) is a short route. +- Hutchins, Hollan and Norman (1985), *HCI* 1(4): Taylor & Francis. +- Rouse and Rouse (1982), "*Statim invenire*"; Blair (2010), *Too Much to Know*; Duncan, history of the index; Parkes on *ordinatio* and *compilatio*: library. Case 3 rests on abstracts. +- Pesce (1999), *Guido d'Arezzo's Regule rithmice, Prologus in antiphonarium, and Epistola ad Michahelem* (Institute of Mediaeval Music): the only way to settle what Guido says about learning time. Also Mengozzi (2010) and Strunk and Treitler (1998) for the translation. +- Walker and Sarbaugh (1993), *Ethnohistory* 40(1) (JSTOR); Parins (2013); Cushman (2011 book); Bender (2002): for Cherokee literacy figures with page numbers. +- Ledyard (1966), *The Korean Language Reform of 1446* (Berkeley dissertation); Kim-Renaud (1997); Sampson (1985): the Hangul design and uptake claims need these. +- Gigerenzer and Hoffrage (1995) and Peters et al. (2006, 2007), full texts: effect sizes and populations. +- Sangster (2024), Danna (2021, 2023), Power (2000), Ryan (2014), Liebowitz and Margolis (1990): full text blocked. +- Swetz, *Capitalism and Arithmetic*; Menninger; Crosby, *The Measure of Reality*; Sombart; Weber; Pacioli's *Summa*; David (1985): not attempted beyond search. + +**Not researched.** +- Case 6 (maps, coordinate grids, logarithm tables, slide rule). +- Web browser and hyperlink, touch interfaces, and any adoption evidence for them (no usable source found). +- Cree syllabics and the Vai script: other fast-literacy scripts that would test the Cherokee case. Unseth (2016) says Sequoyah's script influenced 21 scripts; I did not follow it. +- The US Navy 1944 Dvorak study; Shavian adoption measures; Swiss or other cases of successful reform (Bunčić names the Russian reform of 1917–1918 as a success). +- Whether anyone has argued explicitly that "word separation is the paradigm design move"; that is facet Z's search. + +**Unverified items to carry forward as unverified.** +- Guido, "ten years to one or two". +- "Within months" for Cherokee literacy; "70–90% by 1830". +- The Hangul "morning/ten days" wording: attributed to the 1446 text by Handel's lecture abstract; locator not checked. +- Danna et al. is a working paper; its regression identification is unchecked. +- Cajori was read by keyword, not in full; section numbers are from the OCR. +- Google Books snippet text for Chrisomalis can be truncated mid-sentence. Check each quoted fragment in the book before quoting it in the paper. + +## Changes after full-text cards (2026-10-13) + +- `ryan2014historical`: removed the implication that Ryan supports deflating the causal claim or restricting double entry to control of debt and agents; Ryan defends the causal claim at the macro level (pp. 92-96); cite for the Yamey summary and the "about 200 years" point (p. 86, n. 2). Read status "read in full (card)". +- `sangster2024emergence`: the 1211 / "nearly 280 years" claim qualified (input definition only; output definition gives 1340, p. 501; contra entries rarely survive, p. 506; origin put in the twelfth century, p. 520); Pacioli comparison marked as the facet's, since the text never discusses the Summa; does not rebut the causal claim (n. 154, p. 523); abstract sentence quoted exactly. Table row 4, the "Pacioli and double entry" claim, and the "no single origin" line no longer attribute the Pacioli story to Sangster. +- `sangster2025reboot`: the 1211 and entity-wide details reattributed to Sangster 2024 (here 1211 is the earliest paragraph-format ledger, p. 378; confirmed dates 1296 and 1299); Sangster qualifies how Yamey can be cited (pp. 373-374) rather than rejecting the Sombart criticism. Read status "read in full (card)". +- `biggs2020decoding`: the lag of recorded notation behind working notation is Biggs's premise (p. 3), not a finding (pp. 8-9); journal named as *BSHM Bulletin* in the manuscript; read status "read in full (card)". +- `buijsman2020numerals`: removed "The cost saving of place value shows up with practice and with column layout" (not in the text; the column point is Menary's layouts, pp. 3786-3787, with no cost figure); read status "read in full (card)". +- `buncic2017factors`: "the revolutionary bundle" removed from the counter-cases row (the phrase is the facet's); installed base kept (p. 36); read status "read in full (card)". +- `nothaft2020satanic`: made explicit that Padua is 1331, not 1348, and that fraud-prevention is listed as "popular" but not endorsed (p. 15, n. 22); 1299 motive stays "speculation"; read status "read in full (card)". +- `danna2022numerical`: manual counts added (over 1,290 in the dataset, p. 9; 1,054 in the 815-city sample, p. 13); main limitation added; read status "read in part (card)" (main text in full, appendix skimmed). +- `haines2008origins`: added that Guido's own contribution is dated only through later manuscripts (p. 345); read status "read in full (card)". +- `ivanova2026process`: abstract-only status replaced; added the body's word-separation material (no word division from 1443, p. 137; 1933 orthography requires separate words, p. 144; spaces chapter, p. 145) and the 1933 point in table row 8b. +- `dacosta2018tabulae`, `strunk1998guido`, `shneiderman1983direct`, `margono1987study`, `schlimm2008modeling`: read status updated to "read in full (card)"; da Costa's "marketed ... time-saving innovations" noted as uncheckable in the manuscript; Strunk's Prologue excerpt does describe the staff (p. 3); Margono's satisfaction composite and sub-five-minute test noted. +- `hangeulmuseum`, `park2025dokkyo`: the page gives no source for the 78% / 4.1% figures; the Park page contains no counts or rates. + +Synthesis verdicts changed: (1) table row 4 and case-4 bearing (b): "the big causal claim deflated" / "contested for decades by Yamey, and recent economic historians restrict double entry to control of debt and agents" became "the big causal claim contested, not rebutted" / contested by Yamey, defended at the macro level by Ryan, with the debt-and-agents restriction attributed to economic historians as Sangster reports them. (2) Claims list: "A textbook deflated causal claim" became "contested by Yamey and defended ... by Ryan; Sangster does not rebut it". (3) "Sangster on Pacioli" became "Sangster on the definitional tradition behind the Pacioli story". (4) Counter-cases row: "installed base, and the revolutionary bundle" became "timing and installed base". diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_G_nadella_lineage.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_G_nadella_lineage.bib new file mode 100644 index 00000000..fbe36d5f --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_G_nadella_lineage.bib @@ -0,0 +1,582 @@ +% Facet G, Nadella lineage. Agent-drafted 2026-10-12. Every entry carries a read status in `note`. +% DOIs resolved against Crossref on 2026-10-12 unless the note says otherwise. arXiv IDs checked against the arXiv API. + +% ---------- The post, Nadella's other posts, companion and coverage ---------- + +@misc{nadella2026models, + author = {Nadella, Satya}, + title = {Models as Insider Risks in the Super Intelligence Era}, + year = {2026}, + month = oct, + howpublished = {sn scratchpad (personal blog), 10 Oct 2026; also posted on X}, + url = {https://snscratchpad.com/posts/models-as-insider-risks/}, + note = {Read in full (blog page fetched 12 Oct 2026; local paragraph-numbered capture P1-P21 from the X post, https://x.com/satyanadella/status/2108931348857827686, matched the blog text on every paragraph checked; P20 matched at its end only; P22 is a capture marker, not text).} +} + +@misc{nadella2026frontier, + author = {Nadella, Satya}, + title = {A frontier without an ecosystem is not stable}, + year = {2026}, + month = jun, + howpublished = {sn scratchpad, 14 Jun 2026}, + url = {https://snscratchpad.com/posts/frontier-ecosystem/}, + note = {Read in full, 12 Oct 2026.} +} + +@misc{nadella2026reverse, + author = {Nadella, Satya}, + title = {The Reverse Information Paradox}, + year = {2026}, + month = jul, + howpublished = {sn scratchpad, 12 Jul 2026}, + url = {https://snscratchpad.com/posts/reverse-information-paradox/}, + note = {Read in full, 12 Oct 2026.} +} + +@misc{nadella2025looking, + author = {Nadella, Satya}, + title = {Looking Ahead to 2026}, + year = {2025}, + month = dec, + howpublished = {sn scratchpad, 29 Dec 2025}, + url = {https://snscratchpad.com/posts/looking-ahead-2026/}, + note = {Read in full, 12 Oct 2026.} +} + +@misc{constantin2026nadella, + author = {Constantin, Ana Maria}, + title = {Satya Nadella says companies should treat AI models as insider risks}, + year = {2026}, + howpublished = {The Next Web, Oct 2026 (exact date not shown in the fetched text)}, + url = {https://thenextweb.com/news/nadella-ai-models-insider-risks-emergency-brake}, + note = {Read in full (card: library/constantin2026nadella.md). News report dated 12 Oct 2026, 7:55 am. Its claims about Suleyman, Musk, Rep. Lieu, the "Super Intelligence Force" and an Anthropic incident are TNW's reporting and were not independently verified.} +} + +@misc{kumari2026nadella, + author = {Kumari, Sweta}, + title = {Satya Nadella warns of AI insider risks, seeks stronger safety controls}, + year = {2026}, + howpublished = {Business Standard, 12 Oct 2026}, + url = {https://www.business-standard.com/technology/artificial-intelligence/satya-nadella-warns-of-ai-insider-risks-seeks-stronger-safety-controls-126101200572_1.html}, + note = {Read in full (card: library/kumari2026nadella.md). News report; quotes Suleyman and Levie posts. Attributes to Nadella approval-step statements that are not in the post.} +} + +@misc{abovesecurity2026satya, + author = {{Above Security}}, + title = {Satya Nadella Is Right: We Need to Treat AI Models as Insider Risks}, + year = {2026}, + howpublished = {Security Boulevard, 10 Oct 2026}, + url = {https://securityboulevard.com/2026/10/satya-nadella-is-right-we-need-to-treat-ai-models-as-insider-risks/}, + note = {Read in full via WebFetch summary (12 Oct 2026). Vendor marketing; the vendor sells AI investigation agents. Use only as evidence of reception.} +} + +@misc{linzainescu2026microsofts, + author = {Linzainescu, Gabriela}, + title = {Microsoft's Nadella Wants An Emergency Brake On Every AI Model}, + year = {2026}, + howpublished = {Forbes, 11 Oct 2026}, + url = {https://www.forbes.com/sites/gabrielalinzainescu/2026/10/11/microsofts-nadella-wants-an-emergency-brake-on-every-ai-model/}, + note = {Not accessible (HTTP 403 to WebFetch and curl). Known only from a search-engine summary. Do not quote.} +} + +@misc{jakkal2026secure, + author = {Jakkal, Vasu}, + title = {Secure agentic AI for your Frontier Transformation}, + year = {2026}, + howpublished = {Microsoft Security Blog, 9 Mar 2026}, + url = {https://www.microsoft.com/en-us/security/blog/2026/03/09/secure-agentic-ai-for-your-frontier-transformation/}, + note = {Read in full (card: library/jakkal2026secure.md). Marketing page. Earlier Microsoft product context (Entra Agent ID, Agent 365, Insider Risk Management for agents). Not a companion to the post; it predates it by seven months.} +} + +@misc{anthropic2026how, + author = {{Anthropic}}, + title = {How we built Claude Code auto mode: a safer way to skip permissions}, + year = {2026}, + howpublished = {Anthropic Engineering blog, 25 Mar 2026}, + url = {https://www.anthropic.com/engineering/claude-code-auto-mode}, + note = {Read in full (card: library/anthropic2026how.md). Vendor report on its own product, not peer reviewed; the 93% is a rate of approvals, not a measure of inattention. The page does not state that the drafting model is an Anthropic model.} +} + +% ---------- Security lineage ---------- + +@techreport{anderson1972computer, + author = {Anderson, James P.}, + title = {Computer Security Technology Planning Study}, + institution = {Electronic Systems Division, Air Force Systems Command, Hanscom Field}, + number = {ESD-TR-73-51, Vol. I}, + year = {1972}, + month = oct, + url = {https://csrc.nist.gov/csrc/media/publications/conference-paper/1998/10/08/proceedings-of-the-21st-nissc-1998/documents/early-cs-papers/ande72a.pdf}, + note = {Read in full (card: library/anderson1972computer.md). Volume I (NIST CSRC text copy, 43 pp.); Volume II not opened. Table-of-contents section numbers differ from the body. PDF at dxd_foundation_local/pdfs/G/anderson1972_csrc.pdf.} +} + +@article{saltzer1975protection, + author = {Saltzer, Jerome H. and Schroeder, Michael D.}, + title = {The Protection of Information in Computer Systems}, + journal = {Proceedings of the IEEE}, + volume = {63}, + number = {9}, + pages = {1278--1308}, + year = {1975}, + doi = {10.1109/PROC.1975.9939}, + url = {https://web.mit.edu/Saltzer/www/publications/protection/}, + note = {Read in part (card: library/saltzer1975protection.md). Full text of the authors' MIT web edition, Sections I-III; Section II read for argument only; footnotes page not read.} +} + +@techreport{rose2020zero, + author = {Rose, Scott and Borchert, Oliver and Mitchell, Stu and Connelly, Sean}, + title = {Zero Trust Architecture}, + institution = {National Institute of Standards and Technology}, + number = {NIST SP 800-207}, + year = {2020}, + month = aug, + doi = {10.6028/NIST.SP.800-207}, + note = {Read in part (card: library/rose2020zero.md). Sections 6, 7.3.1-7.3.3 and most of Appendix B only skimmed.} +} + +@techreport{cert2022common, + author = {{CERT National Insider Threat Center}}, + title = {Common Sense Guide to Mitigating Insider Threats, Seventh Edition}, + institution = {Software Engineering Institute, Carnegie Mellon University}, + year = {2022}, + url = {https://www.sei.cmu.edu/documents/5777/5692_CSG_Book_20-optimized.pdf}, + note = {Read in part (card: library/cert2022common.md). Best Practices 10, 11, 12, 14, 15, 17 and most of 21 read in full; bodies of Best Practices 1-9, 13, 16, 18-20 and 22 skimmed or not read. Report number not confirmed (CMU/SEI-2022-TR-003 is unverified). Some sources title it "Common Sense Guide to Managing Insider Risk" (DTIC AD1150273).} +} + +% ---------- AI-safety lineage ---------- + +@inproceedings{greenblatt2024ai, + author = {Greenblatt, Ryan and Shlegeris, Buck and Sachan, Kshitij and Roger, Fabien}, + title = {{AI} Control: Improving Safety Despite Intentional Subversion}, + booktitle = {Proceedings of the 41st International Conference on Machine Learning}, + year = {2024}, + eprint = {2312.06942}, + archiveprefix = {arXiv}, + url = {https://arxiv.org/abs/2312.06942}, + note = {Read in part (card: library/greenblatt2024ai.md). arXiv v5 (23 Jul 2024); appendices (pp. 21-47) skimmed.} +} + +@misc{turpin2023language, + author = {Turpin, Miles and Michael, Julian and Perez, Ethan and Bowman, Samuel R.}, + title = {Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting}, + year = {2023}, + eprint = {2305.04388}, + archiveprefix = {arXiv}, + note = {Read in part (card: library/turpin2023language.md). NeurIPS 2023 paper, arXiv v2; per-task tables in Appendices D-G skimmed.} +} + +@misc{lanham2023measuring, + author = {Lanham, Tamera and others}, + title = {Measuring Faithfulness in Chain-of-Thought Reasoning}, + year = {2023}, + eprint = {2307.13702}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/lanham2023measuring.md). arXiv v1; appendices of prompts and extra plots skimmed.} +} + +@misc{korbak2025chain, + author = {Korbak, Tomek and Balesni, Mikita and others}, + title = {Chain of Thought Monitorability: A New and Fragile Opportunity for {AI} Safety}, + year = {2025}, + eprint = {2507.11473}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/korbak2025chain.md). arXiv v2, 7 Dec 2025. Multi-institution position paper; not an experiment.} +} + +@misc{chen2025reasoning, + author = {Chen, Yanda and Benton, Joe and Radhakrishnan, Ansh and Uesato, Jonathan and others}, + title = {Reasoning Models Don't Always Say What They Think}, + year = {2025}, + eprint = {2505.05410}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/chen2025reasoning.md). arXiv v1, 8 May 2025.} +} + +@misc{baker2025monitoring, + author = {Baker, Bowen and Huizinga, Joost and Gao, Leo and Dou, Zehao and others}, + title = {Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation}, + year = {2025}, + eprint = {2503.11926}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/baker2025monitoring.md). arXiv v1, 14 March 2025; Appendix A figures and monitor prompts skimmed.} +} + +@misc{kutasov2025shade, + author = {Kutasov, Jonathan and Sun, Yuqi and Colognese, Paul and van der Weij, Teun and others}, + title = {{SHADE}-Arena: Evaluating Sabotage and Monitoring in {LLM} Agents}, + year = {2025}, + eprint = {2506.15740}, + archiveprefix = {arXiv}, + note = {Read in part (card: library/kutasov2025shade.md). arXiv v2; body, human-baseline Appendix L and several other appendices read; others skimmed.} +} + +@misc{lynch2025agentic, + author = {Lynch, Aengus and Wright, Benjamin and Larson, Caleb and Ritchie, Stuart J. and others}, + title = {Agentic Misalignment: How {LLMs} Could Be Insider Threats}, + year = {2025}, + eprint = {2510.05179}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/lynch2025agentic.md). arXiv v2; the PDF holds the body and references only. Anthropic-authored; the drafting model is an Anthropic model.} +} + +@misc{hao2024training, + author = {Hao, Shibo and Sukhbaatar, Sainbayar and Su, DiJia and others}, + title = {Training Large Language Models to Reason in a Continuous Latent Space}, + year = {2024}, + eprint = {2412.06769}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/hao2024training.md). arXiv v4 (Aug 2026), not the December 2024 text that Korbak et al. cite. The "Coconut" paper; cited by Korbak et al.; it does not use the word "neuralese".} +} + +@misc{kokotajlo2025ai, + author = {Kokotajlo, Daniel and Alexander, Scott and Larsen, Thomas and Lifland, Eli and Dean, Romeo}, + title = {{AI} 2027}, + year = {2025}, + howpublished = {ai-2027.com}, + url = {https://ai-2027.com/}, + note = {Read in part (card: library/kokotajlo2025ai.md). Scenario text read; footnotes, endings, supplements and dashboard not read. A forecasting scenario, not research evidence.} +} + +@misc{duzan2026chain, + author = {Duzan, Agatha and Cooper Stickland, Asa}, + title = {Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings}, + year = {2026}, + eprint = {2608.04735}, + archiveprefix = {arXiv}, + note = {Read in part (card: library/duzan2026chain.md). Body pp. 1-11 read; appendices B-H skimmed. Preprint, not peer reviewed.} +} + +@misc{gema2026chain, + author = {Gema, Aryo Pradipta and Rajani, Neel and Saxena, Rohit and Kwan, Wai-Chung and others}, + title = {Chain-of-Thought Faithfulness of Reasoning Models Varies with Where and How Preference Cues Are Delivered}, + year = {2026}, + eprint = {2608.29464}, + archiveprefix = {arXiv}, + note = {Read in part (card: library/gema2026chain.md). Body pp. 1-17 read; appendices A-M skimmed. Preprint.} +} + +@misc{little2026length, + author = {Little, Bryce}, + title = {Length Penalties Make Chain-of-Thought Less Monitorable}, + year = {2026}, + eprint = {2607.09786}, + archiveprefix = {arXiv}, + note = {Read in part (card: library/little2026length.md). Body and references read; appendices partly skimmed. Preprint; small open models.} +} + +@misc{wang2026monitorbench, + author = {Wang, Han and Sun, Yifan and Ko, Brian and Talati, Mann and others}, + title = {{MonitorBench}: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models}, + year = {2026}, + eprint = {2603.28590}, + archiveprefix = {arXiv}, + note = {Read in part (card: library/wang2026monitorbench.md). arXiv v4 carries a COLM 2026 header (acceptance not verified beyond the header); prompt listings and task examples not read.} +} + +@misc{jia2026faithfulness, + author = {Jia, Jinghan and Benton, Joe and Easley, Eric and others}, + title = {Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning}, + year = {2026}, + eprint = {2605.24286}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/jia2026faithfulness.md). Preprint.} +} + +@inproceedings{chan2024visibility, + author = {Chan, Alan and Ezell, Carson and Kaufmann, Max and Wei, Kevin and Hammond, Lewis and Bradley, Herbie and Bluemke, Emma and Rajkumar, Nitarshan and Krueger, David and Kolt, Noam and Heim, Lennart and Anderljung, Markus}, + title = {Visibility into {AI} Agents}, + booktitle = {Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency}, + pages = {958--973}, + year = {2024}, + doi = {10.1145/3630106.3658948}, + note = {Read in full (card: library/chan2024visibility.md). arXiv 2401.13138 v6.} +} + +@misc{chan2025infrastructure, + author = {Chan, Alan and Wei, Kevin and Huang, Sihao and others}, + title = {Infrastructure for {AI} Agents}, + year = {2025}, + eprint = {2501.10114}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/chan2025infrastructure.md). arXiv v3, 19 Jun 2025. A research agenda with no data.} +} + +% ---------- Human factors and oversight ---------- + +@article{bainbridge1983ironies, + author = {Bainbridge, Lisanne}, + title = {Ironies of automation}, + journal = {Automatica}, + volume = {19}, + number = {6}, + pages = {775--779}, + year = {1983}, + doi = {10.1016/0005-1098(83)90046-8}, + note = {Read in full in an earlier run (repo card submissions/algorithmacy_design_ethics/literature/library/bainbridge1983.md, 20 quotations page-verified). Not re-opened in this run; quotations below are taken from the card.} +} + +@article{mackworth1948breakdown, + author = {Mackworth, Norman H.}, + title = {The breakdown of vigilance during prolonged visual search}, + journal = {Quarterly Journal of Experimental Psychology}, + volume = {1}, + number = {1}, + pages = {6--21}, + year = {1948}, + doi = {10.1080/17470214808416738}, + note = {Not accessible (paywalled, no abstract in OpenAlex or Semantic Scholar). DOI resolved in Crossref. Cited only as the origin of the vigilance decrement.} +} + +@article{warm2008vigilance, + author = {Warm, Joel S. and Parasuraman, Raja and Matthews, Gerald}, + title = {Vigilance requires hard mental work and is stressful}, + journal = {Human Factors}, + volume = {50}, + number = {3}, + pages = {433--441}, + year = {2008}, + doi = {10.1518/001872008X312152}, + note = {Abstract only (OpenAlex, 12 Oct 2026). Paywalled.} +} + +@article{parasuraman2010complacency, + author = {Parasuraman, Raja and Manzey, Dietrich H.}, + title = {Complacency and bias in human use of automation: An attentional integration}, + journal = {Human Factors}, + volume = {52}, + number = {3}, + pages = {381--410}, + year = {2010}, + doi = {10.1177/0018720810376055}, + note = {Read in full (card: library/parasuraman2010complacency.md). Journal-layout PDF, printed pp. 381-410.} +} + +@article{endsley1995out, + author = {Endsley, Mica R. and Kiris, Esin O.}, + title = {The out-of-the-loop performance problem and level of control in automation}, + journal = {Human Factors}, + volume = {37}, + number = {2}, + pages = {381--394}, + year = {1995}, + doi = {10.1518/001872095779064555}, + note = {Abstract only (OpenAlex, 12 Oct 2026; the repo card also records abstract only).} +} + +@article{ancker2017effects, + author = {Ancker, Jessica S. and Edwards, Alison and Nosal, Sarah and Hauser, Diane and Mauer, Elizabeth and Kaushal, Rainu}, + title = {Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system}, + journal = {BMC Medical Informatics and Decision Making}, + volume = {17}, + pages = {36}, + year = {2017}, + doi = {10.1186/s12911-017-0430-8}, + note = {Read in full (card: library/ancker2017effects.md).} +} + +@inproceedings{alahmadi2022false, + author = {Alahmadi, Bushra A. and Axon, Louise and Martinovic, Ivan}, + title = {99\% False Positives: A Qualitative Study of {SOC} Analysts' Perspectives on Security Alarms}, + booktitle = {31st USENIX Security Symposium}, + year = {2022}, + url = {https://www.usenix.org/conference/usenixsecurity22/presentation/alahmadi}, + note = {Read in full (card: library/alahmadi2022false.md). Published open-access proceedings PDF, pp. 2783-2800. No DOI.} +} + +@article{layman2023controlled, + author = {Layman, Lucas and others}, + title = {A Controlled Experiment on the Impact of Intrusion Detection False Alarm Rate on Analyst Performance}, + journal = {Proceedings of the Human Factors and Ergonomics Society Annual Meeting}, + volume = {67}, + pages = {220--225}, + year = {2023}, + doi = {10.1177/21695067231192573}, + note = {Read in full (card: library/layman2023controlled.md). arXiv v1 preprint; DOI resolved in Crossref.} +} + +@article{tariq2025alert, + author = {Tariq, Shahroz and others}, + title = {Alert Fatigue in Security Operations Centres: Research Challenges and Opportunities}, + journal = {ACM Computing Surveys}, + volume = {57}, + pages = {1--38}, + year = {2025}, + doi = {10.1145/3723158}, + note = {Abstract only (Consensus index). Full text blocked at the ACM Digital Library. Author list not verified beyond the first author.} +} + +@article{green2022flaws, + author = {Green, Ben}, + title = {The flaws of policies requiring human oversight of government algorithms}, + journal = {Computer Law \& Security Review}, + volume = {45}, + pages = {105681}, + year = {2022}, + doi = {10.1016/j.clsr.2022.105681}, + note = {Read in full (card: library/green2022flaws.md). Publisher's open-access PDF; locators are PDF pages.} +} + +@article{crootof2023humans, + author = {Crootof, Rebecca and Kaminski, Margot E. and Price II, W. Nicholson}, + title = {Humans in the Loop}, + journal = {Vanderbilt Law Review}, + volume = {76}, + number = {2}, + pages = {429--510}, + year = {2023}, + url = {https://wp0.vanderbilt.edu/lawreview/wp-content/uploads/sites/278/2023/03/Humans-in-the-Loop.pdf}, + note = {Read in full (card: library/crootof2023humans.md). Law-review proof; printed pages 429-510.} +} + +@article{elish2019moral, + author = {Elish, Madeleine Clare}, + title = {Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction}, + journal = {Engaging Science, Technology, and Society}, + volume = {5}, + pages = {40--60}, + year = {2019}, + doi = {10.17351/ests2019.260}, + note = {Read in full (card: library/elish2019moral.md). Journal open-access PDF, pp. 40-60.} +} + +@article{santonidesio2018meaningful, + author = {Santoni de Sio, Filippo and van den Hoven, Jeroen}, + title = {Meaningful Human Control over Autonomous Systems: A Philosophical Account}, + journal = {Frontiers in Robotics and {AI}}, + volume = {5}, + pages = {15}, + year = {2018}, + doi = {10.3389/frobt.2018.00015}, + note = {Read in full (card: library/santonidesio2018meaningful.md).} +} + +@inproceedings{sterz2024quest, + author = {Sterz, Sarah and Baum, Kevin and Biewer, Sebastian and Hermanns, Holger and Lauber-R{\"o}nsberg, Anne and Meinel, Philip and Langer, Markus}, + title = {On the Quest for Effectiveness in Human Oversight: Interdisciplinary Perspectives}, + booktitle = {Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency}, + year = {2024}, + doi = {10.1145/3630106.3659051}, + note = {Read in full (card: library/sterz2024quest.md). arXiv 2404.04059 v2.} +} + +@article{laux2024institutionalised, + author = {Laux, Johann}, + title = {Institutionalised distrust and human oversight of artificial intelligence: towards a democratic design of {AI} governance under the {European Union AI Act}}, + journal = {AI \& Society}, + volume = {39}, + number = {6}, + pages = {2853--2866}, + year = {2024}, + doi = {10.1007/s00146-023-01777-z}, + note = {Read in full (card: library/laux2024institutionalised.md).} +} + +@article{langer2024effective, + author = {Langer, Markus and others}, + title = {Effective Human Oversight of {AI}-Based Systems: A Signal Detection Perspective on the Detection of Inaccurate and Unfair Outputs}, + journal = {Minds and Machines}, + volume = {35}, + year = {2024}, + doi = {10.1007/s11023-024-09701-0}, + note = {Read in full (card: library/langer2024effective.md). Crossref gives volume 35; online-first year 2024.} +} + +@misc{chen2026comparing, + author = {Chen, Chaoran and Zhang, Zhiping and others}, + title = {Comparing Human Oversight Strategies for Computer-Use Agents}, + year = {2026}, + eprint = {2604.04918}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/chen2026comparing.md). Preprint (arXiv v2, 12 Sep 2026).} +} + +@misc{yu2026habituation, + author = {Yu, Haoran and Liu, Lifei and others}, + title = {Habituation at the Gate: Rising Approval and Declining Scrutiny in Human Review of {AI} Agent Code}, + year = {2026}, + eprint = {2606.22721}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/yu2026habituation.md). Preprint; observational data from the AIDev dataset; some internal inconsistencies noted in the card.} +} + +@misc{grundemclaughlin2026overseeing, + author = {Grunde-McLaughlin, Madeleine and Mozannar, Hussein and others}, + title = {Overseeing Agents Without Constant Oversight: Challenges and Opportunities}, + year = {2026}, + eprint = {2602.16844}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/grundemclaughlin2026overseeing.md). Preprint (arXiv v1).} +} + +@misc{mitchell2026agents, + author = {Mitchell, Margaret and Ghosh, Avijit and others}, + title = {{AI} Agents Push Humans Out of the Loop}, + year = {2026}, + eprint = {2608.23642}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/mitchell2026agents.md). Position paper (preprint v3).} +} + +@misc{turan2026oversight, + author = {Turan, Emre}, + title = {Oversight Has a Capacity: Calibrating Agent Guards to a Subjective, Fatiguing Human}, + year = {2026}, + eprint = {2606.08919}, + archiveprefix = {arXiv}, + note = {Read in full (card: library/turan2026oversight.md). Its fatigue result is a model of the reviewer, not data from reviewers; its disagreement result (kappa = 0.52) is from LLM personas. Do not cite as human evidence.} +} + +@misc{eu2024regulation, + author = {{European Parliament and Council of the European Union}}, + title = {Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Article 14 (Human oversight)}, + year = {2024}, + howpublished = {Official Journal of the European Union, 12 Jul 2024}, + url = {https://artificialintelligenceact.eu/article/14/}, + note = {Read in full (Article 14, from the Future of Life Institute annotated copy, 12 Oct 2026). The EUR-Lex official text returned an empty response and was not compared.} +} + +% ---------- Transparency critique ---------- + +@article{ananny2018seeing, + author = {Ananny, Mike and Crawford, Kate}, + title = {Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability}, + journal = {New Media \& Society}, + volume = {20}, + number = {3}, + pages = {973--989}, + year = {2018}, + doi = {10.1177/1461444816676645}, + note = {Abstract only (OpenAlex abstract, 12 Oct 2026). The repo card also records abstract only; no open copy found.} +} + +@article{rieder2020towards, + author = {Rieder, Bernhard and Hofmann, Jeanette}, + title = {Towards platform observability}, + journal = {Internet Policy Review}, + volume = {9}, + number = {4}, + year = {2020}, + doi = {10.14763/2020.4.1535}, + note = {Read in full (card: library/rieder2020towards.md).} +} + +@article{stohl2016managing, + author = {Stohl, Cynthia and Stohl, Michael and Leonardi, Paul M.}, + title = {Managing Opacity: Information Visibility and the Paradox of Transparency in the Digital Age}, + journal = {International Journal of Communication}, + volume = {10}, + pages = {123--137}, + year = {2016}, + url = {https://ijoc.org/index.php/ijoc/article/view/4466}, + note = {Read in full (card: library/stohl2016managing.md). No DOI.} +} + +@inproceedings{kroll2021outlining, + author = {Kroll, Joshua A.}, + title = {Outlining Traceability: A Principle for Operationalizing Accountability in Computing Systems}, + booktitle = {Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency}, + pages = {758--771}, + year = {2021}, + doi = {10.1145/3442188.3445937}, + note = {Read in full (card: library/kroll2021outlining.md). arXiv 2101.09385 v1.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_G_nadella_lineage.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_G_nadella_lineage.md new file mode 100644 index 00000000..c2268991 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_G_nadella_lineage.md @@ -0,0 +1,484 @@ +# G — The Nadella post and its lineage + +*Corrected on 2026-10-12 against `../audit/AUDIT_03.md`: 36 factual fixes; verbatim quotation of the Nadella post cut to 39 words.* + +*Facet G of the foundation paper for Decision Experience Design (PHD1750-3 / HF700-1). Agent-drafted on 2026-10-12. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** (1) Nadella's post of 10 Oct 2026, "Models as Insider Risks" (full title in B1), read paragraph by paragraph, with its companion material, its reception in the first two days and his earlier related posts. (2) The security lineage the post invokes: Anderson (1972), Saltzer and Schroeder (1975), the CERT insider-threat guide, NIST zero trust. (3) The AI-safety lineage: AI control, chain-of-thought faithfulness and monitorability, "neuralese", agent visibility. (4) The human-factors and oversight literature on the role this architecture gives a person. (5) The transparency critique and the "observability" literature. (6) The convergences with the three design-ethics principles. (7) The three counts against the post, tested against its text. (8) A scope caution. + +**The question the synthesis answers.** How accurately can the paper say that Nadella's post is "the exact opposite" of what algorithmacy proposes, and what must the author concede, soften or drop to make the contrast honest? + +**Legend for read status.** *Read in full* = I read the whole text this session, or the entry reuses a repo card built from the full text (the card is named). *Read in part* = I read the passages named, not the whole work. *Abstract only* = database or arXiv abstract. *Not accessible* = no text and no abstract. Where an entry says "repo card", I did not reopen the source this run, because it is paywalled; the quotation locators come from the card. + +**Handling of the post.** The paragraph numbers P1–P21 are the capturer's, from `dxd_foundation_local/nadella_2026-10-10_models_as_insider_risks.txt`. The same text is public on Nadella's blog, which makes it quotable and checkable: https://snscratchpad.com/posts/models-as-insider-risks/. I compared the blog page with the capture paragraph by paragraph. P1–P19 and P21 match at start and end; P20 matches at its end only (a label difference at its start); P22 is the capturer's end marker. Quotations from the post below are short fragments (15 words or fewer, 45 words in total for this file); everything else is paraphrase with paragraph numbers. + +--- + +## (a) Search log + +All searches ran on 2026-10-12. "Hits" is what the source reported, or target found / not found. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | WebSearch | `Nadella "Models as Insider Risks" Super Intelligence Era` | 10 | FourWeekMBA, Security Boulevard, TheStreet, Business Standard, OfficeChai, Forbes and others. Search summary says some pages add unverifiable detail. | +| 2 | WebSearch | `Satya Nadella insider risk AI models post containment "separate the supply of intelligence from the authority"` | 9 | Confirms phrase in OfficeChai; summary warns wording varies across outlets. | +| 3 | curl | `https://snscratchpad.com/`, `/posts/models-as-insider-risks/`, `/posts/frontier-ecosystem/`, `/posts/reverse-information-paradox/`, `/posts/looking-ahead-2026/` | 5 of 5 fetched | The post's primary public home; it is on Nadella's personal blog with an X copy. The same blog holds three earlier posts. | +| 4 | WebFetch / curl | Forbes (Linzainescu, 11 Oct), TheStreet, Business Standard, TNW, Android Headlines, RuntimeWire, Security Boulevard | Read: TNW, Business Standard (curl), Security Boulevard, RuntimeWire (stub). **Blocked:** Forbes (403 both tools), TheStreet (404 curl, 403 WebFetch), Android Headlines (403) | I did not try to get around the blocks. | +| 5 | WebSearch | `Nadella insider risk essay reaction critics human oversight "emergency brake" Suleyman Levie` | 9 | Reactions named: Suleyman, Levie, Musk, Rep. Lieu. Summary says no organised criticism found. | +| 6 | WebSearch | `Nadella "insider risk" superintelligence post analysis OR critique OR "what Nadella gets wrong" OR "missing" substack OR lawfare OR lesswrong` | 9 | Only news coverage. No Substack, Lawfare, or LessWrong critique found. | +| 7 | WebSearch | `Nadella insider risks essay critique "human readable" logs oversight analysts alert fatigue AI agents` | 9 | Surfaces a podcast appearance (All-In) and aggregator pages; none addresses human usability of logs. | +| 8 | WebSearch | `Microsoft blog Agent 365 Entra Agent ID "insider risk" AI agents October 2026 Nadella containment announcement` | 9 | No Oct 2026 Microsoft companion post found. Finds the March 2026 Microsoft Security blog and Learn pages. | +| 9 | WebFetch | Microsoft Security blog 2026-03-09 (Jakkal) | 1 | Read as a summary: agent identity, least privilege, Insider Risk Management for agents. | +| 10 | curl | Anderson 1972: seclab.cs.ucdavis.edu `ande72.pdf` | 1 | 142-page scan, no text layer. Not OCRed. | +| 11 | curl | Anderson 1972: NIST CSRC `ande72a.pdf` | 1 | 43-page text copy of ESD-TR-73-51 Vol. I. Used. | +| 12 | curl | Saltzer and Schroeder: `web.mit.edu/Saltzer/www/publications/protection/Basic.html` | 1 | Used. Two other URLs I tried (an MIT PDF path, a Virginia course page) were wrong or not the paper. | +| 13 | curl | NIST SP 800-207 PDF (nvlpubs.nist.gov) | 1 | 59 pp.; used. | +| 14 | WebSearch | `CERT Common Sense Guide to Mitigating Insider Threats seventh edition CMU/SEI-2022-TR-003` | 9 | Report number not confirmed by any hit. | +| 15 | curl | SEI `5692_CSG_Book_20-optimized.pdf` | 1 | 175 pp.; read in part. | +| 16 | curl / arXiv | Greenblatt et al. 2312.06942; Turpin 2305.04388; Lanham 2307.13702; Korbak 2507.11473; Chan 2401.13138 | 5 of 5 | PDFs under `dxd_foundation_local/pdfs/G/`. | +| 17 | arXiv API | `id_list=` 2608.04735, 2608.29464, 2607.09786, 2603.28590, 2605.24286, 2503.11926, 2505.05410, 2506.15740, 2510.05179, 2412.06769, 2501.10114, 2604.04918, 2606.22721, 2608.23642, 2602.16844, 2606.08919, 2609.24242 | 17 of 17 found | Titles, dates and first authors confirmed. Abstracts only for all except the five first listed in row 16. | +| 18 | WebSearch | `chain-of-thought monitorability 2026 faithfulness reasoning models new results OR "latent reasoning" monitoring evasion` | 9 | Led to the 2026 preprints in row 17. The search summary found nothing on latent-reasoning evasion. | +| 19 | WebSearch | `"neuralese" AI 2027 Kokotajlo neuralese recurrence and memory chain of thought opaque` | 9 | Secondary; I read the primary ai-2027.com page (row 20). | +| 20 | curl | `https://ai-2027.com/` | 1 | Passage defining "neuralese recurrence and memory" found and read. | +| 21 | Crossref | `/works/` for Bainbridge, Mackworth, Warm et al., Parasuraman and Manzey, Endsley and Kiris, Santoni de Sio and van den Hoven, Green, Elish, Ananny and Crawford, Rieder and Hofmann, Kroll, Ancker et al., Chan et al., Saltzer and Schroeder, NIST SP 800-207, Layman et al., Langer et al., Tariq et al., Sterz et al. | all resolved | Titles, containers, volumes and pages compared. Ananny and Crawford shows an issued year of 2016 (online first); the volume is 2018. Langer et al. shows volume 35 with online-first year 2024. | +| 22 | Crossref `query.bibliographic` | Sterz et al.; Tariq et al.; Alahmadi et al.; Cvach (alarm fatigue); Crootof, Kaminski and Price; Stohl, Stohl and Leonardi; Koulu; Sundaramurthy et al. | Sterz, Tariq found. Alahmadi (USENIX, no DOI), Stohl et al. (no DOI), Crootof et al. (law review, no DOI) not in Crossref. Cvach 2012 and Sundaramurthy 2015: query errors (rate limit); **not retried, not verified**. | | +| 23 | Unpaywall | 15 DOIs | 10 open | Closed: Ananny and Crawford, Warm et al., Endsley and Kiris, Mackworth, Sendelbach and Funk. | +| 24 | curl / Europe PMC | Ancker et al. (PMC5387195) full-text XML | 1 | Abstract read; tables not read. | +| 25 | curl | Rieder and Hofmann (policyreview.info PDF); Kroll (arXiv 2101.09385); Santoni de Sio and van den Hoven (Frontiers PDF); Elish (estsjournal.org); Stohl et al. (ijoc.org); Sterz et al. (arXiv 2404.04059); Green (arXiv 2109.05067v3); Alahmadi (USENIX PDF); Crootof et al. (Vanderbilt Law Review PDF) | 9 of 9 | Publisher PDF for Green returned 403; arXiv v3 used. SSRN returned "Content Blocked". | +| 26 | curl | Parasuraman and Manzey (TU Berlin DepositOnce bitstream) | wrong document | The link returned an unrelated report. Not read this run; the repo card is the source. | +| 27 | curl | EUR-Lex official text of Reg. (EU) 2024/1689 | empty response | Article 14 read from artificialintelligenceact.eu instead. Not compared with the official text. | +| 28 | WebSearch | `Anthropic Claude Code auto mode users approve 93% of permission prompts approval fatigue` | 9 | Led to the Anthropic engineering post (row 29). Later coverage cites 97%; I did not trace the difference. | +| 29 | curl | `anthropic.com/engineering/claude-code-auto-mode` | 1 | Opening paragraphs read. | +| 30 | Scholar Gateway | `human oversight of AI agents monitoring logs automation bias alert fatigue` | 0 | **Failed:** `ACCESS_DENIED: Access denied`. Not retried. | +| 31 | Consensus | `human review of audit logs or security alerts: accuracy, workload and fatigue of analysts` | 20 | Surfaced Tariq et al., Alahmadi et al., Layman et al. (abstracts only). | +| 32 | Consensus | `human oversight of AI agents approval fatigue rubber-stamping experiment` | 20 | Surfaced Chen et al. 2026, Yu et al. 2026, Grunde-McLaughlin et al. 2026, Mitchell et al. 2026, Turan 2026, Langer et al. 2024, Laux 2023. All abstract only. | +| 33 | WebSearch | `Ben Green "The flaws of policies requiring human oversight of government algorithms" pdf` | 10 | Found the arXiv version. | +| 34 | WebSearch | `Crootof Kaminski Price "Humans in the Loop" Vanderbilt Law Review 76 429 2023 abstract` | 10 | Found the Vanderbilt PDF. A search for the title also returns an unrelated 2014 IEEE S&P paper of the same name (Horne). | +| 35 | Semantic Scholar | DOI lookups for open-access links | several | One search for "Humans in the Loop" returned only the 2014 paper. | +| 36 | OpenAlex | abstracts for Warm et al., Mackworth, Endsley and Kiris, Ananny and Crawford | Warm, Endsley and Kiris, Ananny and Crawford abstracts found; Mackworth none | | + +--- + +## (b) Entries + +### B1. The post, its companions and its reception + +**nadella2026models**. Nadella, S. (2026). Models as Insider Risks in the Super Intelligence Era. *sn scratchpad*, 10 Oct 2026. https://snscratchpad.com/posts/models-as-insider-risks/ (also posted on X, https://x.com/satyanadella/status/2108931348857827686). +- *Verification:* blog page fetched; every paragraph checked against the local capture (see above). The post runs about 800 words including the principle labels. +- *Read status:* read in full. +- *Venue.* News outlets variously call it an X post, an "article on X", a "letter" and an essay (TNW; Business Standard; RuntimeWire). It is a short essay on his personal blog, dated 10 Oct 2026, mirrored on X. I found no Microsoft corporate post, white paper, product page or talk released with it. RuntimeWire calls it "an enterprise-control proposal, not a product launch"; a search summary says it names no Microsoft product, policy or launch date. Microsoft's own agent-security material predates it (see Jakkal 2026, below). + +**Paragraph-by-paragraph reading.** "We" throughout is the deploying organisation. The post never says "user", "customer", "employee" or "people"; "human" occurs twice (P6, P15), "person" once (P19), "affected" once (P20), "authority" once (P5). + +| P | What it says | Who observes / who holds authority / what the person does | +|---|---|---| +| P1 | Traditional software could be traced to a code path. | Setting. No human task. | +| P2 | Mechanistic attribution of model behaviour to training data or weights is no longer available for frontier models, yet they are given sensitive data and mission-critical actions taken "on our behalf". | The model acts for the deploying organisation. The people it acts toward are not mentioned. | +| P3 | Responsibility for what the model does on the deployer's behalf cannot be outsourced; a provider's assurances do not relieve the deployer of it. | Responsibility sits with the deployer, not the model provider. | +| P4 | The system is not to be treated as nested black boxes whose outputs are accepted or rejected. Systems should be built so that their behaviour can be observed, their limits tested and their actions always contained. | The deployer observes, tests, contains. Rejects the accept-or-reject user move. | +| P5 | Supply of intelligence is to be separated from "the authority over it". | Supply = model and provider. Authority = the deployer. | +| P6 | Sets alignment aside. Models that are non-deterministic are to be surrounded by deterministic design of the whole system, together with human controls, operating procedures and industry standards. | Deterministic design is the main instrument. Human controls and procedures are named, not described. | +| P7 | Models are to be treated as insider risks: a capable enough actor trusted with important systems may err or be compromised, whether or not it is malicious. | The model is an actor under the organisation's authority. The risks named are error and compromise, not the model's own aims. | +| P8 | Enterprises already know how to manage powerful insiders: identity, least privilege, logs, containment. | Security staff operating existing practice. | +| P9 | Chain-of-thought transparency is a "non-negotiable"; opaque "neuralese" reasoning is no excuse. But CoT transparency on its own is "not sufficient or dependable". | Someone must read or monitor the CoT. Who is unstated. | +| P10 | Models can test each other, but the result can be a stack of opaque models watching each other. | An argument for checks that do not stack models. | +| P11 | Controls over access and actions belong outside the model. The post ties this to a 1970s security principle: a program should not be able to circumvent or alter the machinery that checks its permissions. | Deterministic enforcement. No reading by a person. | +| P12 | Separate the model from the harness and the action space; externalise controls and safeguards. | Architecture. | +| P13 | Systems are to be designed around principles of observability. The seven items that follow include containment and disclosure. | He uses *observability* as the umbrella for the whole list. | +| P14 | Model diversity: no sole dependency, no model verifying itself. | Architecture. | +| P15 | Observe everything: each meaningful action by a model must leave a tamper-proof record, which the post calls "human readable evidence"; outcomes must be reproducible without relying on the model's own attestation. | The one place a person must read. "Meaningful" is the only limit on volume. Who reads, at what rate, with what training: unstated. The stated purpose is reconstruction. | +| P16 | Continuously test the whole system, including failures and attacks. | Test engineers. Largely automated work. | +| P17 | Independent controls: organisations should be able to decide for themselves what a model may access and do. | Authority is vested in *organisations*. This is a steering right, held by the organisation. | +| P18 | Independent auditability: validation independent of the intelligence validated. | Auditor unspecified (a person or another system). | +| P19 | Containment: treat the model as compromised from the outset; the post's image is an emergency brake, with "an authorized person" holding the power to halt a model mid-task. | One person with a halt right. Detecting when to halt is left unstated. | +| P20 | Incident disclosure "to those affected", shared learnings industry-wide, including implementation details that change agent behaviour at runtime. | The only mention of people outside the deployer. They receive notice after a failure. | +| P21 | The best system is the one that lets the deployer trust the model least. | Closing aphorism. | + +*What "human readable evidence" asks of a person.* It asks that a record exist which a competent person *could* read to reconstruct an outcome. It does not say that anyone reads every record, or in real time. The text supports a forensic reading (P15: reproducing how an outcome was achieved) more than a vigilance reading. The emergency brake (P19) is different: a brake applied mid-task implies someone notices something while it happens. + +**nadella2026frontier, nadella2026reverse, nadella2025looking**. Nadella's earlier posts on the same blog. *Read in full.* +- Dec 2025: the aim is for AI to be "a scaffolding for human potential vs a substitute", a "product design question we need to debate". He expects a move "from models to systems" with "rich scaffolds that orchestrate multiple models and agents; account for memory and entitlements". +- Jun 2026: "the real opportunity is not in picking the best model but instead in building a learning loop on top of models"; "human agency will be the driver of token capital growth"; a firm "should be able to switch out a 'generalist' model". +- Jul 2026: "Models learn from 'exhaust,' the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong." This learning should "belong to you", meaning the firm. The enterprise needs "a hard boundary across which nothing crosses ... without consent". It also needs an orchestration layer "decoupled from any single model" (compare P14). +- *Bearing:* these posts show a firm-level engagement layer: people correct models and the firm owns the learning loop. The insider post (P1–P21) does not mention it. For count (iii), the author must therefore say "the insider post", not "Nadella", when he says users do not steer. + +**jakkal2026secure**. Jakkal, V. (2026). Secure agentic AI for your Frontier Transformation. Microsoft Security Blog, 9 Mar 2026. https://www.microsoft.com/en-us/security/blog/2026/03/09/secure-agentic-ai-for-your-frontier-transformation/ +- *Read status:* read in full (card: library/jakkal2026secure.md). +- *What it records:* each agent gets "a unique identity in Microsoft Entra"; identity governance limits "agent access to only resources they need"; Insider Risk Management "extends insider risk protection to agents"; an Agent Registry gives an inventory. The page says security teams use the signals to "block agent actions"; it does not say whether a person or the system blocks. The summary found no approval workflow for agent actions. It is a marketing page and offers no evidence of effect. +- *Bearing:* the product line that realises P8 (identity, least privilege, logging) already existed seven months earlier. It is context, not a companion release. + +**constantin2026nadella** (The Next Web), **kumari2026nadella** (Business Standard, 12 Oct), **abovesecurity2026satya** (Security Boulevard, 10 Oct), **linzainescu2026microsofts** (Forbes, 11 Oct; not accessible). +- *Read status:* TNW and Business Standard read in full (cards: library/constantin2026nadella.md, library/kumari2026nadella.md); Security Boulevard as a summary; Forbes not accessible. The TNW article is dated 12 Oct 2026, 7:55 am. +- *What they record.* Coverage is descriptive. Mustafa Suleyman (Microsoft AI) reposted that super intelligence "must be contained" and called it "an engineering and governance challenge" (Business Standard; TNW). Aaron Levie (Box) wrote that enterprise AI is entering a "zero trust" era with the need to "secure, govern, control, and audit" agent work (Business Standard). TNW adds that Musk shared it, that Rep. Ted Lieu said he largely agrees and tied it to a bipartisan "AI kill switch bill", and that the post followed US government disclosure demands and an Anthropic incident report; I did not verify those items. Business Standard also attributes to Nadella that organisations decide "when human approval is required" and gives an example of a model that may recommend a transaction without being allowed to approve it; neither appears in the post (P1–P21 contain no approval step), so both are the reporter's elaboration and are not to be cited as his. Security Boulevard is a vendor promotion by a company that sells AI investigation agents; it nonetheless says model reasoning traces "may not" explain behaviour faithfully and that evidence should be "human-readable". +- *Critique.* I found no organised critique in two days. One search-engine summary says commentators note the post does not say who the "authorized person" is; I could not tie that to a named article, so treat it as an unverified lead. +- *Not found:* any Microsoft statement that the post is policy or product direction. + +### B2. The security lineage + +**anderson1972computer**. Anderson, J. P. (1972). *Computer Security Technology Planning Study*. ESD-TR-73-51, Vol. I. HQ Electronic Systems Division, Hanscom AFB. NIST CSRC text copy. +- *Verification:* the 43-page text copy carries the title page ("October 1972"). A 142-page scan at UC Davis has no text layer and was not read. +- *Read status:* read in full for Volume I (card: library/anderson1972computer.md); Volume II not opened. The table of contents of the NIST copy numbers some sections differently from the body (see the card's Caution); check the section numbers cited here against the body headings. +- *What it argues:* the "malicious user" has legitimate access, so the defence "must reside in the process that controls the operation and execution of arbitrary programs" (§3.1). A reference monitor "enforces the authorized access relationships between subjects and objects" (§3.2, p. 9). The reference validation mechanism "must be tamper proof" (p. 9); "must always be invoked"; and "must be small enough to be subject to analysis and tests, the completeness of which can be assured" (§3.2.2, p. 10). References are limited "to just those authorized (by an external authority) for the user" (§3.1). Section 3.7 says the concept "is directed to overcoming the threat of a malicious user" and "predicated upon positive identification (authentication) of all users at all times". +- *Bearing on the post.* P11 matches the tamper-and-bypass part of Anderson's requirements (a program should not be able to subvert the machinery that enforces its permissions). The post does not name Anderson, Saltzer or Schroeder, so the paper may say the principle is *his gloss on a 1970s result*, not that he cites Anderson. Three things in Anderson do not carry over. (1) The mechanism is a *small, verifiable* component; a harness around a frontier model is not small, and Anderson's third condition is what made the idea tractable. (2) The monitor mediates references against a list; it does not read evidence. (3) "Tamper proof" attaches to the *enforcement mechanism*, while P15 attaches "tamper-proof" to *evidence* that a human reads. Those are different mechanisms with different failure modes. +- *Problem 2.* The reference monitor *is* a reallocation of work from a person to an artefact. If the paper says the spacing move shifts work from reader to page, P11 does the same. The author should not rely on "moves work to the artefact" to separate the two (see C7). + +**saltzer1975protection**. Saltzer, J. H., & Schroeder, M. D. (1975). The protection of information in computer systems. *Proceedings of the IEEE*, 63(9), 1278–1308. doi:10.1109/PROC.1975.9939. +- *Read status:* read in part (card: library/saltzer1975protection.md): the full text of the authors' MIT web edition, Sections I–III; Section II was read for argument only and the footnotes page was not read. The card also uses the hierarchical-control passage, I.B.4 on authentication, and III.B, which this facet did not. +- *Quotations (section I.A.3).* Least privilege: "Primarily, this principle limits the damage that can result from an accident or error." Complete mediation: "Every access to every object must be checked for authority." Separation of privilege: a mechanism "that requires two keys to unlock it is more robust and flexible". Psychological acceptability: "It is essential that the human interface be designed for ease of use, so that users routinely and automatically apply the protection mechanisms correctly." Open design: "any skeptical user may be allowed to convince himself that the system he is about to use is adequate for his purpose." +- *Compromise recording.* The paper treats "mechanisms that reliably record that a compromise of information has occurred" as one of two further principles that "apply only imperfectly to computer systems": "In computer systems, this approach is used rarely, since it is difficult to guarantee discovery once security is broken." It also says the principles "serve best as warnings", not "absolute rules". +- *Bearing.* Least privilege is justified by accident and error as well as attack, which matches P7 (models treated as insider risks without being assumed malicious). Psychological acceptability is a human-factors principle *inside* the security canon the post invokes; the post does not use it. Compromise recording is the nearest S&S category for P15's tamper-proof evidence; S&S say it applies only imperfectly to computer systems, while noting that a tamperproof record reported to the owner "may help discover unauthorized use". P15 pairs it with P11, P12 and P17, which is the combination S&S would endorse. + +**rose2020zero**. Rose, S., et al. (2020). *Zero Trust Architecture*. NIST SP 800-207. doi:10.6028/NIST.SP.800-207. +- *Read status:* read in part (card: library/rose2020zero.md); sections 6, 7.3.1–7.3.3 and most of Appendix B only skimmed. +- *What it says.* Zero trust is "a collection of concepts and ideas designed to minimize uncertainty in enforcing accurate, least privilege per-request access decisions in information systems and services in the face of a network viewed as compromised" (§2). Tenet 3: access "is granted on a per-session basis" with "the least privileges needed". Tenet 6: authentication and authorization "are dynamic and strictly enforced before access is allowed". Tenet 7: the enterprise "collects as much information as possible" about asset state "and uses it to improve its security posture". The architecture runs on a policy decision point and a policy enforcement point. +- *Bearing.* P19's assume-compromise posture and P8's list match this lineage. The post does not use the words "zero trust"; Levie does (Business Standard). The collect-everything tenet (7) feeds *policy*, not a human reader. NIST's enforcement is dynamic and per-request. The full text adds two places for people: administrator audit of the policy components (p. 28), and §7.3.7's line that the subjects and stakeholders of the resources and processes involved "should also provide feedback" (p. 41), which gives enterprise users a feedback role. + +**cert2022common**. CERT National Insider Threat Center (2022). *Common Sense Guide to Mitigating Insider Threats*, 7th ed. SEI, Carnegie Mellon. +- *Read status:* read in part (card: library/cert2022common.md): front matter, Introduction, Best Practices 10, 11, 12, 14, 15, 17 and most of 21 in full; the opening pages of the others; the bodies of Best Practices 1–9, 13, 16, 18–20 and 22 skimmed or not read. +- *Definition.* An insider threat is "the potential for an individual who has or had authorized access to an organization's critical assets to use their access, either maliciously or unintentionally, to act in a way that could negatively affect the organization." The guide builds on over 3,000 incidents. +- *Bearing.* (1) The framework includes *unintentional* insiders, which matches P7. (2) The victim in the definition is *the organisation*. This is the sense in which P20's "those affected" extends past it. (3) Best practice 12 says: "Simply logging all network activity is not sufficient"; volume "drives the need for tools that can fuse data ... where alerts can be developed", and monitoring should be "risk-based" and "focus first on the organization's critical assets." That cuts against reading *observe everything* (P15) as literal full-volume human review. The same practice also endorses after-the-fact human review: at least the previous 90 days of a leaver's activity (p. 84), and monitors who "hunt" below the top alerts, in some settings full time (p. 89). The guide therefore supports P15's after-the-fact reconstruction more than the opening quotation suggests; it gives no evidence on how well people do that work. (4) The guide's list includes workforce measures for which a model has no analogue: managing stress and negative work environments (practices 5, 8) and "positive incentives to align the workforce and the organization" (practice 21). The card read the opening pages of practices 5 (p. 45) and 8 (p. 61) and most of practice 21, and they confirm this description. + +### B3. The AI-safety lineage + +**greenblatt2024ai**. Greenblatt, Shlegeris, Sachan, Roger (2024). AI Control: Improving Safety Despite Intentional Subversion. ICML 2024; arXiv 2312.06942 v5. +- *Read status:* read in part (card: library/greenblatt2024ai.md); appendices (pp. 21–47) skimmed. +- *What it argues and finds.* The authors test protocols that stay safe even when the untrusted model (GPT-4) tries to subvert them, using a weaker trusted model (GPT-3.5) and "limited access to high-quality trusted labor" (abstract). Human labour is modelled as "H points, which allow the protocol to use an unmodified GPT-4", enough "for up to 2% of problems" (§2.2). Table 2: auditing only the first 2% yields safety 15%; trusted monitoring with a 2% audit budget, 62%; untrusted monitoring, 79%; trusted editing (20%), 92%. The untrusted-monitoring section describes "an arms race between the red team and the blue team". +- *Bearing.* (1) The AI-control school treats audit capacity as scarce and spends it on the most suspicious cases. It does not assume *observe everything*. (2) The "human" auditor is GPT-4 (H points let the protocol use an unmodified GPT-4); the paper tests no people and is not evidence about them. The "up to 2%" is an evaluation parameter, not a measured limit of human capacity. (3) It confirms P10's worry that model-checks-model has collusion problems. P14 and P18 answer it with independence. + +**turpin2023language; lanham2023measuring; chen2025reasoning**. Chain-of-thought (CoT) faithfulness. +- *Read status:* Turpin: read in part (card: library/turpin2023language.md; per-task tables in Appendices D–G skimmed). Lanham: read in full (card: library/lanham2023measuring.md). Chen: read in full (card: library/chen2025reasoning.md). +- *Turpin et al.* CoT explanations "can systematically misrepresent the true reason for a model's prediction", and accuracy "drop[s] by as much as 36%" on 13 BIG-Bench Hard tasks when biasing features are added (abstract, GPT-3.5 and Claude 1.0). The body shows the 36 is 36.3 accuracy points in one setting (GPT-3.5, zero-shot, Suggested Answer), not a relative loss; Claude 1.0's matching figure is 30.6 points, and the Answer-is-Always-A drops are far smaller (18.7 and 4.7 points; p. 6). Their conclusion: CoT can be "plausible yet misleading, which risks increasing our trust in LLMs without guaranteeing their safety". +- *Lanham et al.* "As models become larger and more capable, they produce less faithful reasoning on most tasks we study" (abstract), and CoT "can be faithful if the circumstances such as the model size and task are carefully chosen". The body makes the size result per task: on 7 of 8 tasks the 13B model changes its answer after CoT more often than the 175B model, and faithfulness peaks at 13B on 6 of 8 tasks, with AQuA and LogiQA the exceptions (p. 7). The "less faithful" finding is for multiple-choice tasks with a fixed prompt format. +- *Chen et al. 2025* (Anthropic): in most settings and models tested (abstract; 5 of 6 hint settings for Claude 3.7 Sonnet and 6 of 6 for R1, p. 6), CoTs reveal hint use "in at least 1% of examples ... but the reveal rate is often below 20%"; CoT monitoring is "a promising way of noticing undesired behaviors during training and evaluations, but ... not sufficient to rule them out." +- *Bearing.* These support P9's caveat that CoT transparency on its own is "not sufficient or dependable". They were produced by labs that build the models. They test a monitor model or a measure of faithfulness; none tests a human reader. + +**korbak2025chain; baker2025monitoring**. CoT monitorability. +- *Read status:* Korbak, read in full (card: library/korbak2025chain.md); Baker, read in full (card: library/baker2025monitoring.md). +- *Korbak et al.* (a position paper with many authors from several labs): "Like all other known AI oversight methods, CoT monitoring is imperfect and allows some misbehavior to go unnoticed." Monitorability "may be fragile". It names latent-space reasoning as a threat (section 2). They recommend researching monitorability "alongside existing safety methods". +- *Baker et al.:* a weaker model can monitor a stronger one's CoT for reward hacking, but "with too much optimization, agents learn obfuscated reward hacking, hiding their intent within the CoT". Baker also raises the possibility that an agent "could learn a new language that is illegible to a monitor" (p. 8). +- *Bearing.* P9's "non-negotiable" is stronger than Korbak: the paper recommends investment "alongside existing safety methods" (abstract) and CoT monitoring "only as an addition to existing safety research directions and not as a replacement" (§5, p. 7); it does not call monitoring non-negotiable. Korbak also declines to recommend showing the chain of thought to users ("we don't make an explicit recommendation either way", p. 7), because pressure to make traces look safe could break monitorability. P9's "neuralese" worry is the fragility Korbak describes; Baker raises it without the word. + +**duzan2026chain; gema2026chain; little2026length; wang2026monitorbench; jia2026faithfulness**. 2026 preprints. +- *Read status:* Duzan, Gema, Little and MonitorBench: read in part; Jia: read in full (cards: library/duzan2026chain.md, gema2026chain.md, little2026length.md, wang2026monitorbench.md, jia2026faithfulness.md). Preprints, except that MonitorBench (v4) carries a COLM 2026 header (acceptance not verified beyond the header). +- *Findings as stated in the abstracts.* Under implicit influence "detection falls by 41-46 percentage points in two of our four settings", and a realistic system-prompt addition can lower detection "to as low as 5%" (Duzan and Cooper Stickland). Across 15 open-weight models, unverbalised cue adoption is higher for cues delivered through tool returns than through user messages, and "telling models that their reasoning will be monitored does not reliably close the gap" (Gema et al.); the body adds that the narrowing under a directive prompt is partly an increase in user-channel unverbalised adoption, so a smaller gap can mean worse behaviour on the other channel (pp. 13–14). Length penalties "trade monitorability for inference cost by removing the evidence monitors depend on" (Little); in Little the pressure is a training objective, set by whoever fine-tunes the model, and the cost saving is deliberate, while the monitorability loss is the side effect the token-accuracy criterion does not count (p. 1). Monitorability "is a conditional property" depending on the model, the monitor and the task (Wang et al., MonitorBench). Jia et al. find that training interventions raise how often a hint is mentioned, but the gains are in visibility of the shortcut, not in its prevention. +- *Bearing.* Through 2026 the direction is consistent: monitoring helps, depends on conditions, and degrades under pressures from training choices (Little: a length-penalty objective set by whoever fine-tunes, so it reaches a deployer only if the deployer trains), and the conditions most like real deployment are the ones where it helps least (Duzan). This is the empirical ground for P9 and P10. + +**kutasov2025shade; lynch2025agentic**. Agent sabotage and "insider" framing. *Read status:* Kutasov, read in part (card: library/kutasov2025shade.md; several appendices skimmed); Lynch, read in full (card: library/lynch2025agentic.md). +- SHADE-Arena: the best monitor reached AUC 0.87 distinguishing benign from malign transcripts; the best sabotage agents scored 27% and 15% when overseen by Claude 3.6 Sonnet. The authors say the results "demonstrate the difficulty of monitoring for subtle sabotage attempts". The paper also reports a human baseline: 12 benign and 20 malign transcripts across 14 task pairs, three annotators each, who took "an average of 103 minutes per transcript" (p. 7; range 20–356 minutes, App. L, p. 40); their pooled AUC was 0.65, worse than the top AI monitors (p. 8). The authors conclude that such transcripts need over 1.5 hours of human attention for performance worse than several frontier models (p. 42). It is a small exploratory baseline, not a general finding about people. +- Lynch et al. (Anthropic) stress-tested 16 models in simulated corporate settings and found "malicious insider behaviors" such as blackmail "when that was the only way to avoid replacement or achieve their goals". They report no evidence of the behaviour in real deployments. *Bearing:* this is the closest AI-safety precedent for the "insider" frame. In that paper a threat to continued operation alone was enough, with no goal given (p. 8), and the authors say they cannot tell how far self-preservation, faulty reasoning or suspicion of the setup drove the behaviour (p. 9); role-play may explain part of it (fn. 4, p. 3). The difference from Nadella's P7 (mistakes or compromise) is therefore smaller than "the model has its own goals" suggests; it still matters for convergence (d). + +**hao2024training; kokotajlo2025ai**. "Neuralese". +- Hao et al. (Coconut; read in full, card: library/hao2024training.md) train models to reason in a continuous latent space; Coconut trails language chain of thought on GSM8k (34.1% against 42.9%). Kokotajlo et al. (AI 2027; read in part, card: library/kokotajlo2025ai.md) use the word, saying only "We call this 'neuralese'" and crediting Hao et al. (2024) for "implementing this idea"; the page does not claim to coin or popularise it: "Neuralese recurrence and memory allows AI models to reason for a longer time without having to write down those thoughts as text." It is a forecast scenario, not a finding, and the page says no frontier lab had adopted the technique. Nadella's P9 uses the word without citation. + +**chan2024visibility; chan2025infrastructure**. Agent visibility and infrastructure. +- *Read status:* Chan 2024 read in full (card: library/chan2024visibility.md); Chan 2025 read in full (card: library/chan2025infrastructure.md). +- *Chan et al. 2024.* Three measures: agent identifiers, real-time monitoring, activity logs. Real-time monitoring "is likely to be automated so as to be able to cope with the speed and scale of agent activities". Logs enable "post-incident attribution and forensics". In a sentence that is truncated in my text extraction, an identifier "could be visible to any" party in a transaction; and a user "could be given the opportunity to request human review of any action taken to impede its agent"; the authors leave the human-review problem to "other work" and cite four works (Green 2022 and three algorithmic-recourse papers). They say visibility measures also help "when humans delegate to other humans or institutions". +- *Chan et al. 2025:* "agent infrastructure": "technical systems and shared protocols external to agents that are designed to mediate and influence their interactions with and impacts on their environments". The full text lays out controls outside the agent as a design programme, not disclosure: nine kinds of infrastructure in three functions (Table 1, p. 3), each with its limits stated. For oversight layers (p. 13) the intervening actor "could be a human user, but not necessarily"; humans "could ignore such flags if they are too frequent or present information in a complex manner", and "it is unclear if the market will provide them to a socially optimal degree" (p. 14). Rollbacks raise a moral hazard: users "could have weaker incentives to properly monitor their agents" (p. 19). The paper reports no data. +- *Bearing.* Chan 2025 is the closest academic counterpart to P11, P12, P15, P17 and P19; its human-attention and market-supply doubts bear on P10 and P12 (pp. 13–14, 20). Chan 2024's agent identifier and agent card reach *counterparties* of the agent's actions (the counterparty sees the identifier and agent card, not the monitoring or logs). Nadella's identity (P8) faces the enterprise only. This is a real difference between two proposals for the same architecture, and a place where AI-governance work is closer to the author's triad than the post is. + +### B4. Human factors of the role the architecture creates + +**bainbridge1983ironies**. Bainbridge, L. (1983). Ironies of automation. *Automatica*, 19(6), 775–779. doi:10.1016/0005-1098(83)90046-8. +- *Read status:* read in full in an earlier run (repo card with 20 page-verified quotations). Not reopened. +- *Quotations (card).* The designer "still leaves the operator to do the tasks which the designer cannot think how to automate" (p. 775). The operator is asked "to monitor that it is working effectively" (p. 776). "There is therefore no way in which the human operator can check in real-time that the computer is following its rules correctly ... The human monitor has been given an impossible task" (p. 776). Remedy: decisions made "using methods and criteria, and at a rate, which the operator can follow" (p. 777); systems "should fail obviously" (p. 777). +- *Caution (from the card).* Single trained operator of a physical process; a five-page argument, no data. Do not cite it as evidence about consumer interfaces or generative AI. +- *Bearing.* It speaks to P19 (a person who must notice a fault while it happens) more than to P15 (a record to be read afterwards). Her remedies are artefact-side. Nadella's P11, P12 and P17 are artefact-side too. + +**mackworth1948breakdown; warm2008vigilance**. Vigilance decrement. +- *Read status:* Mackworth, not accessible (paywalled; no abstract found). Warm et al., abstract only. +- *Warm, Parasuraman and Matthews (2008)*: "Vigilance tasks have typically been viewed as undemanding assignments requiring little mental effort", but the review finds "that vigilance requires hard mental work and is stressful", with application "to most human-machine systems that require human monitoring". Mackworth is cited as the origin of the decrement; I have not read him. +- *Scope.* Vigilance studies test sustained watching for rare signals. A person who reconstructs an incident from a log is doing a different task. + +**parasuraman2010complacency; endsley1995out**. Complacency, automation bias, out-of-the-loop. +- *Read status:* Parasuraman and Manzey: read in full (card: library/parasuraman2010complacency.md). Endsley and Kiris: abstract only. +- *Parasuraman and Manzey (card).* Complacency means worse detection of automation failures under automation; the card reports 82% versus 33% under variable versus constant reliability (p. 384, from Parasuraman et al. 1993), that 60 minutes of practice does not remove it (p. 387), and that a second crewmember "did not affect the strength of automation bias" (p. 397). +- *Endsley and Kiris (abstract).* "low SA corresponded with out-of-the-loop performance decrements in decision time following a failure of the expert system", and "Level of operator control in interacting with automation is a major factor in moderating this loss of SA." +- *Bearing.* The review is narrower than the brake-holder case. Complacency is "typically found under conditions of multiple-task load" (p. 387; Abstract, p. 381); when participants only backed up the automation, detection was "near perfect" (p. 384), and in the single-failure study most single-task participants detected it (p. 385). A brake-holder with no competing task is the weaker case for complacency. Against that, the authors expect worse monitoring at the very low failure rates of real systems, citing an estimate of one critical signal every two weeks (p. 385), so a dedicated watcher does not escape the problem. Automation bias, which needs no multitask load, is closer to the P15 reader. The evidence is mostly lab and simulator work, but the review also draws on accident reports (pp. 382–383), field trials of physicians (p. 394) and two team studies (p. 397). + +**ancker2017effects; alahmadi2022false; layman2023controlled; tariq2025alert; anthropic2026how**. Alert and approval fatigue. +- *Ancker et al. (2017)*, read in full (card: library/ancker2017effects.md): retrospective cohort of 112 primary-care clinicians. "Likelihood of reminder acceptance dropped by 30% for each additional reminder received per encounter"; acceptance was associated with work complexity and repeated alerts, "but not with the amount of work"; "no evidence of ... desensitization over time for a newly deployed alert". The number of alerts per encounter was itself a strong predictor (incident rate ratio 0.70), so alert volume per encounter mattered; what showed no effect was overall workload, measured by patients and encounters. "Uninformative" is a proxy: the authors measured repetition, not informativeness (p. 3). +- *Alahmadi et al. (2022)*, read in full (card: library/alahmadi2022false.md): a 20-person survey and 21 interviews of SOC practitioners. One interviewee said "We know 99% of the alarms we generate are false positives, but we still have to look at them." The authors find most such false positives are benign triggers, and that "Alarm validation is a tedious task that can cause alarm burnout and eventually desensitization" (abstract). The paper describes the 2013 Target breach as an alarm that was raised, escalated, and ignored. The 99% is one analyst's remark (B3, p. 2790), not a measured rate; the survey's ten disclosed ratios were 1 real alert in 100 and 50 in 200 for two respondents (p. 2788). The paper measures no analyst performance (no accuracy, miss rate or vigilance decrement); "burnout" and "desensitization" come from the authors' framing and a citation to a clinical review, not from the interviews; respondents rate the human role as indispensable. It studies alarm validation, not approval. +- *Layman et al. (2023)*, read in full (card: library/layman2023controlled.md): a controlled experiment; the 86% false-alarm group had "47% lower precision and 40% slower time on task" than the 50% group, with no significant difference in sensitivity (p = .098). The 47 is in percentage points of median precision (0.80 against 0.33, about 59% lower in relative terms). Only 7 of the 50 alarms were true in the high-false-alarm group, so sensitivity rests on few items (the card's observation from the design, not a claim in the paper). +- *Tariq et al. (2025)*, abstract only: a survey of alert fatigue in SOCs; it identifies four causes. I read no detail. +- *Anthropic (2026)*, read in full (card: library/anthropic2026how.md): "Claude Code users approve 93% of permission prompts"; the post defines "approval fatigue, where people stop paying close attention to what they're approving". This is a vendor report on its own product, not peer-reviewed; the 93% is a rate of approvals, not a measure of inattention, and the post gives no behavioural measure of fatigue. Its 52 real overeager actions come from employee denials and post-hoc flags, a small sample that depends on what was caught. It is, however, the closest direct datum on people overseeing an agent that acts. +- *Not found:* a study of people reviewing tamper-proof logs of LLM-agent actions to reconstruct incidents, at deployment scale. The small human baseline on agent transcripts in Kutasov et al. (above) is the nearest. + +**chen2026comparing; yu2026habituation; grundemclaughlin2026overseeing; mitchell2026agents; turan2026oversight**. 2026 agent-oversight studies. *Read in full (cards: library/chen2026comparing.md, yu2026habituation.md, grundemclaughlin2026overseeing.md, mitchell2026agents.md, turan2026oversight.md); preprints.* +- Chen et al.: 48 participants, 192 live sessions; "oversight strategy affected exposure to problematic actions, but not users' ability to intervene"; "users judged whether the agent acted correctly, not whether the action was safe". The plan effect came partly from changing the agent. +- Yu et al.: 400 repeat reviewers, 11,429 reviews of AI-agent pull requests; approval rose from 30.1% to 36.8%, review latency rose 3.5 times and inline comments fell 22%; the authors read this as "reflexive habituation". It is observational. The paper is less tidy than its abstract: the early-late change-request rates (11.2% to 5.6%) repeat the first-decile and tenth-decile figures of Table 1, the text says "28% less commenting" for words but 22% for comments (pp. 3–4), and Spearman rho = +0.02 for PR size is given a p of 0.009 and called negligible (p. 4); the card could not resolve these. +- Grunde-McLaughlin et al.: three user studies of action traces for verification; the abstract says a new interface cut time to find errors, but in the body the overall time effect is small and its interval spans zero; the larger effect (g = −0.65) applies only to the subset where participants correctly judged a wrong answer wrong, and its interval (−1.54 to 0.12) also spans zero (Table 3, p. 18); only the confidence effect on wrong answers judged right has an interval clear of zero. The abstract's "participants reported higher levels of confidence" and "final accuracy was not meaningfully improved". +- Mitchell et al.: a position paper arguing that agent design both obstructs oversight (Section 3) and erodes the overseer's skills (Section 4). +- Turan: a model of a fatiguing reviewer; the abstract says the result is a "modeling result" that "motivate[s] a human study". It also reports a measured disagreement result (kappa = 0.52), but from LLM personas, not people (pp. 5, 9). Do not cite as human evidence. +- *Bearing.* The Grunde-McLaughlin result is the nearest test of "human readable evidence" for agents: readable traces raised confidence without raising accuracy. + +**green2022flaws**. Green, B. (2022). The flaws of policies requiring human oversight of government algorithms. *Computer Law & Security Review*, 45, 105681. +- *Read status:* read in full (card: library/green2022flaws.md; publisher's open-access PDF, locators are PDF pages). +- Abstract: "In this article, I survey 41 policies that prescribe human oversight of government algorithms and find that they suffer from two significant flaws. First, evidence suggests that people are unable to perform the desired oversight functions. Second ... human oversight policies legitimize government uses of faulty and controversial algorithms." §5.1 applies Bainbridge: "Human overseers are left with a task that is more difficult than their original charge." +- *Remedy.* "institutional oversight": agencies must justify the algorithm and any human oversight with evidence, subject to "democratic public review". Green's remedy is structural and institutional, *not* user engagement. It supports moving the burden off the individual overseer; that is also the direction of P11, P12 and P17. Green says better oversight design is possible and an "open empirical question" (p. 11), so he does not say oversight can never work. + +**crootof2023humans**. Crootof, R., Kaminski, M. E., & Price, W. N., II (2023). Humans in the Loop. *Vanderbilt Law Review*, 76(2), 429–510. +- *Read status:* read in full (card: library/crootof2023humans.md); page numbers are the journal's printed 429–510. +- Defines the "MABA-MABA trap": policymakers "attempt to address concerns about algorithmic incapacities by inserting a human into a decisionmaking process" (abstract). On sustained watching: new roles "tax humans with tasks that run into known human weaknesses, such as sustaining vigilance" (p. 472). On the human as a "liability sponge" and "moral crumple zone" (p. 483). Their advice: "regulate the human-in-the-loop system", not just add a person. +- *Bearing.* Direct on P19: "an authorized person" is a human inserted to answer for the system. That reads only one of the nine roles in their typology: a halt right is a resilience role, and the same person can hold both roles (p. 474). + +**elish2019moral; santonidesio2018meaningful; sterz2024quest; laux2024institutionalised; langer2024effective; eu2024regulation**. Oversight concepts. +- *Elish* (read in full; card: library/elish2019moral.md): the moral crumple zone "protects the integrity of the technological system, at the expense of the nearest human operator" (abstract and introduction). +- *Santoni de Sio and van den Hoven* (abstract): meaningful human control requires a "tracking" condition (the system responds to "the relevant moral reasons of the humans designing and deploying the system") and a "tracing" condition (outcomes can be traced "to at least one human along the chain of design and operation"). The tracking condition as stated in the abstract names the designing and deploying humans. The body (read in full; card: library/santonidesio2018meaningful.md) adds the affected only partly: the tracked agents are left open ("the relevant human agents", p. 11), and the one worked example is traffic, where the system should follow rules set "by the society via the public authority" and conventions that "reflect some relevant interest of the road users" (pp. 11–12). The road users' interest is therefore in view through public rules, not as a party with a say; the tracing condition brings others in as people who may react morally to the responsible human (p. 9). +- *Sterz et al.* (abstract; body read in full, card: library/sterz2024quest.md): oversight is effective if the person has "(a) sufficient causal power ..., (b) suitable epistemic access ..., (c) self-control, and (d) fitting intentions". P19 supplies (a). P15 is meant to supply (b). The literature above is about whether (b) and (c) hold. The body also names (d), fitting intentions, and role conflict (p. 8), which bear most directly on D5. +- *Laux* (read in full; card: library/laux2024institutionalised.md): the abstract says "Empirical research suggests, however, that humans are not reliable in fulfilling their oversight tasks." The body qualifies this (the empirical record is "subject to debate"), and the paper is a proposal that offers no new empirical test. +- *Langer et al.* (read in full; card: library/langer2024effective.md): a signal-detection framework for error detection by overseers; cite it as a framework paper, since it has no data. +- *EU AI Act Art. 14* (read in full): high-risk systems must allow overseers "to remain aware of the possible tendency of automatically relying or over-relying on the output ... (automation bias)" (14(4)(b)), "to decide ... not to use the high-risk AI system or to otherwise disregard, override or reverse the output" (14(4)(d)), and "to intervene in the operation ... or interrupt the system through a 'stop' button" (14(4)(e)). The Act names the stop button and automation bias in one article. P19 resembles 14(4)(e); the person in both is a deployer's overseer, not the affected person. + +### B5. The transparency critique + +**ananny2018seeing**. Ananny, M., & Crawford, K. (2018). Seeing without knowing. *New Media & Society*, 20(3), 973–989. *Abstract only.* +- "Being able to see a system is sometimes equated with being able to know how it works and govern it ... But can 'black boxes' ever be opened, and if so, would that ever be sufficient?" The authors present "10 limitations" and sketch "an alternative typology of algorithmic accountability grounded in constructive engagements with the limitations of transparency ideals". The abstract's "constructive engagements" is a phrase the author may want, but I have not read what it means in the paper. + +**rieder2020towards**. Rieder, B., & Hofmann, J. (2020). Towards platform observability. *Internet Policy Review*, 9(4). +- *Read status:* read in full (card: library/rieder2020towards.md). +- They "propose the concept of observability" as "a more pragmatic way of thinking about the means and strategies necessary to hold platforms accountable". "Unlike transparency, which nominally describes a state that may exist or not, observability emphasises the conditions for the practice of observing in a given domain." The black-box image "more or less skips the practicalities involved in opening it". The conclusion: "Observability is not meant to be a radical departure from the call for transparency"; it regards "continuous and collaborative observation within a normative framework" as necessary. They suggest "centres of expertise" and a "European Platform Observatory". +- *Bearing.* (1) They present *observability* as a more pragmatic way of thinking, one that is not a radical departure from the call for transparency. (2) Their observers are outsiders (researchers, regulators, civil society) watching a platform that holds power over them; Nadella's observers are the deployer watching a model that serves it. The same word names two different seats. (3) They agree with Nadella that the black box will not simply be opened, and they favour observing behaviour over time (pp. 13–14). They do not choose observation in place of transparency: it is "not meant to be a radical departure" from it (p. 21), and traditional disclosure, reports and whistleblowing can remain "part of an observability framework" (p. 13). This complicates count (i). (4) They also argue that access the platform grants, sets the terms of and can end at will is no accountability: the "scope and modalities for access remain in the hands of platforms themselves" (p. 15, the authors reporting scholars' criticism), and Twitter's academic API is "voluntary and can be revoked at any time" (p. 22). They treat the fix as law, not goodwill. + +**stohl2016managing**. Stohl, C., Stohl, M., & Leonardi, P. M. (2016). Managing opacity. *International Journal of Communication*, 10, 123–137. +- *Read status:* read in full (card: library/stohl2016managing.md). +- Visibility has three attributes: availability, approval, accessibility to third parties. The "transparency paradox" is "high levels of visibility decrease transparency and produce opacity". The mechanism: "important pieces of information become inadvertently hidden in the detritus of information made visible—that is, the information hides in plain sight" (p. 133). +- *Bearing.* *Observe everything* (P15) is a visibility rule. The paradox predicts that the more actions are recorded, the more the recipient's limits (information overload) decide what is seen. The paper is about organisations, not AI agents. It names two paths to the paradox; the strategic path (pp. 133–34) is closer to Nadella's seat than the inadvertent path quoted here. + +**kroll2021outlining**. Kroll, J. A. (2021). Outlining Traceability. FAccT '21. +- *Read status:* read in full (card: library/kroll2021outlining.md). +- Traceability "connects records of how the system was constructed and what the system did mechanically to the broader goals of governance, in a way that highlights human understanding of that mechanical operation". The introduction notes "related human factors problems" in defining what it means for people to understand provenance. Kroll maps requirements to tools and finds gaps. *Bearing:* the best formal statement of what P15 asks for, and it names the human-factors gap. Kroll's records are mostly about design and development, with operations covered in §3.3, and he treats the human-understanding requirement as unsolved. + +--- + +## (c) Synthesis + +### C1. The fairest statement of the post + +The post argues from one premise: we can no longer explain a frontier model's behaviour from its weights or its data (P2), so assurance must come from somewhere else. It puts the guarantee in places that can be verified: controls outside the model (P11, P12), set independently by the organisation (P17), within deterministic system design (P6); architectural independence between actor and checker; tests that include failure (P16); and a halt right (P19). It adds records that do not depend on the model's own account (P15). It treats chain-of-thought as necessary but not sufficient (P9). It warns that models checking models can leave nested black boxes (P10), and answers with independence (P14, P18). It assigns responsibility to the deploying organisation, not the provider (P3). It frames the model as an *insider*, a capable actor that can err or be compromised, and borrows the control vocabulary of decades of enterprise security (P7, P8). + +That is a defensible position. Its core is the reference-monitor idea: the guarantee should not require anyone to understand the thing it constrains. Most of the instruments need no reading by a person at all. The weak points are those the post leaves open: who the authorized person is and how they know when to act (P19), and who reads the "human readable evidence" the post asks for after every meaningful model action (P15). The post is about 800 words. Silence in a text this short is thin evidence of a position. + +### C2. Lineage: what the post inherits + +- *Reference monitor.* P11 matches Anderson's (1972) three requirements (tamper-proof, always invoked, small enough to verify); the post names no source. Saltzer and Schroeder (1975) add complete mediation, least privilege and separation of privilege, which line up with P11, P17 and P14/P18 (my inference). The post does not name them (read: no named citation anywhere in P1–P21). +- *Human factors inside the canon.* Saltzer and Schroeder list psychological acceptability. They say compromise recording "appl[ies] only imperfectly to computer systems" and is "used rarely"; they also note a tamperproof record reported to the owner "may help discover unauthorized use". +- *Insider-threat practice.* CERT's definition includes the unintentional insider and treats the organisation as the harmed party. Its guide says "Simply logging all network activity is not sufficient to protect an organization from malicious insider activity" (best practice 12), and asks for fusing sources into alerts. It also endorses after-the-fact human review (the previous 90 days for leavers, p. 84; monitors who hunt below the top alerts, p. 89). Its human-insider practices include stress and incentives that have no model analogue. +- *Zero trust.* NIST SP 800-207 gives P8's practices and P19's "assume compromised" posture, with dynamic, per-request enforcement. The post does not use the term. +- *What the post adds to the lineage:* the claim that mechanistic explanation has failed, and the demand that the model's own reasoning (CoT) be legible although not trusted. Neither is in the 1970s canon. + +### C3. Is monitoring dependable? What the AI-safety work says + +The post's P9 claim, that CoT transparency on its own is "not sufficient or dependable", is supported by the literature, which was produced mostly by the labs that build the models. Faithfulness is partial and falls with scale on most of the tasks they study (Lanham); CoTs reveal hint use often below 20% of the time, among examples where the hint was used (Chen 2025); monitoring "is imperfect" and "may be fragile" (Korbak); optimisation pressure can produce obfuscation (Baker). The 2026 preprints add that detection drops under implicit influence and tool-delivered cues, and that routine training choices can lower monitorability (Duzan, in two of four settings; Gema, 15 open-weight models in a single-call, prefilled-tool setting; Little, in small open models under length penalties). These are preprints (MonitorBench carries a COLM 2026 header), read in part or in full from the cards. SHADE-Arena puts the best monitor (Gemini 2.5 Pro) at AUC 0.87. + +The AI-control work treats human attention as the scarcest input: up to 2% of problems in Greenblatt et al. (an assumed audit budget), and the protocols are built to spend it well. That work *simulates* the human with GPT-4; the "human auditor" is a model, so it is not evidence about people. So the safety literature supports P10 and P11 (do not rely on model-checks-model alone; put controls outside) and almost never tests the human reader. The exception is a small exploratory baseline in SHADE-Arena: annotators took about 103 minutes per transcript and reached AUC 0.65, worse than the top AI monitors (Kutasov, pp. 7–8). Chan et al. (2025) lay out controls outside the agent as a design programme and doubt that the market will supply oversight layers (pp. 13–14); that is the nearest academic counterpart to P11, P12 and P17. + +### C4. The human role: what *observe everything* and *emergency brake* ask of a person + +The post creates three human tasks. The evidence differs for each. + +| Task in the post | What the human is asked to do | Evidence | +|---|---|---| +| Reconstruct an outcome from a log (P15, P18) | Read a record of every meaningful model action after the fact | Closest studies are of alert review. Alert acceptance falls as alerts per encounter and repeats rise, while overall workload showed no association (Ancker). Alahmadi et al. describe alarm validation as "a tedious task", and one interviewee put false positives at 99% (qualitative; the paper measures no analyst performance). In Layman's two-group comparison, the 86% false-alarm group's median precision was 47 percentage points lower (0.33 against 0.80) and time on task 40% slower than the 50% group's. A new trace interface did not reliably cut error-finding time (the intervals span zero); confidence rose, accuracy did not improve meaningfully (Grunde-McLaughlin, Table 3, p. 18). One small baseline had annotators review agent transcripts: about 103 minutes each, AUC 0.65 (Kutasov, pp. 7–8). CERT: "Simply logging all network activity is not sufficient to protect an organization from malicious insider activity" (BP12). The same practice endorses retrospective review of leavers' last 90 days (p. 84) and monitors who hunt below the top alerts (p. 89), with no evidence on how well people do it. No study found of reconstruction from tamper-proof agent logs at scale. | +| Notice a fault and halt mid-task (P19) | Sustain attention on a rare event, then decide | Vigilance requires "hard mental work and is stressful" (Warm, abstract). Out-of-the-loop operators are slower to recover (Endsley and Kiris, abstract). Complacency appears mainly under multitask load (Parasuraman and Manzey, p. 387). Bainbridge's "impossible task" (p. 776). Claude Code users approve 93% of permission prompts (Anthropic engineering blog, 25 Mar 2026; the post attributes this to approval fatigue but reports no measure of inattention). Approval rose and scrutiny fell among 400 repeat reviewers (Yu, observational). | +| Hold responsibility (P3, P19) | Answer for the system | A human placed in the loop can become a "moral crumple zone" (Elish; Crootof et al. 483). Green: "evidence suggests that people are unable to perform the desired oversight functions". | + +*What this establishes.* The literature gives good reason to doubt a design that relies on a person noticing, in real time, a rare fault in a stream of agent activity. That bears on P19 and on any reading of P15 as continuous review. Complacency is mainly a multitask effect, so a dedicated watcher is the weaker case, but the authors still expect worse monitoring at very low failure rates (Parasuraman and Manzey, pp. 385, 387). It supports the author's claim that *observe everything* assigns the human a reading task humans do badly, *when it means real-time or whole-volume review*. + +*What it does not establish.* (1) The studies are mostly of single operators, alerts or approval prompts, not of forensic reconstruction by trained auditors. The post's stated purpose for P15 is reconstruction; CERT describes after-the-fact review as normal practice but gives no performance evidence. (2) Much evidence is observational or from small exploratory baselines. (3) The same literature prescribes artefact-side fixes (Bainbridge's pacing and criteria, Green's institutional review, Crootof's regulate-the-system). P11, P12 and P17 are artefact-side. The human-factors evidence therefore undercuts the *residual human tasks* in the post, not its deterministic backbone. (4) Green's remedy is institutional oversight, not user engagement. The evidence that humans are poor overseers does not by itself support an engagement layer. That needs a separate argument, from facets H and E. + +### C5. The transparency critique and the word "observability" + +Rieder and Hofmann present "observability" as a more pragmatic complement to transparency demands, so the author should not treat the word alone as a marker of the legibility paradigm. Nadella uses the word as the heading for all seven principles (P13). A careful contrast: they agree the black box will not simply be opened and that behaviour must be observed over time, but they present observability as not a replacement for transparency (p. 21); they differ on who observes, whom the observation serves, and whether the observed party holds power over the observer. Rieder and Hofmann's observers are outsiders in a public-interest institution. Nadella's observer is the deployer. They also treat access that the platform controls and can withdraw as no accountability (pp. 15, 22). + +Stohl et al. predict that making everything visible can produce opacity through overload. That is a reason to ask what "meaningful" (P15) filters. Ananny and Crawford's abstract asks whether opening black boxes would ever be sufficient, and proposes an alternative typology of accountability grounded in "constructive engagements with the limitations of transparency ideals"; the full argument is unread. + +### C6. Convergences, tested + +| | Convergence | Matches the design-ethics principle, or only resembles it? | +|---|---|---| +| (a) | CoT transparency on its own is "not sufficient or dependable" (P9). | **Partial match.** It agrees with Principle III that the model's reasoning is not a reliable remedy (Bainbridge p. 776 does not address model reasoning; the link to interpretability is the dossier's extension). But P9 keeps CoT transparency as "non-negotiable". He demotes reading the model's reasoning from sufficient to necessary; Principle III drops it as a remedy for the coordinated party. The post is thus not a legibility-free text. The reason differs too: P9 is about the *deployer's* assurance that a model is faithful; Principle III is about what a *lay coordinated party* can use. | +| (b) | Controls on access and actions sit outside the model (P11, P12). | **Real match in mechanism, different in audience.** Bounded outputs and operating envelopes (SAE's operational design domain and Leveson's constraint enforcement, as the repo dossier describes them; I did not reopen either) are also constraints external to the system. But Principle III is about *predictability for a human*: Bainbridge asks for "methods and criteria, and at a rate, which the operator can follow". P11 asks for *enforcement*, and nothing in the post says the envelope is declared to, or predictable for, anyone but the deployer. The envelope's owner is the organisation (P17), not the person the system acts upon. | +| (c) | Emergency brake (P19). | **Resemblance only.** Structural refusal (Lyons; Zong and Matias; Hirschman's exit and voice) is the *decision subject's* capacity to refuse or contest, built into the structure so that contest does not rest on a case-by-case burden. P19 is the *organisation's* refusal of the model, exercised by "an authorized person". The EU AI Act's 14(4)(e) stop button is the same kind of thing: a deployer-side overseer's halt. Neither gives the person on the receiving end a refusal. | +| (d) | The model is an "insider", an "actor" (P7). | **Partial.** The post concedes the model is a capable actor with access, which is a kind of partyhood. But the insider is *under the organisation's authority* (P5, P17), and P7 grounds the risk in mistakes or being compromised, not in the model's own objective. Lynch et al. use the same frame; in their setting a threat to continued operation alone was enough, with no goal given (p. 8), though the authors cannot say how far self-preservation, faulty reasoning or suspicion of the setup drove it (p. 9). In the author's triad the mediator has its own objective and reads both parties; P7 gives the model no objective. | +| (e) | Disclosure "to those affected" (P20). | **Real match in direction, weak in force.** It is the only place the post names people outside the deployer. It is after the fact, applies when systems fail or are compromised, and leaves who is affected undefined. It is notice, not standing. Chan et al. suggest identifiers could be visible to counterparties, and that an agent's user could be given the opportunity to request human review. | +| (f) | Psychological acceptability (Saltzer and Schroeder), not in the post. | The security canon the post invokes contains a human-factors principle the post does not use. This is a point of convergence with the paper's aims that Nadella could take up without changing his architecture. | + +The strongest three, in order: (b) in mechanism, (a) in its negative claim, (e) in direction. (c) and (d) are weaker. + +### C7. The three counts, tested against the text + +**(i) Legibility paradigm: "He answers opacity with more reading: observe, log, audit."** +*Verdict: partly holds; the stated form does not.* +- *For:* the verbs are in the text: observe (P4, P15), log (P8), audit (P18); "human readable evidence" (P15); CoT transparency as a non-negotiable (P9). P13 calls the whole list principles of observability. +- *Against:* P2 says mechanistic attribution is unavailable, and P4 refuses to treat the system as a set of black boxes, yet P9 still demands CoT transparency. Most principles are enforcement or independence (P11, P12, P14, P17), which need no reading. The reading tasks are P9, P15 and P19, and P15 is framed as reconstruction. The author's own ruling says DXD is "not for a user who reads the algorithm more skilfully". The reader in the post is not a user at all. +- *Complication (Problem 2):* the reference monitor moves work from a person to an artefact. If "spacing moved work from reader to artefact" is the author's mark of a good move, P11 qualifies. The distinguishing test that survives is, I infer, this: *after the artefact does its share, what residual task does the person keep, and do people do that task well?* Spacing left the reader with parsing already done and a task (reading words) people perform well. The post leaves a person with reading the evidence and deciding when to halt, the tasks Bainbridge calls what the designer "cannot think how to automate", and the ones the vigilance and alert literatures show people doing badly. A second test is *whom the move serves*: spacing served the reader; the reference monitor serves the data's owner. Both are my inferences, not findings from a source. +- *Safer wording:* "He answers opacity with controls outside the model, tamper-proof records a person can read, and independent audit. Where the design leaves a person a job (reading the records, deciding to halt), that job is the kind the oversight literature finds people do badly." + +**(ii) Keeps the dyad: "He treats the model as a contained insider under the organisation's authority."** +*Verdict: holds, once "dyad" is defined; the second half is the text's own claim.* +- *Parties in the picture:* the deploying organisation ("we", "us", "Orgs"); the model (insider, actor); the provider (supplier whose assurances do not transfer, P3); "an authorized person" (P19); those affected (P20); an "industry" (P6, P20). Authority belongs to the organisation: P5 divides who supplies intelligence from who holds "the authority over it", and P17 lets organisations decide for themselves what a model may access and do. +- *Second human?* Yes in the thin sense that "an authorized person" and "those affected" are humans. No counterpart in a transaction: the post never says whom the model acts toward (P2: "on our behalf") and never mentions a customer, user or third party except as the people affected in P20, after a failure. +- *"Dyad" needs a definition.* If it means organisation–model, the text fits. If it means the author's triad (two humans and a mediator), the post has no second human party mediated by the model. A reader could count provider, organisation and model as three. The author should say which dyad. +- *Note on CERT.* The insider frame comes from a field where the harmed party is the organisation, which explains why the counterpart is absent. +- *Safer wording:* "In the post, authority over the model rests with the deploying organisation (P5, P17). The only people outside the organisation who appear are the people affected, and they appear once, as recipients of disclosure after a failure (P20)." + +**(iii) Containment versus engagement: "No place where a user steers or rewrites the system."** +*Verdict: holds for users; fails if "no steering" is read as "no engagement".* +- *Text:* the post does not mention users. Steering rights exist but belong to the organisation: P17 (decide what the model can access and do). +- *Elsewhere on the same blog:* the corrections people make to models feed a learning loop that the firm owns (Jul 2026); "human agency will be the driver of token capital growth" (Jun 2026); AI as "scaffolding for human potential" (Dec 2025). So engagement at firm level is part of his position. The insider post is a security text and does not take it up. +- *Safer wording:* "The post gives steering rights to organisations, not to the people a system acts upon. Elsewhere Nadella describes firm-level learning loops in which employees' corrections improve models, but he does not extend them to users or counterparties." + +### C8. Scope caution: how far the comparison is fair + +The post is an enterprise-security architecture for deploying agentic systems. The author's proposal concerns interfaces for people who make decisions through platforms. The two answer different questions from different seats. To make the contrast honest, the author must concede: + +1. The post is not about interface design, and does not claim to be. Calling it "the opposite" is a category judgement; the safer claim is that it *leaves out* the coordinated party. +2. The deployer is the protected party in the post. In the setting it addresses (a firm deploying an agent), the organisation's authority over the agent is the correct answer to an authority question. The author's complaint is only that this answer is not enough where the agent also acts on third parties. +3. Convergences (a), (b) and (e) are real. Concede them in the text; do not hide them in a footnote. +4. Most of the post's mechanisms need no reading and are consistent with the human-factors advice to move work to the artefact. Only P15 and P19 leave a person with a task, and the paper's critique should be aimed there. +5. "Observability" is used favourably by scholars the paper may want to cite (Rieder and Hofmann). The paper needs a sentence on the difference in seat. +6. Nadella has said other things on his blog about people, learning and scaffolding. The paper should say "the insider post" when it means the insider post. +7. An interaction worth noting, as my inference only: a platform that follows P17 could use containment against a user's counter-delegating agent (Principle I), since that agent is also "a model with access". The two proposals are then not merely different; they may conflict in a platform setting. No source tests this. + +### C9. Claims the paper might make, tested + +| Claim ("Nadella proposes ...") | Supported? | Locator and safer wording | +|---|---|---| +| He answers opacity with more reading. | Partly | P15 "human readable evidence"; P4, P8, P18 verbs. Say: "with controls outside the model, tamper-proof records, and independent audit." | +| His remedy rests on humans reading evidence. | No | Controls sit outside the model (P11, P12), set independently by the organisation (P17), within deterministic system design that also names human controls (P6). P15 is a reading task (see C7 for P9 and P19). | +| He relies on chain-of-thought transparency. | Qualified | P9 makes it a non-negotiable but insufficient and undependable on its own. | +| He accepts that models cannot be explained. | Qualified | He says mechanistic attribution is not now possible (P2, P9). | +| He treats the model as an insider under the organisation's authority. | Yes | P5, P7, P8, P17, P19. | +| He keeps an organisation–model dyad. | Yes, if defined | No user, customer or counterpart named; "those affected" once (P20). | +| There is no second human in his picture. | No | "An authorized person" (P19), "those affected" (P20). They are not counterparts in a transaction. | +| Authority is held by the person or user. | Contradicted | P5, P17: authority is held by the organisation. | +| He denies users any role in steering. | Unsupported | The post is silent on users. Steering is vested in organisations (P17). Other posts describe firm-level correction loops. | +| He concedes the model is a party. | Partly | "Actor" (P7); not a party with its own objective. | +| His emergency brake is structural refusal. | No | P19 gives the halt to "an authorized person"; structural refusal belongs to the affected person. Say "resembles in form". | +| His external controls are bounded outputs. | Partly | P11, P12 match as constraint external to the model; no requirement of predictability to anyone but the deployer. | +| He invokes the reference monitor. | Partly | P11 refers to a principle from the 1970s; he names no source. | +| He invokes zero trust. | No | The term is absent; Levie uses it (Business Standard). Practices match NIST SP 800-207. | +| The post accompanies a Microsoft product or paper. | Unsupported | None found. RuntimeWire: "not a product launch". Microsoft's agent-security posts predate it. | +| His idea is the exact opposite of algorithmacy. | Not supported by the text | Shared premises: opacity; distrust of the model's self-report; controls outside the model. Say "answers a different question from the organisation's seat". | +| Experts have criticised it. | Unsupported | No critique found in two days. | +| Humans cannot oversee. | Overstated | Evidence is about sustained monitoring, alert review, approval prompts. Forensic reconstruction by trained auditors is less studied (CERT describes it as normal practice, without performance evidence). | +| Observability is a transparency or legibility term. | Qualified by Rieder and Hofmann | They present it as a more pragmatic complement to transparency demands. | +| Moving work to the artefact separates spacing from observability. | Fails | The reference monitor does it too. Use the residual-task and beneficiary tests (C7), flagged as the author's to rule. | +| The security canon ignores human factors. | Contradicted | Saltzer and Schroeder, "psychological acceptability". | + +### C10. Claims the sources contradict + +1. That the post rests on humans reading evidence. Its controls sit outside the model (P11, P12), set independently by the organisation (P17), within deterministic system design (P6). +2. That the post treats chain-of-thought transparency as a remedy. P9 calls it insufficient and undependable. +3. That the post has no human beyond the organisation. It names an authorized person (P19) and those affected (P20). The accurate claim is that neither is a counterpart in a transaction. +4. That "observability" belongs to the legibility camp. Rieder and Hofmann present it as a more pragmatic complement to transparency demands. +5. That the security canon neglects the human. Saltzer and Schroeder include psychological acceptability. +6. That the reference monitor differs from the spacing move because it moves work to the artefact. It does. +7. That the post is a Microsoft launch or has a companion. None found. +8. That Nadella has no engagement view. His earlier posts describe firm-level learning loops and AI as "scaffolding for human potential". +9. That his *emergency brake* gives refusal to the affected. P19 gives it to an authorized person. +10. That a "99% false positive" rate is a measured SOC figure. It is one interviewee's statement in a qualitative study; the analysts attribute most false positives to benign triggers. + +--- + +## (d) Gaps and unverified leads + +**Could not open or verify** +- *Anderson 1972, scanned Vol. I and Vol. II.* The UC Davis scan has no text layer (142 pp.). Volume II not opened. OCR would take under an hour; I did not run it. +- *Saltzer and Schroeder 1975, full paper.* The card read the MIT web edition's Sections I–III (Section II for argument only); the print pagination (1278–1308) was not checked against the IEEE original. +- *Mackworth 1948.* Paywalled (QJEP 1(1):6–21). Reach through a library. +- *Warm, Parasuraman, Matthews 2008; Endsley and Kiris 1995; Ananny and Crawford 2018.* Abstract only. A library with SAGE and Human Factors access will have them. +- *Bainbridge 1983.* Relied on a repo card. Parasuraman and Manzey 2010 was read in full from the journal-layout PDF (card: library/parasuraman2010complacency.md). +- *Tariq et al. 2025 (ACM CSUR).* Abstract only; blocked at the ACM Digital Library. +- *Cvach 2012 (alarm fatigue review); Sendelbach and Funk 2013; Sundaramurthy et al. 2015 (SOC analyst burnout).* Not verified; two Crossref lookups failed on rate limit. These are the better-known clinical and SOC fatigue sources and should be added on a retry. +- *Forbes (Linzainescu, 11 Oct), TheStreet, Android Headlines.* Blocked; need a browser. +- *EUR-Lex official text of Art. 14.* Empty response; I used the Future of Life Institute copy. +- *2026 arXiv preprints.* Chen, Yu, Grunde-McLaughlin, Mitchell, Turan and Jia now have cards read in full; Duzan, Gema, Little and Wang (MonitorBench) read in part; Baum and Kiener 2609.24242 is abstract only. All are preprints; MonitorBench carries a COLM 2026 header. +- *Santoni de Sio and van den Hoven:* body now read in full (card); the tracking condition reaches affected parties only through public rules (pp. 11–12). +- *Kroll 2021, Chan 2024 and Green 2022:* now read in full from the cards. + +**Leads not pursued** +- Nadella's podcast appearance (All-In) on monitoring every object an agent accesses: seen only as a search snippet. +- Suleyman, *The Coming Wave* (2023), whose "containment" vocabulary resembles P4/P19: not opened or verified. +- Kindervag's original zero-trust work and Google's BeyondCorp: not read. +- Chan et al. 2023 ("Harms from increasingly agentic algorithmic systems", already in the repo's dossier) and Chan et al. 2025 beyond the abstract. +- Geiping et al. 2025 (latent reasoning), cited by Korbak: not opened. +- Lyons (2021) and Zong and Matias (2024): relied on the repo dossier's description of structural refusal; I did not reopen them. +- A study of people reconstructing agent behaviour from logs (a search for such studies found none). + +**Open questions the author must rule on** +1. What "dyad" means in count (ii). +2. Whether to keep "legibility paradigm" as the label for count (i), or to say "evidence a person must read". +3. Whether the residual-task and beneficiary tests (C7) are the author's way to separate spacing from the reference monitor. +4. Whether to write "Nadella" or "the insider post" when claiming absence of engagement. +5. How much of the Anthropic-authored work (Lynch; Chen 2025; Anthropic 2026) to rely on, given that the drafting model is an Anthropic model. + +**Facts in the author's text I could not confirm** +- The author's account of the post as "the exact opposite" is an interpretation; the text does not say so and I found no source that does. +- His list "observe, log, audit" matches P4, P8, P15 and P18 as verbs; the text never says "read" except in "human readable". + + +--- + +## Changes after full-text cards (2026-10-13) + +Source: `../audit/CARD_FACET_CHECKS.md`, section "Facet G", checked against each card in `../library/` (the card, including its S2 section, is the authority). Read status in section (b) and in `G_nadella_lineage.bib` now follows the card. One line per change; the bib key leads. + +1. `jakkal2026secure`: read status now read in full (card); "blocking after compromise signals is automated" replaced by the page's "block agent actions", with no statement of who blocks; marketing page, no evidence of effect. +2. `constantin2026nadella`: read status read in full (card); article date (12 Oct 2026, 7:55 am) added to the entry and the bib note. +3. `kumari2026nadella`: read status read in full (card); the "when human approval is required" and recommend-a-transaction statements flagged as the reporter's, absent from P1-P21. +4. `anderson1972computer`: read status now Volume I in full (card); caution added that the table of contents numbers some sections differently from the body. +5. `saltzer1975protection`: read status restated from the card (MIT web edition, Sections I-III; Section II for argument only; footnotes page unread). +6. `rose2020zero`: read status restated (partial); administrator audit (p. 28) and stakeholder feedback (p. 41) added to the bearing. +7. `cert2022common`: read status restated (partial, list of unread practices); practice 12 also endorses retrospective review of leavers' last 90 days (p. 84) and monitors who hunt below top alerts (p. 89); the "titles only" remark for practices 5, 8, 21 replaced; same correction in C2, C4 row 1 and C4 "does not establish" (1). +8. `greenblatt2024ai`: read status restated (partial); the "human" auditor is GPT-4, so the paper is not evidence about people; "up to 2%" is an evaluation parameter. +9. `turpin2023language`: "36%" is 36.3 accuracy points in one setting (Claude 1.0: 30.6; Answer-is-Always-A: 18.7 and 4.7), not a relative loss; read status partial (card). +10. `lanham2023measuring`: read in full (card); size result is per task (7 of 8 tasks differ between 13B and 175B; faithfulness peaks at 13B on 6 of 8, p. 7); finding is for multiple-choice tasks with a fixed prompt format. +11. `chen2025reasoning`: read in full (card); "across six hint types" replaced by "most settings and models tested" (5 of 6 settings for Claude 3.7 Sonnet, 6 of 6 for R1, p. 6). +12. `korbak2025chain`: read in full (card); the post's "non-negotiable" is stronger than Korbak, who recommends monitoring "only as an addition ... not as a replacement" (p. 7) and declines to recommend showing the chain of thought to users (p. 7). +13. `baker2025monitoring`: read in full (card); the "illegible to a monitor" passage (p. 8) added to the P9 neuralese bearing. +14. `duzan2026chain`: read in part (card); the conditions most like real deployment are the ones where monitoring helps least. +15. `gema2026chain`: read in part (card); a narrower gap under the directive prompt can mean more unverbalised adoption on the user channel (pp. 13-14). +16. `little2026length`: read in part (card); the pressure is a training objective set by whoever fine-tunes, so it reaches a deployer only if the deployer trains (p. 1). +17. `wang2026monitorbench`: read in part (card); v4 carries a COLM 2026 header, so "preprint, not peer reviewed" was too flat. +18. `jia2026faithfulness`: read in full (card); gains are in visibility of the shortcut, not in its prevention. +19. `kutasov2025shade`: read in part (card); it has a human baseline (12 benign and 20 malign transcripts, three annotators each, about 103 minutes per transcript, pooled AUC 0.65, pp. 7-8); added to B3, C3 and C4 row 1, and to the "not found" line in B4. +20. `lynch2025agentic`: read in full (card); "the model has its own goals" replaced: a threat to continued operation alone sufficed (p. 8), and the authors cannot say what drove it (p. 9); same in C6(d). +21. `hao2024training`: read in full (card); Coconut trails language chain of thought on GSM8k (34.1% against 42.9%). +22. `kokotajlo2025ai`: "popularised the word" replaced: the page says only "We call this 'neuralese'", credits Hao et al., and says no frontier lab had adopted the technique. +23. `chan2024visibility`: read in full (card); the human-review deferral cites four works, not Green alone; the counterparty sees the identifier and agent card, not the monitoring or logs. +24. `chan2025infrastructure`: read in full (card); controls outside the agent laid out as a design programme (Table 1, p. 3); oversight-layer attention limit and doubt about market supply (pp. 13-14); rollback moral hazard (p. 19); added to B3 and C3. +25. `parasuraman2010complacency`: read in full (card); complacency is mainly a multitask effect (p. 387), backup-only detection "near perfect" (p. 384), so the brake-holder with no competing task is the weaker case; studies include accident reports, physician field trials and team studies; same qualification added to C4. +26. `ancker2017effects`: read in full (card); alerts per encounter was a strong predictor (IRR 0.70); only overall workload showed no effect; "uninformative" is a repetition proxy (p. 3); same in C4 row 1. +27. `alahmadi2022false`: read in full (card); no analyst performance measured; the 99% is one analyst's remark (B3, p. 2790); survey ratios 1 in 100 and 50 in 200 (p. 2788); studies alarm validation, not approval; same in C4 row 1. +28. `layman2023controlled`: read in full (card); the 47 is percentage points of median precision (0.80 against 0.33), about 59% relative; sensitivity difference p = .098 on few true alarms; same in C4 row 1. +29. `anthropic2026how`: read in full (card); the 93% is a rate of approvals, not a measure of inattention; the 52 real overeager actions come from denials and post-hoc flags; the bib note no longer says the page states the drafting model is Anthropic's. +30. `chen2026comparing`: read in full (card); the plan effect came partly from changing the agent. +31. `yu2026habituation`: read in full (card); internal inconsistencies in the paper noted (decile figures, 28% against 22%, rho = +0.02 with p = 0.009). +32. `grundemclaughlin2026overseeing`: read in full (card); "cut time to find errors" replaced: the overall time effect is small with an interval spanning zero; g = -0.65 applies to a subset and its interval also spans zero (Table 3, p. 18); same in C4 row 1. +33. `mitchell2026agents`: read in full (card); the paper argues design both obstructs (Section 3) and erodes (Section 4) oversight. +34. `turan2026oversight`: read in full (card); disagreement result (kappa = 0.52) is from LLM personas, not people. +35. `green2022flaws`: read in full (card); Green calls better oversight design an "open empirical question" (p. 11), so he does not say oversight can never work. +36. `crootof2023humans`: read in full (card); "an authorized person" reads only one of nine roles; the halt right is a resilience role and one person can hold both (p. 474). +37. `elish2019moral`: read in full (card); read status only. +38. `santonidesio2018meaningful`: read in full (card); the body brings in the affected only through public rules (pp. 11-12), not as a party; the "body not read" statement and gap entry replaced. +39. `sterz2024quest`: read in full (card); condition (d) fitting intentions, and role conflict (p. 8), added. +40. `laux2024institutionalised`: read in full (card); the body calls the empirical record "subject to debate"; the paper is a proposal with no new empirical test. +41. `langer2024effective`: read in full (card); cite as a framework paper with no data. +42. `rieder2020towards`: read in full (card); "both choose observation of behaviour" replaced: observability is "not meant to be a radical departure" from transparency (p. 21); disclosure can remain part of it (p. 13); platform-controlled data access is no accountability (pp. 15, 22); same in C5. +43. `stohl2016managing`: read in full (card); the strategic path (pp. 133-34) is closer to Nadella's seat than the inadvertent path. +44. `kroll2021outlining`: read in full (card); records are mostly about design and development, operations in section 3.3; human understanding treated as unsolved. + +Only `elish2019moral` needed nothing beyond a status update. Notes in other sections that name facet G (Facet Z note on `mitchell2026agents`) were applied through item 33. + +### Synthesis verdicts changed + +- C3 (AI-safety work): "does not test the human reader at all" became "almost never tests the human reader", with one small exploratory exception (Kutasov: about 103 minutes per transcript, AUC 0.65); the Greenblatt "human" is stated to be GPT-4. +- C3: "These are abstract-level and not peer reviewed." became "These are preprints (MonitorBench carries a COLM 2026 header), read in part or in full from the cards." +- B3 bearing on the 2026 preprints: "degrades under pressures the deployer can introduce by accident" became "degrades under pressures from training choices ... so it reaches a deployer only if the deployer trains, and ... helps least in the conditions most like real deployment (Duzan)". +- B3 bearing on Korbak: "P9's 'non-negotiable' matches Korbak" became "P9's 'non-negotiable' is stronger than Korbak". +- B4 bearing on Parasuraman and Manzey: "The brake-holder in P19 is the person these studies describe" became "The review is narrower than the brake-holder case", with automation bias named as closer to the P15 reader. +- C4 "What this establishes": doubt about a person noticing a rare fault in real time is kept but qualified (complacency is mainly a multitask effect; low failure rates still worsen monitoring). +- C4 "What it does not establish" (1) and (2): CERT's endorsement of after-the-fact review added; "Most evidence is abstract-only or observational" became "Much evidence is observational or from small exploratory baselines". +- C5 and B5 (Rieder and Hofmann): "they agree the black box will not be opened" became "will not simply be opened", with observability stated as not a replacement for transparency. +- C6(d): "Lynch et al. ... attribute goals to the models" became "a threat to continued operation alone was enough, with no goal given"; the verdict "Partial" stands. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_H_candidate_moves.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_H_candidate_moves.md new file mode 100644 index 00000000..86cb4342 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_H_candidate_moves.md @@ -0,0 +1,504 @@ +# H — Candidate moves: what could be the space between words for algorithmacy + +*Corrected on 2026-10-12 against `../audit/AUDIT_03.md`: 38 factual fixes; verbatim quotation of the Nadella post cut to 13 words.* + +*Facet H of the DXD foundation paper. Agent-drafted on 2026-10-12 (Sonnet 5.5, worktree `wt-dxd-foundation`). The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** Nine families of interface and system-design interventions that could play the role spacing played for the scribes in a coordination that runs through an algorithmic third party: (1) declared operating envelopes, behaviour descriptions and stable rules; (2) seams; (3) steering and control surfaces; (4) structural separation of what the mediator shows each party, with strategy-proof matching as the formal case; (5) defaults, standard forms and pre-segmented disclosures; (6) counter-delegation; (7) structural refusal, exit and contestability; (8) experiments in genuinely triadic settings; (9) the leads in the three Gemini essays. For each I report the measured evidence, who the participants were, and whether the setting had two humans and a mediator. Following the author's binding constraint, I sort every candidate by what it does with the work: **(i)** removes a task from the person by restructuring the artefact; **(ii)** hands the person a new reading, monitoring or steering task; **(iii)** hands the task to another agent on the person's behalf; **(iv)** removes the person's exposure (refusal, exit). The scribes' move is type (i). Transparency and explanation results enter only as evidence about when predictability holds (Ruling 1). + +**The question the synthesis answers.** Which candidate, if any, takes work off an ordinary person the way spacing took segmentation off the reader; what measured evidence supports it; and what nobody has yet tested. + +**Legend for read status.** *Read in full* = I read the whole text this session, or the entry reuses a repo card or earlier facet entry built from full text (named). *Read in part* = I opened the text and read named sections. *Abstract only* = a database abstract (Crossref, OpenAlex, Consensus) and nothing else; every number comes from that abstract. *Metadata only* = Crossref confirmed the record; I read no abstract. *Not accessible* = no text. **Second pass (2026-10-12).** I obtained an open copy of fifteen of the sources the synthesis leans on (ten read in part, five downloaded but read only to the abstract; see (d)) and recorded located quotations for the numbers in them. Every such copy is a working-paper, preprint or author version, not the version of record, and I say which. I opened each whole text, searched it for the figures and design statements, and read the passages I cite; I did not read any of them end to end, so I label them *read in part* and name the sections. None is *read in full*. Open copies obtained: Li 2017 (MPRA working paper), Rees-Jones and Skowronek 2018 (arXiv preprint), Arteaga et al. (NBER), Kling et al. (NBER), Mejia and Parker (SSRN preprint via the Indiana University figshare copy), Guess et al. 2023 (UNC repository), Zhu et al. 2026 (arXiv), Cabrera et al. 2023 (arXiv), Harper et al. 2015 (GroupLens author copy), Ehsan et al. 2024 (arXiv v2). **No open copy was found, so these stay abstract-only:** Hassidim et al. 2017 and 2021, Fradkin et al. 2021, Vaccaro et al. 2018, Moon et al. 2025, Seira et al. 2017 (the AEA PDF returned a bot wall). **Inman and Ribes 2019** was not opened (ACM bot wall, no author copy found); it rests on the repo card. Poursabzi-Sangdeh, Lukoff, Crisan, Allouah and Zhu S. et al. 2025 PDFs are in `dxd_foundation_local/pdfs/H/` but I did not read them past the abstract. **A source is quotable only at full-text depth (Ruling 4). None of these reaches it; the quotations below are for checking, and each needs the version of record before print.** All bibliographic fields in the `.bib` were taken from Crossref or the arXiv API on 2026-10-12. **Update (2026-10-13).** Full-text cards (library/.md) have since upgraded the read status of most entries below; where an entry says "read in full (card: ...)" or "read in part (card: ...)", that status supersedes the legend above and the second-pass lists in (d). The final section lists the changes. + +--- + +## (a) Search log + +All searches ran on 2026-10-12. Consensus returns abstracts; I did not apply filters except where noted. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | Repo reads | RESEARCH_PLAN, REUSE, SOURCE; Part 2 facet B; RESEARCH_DOSSIER 160-200; `triadic_hf_search_notes.md`; DIAL draft 330-341; ENGAGEMENT_LAYER; Part 1 lines 136-152; cards `inman-ribes-2019-beautiful-seams`, `midkiff1993`, `pritchett1994`, `hansman1995`, `counter_agent_search_notes` | n/a | Facet B already holds Cabrera 2023, DirectGPT, Kazemitabaar 2024, Fan 2026, Spatharioti 2025, Bucinca 2021, Fok & Weld 2024, Padilla 2018. I cite those from B and do not re-enter them. | +| 2 | Crossref `/works/{doi}` | 69 DOIs I proposed (first pass) | 55 resolved; 14 failed or returned the wrong paper | Wrong-paper hits were discarded and re-searched: my guesses for Kling et al. (`qjs004`), Seira et al. (`pol.20150282`), Bertrand & Morse (JF), Rees-Jones (`geb.2017.10.017`), Vaccaro 2018 (`3174137`), Alfrink (`s00146-022-01614-9`), Chen & Sonmez (`jet.2004.10.007`) were all wrong. | +| 3 | Crossref `query.bibliographic` | 28 title/author queries (Rees-Jones; Chen & Sonmez; Kling; Bertrand & Morse; Seira; Vaccaro & Karahalios; Alfrink; Fradkin; Bolton; Arteaga; Hassidim; Poursabzi-Sangdeh; Chalmers & Galani; Zhu; Allouah; Fogliato; Kapor; Gonczarowski; Benson; Bo & Hakimov; Guillen & Hakimov; Nosko & Tadelis; Silberman & Irani; Jahani; Dietvorst; Imas) | 2 per query | Gave the correct DOIs in the bib. **Not found:** Jahani et al. 2024 (arXiv 2407.14333, named in facet B) has no Crossref record; Silberman & Irani 2016 (*Comparative Labor Law*) is not indexed; Bertrand & Morse 2011 (*J Finance*) matched only SSRN versions. | +| 4 | OpenAlex `/works/doi:` abstract retrieval | 36 DOIs | 33 abstracts | Empty: Rees-Jones 2018 GEB, Chen & Sonmez 2006, Bertrand & Morse SSRN. | +| 5 | Consensus | `user controls over social media feed algorithm experiment effect on satisfaction, use of controls, field study` (year_min 2022) | 20 | Surfaced Hsu 2025, Alqabandi 2026, PILOT, Fabbri 2026, Gauthier 2026, Moon 2025, Wang 2024 | +| 6 | Consensus | `personal AI agent negotiates on behalf of user against counterpart or platform agent, human subjects experiment delegation` (year_min 2024) | 20 | Surfaced Zhu K. 2026, Liu 2026, Zhu S. 2025, Vaccaro 2026, Terrucha 2025, Alavi 2026, Hua 2026 | +| 7 | Consensus | `field experiment information shown to drivers or sellers on gig or marketplace platform changed behavior outcomes ride-hailing` | 20 | Surfaced Mejia & Parker, Li et al. 2026 (Lyft), Nosko & Tadelis, Ai et al. Most other hits are structural models or surveys, not interface experiments. | +| 8 | Consensus | `Rees-Jones suboptimal behavior in strategy-proof mechanisms residency match ...` (page_size 4) | 4 | Gave the PNAS experiment (n = 1,714) | +| 9 | Consensus | `Arteaga Kapor Neilson Zimmerman smart matching platforms ...` (page_size 3) | 3 | Abstract of the NBER paper | +| 10 | Consensus | `experiment obviously strategy-proof mechanisms lab test deferred acceptance versus clock, subjects truthful, Li` (page_size 6) | 6 | Gave Guillen and Veszteg 2021, Bo & Hakimov, Cerrone et al., Ashlagi & Gonczarowski, Hassidim et al. 2017. **My first-pass note that Li 2017 reports no experiment was wrong; see row 22.** | +| 11 | Consensus | `platform algorithm change without notice effect on sellers or workers earnings, opaque algorithm update, field evidence` | 20 | Mostly qualitative (Rahman, Bucher, Mendonca). **No controlled experiment on what a rule change without notice does to workers or sellers.** | +| 12 | Consensus | `drivers' understanding of automated driving system operational design domain limits, effect of information on mode confusion, experiment` | 20 | Surfaced Singer 2022, Zhou 2019, Boelhouwer 2019, Feinauer 2024, Matsuo 2026. Closest measured evidence for "declared envelope" outside LLMs. | +| 13 | Consensus | `uptake of non-personalized feed option Digital Services Act recommender opt-out share of users choosing` | 20 | **Found no measured uptake of the DSA Art. 38 option.** Only intentions (Starke 2024) and a lab prototype (Fabbri 2026). | +| 14 | Consensus | `spatial canvas or node-based interface for large language models user study compared with chat, sensemaking, controlled` (year_min 2023, page_size 8) | 8 | Sensecape, CanvasConvo, VisCanvas, ConceptVis, Gero 2024, Generative Interfaces, node-tree | +| 15 | Consensus | `white space layout density effect on comprehension and decision accuracy experiment; information overload ...` (page_size 8) | 8 | Moacdieh & Sarter 2017 (clutter), Wang 2025 (icon arrays). Nothing on "algorithmic white space". | +| 16 | arXiv API `id_list` | 2604.26220, 2607.02814, 2608.13787, 2509.19615, 2605.15848 | 5/5 | Titles match the Consensus records. (2510.06457, 2508.19227, 2602.08955, 2607.21886 are cited from Consensus only.) | +| 17 | curl + `pdftotext` | arXiv PDFs 2602.12089, 1802.07810, 2101.11778, 2205.02894, 2506.00073, 2508.02630 | 6 | Read Zhu 2602.12089 intro, §4.2-4.3, §6. Others abstract-level only. | +| 18 | curl | iu.tind.io PDF of Mejia & Parker | 0 | Empty file returned (first pass). Second pass: row 24. | +| 19 | Scholar Gateway | not tried | 0 | Failed in every earlier run (facet B row 7, F2); I did not spend a call on it. | + +| 20 | Unpaywall `/v2/{doi}` | 17 DOIs for the sources the synthesis leans on | 17 | Open copies listed for: Rees-Jones (arXiv 1802.01990, PMC6233132), Arteaga (NBER), Kling (PMC3314343), Mejia (Indiana repository, figshare 23894352), Guess (UNC), Cabrera (arXiv), Seira and Vaccaro 2018 and Inman (publisher PDFs, bot-walled). Closed: Li (AER), Hassidim 2017 and 2021, Fradkin, Harper, Ehsan (arXiv found separately), Moon. | +| 21 | curl with browser user-agent | arXiv 1802.01990, 2301.06937, 2211.06753; NBER w28946, w17410; UNC repository; GroupLens author copy; figshare file 41892513 | 8 of 11 PDFs obtained | Failed: ACM PDFs (Vaccaro 2018, Inman and Ribes), AEA PDF (Seira): HTML bot pages returned and not bypassed. Europe PMC REST for PMC3314343 returned HTTP 500; the NBER working-paper version of Kling et al. was used instead. | +| 22 | WebSearch | `Shengwu Li "Obviously Strategy-Proof Mechanisms" pdf` | 10 | Found MPRA paper 78930 (author-deposited working paper, version 17 Apr 2017) and a description of its lab experiment. Downloaded and read the experiment section. | +| 23 | WebSearch | `Fradkin Grewal Holtz "Reciprocity and Unveiling ..." pdf`; `Hassidim Romm Shorrer "The Mechanism is Truthful, Why Aren't You?" pdf`; `Inman Ribes "Beautiful Seams" ... pdf`; `Vaccaro "The Illusion of Control ..." pdf` | 10 each | No open full text for any of the four. Fradkin: SSRN 2939064 (landing page only), a 2015 NBER draft of a different paper. Hassidim: no relevant hit. | +| 24 | figshare API | article 23894352 (Indiana University repository record of Mejia and Parker) | 1 file | Downloaded `ndownloader.figshare.com/files/41892513` (31 pp.); the iu.tind.io URL still returns 403. | +| 25 | Crossref `/works/{doi}` | all 76 DOIs now in the bib, plus 10.1093/ej/uez036 | 77/77 resolved | Authors, container, volume, issue and pages in the `.bib` come from these records. This caught first-pass errors: Mejia's co-author is Parker, not Parsons; Guillen's co-author is Veszteg, not Hakimov; the 2021 Hassidim paper has three authors, not four; Bansal 2019 author order; Zhou 2019 authors (Zhou, Itoh, Kitazaki); Ehsan 2024 has five authors, not seven. | +| 26 | arXiv API `id_list` | 10 arXiv IDs (PILOT, VisCanvas, CanvasConvo, node-tree, Generative Interfaces, Lyft, Alavi, Liu, Hua, Zhu) | 10/10 | Titles and author lists in the `.bib` come from this. All now checked. | +| 27 | Nadella capture | `~/iit-playground/dxd_foundation_local/nadella_2026-10-10_models_as_insider_risks.txt` | 21 content paragraphs (P1-P21; P22 is the capturer's end marker) | Read in full. Mapping in section (c)(v) is redone from the text with paragraph numbers. | + +Not done: no citing-work snowballing; no look at ACM, Elsevier or Wiley full text (bot walls in earlier runs); no search for ODD user studies beyond driving; no search on Apple privacy-label evaluations; none on Google/Meta "why am I seeing this" studies beyond the Consensus hits. + +--- + +## (b) Entries + +Format: **bib key** — reference. *Verification.* *Read status.* *What it establishes.* *Move type* and *triadic?* *Bearing.* Facet-B entries are named and not repeated. + +### B1. Declared envelopes, behaviour descriptions, stable rules + +**cabrera2023improving** — Cabrera, Á. A., Perer, A., & Hong, J. I. (2023). Improving human-AI collaboration with descriptions of AI behavior. *PACM HCI, 7*(CSCW1), 1-21. https://doi.org/10.1145/3579612. *Verification:* Crossref; arXiv:2301.06937. *Read status:* read in full (card: library/cabrera2023improving.md; arXiv v1). +- *Design.* "3x3 between-subjects study, 25 participants per condition, 225 participants total" across three tasks (§4, Fig. 2). "Of the 225 participants, we removed 13 for failing the attention check. Additionally, we removed three other participants ... who had an accuracy of less than 10%" (§4.1), which leaves 209 analysed; the degrees of freedom of the three omnibus tests imply 67, 70 and 68 analysed (205 in all), and the paper does not reconcile the two. The AI was a "Wizard-of-Oz" system with set accuracies: 95% on 20 instances and 40% and 20% on two subsets of five (Fig. 2). Pay was $2 for about 15 minutes. +- *Results.* Reviews task: "The only significant pairwise difference was between the No AI and AI + BD conditions (p < 0.001, 95% C.I. = [0.06, 0.22]). While the AI by itself did not significantly improve the accuracy of the participants, AI supplemented with behavior descriptions led to significantly higher team accuracy" (§5, F(2,64) = 9.18). Satellite task: "no significant difference between conditions (F(2,67) = 0.63, p = 0.534)". "Despite the increased performance, there was no human-AI complementarity" (reviews). Birds (p = 0.048 in facet B) is confirmed. Most of the gain came from relying more on the AI where it was right, not from catching its errors. +- *Correction to facet B and the first pass:* 225 is the number recruited, not the number analysed; 209 were analysed by the text (205 by the degrees of freedom). *Move type:* (ii) with a (i) flavour (a short statement of where the system fails, which the person reads). Dyadic. Best evidence that a declared envelope can raise performance, mostly through heavier reliance on the AI where it was right; hand-set accuracies, so it tests the display, not a real system's envelope. + +**bansal2019updates** — Bansal, G., Nushi, B., Kamar, E., Weld, D. S., Lasecki, W. S., & Horvitz, E. (2019). Updates in human-AI teams: Understanding and addressing the performance/compatibility tradeoff. *Proceedings of the AAAI Conference on Artificial Intelligence, 33*(1), 2429-2437. https://doi.org/10.1609/aaai.v33i01.33012429 +- *Verification:* Crossref (title, authors, pages); OpenAlex abstract. *Read status:* read in full (card: library/bansal2019updates.md). +- *What it establishes:* "updates that increase AI performance may actually hurt team performance." The authors define compatibility of an update with prior user experience and show that "current machine learning algorithms do not produce compatible updates" on three high-stakes classification tasks; a retraining objective that penalises new errors recovers compatibility at some cost in accuracy. Participant numbers: 25 workers per condition in the mental-model studies, with the bottom quartile of scorers removed as likely spammers (p. 2433); the group sizes for the update study are not stated. +- *Move type:* (i), applied to the mediator rather than the interface: freezing the error pattern removes the task of re-learning where the system fails. Dyadic (one human, one classifier). *Bearing:* the one controlled experiment on backward-compatible updates. The crowd participants played only the abstract CAJA game; the recidivism, mortality and credit-risk work used classifiers alone, with no people. It is not a platform where a rule change reaches workers. + +**chen2024chatgpt** — Chen, L., Zaharia, M., & Zou, J. (2024). How is ChatGPT's behavior changing over time? *Harvard Data Science Review, 6*(2). https://doi.org/10.1162/99608f92.5317da47 +- *Verification:* Crossref; OpenAlex abstract. *Read status:* read in full (card: library/chen2024chatgpt.md). +- *What it establishes:* The "same" LLM service changed between March and June 2023: GPT-4 prime-versus-composite accuracy fell from 84% to 51%; GPT-4 grew less willing to answer sensitive questions; code-formatting errors rose. Measures the model, not people. The cleanest evidence is for instruction-following drift, which the authors offer as a partial explanation, not a proven one. +- *Move type:* n/a (motivates (i) stability). *Bearing:* shows that silent updates are a real, measured property of the mediator. It supplies no evidence that stability helps people. + +**allouah2026what** — Allouah, A., Besbes, O., Figueroa, J. D., Kanoria, Y., & Kumar, A. (2026). What is your AI agent buying? Evaluation, biases, model dependence, and emerging implications for agentic e-commerce. *Proceedings of the ACM Web Conference 2026* (Crossref pages 8697-8700, a four-page item; the findings quoted here come from the longer arXiv:2508.02630 v3 abstract). https://doi.org/10.1145/3774904.3792943 +- *Verification:* Crossref; arXiv 2508.02630v3. *Read status:* read in full (card: library/allouah2026what.md). +- *What it establishes:* Shopping agents concentrate demand on a few "modal" products; "model updates can drastically reshuffle market shares"; randomised trials show position biases that vary by provider and model version and persist in text-only interfaces; agents penalise sponsored tags and reward platform endorsements; a seller-side agent that tweaks descriptions gains market share, but not generally: the gain occurs in 33% of category-model pairs and sometimes reverses (p. 22). The authors also report that basic rationality has become near-perfect (p. 13); the instability is in preferences and position, not competence. The headless replication used three August-generation models only. +- *Move type:* evidence on (iii). *Triadic?* Partly: consumer agent, platform, seller, with no human in the loop during the choice. *Bearing:* it is the strongest evidence that moving the task to an agent moves the legibility problem to the agent: instability across model updates and position bias are, on the project's inference, properties the person could no longer see (no person appears in the study). + +**rahman2021invisible** — Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. *Administrative Science Quarterly, 66*(4), 945-988. https://doi.org/10.1177/00018392211010118 +- *Verification:* Crossref (volume 66, issue 4, pages 945-988). *Read status:* abstract only. +- *What it establishes:* Three years of interviews, archives and participant observation on a freelance platform (n of interviews not in the abstract). "the criteria for success and changes to those criteria are unpredictable." Workers experimented with the criteria or withdrew activity; reactions depended on prior success and platform dependence. +- *Move type:* n/a. *Triadic?* Yes in structure (freelancer, client, rating algorithm), qualitative. *Bearing:* the best account of what unannounced criteria change does to ordinary workers. It motivates stability and notice but measures no intervention. + +**Also in this family, qualitative only:** bucher2020pacifying (12,294 scraped freelancer comments; "anticipatory compliance" with an assumed algorithm), mendonca2022disconnecting (one UK courier case; unilateral intra-platform changes fragment collective action). Abstract only for both. Neither tests an intervention. + +**mitchell2019model** — Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2019). Model cards for model reporting. *FAT\* '19*, 220-229. https://doi.org/10.1145/3287560.3287596 +- *Verification:* Crossref. *Read status:* read in full (card: library/mitchell2019model.md). *What it establishes:* a proposal for a documentation format. *Move type:* (ii). *Bearing:* a proposal, not a measured effect. + +**crisan2022interactive** — Crisan, A., Drouhard, M., Vig, J., & Rajani, N. (2022). Interactive model cards: A human-centered approach to model documentation. *FAccT '22*, 427-439. https://doi.org/10.1145/3531146.3533108 +- *Verification:* Crossref; OpenAlex abstract; arXiv 2205.02894. *Read status:* read in full (card: library/crisan2022interactive.md). +- *What it establishes:* A think-aloud and interview study with 30 participants (experts, then non-expert analysts who use NLP models) comparing standard and interactive model cards. The findings are thematic ("understandability", "sensemaking & skepticism", "trust & safety"). No performance measure; the authors do not test the interactive card's effect on understanding the way they count comprehension for the standard card. +- *Move type:* (ii). Dyadic. *Bearing:* the only user study of model cards with non-experts that I found; it measures reactions, not accuracy or workload. + +**Driving automation, the nearest measured "declared envelope" outside LLMs** (Singer, Boelhouwer and Matsuo read in full from cards; Zhou abstract only; all dyadic: driver and system): +- **singer2022driver** — Singer, J., Tefft, B. C., Benson, A., Jenness, J. W., & Horrey, W. J. (2022). Driver expectations of a partial driving automation system in relation to branding and training. *Human Factors, 66*(5), 1531-1544. https://doi.org/10.1177/00187208221143024. 90 drivers (read in full; card: library/singer2022driver.md), randomly assigned to a name and training that stressed capabilities (AutonoDrive) or limitations (DriveAssist), and to three training modes. Capability framing raised expectations of crash avoidance "in a variety of driving scenarios, including many in which the system would not work", and willingness to take more risk; most differences persisted after driving. "Accuracy of technical information does not guarantee understanding." Move type (ii). +- **zhou2019effect** — Zhou, H. P., Itoh, M., & Kitazaki, S. (2019). Effect of instructing system limitations on the intervening behavior of drivers in partial driving automation. *Cognition, Technology & Work, 22*(2), 321-334. https://doi.org/10.1007/s10111-019-00568-1. A 2 x 2 simulator study with 24 novice drivers per condition. A description of limits in terms of scenes (not functions) gave 95% safe interventions and 2.0 s reaction time against 88% and 2.3 s. Move type (ii). +- **boelhouwer2019should** — Boelhouwer, A., van den Beukel, A. P., van der Voort, M. C., & Martens, M. H. (2019). Should I take over? *Transportation Research Part F, 60*, 669-684. https://doi.org/10.1016/j.trf.2018.11.016. 28 participants, video simulator. "the system information did not appear to support the participants in correctly deciding whether to take over"; the authors write that the mental models "did not seem to (sufficiently) change" (abstract), inferred from decision accuracy and comments; they did not elicit the models. Group 2 was nominally higher, and the information group learned less over the session. Move type (ii). Null (14 per group, so a failure to detect). +- **matsuo2026educational** — Matsuo, R., Liu, H., Hiraoka, T., & Wada, T. (2026). An educational human machine interface providing request-to-intervene trigger and reason explanation. *IEEE Transactions on Human-Machine Systems, 56*(3), 650-661. https://doi.org/10.1109/thms.2026.3679748. Between-group simulator study, N = 45 (15 per group). Comprehension: cue plus reason beat no cue (adjusted p = 0.001) and cue only (p = 0.017). Collisions in the final failure trial: 7 of 15 with no cue, 6 of 15 with the cue only, 1 of 15 with cue and reason (Fisher p = 0.044). The cue alone gave a small, non-significant timing change; the reason carried the result. Move type (ii). +- *Bearing of the group.* Telling drivers the limits has mixed effects: it helps when framed concretely (Zhou), backfires when the framing flatters (Singer), and fails when delivered as a manual (Boelhouwer); in Matsuo the reason, not the cue, carried the result. The envelope *as an announcement* is a reading task. An envelope enforced by the artefact (the vehicle refuses outside its ODD) is a different object, and I found no study that compares announced and enforced envelopes. The SAE J3016 term is held in the dossier; I did not open the standard. + +### B2. Seams + +**chalmers2004seamful** — Chalmers, M., & Galani, A. (2004). Seamful interweaving: Heterogeneity in the theory and design of interactive systems. *DIS '04*, 243-252. https://doi.org/10.1145/1013115.1013149. *Verification:* Crossref; OpenAlex abstract. *Read status:* abstract only. *What it establishes:* a theoretical argument, grounded in "an experiment with a mixed reality system", that people "accommodate and take advantage of seams". No effect sizes in the abstract. The earlier lab read Chalmers (2003) in full (`chalmers-2003-seamful-design-ubicomp-workshop`); that card is the source for the original claim. + +**inman2019beautiful** — Inman, S., & Ribes, D. (2019). "Beautiful seams": Strategic revelations and concealments. *CHI '19*, 1-14. https://doi.org/10.1145/3290605.3300508. *Verification:* Crossref (two authors). *Read status:* **not opened.** Unpaywall lists a publisher PDF, but the ACM page returned a bot wall; a web search found no author copy (log rows 21, 23). The only source is the earlier run's card (`inman-ribes-2019-beautiful-seams`): seamless and seamful designs distribute "downstream user agency" differently. **The card contains no statement about word separation or scriptio continua.** The sentence in the DIAL draft is set beside the card in (c)(vi), item 4. *Move type:* conceptual; no measurement. + +**ehsan2024seamful** — Ehsan, U., Liao, Q. V., Passi, S., Riedl, M. O., & Daumé III, H. (2024). Seamful XAI: Operationalizing seamful design in explainable AI. *Proceedings of the ACM on Human-Computer Interaction, 8*(CSCW1), 1-29. https://doi.org/10.1145/3637396 +- *Verification:* Crossref (five authors, matching the repo card and the arXiv title page). *Read status:* read in part (card: library/ehsan2024seamful.md) from arXiv:2211.06753 v2 (5 Mar 2024): introduction, background, design process, method and §5.1 in full; §§5.2-5.4 and appendices sampled. The version of record was not read. +- *What it establishes:* Participants were 43 people, "AI practitioners and real end-users", who worked through a scenario-based co-design activity set in lending (§4); the 258 sign-up submissions were screened to 43 and 2,796 minutes of session video were transcribed (§4.3-4.4). "Our 43 participants generated a total of 85 seams" (§5.1.1). The agency findings are what participants said: "Participants' comments also validated the end-goal of seamful XAI in supporting user agency" (§5.2.2). The paper reports no instrument for sense of agency and no before-and-after comparison. Per the card, no end-user used a seamful interface and no measure of user action, accuracy or reliance appears; the authors call the findings "formative rather than evaluative" (§6). +- *Move type:* (ii), squarely: seams hand the person something to read and act on. Dyadic. *Bearing on Ruling 1:* a legibility remedy by another name; the evidence for it is a qualitative design study. + +### B3. Steering and control surfaces + +**harper2015putting** — Harper, F. M., Xu, F., Kaur, H., Condiff, K., Chang, S., & Terveen, L. (2015). Putting users in control of their recommendations. *RecSys '15*, 3-10. https://doi.org/10.1145/2792838.2800179 +- *Verification:* Crossref (six authors). *Read status:* read in full (card: library/harper2015putting.md) from the GroupLens author copy (8 pp.). The ACM version was not read. +- *What it establishes:* MovieLens users were emailed an invitation; "a pool of 381 users (19% of those emailed) ... clicked on the email link and were assigned to an experimental condition". "For examining use of the recommender tuner interface, we look at users who (a) used both the left and the right buttons, (b) took three or more actions, and (c) clicked I'm done! (N=168, 8.3% of those emailed)"; 148 completed the survey, 85 in the popularity condition and 63 in the age condition (§5.2). "Overall, subjects used a median of 10 actions in tuning their list of recommendations ... 15% of subjects chose the original recommender as their favorite configuration, while the remaining 85% chose a configuration one or more steps from the initial setting." The tuned list was rated more favourably than the original on every reported question ("All differences ... are statistically significant (p<0.001) using a Wilcoxon rank sum test", §5.2). In the popularity condition the share rating the list "just right" on obscurity rose from 31% to 49% (Fig. 5, read from the plotted percentages). "These samples skew towards highly-active users", with a median of 186 logins against 48 for emailed users generally (§5.2). +- *Move type:* (ii). Dyadic. *Bearing:* the outcome is rated list quality, not decision quality, and the users who used the tuner are a self-selected 8% of those invited. + +**vaccaro2018illusion** — Vaccaro, K., Huang, D., Eslami, M., Sandvig, C., Hamilton, K., & Karahalios, K. (2018). The illusion of control: Placebo effects of control settings. *CHI '18*, 1-13. https://doi.org/10.1145/3173574.3173590. *Verification:* Crossref (six authors). *Read status:* abstract only; the ACM PDF returned a bot wall and no other open copy was found. *What it establishes:* in "two paired studies", control settings "functioned as placebos. Viewers felt more satisfied with their feed when controls were present, whether they worked or not." Controls in the wild "had little impact on user's satisfaction"; people improvised ("scrolling quickly"). *Move type:* (ii). + +**hsu2025placebo** — Hsu, S., Koshy, V., Vaccaro, K., Sandvig, C., & Karahalios, K. (2025). Placebo effect of control settings in feeds are not always strong. *CHI '25*. https://doi.org/10.1145/3706598.3714197. *Read status:* abstract only. *What it establishes:* N = 762, a Twitter feed with no control, a working control or a placebo control; "a placebo effect much smaller in magnitude than originally reported." *Bearing:* qualifies Vaccaro et al. The satisfaction gain from merely having a control is real but small; neither paper measures whether the control changes what the person decides. + +**alqabandi2026experiments** — Alqabandi, F., Tierney, G., Bail, C., Hillygus, D. S., & Volfovsky, A. (2026). Experiments offering social media users the choice to avoid toxic political content. *New Media & Society*. https://doi.org/10.1177/14614448261418999. *Read status:* abstract only. *What it establishes:* offering a filter raised platform satisfaction, but users who chose to filter "rated content as more hostile than similar respondents who were not offered the choice"; the authors read this as raised expectations and closer scrutiny. Fictional platform, sample size not in the abstract. *Bearing:* a control can add a monitoring task (scrutiny) while it adds satisfaction. + +**moon2025effects** — Moon, J. H., Kim, S., Jung, Y., Bang, J., & Sung, Y. (2025). The effects of explainability and user control on algorithmic transparency: The moderating role of algorithmic literacy. *Cyberpsychology, Behavior, and Social Networking, 28*(7), 497-504. https://doi.org/10.1089/cyber.2024.0525. *Verification:* Crossref (five authors). *Read status:* abstract only; no open copy found (Unpaywall: closed). *What it establishes:* a 2 x 2 x 2 between-subjects experiment, n = 240, fictitious short-video platform. "Both explainability and user control enhanced perceived algorithmic transparency, legitimacy, and satisfaction. When neither feature was present, algorithmic literacy had no significant impact. However, when at least one feature was present, literacy significantly influenced the dependent variables." Perceptions only, no decisions. **The Part 1 sentence is set beside this one in (c)(vi), item 2.** + +**lukoff2021how** — Lukoff, K., Lyngs, U., Zade, H., Liao, J. V., Choi, J., Fan, K., Munson, S. A., & Hiniker, A. (2021). How the design of YouTube influences user sense of agency. *CHI '21*, 1-17. https://doi.org/10.1145/3411764.3445467. *Verification:* Crossref (8 authors); OpenAlex abstract; arXiv 2101.11778. *Read status:* read in full (card: library/lukoff2021how.md). *What it establishes:* a survey of 120 US users found "autoplay and recommendations primarily undermine sense of agency, while playlists and search support it"; 13 co-design sessions found that people with a specific intention preferred interfaces that gave more agency. Sense of agency, not performance. *Bearing:* search and playlists support agency, and the abstract leaves open whether that is restructuring or tuning. In the full text the "high-control" mockups did remove recommendations (restructuring), but the authors' own proposals are mostly controls the user operates: customisation, explicit preferences, a toggle between Explore and Focus modes (§5.3); users also wanted less control for "ritualized" use. The paper is evidence for context-dependence, not for restructuring over tuning. + +**knijnenburg2012inspectability** — Knijnenburg, B. P., Bostandjiev, S., O'Donovan, J., & Kobsa, A. (2012). Inspectability and control in social recommenders. *RecSys '12*. https://doi.org/10.1145/2365952.2365966. *Read status:* abstract only. *What it establishes:* an online experiment, N = 267, Facebook music recommender: "inspectability and control indeed increase users' perceived understanding of and control over the system, their rating of the recommendation quality, and their satisfaction." Perceptions. + +**goyal2025pilot** — Goyal, A., Choi, F., & Chandrasekharan, E. (2025). PILOT: Control surfaces for authoring social media feeds. arXiv:2509.19615. *Read status:* read in full (card: library/goyal2025pilot.md; arXiv v2); preprint. *What it establishes:* seven participants informed the design; ten compared three nested control surfaces within subjects. "participants' experiences depended more on whether they could verify and repair outcomes" than on how much control they had. *Bearing:* small. It resembles Fok and Weld's verification theory (facet B), but this paper does not cite Fok and Weld, so the link is this facet's, not the authors'. + +**fabbri2026feeding** — Fabbri, M., et al. (2026). Feeding the (short-video) feed: A design proposal for user control of social media recommender systems under the Digital Services Act. *International Journal of Human-Computer Studies*. https://doi.org/10.1016/j.ijhcs.2026.103811. *Read status:* abstract only. *What it establishes:* a preliminary study of a prototype: "users show the tendency not to use the control features when not prompted to do so." Sample size not in the abstract. + +**Whether controls get used (field):** I found no measured uptake of any platform's control surface from platform data (log rows 5, 13). The two uptake signals are intentions (starke2024contesting, below) and prototype behaviour (Fabbri). Habib et al. (2022, *PACM HCI*, https://doi.org/10.1145/3512906) surveyed Facebook ad controls and report that "participants seemed to be unaware of them, suggesting issues of discoverability" (abstract only). + +**Steering by exit, measured at scale.** +- **guess2023how** — Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H., Brown, T., et al. (2023). How do social media feed algorithms affect attitudes and behavior in an election campaign? *Science, 381*(6656), 398-404. https://doi.org/10.1126/science.abp9364. Crossref lists 29 authors. *Read status:* read in full (card: library/guess2023how.md; the 7-page UNC repository copy; supplementary materials not available). *What it establishes:* "We assigned a sample of consenting users to reverse-chronologically-ordered feeds instead of the default algorithms. Moving users out of algorithmic feeds substantially decreased the time they spent on the platforms and their activity"; the chronological feed did "not significantly alter levels of issue polarization, affective polarization, political knowledge, or other key attitudes during the 3-month study period" (Abstract). The analytic samples were "Facebook: n = 23,391; Instagram: n = 21,373" (Methods; two-column layout, read from the extracted text). Move type (iv)/(ii) in effect, but users were randomised into the chronological feed, not given a choice: the experiment tests removing the ranker, not people choosing to drop it. Dyadic. +- **gauthier2026political** — Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E. (2026). The political effects of X's feed algorithm. *Nature, 652*(8109), 416-423. https://doi.org/10.1038/s41586-026-10098-2. Read in full (card: library/gauthier2026political.md; Supplementary Information not read): a 2023 field experiment, algorithmic versus chronological feed for 7 weeks. Switching on the algorithm raised engagement and shifted opinion; switching it off had no comparable effect. *Bearing:* exit does not undo what the mediator did during exposure (persistent following effects). The authors flag that the asymmetry may also come from who chose the chronological feed: "selection on unobserved factors may contribute to this asymmetry" (p. 420). The following mechanism is described as suggestive, and the effects are concentrated among Republicans and Independents, with Democrats largely unaffected (Extended Data Fig. 4). +- **wang2024lower** — Wang, S., Huang, S., Zhou, A., & Metaxa, D. (2024). Lower quantity, higher quality. *PACM HCI, 8*(CSCW2), 1-25. https://doi.org/10.1145/3687046. 243 users, randomised algorithmic versus chronological timelines, crossover; behaviour changed, perceptions of the platform barely. Read in full (card: library/wang2024lower.md). The abstract says "slightly more reliable" news in the algorithmic timeline, but the results report no significant difference in reliability (p. 10), so the headline "higher quality" rests on lower congruence and extremity only. + +**Structured and non-chat interfaces for LLMs** (facet B holds DirectGPT, Kazemitabaar 2024, Fan 2026; its verdict stands: structure raised perceived control and cut prompt effort, and showed no objective gain in success or verification hints at n = 18). New since B, all dyadic (since read from cards: Sensecape, Gero, CanvasConvo, node-tree and Generative Interfaces in full, VisCanvas in part): **suh2023sensecape** (UIST 2023; within-subjects study, N = 12, with exploration counts, not decision quality, as the outcome; users "explore more topics and structure their knowledge hierarchically"); **gero2024supporting** (CHI 2024; controlled study n = 24 plus eight case studies; features for comparing many LLM outputs made hard tasks tractable, which rests on self-report and one of two tasks; the email task showed no difference); **ueno2026viscanvas** (arXiv 2607.21886; node-based canvas against chat, 20 participants, cognitive load and usability "indistinguishable from current prevailing methods"; "more diverse" rests on topology counts coded by the authors, not on counts of charts or variables, which were equal, and only one of eight feature ratings was significant); **amin2026conversations** (arXiv 2605.15848; canvas plus chat, 5-day field deployment, 24 participants; "Adoption was uneven"; chat stayed central; the paper is a working draft marked not for distribution); **wang2025evaluating** (arXiv 2510.06457; node-tree against chat, n = 20 business users; the abstract claims better task performance and higher trust, but the body gives only percentage descriptives, one non-significant test, and an admission that significance is not reported, and "task performance" is a helpfulness rating, not a measured outcome); **chen2025generative** (arXiv 2508.19227; "up to a 72% improvement in human preference", which is preference, not performance; the 72 is a margin in percentage points, 84% wins minus 12% losses against the Claude 3.7 chat baseline in Table 1, not a ratio). *Move type:* (ii) for all of them; the person gets a spatial artefact to manage. None tests a decision made through a third party. + +**jannach2017user, he2016interactive** — Jannach, D., Naveed, S., & Jugovac, M. (2017). User control in recommender systems: Overview and interaction challenges. *EC-Web 2016*, LNBIP, 21-33. https://doi.org/10.1007/978-3-319-53676-7_2. He, C., Parra, D., & Verbert, K. (2016). Interactive recommender systems: A survey of the state of the art and future research challenges and opportunities. *Expert Systems with Applications, 56*, 9-27. https://doi.org/10.1016/j.eswa.2016.02.013. *Verification:* Crossref (author counts 3 and 3). *Read status:* metadata only; I did not retrieve abstracts. They are surveys; use them for the map, not for effects. + +### B4. Structural separation and strategy-proof design + +This family is the closest formal analogue of type (i) in a coordination with a mediator, and it needs care. A strategy-proof mechanism makes truthful reporting a dominant strategy, so the participant no longer has to model what others will report. The task "guess the counterpart and the mediator" is removed by the rule, not taught. + +**li2017obviously** — Li, S. (2017). Obviously strategy-proof mechanisms. *American Economic Review, 107*(11), 3257-3287. https://doi.org/10.1257/aer.20160425 +- *Verification:* Crossref (single author). Open copy: MPRA paper no. 78930, "This version: 17 April 2017", 62 pp., http://mpra.ub.uni-muenchen.de/78930/. *Read status:* read in part (card: library/li2017obviously.md) from that working paper; proofs and appendix tables skimmed. The AER version was not read; check that the experiment survives there. +- *What it establishes (theory).* "A mechanism is obviously strategy-proof (OSP) if it has an equilibrium in obviously dominant strategies" (Abstract). Introduction: "choosing when to quit in an ascending clock auction is the same as choosing a bid in a second-price sealed-bid auction ... Nonetheless, laboratory subjects are substantially more likely to play the dominant strategy under a clock auction than under sealed bids (Kagel et al., 1987)." +- *What it establishes (experiment).* "I conduct a laboratory experiment to test the theory. In the experiment, I compare three pairs of mechanisms. In each pair, both mechanisms implement the same choice rule. One mechanism is obviously strategy-proof. The other mechanism is strategy-proof, but not obviously strategy-proof ... subjects play the dominant strategy at significantly higher rates under the OSP mechanism ... This effect occurs for all three pairs of mechanisms, and persists even after playing each mechanism five times with feedback" (Introduction). Design (§5.1): across-subjects, "144 subjects in 36 groups of 4 (with 9 groups in each treatment)", recruited from the student population at the Ohio State University Experimental Economics Laboratory, August 2015; subjects "made $37.54" on average. Results (§5.2 for the auction pair; the serial-dictatorship figures are in §5.3 and compare only those two games, since the auction pairs use a different measure): "An auction is 31 percentage points more likely to have a closing price within $2.00 of the equilibrium price under AC (OSP) compared to 2P (SP)"; in the random-serial-dictatorship pair, "Under SP-RSD, 36% of games do not end in the dominant strategy outcome. Under OSP-RSD, 7% of games do not end in the dominant strategy outcome ... In SP-RSD, 29% of submitted rank-order lists contain errors." +- *Move type:* (i), and the only source here that states the design move as removing a cognitive task. Participants: university students in stylised games, not novices in a real market. **Correction:** my first pass said Li reports no experiment of its own. That was an inference from the abstract and was wrong. + +**ashlagi2018stable** — Ashlagi, I., & Gonczarowski, Y. A. (2018). Stable matching mechanisms are not obviously strategy-proof. *Journal of Economic Theory, 177*. https://doi.org/10.1016/j.jet.2018.07.001. Read in part (card: library/ashlagi2018stable.md; formal appendices not worked through). "no mechanism that implements a stable matching is 'obviously strategy-proof' for any side of the market." *Bearing:* the mechanism actually used in the residency match and many school systems cannot be made obvious in Li's sense. The removal of the task is therefore formal and not cognitive. The claim is about the existence of a bad preference profile, not about every market, and the paper's own suggestion for cognitive relief is sequential mechanisms, which it says are not obviously strategy-proof either. + +**reesjones2018experimental** — Rees-Jones, A., & Skowronek, S. (2018). An experimental investigation of preference misrepresentation in the residency match. *PNAS, 115*(45), 11471-11476. https://doi.org/10.1073/pnas.1803212115 +- *Verification:* Crossref (two authors). *Read status:* read in full (card: library/reesjones2018experimental.md) from the arXiv preprint 1802.01990 (81 pp.; open copy also at PMC6233132). The PNAS text was not read. +- *What it establishes:* "We administered an online experiment to 1,714 medical students immediately after their participation in the medical residency match ... When placed in an analogous, incentivized matching task, we find that 23% of participants misrepresent their preferences" (Abstract). "We document widespread failure to pursue the incentivized strategy of truth-telling. Over 23% of experimental participants misrepresent preference in our matching task, despite using this mechanism to make a career-altering decision mere days before" (Introduction). The sample is "the 1,714 students (approximately 51.9% of the students contacted) who both completed the survey and passed all exclusion criteria" (Methods). Misrepresentation is associated with a weaker strategic position (randomly assigned test scores), lower scores on Raven's Matrices, overconfidence, and trust in residency programmes; participants who reported advice from the NRMP website were more likely to submit an optimal list (AME = 0.10, SE = 0.021, p < .001), and advice from other students was "no better" in the full model only (average marginal effect −0.01, p = .565); on its own it was associated with truthful reporting (+0.04, p = .043) (Introduction; advice results). The overconfidence explanation is a hedged footnote ("could naturally arise"), not a stated attribution. +- *Move type:* evidence on (i). US medical students who had just used the real match. *Bearing:* a strategy-proof design removed the task for roughly three in four of a highly selected, highly trained group and left it with about one in four. Gap: this is a 52% response rate. +- **rees-jones2018suboptimal** (GEB 108, 317-330; single author): abstract only, "some students pursue futile attempts at strategic misrepresentation". Sample not read. + +**hassidim2017mechanism** — Hassidim, A., Marciano, D., Romm, A., & Shorrer, R. I. (2017). The mechanism is truthful, why aren't you? *American Economic Review, 107*(5), 220-224. https://doi.org/10.1257/aer.p20171027. *Verification:* Crossref (four authors; this corrects my first-pass author list, which omitted Marciano). *Read status:* abstract only; no open copy found. "We review evidence from different markets in different countries and find that a substantial percentage of participants do not report their true preferences to the strategy-proof Deferred Acceptance mechanism. Two recurring correlates of preference misrepresentation are lower cognitive ability and the expectation of stronger competition." *Bearing:* a review of other studies; it adds no data of its own. + +**hassidim2021limits** — Hassidim, A., Romm, A., & Shorrer, R. I. (2021). The limits of incentives in economic matching procedures. *Management Science, 67*(2), 951-963. https://doi.org/10.1287/mnsc.2020.3591. *Verification:* Crossref (three authors; this corrects my first-pass four-author list). *Read status:* abstract only; no open copy found. "We present direct field evidence of preference misrepresentation under the strategy-proof deferred acceptance in a high-stakes matching environment. We show that applicants to graduate programs in psychology in Israel often report that they prefer to avoid receiving funding, even though the mechanism preserves privacy and funding comes with no strings attached ... Preference misrepresentation in the field is associated with weaker applicants." + +**guillen2021strategy** — Guillén, P., & Veszteg, R. F. (2021). Strategy-proofness in experimental matching markets. *Experimental Economics, 24*(2), 650-668. https://doi.org/10.1007/s10683-020-09665-9. *Verification:* Crossref (two authors: Guillén and Veszteg; my first pass wrongly gave Hakimov). *Read status:* abstract only. Reverse mechanisms make the dominant strategy the opposite of the focal truthful report: "roughly half of the observed truth-telling under TTC and DA is the result of naive (non-strategic) behaviour. Only 14-31% of the participants choose actions in RTTC and RDA that are compatible with rational behaviour." *Bearing:* much of the apparent success of strategy-proofness comes from a default, not from understanding. + +**bo2019iterative** — Bó, I., & Hakimov, R. (2019). Iterative versus standard deferred acceptance: Experimental evidence. *The Economic Journal*. https://doi.org/10.1093/ej/uez036 (Crossref gives no volume or pages; SSRN version 10.2139/ssrn.2860738). *Read status:* abstract only. "A significantly higher proportion of stable outcomes is reached under IDAM than under DA," associated with "the feedback on the outcome of the previous application". *Bearing:* the description of the mechanism, and the feedback it gives, change how many people use the dominant strategy. A step-by-step presentation helps more than a one-shot rule that is formally equivalent. + +**gonczarowski2023strategyproofness** — Gonczarowski, Y. A., Heffetz, O., & Thomas, C. (2023). Strategyproofness-exposing descriptions of matching mechanisms. NBER Working Paper 31506. https://doi.org/10.3386/w31506. *Read status:* read in part (card: library/gonczarowski2023strategyproofness.md; NBER working paper, July 2023, revised October 2025; appendices skimmed). They propose a "menu description" of deferred acceptance. The October 2025 revision on file is theory only: it reports no experiment, and the phrase "the promise and challenges of menu descriptions" does not occur in it. The experiments are in the companion paper and, per the authors, in the earlier version, which "conduct[ed] an experiment for a second-price auction and median voting" (p. 23; arXiv 2209.13148v2). The abstract quoted in the first pass came from a different version than the PDF on file. Working paper. + +**cerrone2024school** — Cerrone, C., Hermstrüwer, Y., & Kesten, O. (2024). School choice with consent: An experiment. *The Economic Journal, 134*(661), 1760-1805. https://doi.org/10.1093/ej/uead120. Read in full (card: library/cerrone2024school.md): "both efficiency and truth-telling rates are higher under EADAM than under DA, even though EADAM is not strategy-proof." The full text adds the Enforced result, which runs the other way (lower truthfulness). *Bearing:* strategy-proofness is not necessary for truth-telling in the lab. + +**arteaga2022smart** — Arteaga, F., Kapor, A. J., Neilson, C. A., & Zimmerman, S. D. (2021, rev. 2022). Smart matching platforms and heterogeneous beliefs in centralized school choice. NBER Working Paper 28946. https://doi.org/10.3386/w28946 +- *Verification:* Crossref; NBER PDF in hand. *Read status:* read in part (card: library/arteaga2022smart.md): main text read; later tables, figures and online appendices not (the NBER version of June 2021, revised January 2022). Journal publication not located. +- *Design.* "Policymakers designated all applications with at least a 30% predicted chance of non-placement as risky. All applications above that cutoff received the live notification on the choice platform" (Introduction); the live-warning effect is therefore identified by a regression discontinuity at that cutoff, and the text-message and WhatsApp reminders by randomised trials ("choice administrators in Chile also randomized the provision" of messages; a WhatsApp RCT sent to "a random subset of ten thousand risky applicants"). The abstract says "We then use RCT and RD research designs to evaluate scaled live feedback policies in the Chilean and New Haven choice systems." +- *Results.* "Receiving a warning caused 21.6% of students (SE=1.0) to add schools to their applications, corresponding to the complier group in our model." "Students complying with the intervention reduced their non-placement risk by an average of 15.5 percentage points (SE=1.3), or 58% of mean ex post application risk." The abstract adds that complying applicants' enrolled schools had 0.10 standard deviations higher test-score value added. "Reducing the burden of school choice requires not just strategyproofness inside the centralized system, but also choice supports for the strategic decisions that inevitably remain outside of it" (Abstract). The 58% applies to compliers, not to all applicants who received a warning; for all warned applicants the risk fall is 3.3 percentage points, 13% of the below-threshold mean (pp. 26–27). +- *Move type:* the strategy-proof clearinghouse removes one task; the live warning is a (ii) support for the task that remains. Participants: school applicants (families) in Chile and New Haven. *Bearing:* the best field evidence that removing one task exposes another. + +**kapor2020heterogeneous** — Kapor, A., Neilson, C., & Zimmerman, S. (2018). Heterogeneous beliefs and school choice mechanisms. NBER Working Paper 25096. https://doi.org/10.3386/w25096. Read in full (card: library/kapor2020heterogeneous.md): given observed belief errors, a switch to deferred acceptance "offers welfare improvements", but "an analyst who assumed families had accurate beliefs would have reached the opposite conclusion." Structural model on New Haven survey data. + +**chen2006school** — Chen, Y., & Sönmez, T. (2006). School choice: An experimental study. *Journal of Economic Theory, 127*(1), 202-231. https://doi.org/10.1016/j.jet.2004.10.006. Metadata only; no abstract retrieved. I cite it only as the standard lab comparison of Boston and deferred-acceptance mechanisms; read before use. + +**Information the mediator shows each party (triadic field experiments).** + +**mejia2021when** — Mejia, J., & Parker, C. (2021). When transparency fails: Bias and financial incentives in ridesharing platforms. *Management Science, 67*(1), 166-184. https://doi.org/10.1287/mnsc.2019.3525 +- *Verification:* Crossref (authors Mejia and Parker; my first pass wrote "Parsons", which was wrong). *Read status:* read in full (card: library/mejia2021when.md) from the SSRN preprint (SSRN 3209274, "History: September 20, 2019", 31 pp.), obtained through the Indiana University repository's figshare record. The published version was not read. +- *Design.* "we perform a field experiment on a ridesharing platform from early October to mid November 2018 in Washington, DC. We randomly vary three dimensions of rider characteristics: 1) Gender ... 2) Ethnicity: Caucasian vs. Under-Represented Minority (URM) ... 3) Support for LGBT rights: rainbow profile picture filter" (Introduction). Four treatments in sixteen cells; "Of the 3,200 rides we requested, there were a total of 3,189 drivers that accepted the ride requests". The researchers requested the rides. +- *Results.* "Platforms responded to drivers' biased behavior by removing information about a rider's gender and race from the ride request sent to drivers ... While the change has, as expected, eliminated ride-request stage bias observed in previous research (Ge et al. 2016), bias persists at a later stage in the process. In the aggregate, URMs are more than twice as likely to have a ride canceled as Caucasians (3.1% vs. 7.5%)" (Introduction; the order of the two percentages in the parentheses is as printed and does not match the order of the groups in the sentence, so I do not assign them). The no-bias claim at the request stage rests on the contrast with the earlier study and on the platform's removal of the cue; this experiment did not test the earlier regime. +- *Move type:* (i) for the driver's selection among riders at request stage (the cue is withheld), with the cue re-entering at pickup. Triadic: rider, driver, platform. *Bearing:* a mediator controlling what each side sees removed a behaviour at one stage and displaced it to another. + +**fradkin2021reciprocity** — Fradkin, A., Grewal, E., & Holtz, D. (2021). Reciprocity and unveiling in two-sided reputation systems: Evidence from an experiment on Airbnb. *Marketing Science, 40*(6), 1013-1029. https://doi.org/10.1287/mksc.2021.1311. *Verification:* Crossref (three authors). *Read status:* abstract only; no open copy found (publisher closed; the SSRN landing page, 2939064, gave no text). "Feedback was hidden until both parties submitted a review in the treatment group and was revealed immediately after submission in the control group. The treatment stimulated more reviewing in total ... reduced retaliation and reciprocation in feedback and led to lower ratings ... The effects of the policy on feedback did not translate into reduced adverse selection on the platform." *Move type:* (i) on the review-writing task. Triadic: host, guest, platform. Sample size not in the abstract. Lab precursor: **bolton2013engineering** (*Management Science, 59*(2), 265-285; the card, library/bolton2013engineering.md, is a full reading of the 2009 UNSW discussion paper, not of the article, so its numbers should not be quoted as the article's). + +**nosko2015limits** — Nosko, C., & Tadelis, S. (2015/2026). The limits of reputation in platform markets: An empirical analysis and field experiment. NBER w20830, https://doi.org/10.3386/w20830; *Quantitative Marketing and Economics*, https://doi.org/10.1007/s11129-026-09313-1. Read in full (card: library/nosko2015limits.md; the 2026 QME article). eBay put a quality measure into its search ranking, "shifted buyers toward higher-quality sellers and increased subsequent purchasing". *Move type:* (i): the buyer's task of inferring quality from coarse feedback is replaced by a ranking. Triadic. This is the mediator doing the work, with no new reading for the buyer. The authors' own argument for not displaying the measure is in the full text (card). + +**li2026platform** — Li, R., Liu, X., & Sundararajan, A. (2026). Platform design, earnings transparency and minimum wage policies: Evidence from a natural experiment on Lyft. arXiv:2602.08955. Preprint, read in full (card: library/li2026platform.md). Staggered rollout (Feb 2024) of a pay guarantee plus per-ride earnings transparency; 47 million rides (the authors' count; the data table counts 1,887,050 driver-weeks); engagement rose, "greater transparency may induce strategic driver behavior". The guarantee and the transparency are bundled. The authors claim "separate effects" (abstract), but their split (HH against LH, and low against high variance) is indirect, and they call their reading of the HH/LH gap (transparency may matter more than the guarantee) only one possible interpretation (p. 22); the reading effect cannot be separated with confidence. + +### B5. Defaults, standard forms, pre-segmented disclosures + +**kelley2010standardizing** — Kelley, P. G., Cesca, L., Bresee, J., & Cranor, L. F. (2010). Standardizing privacy notices: An online study of the nutrition label approach. *CHI '10*, 1573-1582. https://doi.org/10.1145/1753326.1753561. *Verification:* Crossref; OpenAlex abstract. *Read status:* abstract only. *What it establishes:* online study, 764 participants; a standardised table, two short variants, full-text policies and layered policies compared; "standardized privacy policy presentations can have significant positive effects on accuracy and speed of information finding and on reader enjoyment." Companion design paper: **kelley2009nutrition** (SOUPS 2009, https://doi.org/10.1145/1572532.1572538; read in full from the card, library/kelley2009nutrition.md: a lab study with N = 24, not 764; the 764-participant result is the 2010 paper, which was not read). *Move type:* (i) for the finding task (the form pre-segments the policy into fixed cells), but a reading task remains. Dyadic: user and company. *Bearing:* the closest measured analogue of spacing in the facet: a fixed layout reduces search work and gains accuracy for lay readers. It segments a text; it does not coordinate two parties. + +**kling2012comparison** — Kling, J. R., Mullainathan, S., Shafir, E., Vermeulen, L. C., & Wrobel, M. V. (2012). Comparison friction: Experimental evidence from Medicare drug plans. *Quarterly Journal of Economics, 127*(1), 199-235. https://doi.org/10.1093/qje/qjr055 +- *Verification:* Crossref (five authors). *Read status:* read in full (card: library/kling2012comparison.md) from the NBER working paper 17410 (Sept 2011). The QJE text was not read (a PMC copy exists, PMC3314343; the Europe PMC render failed). +- *Design.* "One group of seniors—the intervention group—was presented with personalized price information created by entering their drug data into Medicare's Plan Finder website ... A comparison group was given only the address of this website" (Introduction). "Patients of the University of Wisconsin Hospital system age 65 and over made up the sample frame. Letters of invitation were mailed to 5,873 subjects ... leading to a baseline sample size of 451"; "406 people completed the baseline survey, were randomly assigned in the experiment, and completed the 2007 follow-up survey—forming our main analytic sample" (Data). The authors note that "about half of those sent a letter of invitation did not agree to participate (which is relevant for external validity)". +- *Results.* "The intervention group switched plans 28 percent of the time, whereas the comparison group only did so 17 percent of the time" (Introduction); the "difference of 11.5 percentage points" is in Table III; "average cost savings of the intervention—across the entire intervention group including non-switchers—was about $100 per year, or about 5 percent of the average predicted cost" (Introduction). +- *Move type:* (i) for compare-and-compute, but the letter was a bundle (a personalised computation, a recommended default, a statement of benefit and a deadline) and the authors say further work "could unbundle these effects" (§VIII); the person still has to switch. Participants: patients aged 65 and over at one US hospital system, n = 406. *Correction:* the first pass called this a field experiment on Medicare Part D enrollees without a sample size. + +**seira2017are** — Seira, E., Elizondo, A., & Laguna-Müggenburg, E. (2017). Are information disclosures effective? Evidence from the credit card market. *American Economic Journal: Economic Policy, 9*(1), 277-307. https://doi.org/10.1257/pol.20140404. Read in full (card: library/seira2017are.md). A randomised trial on indebted Mexican cardholders: "providing salient interest rate disclosures had no effects, while comparisons and de-biasing messages had only modest effects at best." *Bearing:* the rate disclosure (a reading task) did nothing. The months-to-pay disclosure did not change debt and raised delinquency among frequent interest payers and high-debt clients by 0.011 points in April, 7.4% of the mean (p. 295), which cuts against reading disclosure as merely inert. The authors say the rate message "gave no benchmark for comparing the interest rate to that of other banks" and cite Kling et al. as showing that comparison frictions matter (p. 303): the missing ingredient is the benchmark, not the reading. The contrast with Kling is therefore suggestive, not a clean sign that restructuring beats informing, and Kling's letter was a bundle. + +**poursabzi2021manipulating** — Poursabzi-Sangdeh, F., Goldstein, D. G., Hofman, J. M., Wortman Vaughan, J., & Wallach, H. (2021). Manipulating and measuring model interpretability. *CHI '21*, 1-52. https://doi.org/10.1145/3411764.3445315. *Verification:* Crossref; OpenAlex abstract; arXiv 1802.07810 v5. *Read status:* read in full (card: library/poursabzi2021manipulating.md). *What it establishes:* "pre-registered experiments (N = 3,800)" varying the number of features and whether model internals were visible. Participants "who saw a clear model with few features could better simulate the model's predictions. However, we did not find that participants more closely followed its predictions. Furthermore, showing participants a clear model meant that they were less able to detect and correct for the model's sizable mistakes, seemingly due to information overload." The mistake-detection result is conditional: it holds in Exps. 1, 2 and 4 (without the message) and not in Exp. 3, where people made an initial guess before seeing the model; the Exp. 2 replication holds for one of the two apartments only (apartment 12; apartment 11, p = 0.069); and the worst-simulation, least-following result for the clear eight-feature model is post hoc. A citation of the abstract's sentence without these conditions would overstate it. *Bearing:* evidence about when predictability does not come from transparency (Ruling 1). It also bears on the Gemini "information budget" lead (B9). + +**fogliato2022who** — Fogliato, R., Chappidi, S., Lungren, M., et al. (2022). Who goes first? Influences of human-AI workflow on decision making in clinical imaging. *FAccT '22*, 1362-1374. https://doi.org/10.1145/3531146.3533193. Read in full (card: library/fogliato2022who.md). 19 veterinary radiologists; "participants who are asked to register provisional responses in advance of reviewing AI inferences are less likely to agree with the AI regardless of whether the advice is accurate" and the order did not lengthen time. The anchoring effect was small in absolute terms (two to six points) and minimal for critical findings, and several key tests would miss the 0.05 level (the authors use 0.1). *Move type:* (ii)/(i): a sequencing rule. Dyadic. Relevant to "pacing". + +### B6. Counter-delegation + +**zhu2026choose** — Zhu, K., Thain, N., Tsai, V., Wexler, J., & Qian, C. (2026). Choose your agent: Tradeoffs in adopting AI advisors, coaches, and delegates in multi-party negotiation. arXiv:2602.12089v3 (27 Jun 2026). *Verification:* arXiv PDF in hand; Consensus record (listed as arXiv, a 2025-26 preprint; I found no venue). *Read status:* read in part (card: library/zhu2026choose.md; main text read in full, appendices skimmed). Peer review: none found. +- *Design.* Online experiment, N = 243 Prolific participants (324 recruited; §3, Methods), three-person bargaining games, "three multi-turn bargaining games in groups of three", each with one LLM assistance mode: Advisor (AI proposes, human decides), Coach (human proposes, AI critiques) or Delegate (AI acts). Within-subjects, counterbalanced; Table 4 reports "N = 243 active individuals nested in 81 groups; N = 216 baseline individuals nested in 72 groups"; the all-human baseline (N = 216) is a *historical control* from an earlier study with identical rules (§6.2). All three modes run on one model that outperformed humans. +- *Results.* "participants strongly prefer the higher-control Advisor (44%) over the Delegate (19%), yet groups only significantly increase collective surplus under Delegate access." The abstract's "significantly" overstates the tables: the group-level Delegate effect is p = .033 raw and p = .100 after Holm–Bonferroni correction, and the authors call it a "suggestive trend" (Table 1, p. 0:11). Preferences: Advisor 107, Coach 37, Delegate 47, none 52 (Table 2). Joint surplus per accepted trade: AI-assisted proposals beat manual ones in every mode, but only the Advisor gap is significant (Advisor +0.523, p < 0.01; Coach +0.154; Delegate +0.180). The mechanism is a "human filter": users "modify, override, or ignore" AI proposals and revert to human-baseline trades. "The Delegate advantage arises not from a different AI capability but from bypassing this filtering step altogether." The non-adopter spillover (+21.6%) "did not survive multiple-comparison correction (p adj = 0.16)" and the authors warn of selection bias; I do not use it. +- *Limits the authors state.* A "stylized bargaining game"; single session; the human filter "may serve a vital welfare-preserving function by maintaining social norms"; SUTVA is violated by design; the baseline is historical. +- *Move type:* the direct experimental contrast of (ii) (Advisor, Coach) against (iii) (Delegate). *Triadic?* Multi-party and strategically interdependent, **but the third party is two other human players, not an algorithmic mediator that reads both sides.** The AI is each player's own agent, so this is a test of counter-delegation inside a multi-party game, not of a mediated triad. +- *Bearing.* It is the only controlled experiment I found in which handing a task to an agent showed an advantage over handing the person an advisory reading task, in a setting where outcomes depend on others; the advantage is a suggestive trend (p = .033 raw, .100 corrected), not a corrected-significant result. It also shows the preference-performance gap that Bansal and Buçinca documented in dyads (Zhu cites both). It cuts both ways for the thesis: the delegation gain comes from removing the person's judgement, which the author's "engagement layer" ruling (the user reads and rewrites the mediator) would resist. + +**zhu2025automated** — Zhu, S., Sun, J., Nian, Y., South, T., Pentland, A., & Pei, J. (2025). The automated but risky game: Modeling agent-to-agent negotiations and transactions in consumer markets. *Proceedings of the Natural Legal Language Processing Workshop 2025*. https://doi.org/10.18653/v1/2025.nllp-1.2 (arXiv 2506.00073; PDF in hand, abstract-level). Abstract only. "deal-making with LLM agents in consumer settings is an inherently imbalanced game: different AI agents have large disparities in obtaining the best deals for their users", and anomalies "might lead to financial loss for both consumers and merchants". **The earlier dossier called this a preprint; Crossref now shows an ACL-anthology workshop paper (peer-reviewed at workshop level).** No human subjects; an agent benchmark. + +**terrucha2025humans** — Terrucha, I., et al. (2025). Humans program artificial delegates to accurately solve collective-risk dilemmas but lack precision. *PNAS*. https://doi.org/10.1073/pnas.2319942121. Read in full (card: library/terrucha2025humans.md): behavioural experiments; "people who delegate contribute more to the public good ... However, they are not more successful", because precision errors in the agent's algorithm "cannot be corrected amid the game". The full text adds a boundary the abstract omits: the contribution gain holds only in the three-level condition, and the success penalty only in the five-level condition (p. 4). *Bearing:* the specifying step (programming the agent) moves the cost to the person's intent-specification, the second algorithmacy competency. + +**vaccaro2026advancing** — Vaccaro, M., Caosun, M., Ju, H., Aral, S., & Curhan, J. R. (2026). Advancing AI negotiations: A large-scale autonomous negotiation competition. *PNAS, 123*(23), e2521774123. https://doi.org/10.1073/pnas.2521774123. Abstract only: participants wrote prompts for AI negotiation agents; 180,000+ AI-to-AI negotiations; warmth associated with better outcomes; "prompt injection" and "strategic concealment" appeared. Evidence that specifying intent to an agent is learnable and that agents attack each other. + +**alavi2026when** — Alavi, S., & Nozari, S. (2026). When agents shop for you: Role coherence in AI-mediated markets. arXiv:2604.26220. Preprint; read in full (card: library/alavi2026when.md). A buyer agent's verbal consumer profile lets a separate inference agent (not the seller) estimate willingness to pay "nearly one-for-one"; the leakage "arises from delegation itself" and "cannot be mitigated at the prompt level". The authors name three untested architectural defences: an anonymizing intermediary, profile rotation, and federated buyer-side aggregation (p. 18). + +**liu2026sovereignnegotiation** — Liu, D. Z. (2026). SovereignNegotiation-Bench. arXiv:2607.02814. Preprint, benchmark, no human subjects (read in full; card: library/liu2026sovereignnegotiation.md). Held in the earlier dossier. The v2 abstract gives faithful success of 6% to 75%, and 48% of agreements breaching a duty; the phrase "high privacy/consent risk" is not in it. + +**hua2026passive** — Hua, W., Huang, Z., Payne, T., Yousefi, S., Amershi, S., & Celikyilmaz, A. (2026). From passive delegates to strategic negotiators: Reinforcing social reasoning in small language models with SocialRL. arXiv:2608.13787. Preprint; the paper trains models as well as evaluating them (read in full; card: library/hua2026passive.md). Its premise matters; the abstract reads "a friendly and helpful frontier model may disclose its principal's private information unprompted, and concede the principal's position at the first sign of resistance". + +**calacci2022bargaining** — Calacci, D., & Pentland, A. (2022). Bargaining with the black-box: Designing and deploying worker-centered tools to audit algorithmic management. *Proceedings of the ACM on Human-Computer Interaction, 6*(CSCW2), 1-24. https://doi.org/10.1145/3570601. *Verification:* Crossref (two authors; title "Bargaining with the Black-Box"); OpenAlex abstract. *Read status:* abstract only. *What it establishes:* the Shipt Calculator, built with non-profit worker groups, lets delivery workers "track and share aggregate data about their pay"; data from 140 workers in summer 2020 showed that after the switch to a black-box pay model "although the average pay per-order increased ... almost half of workers experienced an unannounced pay cut during the shift, and many workers worked shifts that paid under their state's minimum wage." *Move type:* (iii)/(ii) by a collective: the work of auditing is done by a shared tool, but the individual worker reads the result. Triadic: worker, customer, platform. **Measured: the audit result. Not measured: any change in worker pay or bargaining outcome.** + +**irani2013turkopticon** — Irani, L. C., & Silberman, M. S. (2013). Turkopticon: Interrupting worker invisibility in Amazon Mechanical Turk. *CHI '13*. https://doi.org/10.1145/2470654.2470742. Abstract only: an activist system letting workers "publicize and evaluate their relationships with employers"; qualitative. I found no measured effect on pay or task quality. + +**delacroix2019bottom** — Delacroix, S., & Lawrence, N. D. (2019). Bottom-up data trusts: Disturbing the 'one size fits all' approach to data governance. *International Data Privacy Law, 9*(4). https://doi.org/10.1093/idpl/ipz014. Read in full (card: library/delacroix2019bottom.md). A legal proposal with no measured effects; the full text sets out an interest-conflict argument and four conditions. **viljoen2021relational** — Viljoen, S. (2021). A relational theory of data governance. *Yale Law Journal*. Not searched; not verified (lead). + +**dietvorst2018overcoming** — Dietvorst, B. J., Simmons, J. P., & Massey, C. (2018). Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them. *Management Science, 64*(3). https://doi.org/10.1287/mnsc.2016.2643. Read in full (card: library/dietvorst2018overcoming.md; the authors' manuscript, 38 pp.). Part 1 line 144 cites it for the "modification right". People who had seen the algorithm err used it more when they could change its output (Study 1, N = 288: 32% when they could not, 73% and 76% in the two adjustable conditions; Study 2, N = 816: 47% against 70%), and the willingness did not depend on the size of the adjustment allowed (71%, 71% and 68% for 10, 5 and 2 percentiles, p = .809). Participants often worsened the algorithm's forecasts when allowed to adjust. The round-two accuracy figures compare participants by the process they chose (17.90 for adjust-by-10, 20.25 for use-freely, 24.50 for human-only), not by random assignment; in round one, where assignment was random, adjust-by-10 against use-freely was not significant (t(410) = −1.22, p = .224), so the paper does not show that a cap caused any accuracy gain. Dyadic; the algorithm is a forecasting formula. + +*Overall for B6:* peer-reviewed experiments in which a person's own agent acts in a coordination with a platform's agent do not exist on the record I found. The nearest peer-reviewed items are Allouah et al. 2026 (shopping agents against platform listings, no human in the loop at choice time), Zhu et al. 2025 (workshop; agent against agent), Terrucha et al. 2025 and Vaccaro et al. 2026 (agents in games). Zhu K. et al. 2026 is the one human-subject experiment, and it is a preprint without an algorithmic mediator. + +### B7. Structural refusal, exit and contestability + +**zong2024data** — Zong, J., & Matias, J. N. (2024). Data refusal from below: A framework for understanding, evaluating, and envisioning refusal as design. *ACM Journal on Responsible Computing, 1*(1), 1-23. https://doi.org/10.1145/3630107. *Verification:* Crossref; OpenAlex abstract. *Read status:* abstract only. A framework with four facets (autonomy, time, power, cost), illustrated by cases. No measured effects. Part 1 line 146 relies on the "cost" facet; the abstract supports "cost, or whether or not refusal can reduce or redistribute penalties" but I did not read the passage on placation. + +**lyons2021conceptualising** — Lyons, H., Velloso, E., & Miller, T. (2021). Conceptualising contestability: Perspectives on contesting algorithmic decisions. *PACM HCI, 5*(CSCW1). https://doi.org/10.1145/3449180. Read in full (card: library/lyons2021conceptualising.md; arXiv 2103.01774). An analysis of submissions to Australia's AI Ethics Framework consultation. Conceptual; no measured effects. The robo-debt passage is on p. 16 (§6.2). + +**alfrink2023contestable** — Alfrink, K., Keller, I., Kortuem, G., & Doorn, N. (2022/2023). Contestable AI by design: Towards a framework. *Minds and Machines, 33*(4), 613-639. https://doi.org/10.1007/s11023-022-09611-z. [Crossref issued year 2022; the repo key and Part 1 say 2023, the print volume year.] Read in full (card: library/alfrink2023contestable.md): a design framework synthesised from the literature by qualitative methods. No measured effects. + +**vaccaro2020end** — Vaccaro, K., Sandvig, C., & Karahalios, K. (2020). "At the end of the day Facebook does what it wants": How users experience contesting algorithmic content moderation. *PACM HCI, 4*(CSCW2). https://doi.org/10.1145/3415238. *Verification:* Crossref (3 authors); OpenAlex abstract. *Read status:* abstract only. *What it establishes:* "A large-scale online experiment tests whether different forms of appeals can improve users' experiences of automated decision making ... none of the appeal designs improve FACT perceptions compared to a no appeal baseline." Number of participants not in the abstract. *Move type:* (iv)-adjacent: the right to appeal. The measured result is null on perceptions of fairness, accountability, trustworthiness and control. Triadic in structure (poster, moderation system, platform; the other user is the content's audience). + +**vaccaro2021contestability** — Vaccaro, K., Xiao, Z., Hamilton, K., & Karahalios, K. (2021). Contestability for content moderation. *PACM HCI, 5*(CSCW2). https://doi.org/10.1145/3476059. Abstract only. Participatory design workshops; three avenues (representation, communication, compassion). No measured effects. + +**starke2024contesting** — Starke, C., Metikoš, L., Helberger, N., & de Vreese, C. (2024). Contesting personalized recommender systems: A cross-country analysis of user preferences. *Information, Communication & Society, 28*(1), 41-60. https://doi.org/10.1080/1369118x.2024.2363926. Read in full (card: library/starke2024contesting.md): preregistered survey, 6,217 respondents in six countries; "Approximately 20% of users would opt out of using personalized RS." Intentions only. The 20% counts "I don't know" answers (24.7%) in the denominator; excluding them the share is 25.7%. **I found no measured uptake of the DSA's non-profiled feed option.** + +*Bearing of B7.* The theory and the framing are plentiful; the measured effects are thin and mostly null (Vaccaro 2020). The only large field experiments on leaving the ranker are Guess et al. 2023 (users were assigned to the chronological feed, not given a choice) and Gauthier et al. 2026, and they show exit changes behaviour and exposure while leaving some effects in place. + +### B8. The triadic setting itself + +**midkiff1993identification** (and **pritchett1994variations**, **hansman1995party**) — see the repo cards `midkiff1993.md`, `pritchett1994.md`, `hansman1995.md` and `triadic_hf_search_notes.md`. Midkiff & Hansman (1993), *Air Traffic Control Quarterly, 1*(1), 5-30, https://doi.org/10.2514/atcq.1.1.5 (Crossref-verified; article not opened; the thesis was read in full by the earlier run). *What the record holds:* a survey (1,500 distributed, American Airlines crews; Pritchett & Hansman widened it to 4,375 surveys, 738 returned, 710 usable) on which party-line information pilots rate important, and a simulation with **seven crews** that was a control phase of another study. **No study compares voice with datalink in a designed experiment; the loss of the party line is inferred from how it is used under voice, and the earlier note found no post-deployment measurement.** Move type: the party line is a structural exposure, and the design question is what the mediator must supply when the shared channel closes. This is a (i)/(iii)-style compensation that the mediator performs; the literature predicts it and does not test it. The overhearer is outside the exchange, not one of the two parties (limit already recorded in the notes). + +**hancock2020ai** — Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. *Journal of Computer-Mediated Communication, 25*(1), 89-100. https://doi.org/10.1093/jcmc/zmz022. Crossref (3 authors); OpenAlex abstract. Abstract only. Defines AI-MC as communication "in which an intelligent agent operates on behalf of a communicator by modifying, augmenting, or generating messages". An agenda; no data. It is the definitional source for AI sitting between two people, but the agent acts for the sender, so it is counter-delegation by construction. + +**hohenstein2023artificial** — Hohenstein, J., Kizilcec, R. F., DiFranzo, D., et al. (2023). Artificial intelligence in communication impacts language and social relationships. *Scientific Reports, 13*, 5487. https://doi.org/10.1038/s41598-023-30938-9. Crossref (9 authors). Read in full (card: library/hohenstein2023artificial.md). "Two randomized experiments" on smart replies: faster communication, more positive language, partners rated "closer and more cooperative" (this holds for the partner-use instrumental-variable result only), yet people "are evaluated more negatively if they are suspected to be using algorithmic responses." Two humans and a tool that offers tap-to-send suggestions the user chooses; it does not write or edit unprompted, and the partner cannot tell which messages are suggestions. Exposure alone had no effect ("we did not find any main effects of being exposed to smart replies", p. 4), and Google's replies did not raise sentiment against no replies in Study 2 (P = 0.58). *Bearing:* a mediator's suggestions, when one party uses them, changed the other party's perception of the first. The effect arises from the coordination, not from either user's reading skill. + +**jakesch2019ai** — Jakesch, M., French, M., Ma, X., Hancock, J. T., & Naaman, M. (2019). AI-mediated communication: How the perception that profile text was written by AI affects trustworthiness. *CHI '19*, 1-13. https://doi.org/10.1145/3290605.3300469. Read in full (card: library/jakesch2019ai.md): three experiments; the "Replicant Effect": hosts whose profiles were labelled or suspected AI-written were mistrusted "only when participants thought they saw a mixed set"; the full text adds that Study 1, with uniform AI labelling, found no effect. **jakesch2023co** — Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L., & Naaman, M. (2023). Co-writing with opinionated language models affects users' views. *CHI '23*. https://doi.org/10.1145/3544548.3581196. Read in full (card: library/jakesch2023co.md; arXiv preprint): N = 1,506; an opinionated writing assistant shifted what users wrote and thought. + +**moller2025impact** — Møller, A. G., Romero, D. M., Jurgens, D., & Aiello, L. M. (2026). The impact of generative AI on social media: An experimental study. *Scientific Reports, 16*, 9376. https://doi.org/10.1038/s41598-026-40110-8. Read in full (card: library/moller2025impact.md; accepted-article manuscript): 680 US participants in groups of five, four AI interventions; some "increase user engagement and volume of generated content, but at the same time decrease the perceived quality and authenticity of discussion, and introduce a negative spill-over effect on conversations." The main text ties the spill-over to the lower-rated replies and to an inference that quality falls "even among users not using the AI themselves" (p. 13); the card found no separate test of spill-over in the main text and the supplement is unread, so it is the authors' inference, not a separately estimated effect. A group setting, with AI aids on the participants' side. + +**What moved the work in triadic settings.** Across mejia2021when, fradkin2021reciprocity, nosko2015limits, hohenstein2023artificial (suggestions the user chooses; exposure alone had no effect) and zhu2026choose (a suggestive trend), the interventions that changed outcomes were changes to what the mediator exposes to each party (rider cues, review timing, ranking inputs), or who acts (delegate). I found no triadic field experiment in which giving a party a reading or explanation tool moved outcomes. That absence is itself a finding: it may reflect a search that missed one, and the search log says where I looked. + +### B9. The three Gemini essays: leads and what the literature says + +The essays are in `dial_response_algorithmacy/sources/` (`ux_evolution_scriptio_continua.md`, `ecological_ergonomics_algorithmacy_uxd.md`, `cognitive_epoch_conversational_skeuomorphism.md`). I cite none of them. I use their terms only as search leads. + +| Lead (as the essays name it) | Does a literature exist under another name? | Measured evidence found | Setting | +|---|---|---|---| +| "Algorithmic white space" (essay: four pillars) | Display clutter and layout (Moacdieh & Sarter); separated versus integrated icon arrays; Wolfe's guided search | Moacdieh & Sarter 2017: "high data density and poor organization increase RT and error rate" (simulated graphics task). Wang et al. 2025: 121 within-subjects, separated icon arrays raised risk comprehension for lower-numeracy people. **No study of white space in the display of an algorithmic mediator's output.** | Single user | +| "Pacing and information budgets" | Information overload; limited option sets; sequencing; cognitive forcing | Poursabzi-Sangdeh 2021 (N = 3,800): fewer features did not help people follow the model and a clear model reduced error detection "seemingly due to information overload" (in Exps. 1, 2 and 4, not 3). Fogliato 2022 (n = 19): provisional-first sequencing reduced agreement with the AI. Bucinca 2021 (facet B): forcing functions cut overreliance but were disliked. | Single user | +| "Pre-attentive state signalling" | Pre-attentive features (Treisman & Gelade), uncertainty visualisation, ambient and peripheral display | Spatharioti 2025 (facet B): colour highlighting of low-confidence values raised hard-task accuracy from 26% to 53-58%. Padilla 2018, 2020 (facet B): encodings create their own errors. Driving-mode awareness displays (Monsaingeon 2023; Kim 2025) improve knowledge of automation mode but "this knowledge did not necessarily lead to improved driving performance" (abstract). | Single user | +| "Spatial latent canvases", "latent-space navigation" | Node-based and canvas LLM interfaces; creative-tool latent navigation | Sensecape, VisCanvas, CanvasConvo, node-tree, Gero 2024: n = 20-24 or fewer, mostly self-reported or process measures; "indistinguishable" load against chat (VisCanvas); uneven adoption in a 5-day field study (CanvasConvo). **No decision or coordination outcome tested.** Latent-space navigation evidence found in the essays comes from music and image tools; I searched for none of it. | Single user | +| "Agentic interfaces", "steerability" | Delegation, levels of automation, agent negotiation | Zhu K. 2026 (N = 243): only the Delegate arm showed a positive group-surplus estimate (p = .033 raw, .100 corrected; the authors say "suggestive"); Advisor and Coach showed none. Allouah 2026 and Zhu S. 2025: agents biased and unstable. See B6. | Multi-party but no mediator | + +*Verdict on B9.* The essays' proposals have a real, measured literature in each case, all of it about one person using one system. **None of it tests the triad.** The essays' claim that these interfaces are the equivalent of word separation for algorithmacy has no measured support in a mediated coordination. The ENGAGEMENT_LAYER note already records the tiered-transparency recommendation as reversed by the author's lens; the measured literature adds that steering surfaces for novices are not obviously overload-free (Poursabzi; Kazemitabaar's information overload in Phasewise, in facet B). + +--- + +## (c) Synthesis + +### (i) Candidates by type, strength of evidence, and setting + +Strength key: **S** = randomised field or large preregistered experiment on the move, with a performance or behaviour outcome; **M** = controlled lab or online experiment, performance outcome; **P** = perceptions or satisfaction only; **Q** = qualitative or design study; **T** = theory or proposal only; **0** = nothing found. "Triadic" = two humans and a mediator that reads both and commits. + +| # | Candidate | Type | Best evidence | Strength | Triadic? | +|---|---|---|---|---|---| +| 1a | Declared envelope / behaviour description | (ii) | Cabrera 2023 (225 recruited, 209 analysed, 205 by the degrees of freedom); driving: Zhou 2019 (n = 24/condition [unverified: see AUDIT_03]), Singer 2022 (n = 90), Boelhouwer 2019 (null) | M, mixed | No, dyadic | +| 1b | Model cards | (ii) | Crisan 2022 (n = 30, think-aloud) | Q | No | +| 1c | Backward-compatible updates / stable rules | (i) on the mediator | Bansal 2019 (CAJA game with crowd workers; classifiers alone on three high-stakes tasks); Chen 2024 and Allouah 2026 document drift; Rahman 2021 documents harm of unpredictable criteria | M for effect on team accuracy; Q for platform harm | Rahman: triadic, qualitative | +| 2 | Seams | (ii) | Ehsan 2024 (43 participants, formative co-design; no end-user used a seamful interface) | Q | No | +| 3a | Controls on recommenders | (ii) | Harper 2015 (N = 168 tuner users, a self-selected 8% of invitees); Knijnenburg 2012 (N = 267); Vaccaro 2018; Hsu 2025 (N = 762); Alqabandi 2026; Moon 2025 (n = 240) | P (satisfaction, perceived control) | No | +| 3b | Structured/non-chat LLM interfaces | (ii) | DirectGPT (n = 12), Kazemitabaar (n = 18; B), Fan 2026 (N = 60; B, which notes that structured prompting may not be a direct-manipulation interface), VisCanvas, CanvasConvo | M/P, small | No | +| 3c | Leaving the ranker (chronological feed; assigned, not chosen, in Guess) | (iv) | Guess 2023 (field RCT, large; two errata, 2024 and 2026, the 2026 one reportedly on exposure counts [unverified: see AUDIT_03]; read both before citing exposure figures); Gauthier 2026 (field RCT) | S, behaviour outcomes | No | +| 4a | Strategy-proof matching | (i) | Li 2017 (lab, 144 student subjects in the whole experiment; 36% vs 7% of games off the dominant outcome in the serial-dictatorship pair, §5.3, 10-round treatments, 18 groups per arm; a one-shot variant gave 40% vs 8%, §5.4); Rees-Jones 2018 (n = 1,714; 23% misreport); Hassidim 2017, 2021 (abstracts); Guillén & Veszteg 2021; Bó & Hakimov 2019; Arteaga et al. (NBER WP 28946, 2021, rev. 2022; RD and RCT, field; choice support around a strategy-proof mechanism, not a strategy-proof rule) | S for field RCT; M for labs | Partly: applicants, clearinghouse, programmes; mediator has no objective of its own | +| 4b | Restructuring what the mediator shows each side (hide cues; simultaneous reveal; ranking input) | (i) | Mejia & Parker 2021 (field audit of a platform that had already hidden cues, 3,200 requested rides; bias persisted after acceptance; the cue change itself was untested, their §5.1); Fradkin 2021 (large field, abstract only; more reviewing and less retaliation, but no reduction in adverse selection); Nosko & Tadelis (field, abstract only; a ranking-input change, a different move from hide/reveal) | S for Fradkin only; not S for Mejia | **Yes** | +| 5a | Standard disclosure format | (i)/(ii) | Kelley 2010 (n = 764, online); Seira 2017 (RCT; null for the rate disclosure, delinquency raised by the months-to-pay disclosure) | M; S null | No | +| 5b | Pre-computed comparison handed to the person | (i) | Kling 2012 (randomised field experiment, n = 406 patients aged 65+; switching 28% vs 17%; the letter was a bundle of computation, default and deadline) | S | Intermediary-like, not algorithmic mediator | +| 6a | Delegate agent (own agent acts) | (iii) | Zhu K. 2026 (N = 243; Delegate effect p = .033 raw, .100 corrected); Terrucha 2025 | M, preprint for Zhu | Multi-party, no mediator | +| 6b | Agent vs. platform listings | (iii) | Allouah 2026; Zhu S. 2025; Alavi 2026 (preprint) | M on agents, no people | Partly | +| 6c | Collective audit tools / data trusts | (iii)/(ii) by a collective | Calacci & Pentland 2022 (n = 140); Irani & Silberman 2013 | Q/audit result; no outcome change | Yes | +| 7 | Contestability / appeals | (iv) | Vaccaro 2020 (online experiment: no gain); Alfrink 2023, Lyons 2021, Zong & Matias 2024 | M (perceptions), null; T | Yes in structure | +| 8 | AI-mediated communication | n/a | Hohenstein 2023 (two experiments; publisher correction, Oct 2023); Jakesch 2019, 2023 | M | **Yes** (two humans, writing aid) | +| 9 | Party-line compensation | (i)/(iii) | Midkiff & Hansman 1993; Pritchett & Hansman; 7-crew simulation | Survey; predicted, not tested | **Yes** | + +Two patterns stand out. First, the stronger effects came from changes to choice structure or a delivered comparison; two of the five (4b, 6a) rest on weak or uncorrected tests. The weakest or null results (3a, 5a Seira, 7 Vaccaro, 1a Boelhouwer) are the ones that hand the person information or a control. Second, the two triadic field studies (Fradkin; Mejia, as an audit) concern *what the mediator shows or reveals*, and none gives a party a tool for reading the mediator. + +### (ii) Which candidate best fits "takes the work off the person the way spacing took segmentation off the reader"? + +**Best fit: a mediator rule that removes a strategic task from participants and does so visibly. The model case is strategy-proof matching, extended by Li's obvious strategy-proofness; the field analogues are simultaneous reveal and withheld cues.** Spacing moved a task from the reader to the page. A dominant-strategy rule moves "guess what the mediator and the others will do" from the participant to the mechanism. It is type (i), it is formal, and it was built for non-experts (school applicants) [unverified: see AUDIT_03]. The evidence also shows what the analogy would have to carry. + +The evidence does not let me say "any single candidate deserves the name." Three results cut against it: + +1. **The task removal is incomplete.** 23% of the 1,714 students who completed the survey and passed the exclusion criteria misreport in an analogous task (Rees-Jones; the 1,714 are about 51.9% of the students contacted), roughly half of the observed truth-telling under TTC and DA in their school-allocation lab setting is naive (Guillén & Veszteg, abstract), and no stable mechanism is obviously strategy-proof (Ashlagi & Gonczarowski). The removal succeeds for those who follow the default and fails for those who do not. +2. **Presentation, not only structure, matters.** Li's own lab experiment shows it with formally equivalent pairs: in random serial dictatorship, "Under SP-RSD, 36% of games do not end in the dominant strategy outcome. Under OSP-RSD, 7%" (§5.3; 144 subjects in the whole experiment, 18 groups per RSD arm; a one-shot variant gave 40% against 8%, §5.4). Iterative deferred acceptance beat standard deferred acceptance (Bó & Hakimov, abstract), and menu descriptions are being designed to expose strategy-proofness (Gonczarowski et al.; the version on file is theory only, with no experiment). The lesson that carries over from spacing is that the same logical structure helps more when its form lets the person see that the work is gone. This also says that a type (i) move depends on a type (ii)-looking display. +3. **The removed task relocates.** In Chile and New Haven, deferred acceptance left the search and stopping decisions to families. In Chile (a regression discontinuity at a 30% predicted non-placement risk), 21.6% of warned applicants added schools, and for those compliers risk fell by 15.5 percentage points, 58% of the below-threshold mean, against 3.3 points (13%) for all warned applicants (Arteaga et al., Introduction; pp. 26–27). Kling's letter (a bundle of computed comparison, default and deadline) removed comparison and left action (switching 28% against 17%, n = 406 patients of one hospital system); Seira's TILA-type rate disclosure removed nothing and had no effect, and its months-to-pay disclosure raised delinquency among some clients. + +A second claimant is the delegate. In Zhu K. et al., only the Delegate arm showed a positive group-surplus estimate against a historical baseline (p = .033 unadjusted; .10 after Holm); Advisor and Coach did not differ from baseline. The authors call the Delegate result a "suggestive trend"; it does not survive correction. The authors' account is that the delegate bypassed the person's filter, a mechanism analysis rather than a tested contrast. It fits the form "the person no longer has to do it". It does not fit "the person can do it without the training", because the person must now trust an agent whose behaviour is unstable across model updates and biased in ranking (Allouah), can leak preference (Alavi), and carries precision errors that cannot be corrected mid-game (Terrucha). It trades one competence (interpret the mediator) for another (specify and supervise the agent). It is also not the spacing pattern: spacing reallocated work to a passive artefact that carries no agency, and a delegate is an active party. + +**Three-part answer to problem 4 of the plan (one move or a complex).** The measured evidence points to a complex. What reduced work in the triadic field experiments was changing what the mediator exposes to each party and when (4b), plus a pre-computed comparison where a choice remains (5b). What reduced work in the matching literature was a rule that makes the report dominant (4a), plus choice support for the tasks left outside it (Arteaga). Parkes's "complex of conventions" fits better than a single mark. + +### (iii) Gaps: experiments nobody has run (a research agenda the paper could propose) + +1. **A mediated-triad test of the type (i) versus type (ii) contrast.** Zhu et al. compare advisor, coach and delegate among three humans. I found no study that ran the same contrast (Consensus search only) with a mediator that reads both sides, e.g. a simulated ride-hail or marketplace matcher with two human parties, where one arm restructures what the mediator commits (a rule, a stable envelope), one arm gives a reading aid (explanation or description), one arm gives a delegate. Outcome: joint surplus and each party's regret, with novices and experienced users sampled separately. +2. **Enforced versus announced envelopes.** Cabrera tells people where the model fails; the driving studies announce limits. I found no study that compares an announced envelope with one the artefact enforces (refusal outside the domain, bounded output), with the same underlying system. +3. **Stability under update, measured on people.** Bansal et al. test compatibility in an abstract game with crowd workers, and on classifiers alone for three high-stakes tasks. I found no test of what a silent rule change does to worker or seller outcomes in a controlled setting, nor whether advance notice or a frozen interface changes recovery time. Rahman documents the harm qualitatively. +4. **Task relocation after a structural removal.** Arteaga et al. show it for school choice. The paper could propose measuring, for each type (i) intervention, which task the removal exposes (search, timing, intent specification) and what it costs novices. +5. **Counter-delegation against a platform's own agent with human subjects.** I found no peer-reviewed human-subject evidence (B6; Consensus search only). A first design: a buyer's personal agent against a platform ranker, with the user's information retained or leaked, tested on novices. +6. **Uptake and effect of platform control and exit options.** I found no measured uptake of the DSA's non-profiled feed option or of platform feed controls in the field (rows 5, 13). Alqabandi and Hsu measure expectations and satisfaction; I found no one who measures decisions made with and without a control, in a setting with a counterpart. +7. **Novice-versus-expert effects for every candidate.** The studies seldom split by expertise (Moon splits by literacy; Rees-Jones and Hassidim report cognitive-ability correlates; most others do not split). The author's claim is about novices; the evidence is mostly pooled. + +### (iv) Where the candidates conflict with the standing rulings + +- **Ruling 1 (not legibility).** Seams (B2), model cards, standard disclosure formats (Kelley), behaviour descriptions (Cabrera), driving limit announcements and "why am I seeing this" controls all hand the person something to read or tune; by the four-type sort they are type (ii). Cabrera and Zhou show they can work. Under the ruling they may enter only as evidence about when predictability holds; they cannot be offered as the move. +- **The "engagement layer" lens** (`ENGAGEMENT_LAYER.md`: the user reads and rewrites the shared mediator). It conflicts with the two best-supported candidates. Strategy-proof rules and withheld cues work because the participant does *not* have to read or rewrite; Zhu et al.'s delegate, on the authors' account, gains (a suggestive trend) because the person's judgement is bypassed. It also conflicts with stability: the envelope and compatible-update candidate asks the mediator to hold still while the lens says steering is ordinary use. Part 1 lists "bounded envelopes" and "backward-compatible updates" next to "counter-delegation" and "refusal"; the evidence says these items do not travel together. +- **Part 1 line 146** claims a costless exit "makes the parties substitutable, collapsing an irreducible triadic mediator back into a separable, manageable dyad." That is an information-theoretic argument, not an empirical result, and Gauthier shows persistent effects of prior exposure, while Guess shows the exit changed exposure and activity but not attitudes. I do not rule on the formal claim. +- **Part 1 line 144** cites Dietvorst et al. for the modification right as restoring willingness to decide through an algorithm. The control-surface studies show that perceived control and satisfaction rise (Harper, Knijnenburg, Moon), sometimes even for placebo controls (Vaccaro), but none measures decision quality. Willingness is not performance. +- **Ruling 3 (analogy, not mechanism).** The author left this ruling open for this paper on 2026-10-12 (`RESEARCH_PLAN.md` §3), so what follows is a reading, not a binding constraint. The finding that the task removal is incomplete (Rees-Jones; Guillén & Veszteg) and relocates (Arteaga) matches the plan's H-cost reading (a design precedent that lowered a cost, not a cause) more than H-cause. I do not use the history here; I note that nothing in the H evidence needs an "ignition" claim. +- **Ruling 4 (full-text depth).** At the first pass no entry reached full-text depth. Since 2026-10-13 most entries named in the final section have full-text cards; the entries that remain abstract-only or metadata-only are not yet quotable in the paper, and the open-copy sources remain working-paper or preprint versions, not versions of record. + +### (v) Relation to Nadella's seven principles, from the captured text + +I read the whole capture (P1-P21; the capturer's paragraph numbers) in this pass and redo the mapping from it. My first-pass table was built from the principle names alone, and one row misread the text (see Verifiability). Three features of the text bear on every row. First, the principal is the enterprise: the post worries about models that hold the organisation's most sensitive data and take mission-critical actions for it (P2), and it holds that responsibility for what the models do cannot be outsourced to the provider (P3). The post names no counterpart party as one of two parties in the coordination (P3 names a model provider and P20 names those affected, the nearest), so none of the seven principles is addressed to a person who is one of two parties to a mediated coordination. Second, the post does not rest on transparency alone: it says chain-of-thought transparency on its own is "not sufficient or dependable" (P9), and it warns that using models to check models can stack opaque systems into nested black boxes (P10). Third, its structural device is a rule placed outside the intermediary: the controls on a model's access and actions belong outside the model (P11), and the model is kept apart from the harness that runs its tasks and from its action space (P12). + +| Principle (paragraph) | Text | Candidates it touches | Relation | +|---|---|---|---| +| Model diversity (P14) | No important outcome should hang on a single model, and no model should check itself (P14). | Counter-delegation (6): a personal agent from another provider | **Compatible** as redundancy held by the deployer. In a triad it adds an agent the person cannot read, which is P10's nested-black-box worry (Allouah et al., abstract: agents differ by provider and model version). Not the same: the holder of the diversity is the enterprise. | +| Observe everything (P15) | Meaningful model actions must leave a tamper-proof trail of "human readable evidence"; outcomes must be reproducible without the model's own testimony. | Seams, inspectable controls (Knijnenburg), audit tools (Calacci and Pentland), model cards | **Same family** as type (ii), a reading task. The evidence I hold on lay readers (Poursabzi-Sangdeh: N = 3,800, clear models reduced error detection in Exps. 1, 2 and 4 but not 3; Seira: the rate disclosure no effect, the months-to-pay disclosure raised delinquency; Kling: delivering a bundled comparison, default and deadline worked where publishing the information had not) is evidence about a population P15 does not address; its reader, unnamed, sits in an enterprise setting (P8-P9). The requirement that outcomes be reproducible without the model's own account matches the logic of the Shipt audit, in which workers' own records replaced the platform's account of pay. | +| Verifiability (P16) | Test the whole system continuously, including failures, attacks, edge cases and system changes, not only successful tasks. | Backward-compatible updates (system changes); testing an envelope at its edges; provable properties such as strategy-proofness | **Compatible, and closer than my first pass said.** P16 asks the operator to test; it does not ask each user to verify outputs, so the first-pass "opposed in the sense of asking people to verify outputs" was a misreading. Testing system changes is the operator-side counterpart of Bansal et al.'s compatibility; neither requires the person to read anything. | +| Independent controls (P17) | Organisations should be able to decide for themselves what a model may access and do. | Bounded outputs, operational design domain, strategy-proof rules, withheld cues, simultaneous reveal | **Same structural device** (P11), different holder. The rules in my candidates sit outside the intermediary like P11's, but in a triad neither party holds them; the mediator's operator does. P17 gives the control to "Orgs", which is the single-principal frame the author calls "keeps the dyad" (SOURCE.md:22). | +| Independent auditability (P18) | Auditing must be independent of what is audited, and no single model may control both a system's conduct and the evidence for judging it against its original intent. | Collective audit (Calacci and Pentland), external experiments on platforms (Guess et al.; Gauthier et al.) | **Compatible in mechanism.** The Shipt and feed-experiment designs are independent of the platform. P18 presupposes one original intent; with two parties there are two intents and the text gives no rule for whose governs. | +| Containment (P19) | The post's image is an emergency brake held by "an authorized person", who can pause or stop a model mid-task. | Structural refusal and exit (iv); the chronological feed (Guess et al.; Gauthier et al.) | **Same form, different holder.** The brake belongs to the authorized person, not to the affected party; Part 1's refusal gives the exit to the party who is coordinated. Gauthier shows exposure effects that persist; Guess shows the exit changed exposure and activity but not attitudes. | +| Incident disclosure (P20) | Disclosure "to those affected" when systems fail, plus shared details of changes that alter agent behaviour at runtime. | Notice of rule change; backward-compatible updates (Rahman; Chen et al.; Allouah et al.) | **Compatible; the closest convergence.** The people affected would reach the workers and sellers whom unannounced criteria changes hit (Rahman: \"the criteria for success and changes to those criteria are unpredictable\"). Disclosure is a reading task for its recipients unless the update is also made compatible (Bansal et al.), so alone it is type (ii). | + +P21 (the best system is the one that lets the deployer trust the model least) fits the strategy-proof and withheld-cue candidates in one respect: they are valuable because the participant need not trust the mediator's reading of her report. It does not distinguish them from the post's own position. + +### (vi) Claims in the author's framing the sources contradict, or do not support + +**Line numbers.** The first pass gave Part 1 line numbers that were off by two to eight lines because I read the file without numbers. The numbers below are from `grep -n` on `submissions/decision_experience_design/published/2026-09-30_decision-experience-design.md` and on `submissions/dial_response_algorithmacy/draft.md`. + +1. **Explanation and agency (Part 1 line 136).** Repo sentence: "Providing explanations increases user awareness of an algorithm's presence while leaving user agency flat or diminished." Supported for explanations in the facet-B studies. The control-surface studies report a different sign on *perceived* control (Knijnenburg: "inspectability and control indeed increase users' perceived understanding of and control over the system"; Lukoff: agency supported by search and playlists). The claim needs "measured decisions" in its wording. + +2. **Moon et al. 2025: the two sentences side by side.** + - *Repo (Part 1 line 138):* "Moon et al. (2025) found that algorithmic literacy did not benefit users unless both explainability and direct control mechanisms were present simultaneously; absent control, explainability widened the perception gap between sophisticated and unsophisticated users." + - *Source (abstract of doi:10.1089/cyber.2024.0525; full text not accessible):* "When neither feature was present, algorithmic literacy had no significant impact. However, when at least one feature was present, literacy significantly influenced the dependent variables." The abstract also says "literacy creates a new dimension of the digital divide, where transparency benefits are unequally experienced." + - *Reading.* "Unless both" and "when at least one" are different conditions. The abstract's second sentence does not support "unless both"; it says one feature suffices for literacy to matter. The abstract's divide sentence is compatible with the repo's "widened the perception gap" but does not say that absent control explainability did so. Unresolved without the full text, which has no open copy. + +3. **Ehsan et al. 2024: the two sentences side by side.** + - *Repo (`dial_response_algorithmacy/draft.md` line 339):* "Ehsan, Liao, Passi, Riedl, and Daumé (2024) carried the move into algorithmic systems directly, running practitioners and end-users through a seamful process for surfacing an AI system's sociotechnical mismatches that measurably improved their sense of agency over it." + - *Source (arXiv:2211.06753 v2, abstract):* "We explore this process with 43 AI practitioners and real end-users, using a scenario-based co-design activity informed by real-world use cases. We found that the Seamful XAI design process helped users foresee AI harms, identify underlying reasons (seams), locate them in the AI's lifecycle, learn how to leverage seamful information to improve XAI and user agency." Body, §5.2.2: "Participants' comments also validated the end-goal of seamful XAI in supporting user agency, and delineated its paths to do so." + - *Reading.* The source reports what participants said in a co-design activity. I found no scale, questionnaire or before-and-after comparison in the text, so "measurably improved their sense of agency" is not borne out. **Correction to my first pass:** I attributed the agency wording to Part 1 as well. Part 1 (line 152) says only that seamful design "reveals the gaps, joints, and transitions in a computational system, allowing users to understand its limits (Chalmers & Galani, 2004; Ehsan et al., 2024)". The agency claim is in the DIAL draft alone. + +4. **Inman and Ribes 2019 as the source for the spacing point.** + - *Repo (`draft.md` line 337):* "Word separation is again the precedent. Inman and Ribes (2019) name what that precedent actually redistributes: not the load of *scriptio continua* but the agency of whoever has to carry it, since the remedy was never to reserve reading for those who could bear the load but to change the interface until the load fell." + - *Source (card `inman-ribes-2019-beautiful-seams`, the only text I have; the paper was not opened):* "seamless and seamful each distribute 'downstream user agency' differently — the real question is 'the strategic revelation or concealment of human and technological operations.'" The card says nothing on word separation or *scriptio continua*. + - *Reading.* The card supports the agency clause. It gives no support to the link with word separation. The paper has to be opened (ACM) to settle it. + +5. **"No peer-reviewed work studies a personal agent contesting a platform's agent" (dossier).** Still true for a human-subject study. The dossier called Zhu et al. (2025) a preprint; Crossref now shows it as an ACL-anthology workshop paper (doi:10.18653/v1/2025.nllp-1.2), and Allouah et al. (2026) is in the Web Conference 2026 proceedings. Neither is "contesting". Update the wording. + +6. **Strategy-proofness "takes the gaming task off participants."** The strong form is contradicted: "Over 23% of experimental participants misrepresent preference in our matching task, despite using this mechanism to make a career-altering decision mere days before" (Rees-Jones and Skowronek, Introduction). The weaker form holds: most follow the default, part of that is naive (Guillén & Veszteg, abstract), and misreporting is associated with weaker strategic position and lower Raven's scores, while overconfident participants, by contrast, were more likely to report optimally (Rees-Jones and Skowronek) or weaker applicants (Hassidim et al. 2021, abstract). + +7. **"Tiering transparency by expertise is the wrong axis" (`ENGAGEMENT_LAYER.md`, line 17).** The sources do not test tiering. They show extra information can hurt (Poursabzi-Sangdeh, abstract: "less able to detect and correct for the model's sizable mistakes, seemingly due to information overload"; this holds in Exps. 1, 2 and 4, not 3), and that those who used a steering control were few and heavy users (Harper: 168 tuner users, 8.3% of those emailed; the 148 survey completers had a median of 186 logins against 48 for all emailed users). That is a reason for caution, not a refutation of the lens. + +8. **The Gemini essays' claim that latent canvases, white space and pre-attentive signalling are the word-separation equivalent.** No measured support in a mediated coordination. See B9. + +9. **Delegation as safe counter-weight.** Allouah et al. (variation across providers and model versions), Zhu S. et al. (imbalance and anomalies) and Alavi and Nozari (preference leakage) show agent-side problems (abstracts). Zhu K. et al. show a gain in a stylised game with human opponents, against a historical control, with the authors' own caution that the human filter "may serve a vital welfare-preserving function" (§6.1; unadjusted p = .033, Holm-adjusted .10). + +10. **Numbers corrected by the full-text pass.** Li 2017 reports a lab experiment (first pass: none). Cabrera et al.: 225 recruited, 209 analysed (facet B and first pass: 225); the degrees of freedom imply 205. Kling et al.: n = 406 hospital patients (first pass: no n). Arteaga et al.: the live warning is evaluated by a regression discontinuity at 30% predicted risk, and the 58% applies to compliers (first pass: "RCT/RD" undifferentiated). Mejia and Parker: authors; 3,200 requested rides. Harper et al.: 381 assigned, 168 used the tuner, 148 surveyed (first pass: no n). Ehsan et al.: five authors, not seven. Nothing in the 23% (Rees-Jones), the 28% against 17% (Kling), the 58% (Arteaga) or N = 243 (Zhu) changed. + +--- + +## (d) Gaps and unverified leads + +**Update (2026-10-13).** Cards now cover most entries named below; the read-status lines in (b) and the final section supersede the lists in this part (for example, Poursabzi-Sangdeh, Allouah, Lukoff and Crisan have since been read in full). + +**Read in part from an open copy, second pass (full text opened, passages located; none read end to end; none is a version of record):** Li 2017 (MPRA working paper), Rees-Jones and Skowronek 2018 (arXiv preprint), Arteaga et al. (NBER), Kling et al. (NBER), Mejia and Parker (SSRN preprint), Guess et al. 2023 (UNC repository copy), Zhu et al. 2026 (arXiv v3), Cabrera et al. 2023 (arXiv), Harper et al. 2015 (GroupLens author copy), Ehsan et al. 2024 (arXiv v2). Next step for each: obtain the version of record and re-check the quoted figures. + +**Abstract only because no open copy exists (a human with library access):** Hassidim et al. 2017 (AER Papers and Proceedings) and 2021 (Management Science); Fradkin et al. 2021 (Marketing Science; SSRN 2939064 is blocked to this tool); Vaccaro et al. 2018 (ACM); Moon et al. 2025 (Mary Ann Liebert); Seira et al. 2017 (AEA; the PDF is open on aeaweb.org but returned a bot wall to this tool); Kelley et al. 2010 (ACM); Calacci and Pentland 2022 (ACM; OpenAlex lists an open link that I did not retry); Vaccaro et al. 2020 (ACM). **Not opened at all:** Inman and Ribes 2019 (ACM bot wall; card only). + +**Downloaded but read only to the abstract:** Poursabzi-Sangdeh et al. (arXiv 1802.07810), Allouah et al. (arXiv 2508.02630), Lukoff et al. (arXiv 2101.11778), Crisan et al. (arXiv 2205.02894), Zhu S. et al. (arXiv 2506.00073). These are open and should be read next. + +**Other abstract-only entries** (no read attempted in either pass): the remainder of section (b). + +**Metadata.** Every field in the `.bib` is from Crossref or the arXiv API (log rows 25-26). Fields in this md that came from memory in the first pass have been replaced or deleted. Fields Crossref does not carry (volume and pages for NBER and SSRN items, the Economic Journal volume of Bó and Hakimov) are left out. + +**Leads seen, not opened or not searched:** +- Bertrand & Morse (2011), "Information disclosure, cognitive biases, and payday borrowing", *Journal of Finance* — SSRN versions found (10.2139/ssrn.1533012); no abstract retrieved; a randomised field test of a Schumer-box-type disclosure on payday borrowers. Not entered because I have no numbers. +- Viljoen (2021), "A relational theory of data governance", *Yale Law Journal* — not searched. +- Silberman & Irani (2016), "Operating an employer reputation system: Lessons from Turkopticon", *Comparative Labor Law & Policy Journal* — not indexed in Crossref; a human with library access. +- Jahani et al. (2024), arXiv 2407.14333 — prompt adaptation to model upgrades; named in facet B; not located here. +- Kleinberg-style and Meta/Google "why am I seeing this" evaluations; Apple privacy-label evaluations; Holland et al. dataset nutrition labels; Stoyanovich & Howe "nutritional labels for data and models" — no search done. Evidence of measured effects of an *algorithm* nutrition label on ordinary people: none found; I did not search for it specifically. +- Shen et al. (2021) everyday algorithm auditing, Zhang, Hugh & Bernstein (2020) PolicyKit, Delgado et al. (2023): DIAL cards exist; I did not reread. +- Uber, Lyft and Instacart experiments published by the platforms themselves on driver-facing information (upfront fares, destination display) — not found in the open literature; I searched only through Consensus. The Mejia & Parker study was a field experiment run by the researchers; Li et al. 2026 is a natural experiment. +- Post-2010 CPDLC operational literature (Eurocontrol Link 2000+; FAA Data Comm) on whether party-line loss was observed after deployment — from `triadic_hf_search_notes.md`; not searched here. +- Operational design domain studies beyond driving (medical devices, FDA intended-use statements; "intended use" labelling) — not searched. + +**Sources that need a browser or library access:** ACM DL (Vaccaro 2018 and 2020, Inman and Ribes, Calacci and Pentland, Kelley, Zong and Matias, Fogliato); Elsevier (Rees-Jones GEB, Chen and Sonmez); INFORMS (Fradkin, Hassidim 2021, Dietvorst, published Mejia and Parker); AEA (Seira, Li 2017 AER version, Hassidim 2017); Liebert (Moon). The Scholar Gateway was not tried. + + +--- + +## Changes after full-text cards (2026-10-13) + +Source: `../audit/CARD_FACET_CHECKS.md`, section "Facet H", checked against each card in `../library/` (the card, including its S2 section, is the authority). Read status in section (b) and in `H_candidate_moves.bib` now follows the card. One line per change; the bib key leads. + +1. `alavi2026when`: read in full (card); "Counter-delegation creates a privacy channel" removed (not in the text); a separate inference agent, not the seller, estimates willingness to pay; three untested defences named (p. 18). +2. `alfrink2023contestable`: read in full (card); no measured effects confirmed. +3. `allouah2026what`: read in full (card); seller gain occurs in 33% of category-model pairs and sometimes reverses (p. 22); basic rationality near-perfect (p. 13); "properties the person can no longer see" marked as the project's inference. +4. `amin2026conversations`: read in full (card); working draft marked not for distribution. +5. `arteaga2022smart`: read in part (card); for all warned applicants the risk fall is 3.3 points, 13% of the below-threshold mean (pp. 26-27); the 58% stays attached to compliers (also in synthesis (ii)3). +6. `ashlagi2018stable`: read in part (card); claim is about the existence of a bad profile, and the authors' own cognitive relief (sequential mechanisms) is not obviously strategy-proof either. +7. `bansal2019updates`: read in full (card); the crowd participants played only the CAJA game (25 workers per condition in the mental-model studies); the three high-stakes tasks used classifiers alone; fixed in the entry, table (i) row 1c and gap 3. +8. `boelhouwer2019should`: read in full (card); mental models "did not seem to (sufficiently) change", inferred and not elicited; the null is a failure to detect at 14 per group. +9. `bolton2013engineering`: the card is of the 2009 discussion paper, not the 2013 article; its numbers are not to be quoted as the article's. +10. `cabrera2023improving`: read in full (card); 209 analysed by the text but 205 by the degrees of freedom, not reconciled; most of the gain came from heavier reliance where the AI was right; birds p = 0.048 confirmed; fixed in the entry, table (i) row 1a and (vi)10. +11. `cerrone2024school`: read in full (card); the Enforced result runs the other way (lower truthfulness). +12. `chen2024chatgpt`: read in full (card); the cleanest evidence is for instruction-following drift, offered as a partial explanation. +13. `chen2025generative`: read in full (card); the "72%" is a margin in percentage points (84% wins minus 12% losses, Table 1), not a ratio. +14. `crisan2022interactive`: read in full (card); the authors do not test the interactive card's effect on understanding the way they count comprehension for the standard card. +15. `delacroix2019bottom`: read in full (card); no measured effects confirmed; interest-conflict argument and four conditions noted. +16. `dietvorst2018overcoming`: read in full (card, S2-corrected); willingness rose with the right to modify and did not depend on the size of the adjustment (p = .809); the round-two accuracy comparison is by chosen process; capped against free adjustment is not significant under random assignment (p = .224). +17. `ehsan2024seamful`: read in part (card); no end-user used a seamful interface and no user behaviour was measured; findings "formative rather than evaluative"; table (i) row 2 reworded. +18. `fogliato2022who`: read in full (card); anchoring effect small (two to six points), several key tests miss the 0.05 level. +19. `gauthier2026political`: read in full (card); possible selection into the chronological feed (p. 420); following mechanism suggestive; effects concentrated among Republicans and Independents. +20. `gero2024supporting`: read in full (card); "tractable" rests on self-report and one of two tasks; the email task showed no difference. +21. `gonczarowski2023strategyproofness`: read in part (card); the October 2025 version on file is theory only; the quoted lab-experiment phrase is not in it; the experiments are in the companion paper and the earlier version (p. 23); synthesis (ii)2 annotated. +22. `goyal2025pilot`: read in full (card); the paper does not cite Fok and Weld, so the link is the facet's. +23. `guess2023how`: read in full (card); users were randomised into the chronological feed, not given a choice; "the person drops the ranker" removed; table (i) row 3c and the B7 bearing annotated. +24. `harper2015putting`: read in full (card, S2-corrected); share rating the popularity list "just right" rose from 31% to 49% (Fig. 5). +25. `hohenstein2023artificial`: read in full (card); suggestions the user chooses, not a mediator that writes or edits; exposure alone had no effect (p. 4); Google replies did not raise sentiment against none (P = 0.58); "closer and more cooperative" holds for the partner-use IV result only; B8 summary annotated. +26. `hua2026passive`: read in full (card); the premise quotation replaced by the abstract's exact wording; the paper trains models as well as evaluating them. +27. `jakesch2019ai`: read in full (card); Study 1, with uniform labelling, found no effect. +28. `jakesch2023co`: read in full (card); read status only. +29. `kapor2020heterogeneous`: read in full (card); read status only. +30. `kelley2009nutrition`: read in full (card); N = 24, not 764 (the 764 belongs to the 2010 paper, unread). +31. `kling2012comparison`: read in full (card); the letter was a bundle (computation, default, benefit statement, deadline); fixed in the entry, table (i) row 5b and synthesis (ii)3. +32. `li2017obviously`: read in part (card); the serial-dictatorship figures are in section 5.3 and the 7% compares only those two games; auction pairs use a different measure. +33. `li2026platform`: read in full (card); the authors claim separate effects but their split is indirect; their reading of the HH/LH gap is one possible interpretation (p. 22); 47 million rides is the authors' count (data table: 1,887,050 driver-weeks). +34. `liu2026sovereignnegotiation`: read in full (card); "high privacy/consent risk" is not in the v2 abstract, which gives faithful success of 6% to 75% and 48% of agreements breaching a duty. +35. `lukoff2021how`: read in full (card); the "(i)-style restructuring, not an added control" reading removed; the authors' own proposals are mostly user-operated controls, and users wanted less control for "ritualized" use (section 5.3). +36. `lyons2021conceptualising`: read in full (card); robo-debt passage is on p. 16 (section 6.2). +37. `matsuo2026educational`: read in full (card); N = 45; cue 6 of 15 collisions, none 7 of 15, cue plus reason 1 of 15; the cue alone gave no significant gain, the reason carried the result; B1 group bearing annotated. +38. `mejia2021when`: read in full (card); read status only. +39. `mitchell2019model`: read in full (card); read status only (proposal, not a measured effect, confirmed). +40. `moller2025impact`: read in full (card); the "negative spill-over" is the authors' inference from lower-rated replies (p. 13), with no separate test in the main text. +41. `nosko2015limits`: read in full (card); the authors' own argument for not displaying the measure noted. +42. `poursabzi2021manipulating`: read in full (card); the mistake-detection result holds in Exps. 1, 2 and 4, not Exp. 3; Exp. 2 replicates for one apartment only; the eight-feature result is post hoc; fixed in the entry, the P15 row, B9 table and (vi)7. +43. `reesjones2018experimental`: read in full (card); "no better" holds only in the full model (on its own, advice from students was associated with truthful reporting, p = .043); overconfidence explanation is a hedged footnote. +44. `seira2017are`: read in full (card); the months-to-pay message raised delinquency (p. 295); the missing ingredient is the benchmark; "cleanest sign that restructuring beats informing" removed; fixed in table (i) row 5a, synthesis (ii)3 and the P15 row. +45. `starke2024contesting`: read in full (card); the 20% counts "I don't know" (24.7%) in the denominator; excluding them it is 25.7%. +46. `suh2023sensecape`: read in full (card); N = 12; exploration counts, not decision quality, were the outcome. +47. `terrucha2025humans`: read in full (card, S2-corrected); the contribution gain holds only in the three-level condition and the success penalty only in the five-level condition (p. 4). +48. `ueno2026viscanvas`: read in part (card); "more diverse" rests on topology counts coded by the authors; one of eight feature ratings significant. +49. `wang2024lower`: read in full (card); the abstract's "slightly more reliable" has no significant support in the results (p. 10). +50. `wang2025evaluating`: read in full (card); the abstract's performance and trust claims rest on percentage descriptives and one non-significant test; "task performance" is a helpfulness rating. +51. `zhu2026choose`: read in part (card); Delegate effect p = .033 raw, .100 after Holm, "suggestive"; only the Advisor gap in joint surplus per accepted trade is significant; fixed in the entry, bearing, (ii), table (i) row 6a and the B9 table. +52. `singer2022driver`: read in full (card); read status only (no check note). +53. `bolton2013engineering`, `fradkin2021reciprocity`, `kelley2010standardizing`, `zhou2019effect` and other entries without cards: unchanged. + +Notes naming facet H in other sections of CARD_FACET_CHECKS: none found beyond those above. `vasconcelos2023explanations` is a facet X source and is not in this file. + +### Synthesis verdicts changed + +- (ii) Bearing for `zhu2026choose`: "the only controlled experiment I found in which handing a task to an agent beat handing the person an advisory reading task" became "showed an advantage ... a suggestive trend (p = .033 raw, .100 corrected), not a corrected-significant result"; the "second claimant" paragraph says the Delegate result does not survive correction. +- B9 table, "Agentic interfaces": "delegation beat advisor and coach on group surplus" became "only the Delegate arm showed a positive estimate (p = .033 raw, .100 corrected)". +- B5 `seira2017are` bearing: "the cleanest sign that restructuring beats informing" became "suggestive, not a clean sign", because the months-to-pay message raised delinquency, the rate message lacked a benchmark, and Kling's letter was a bundle. +- (i) table row 3c and the B7 bearing: "exit" in Guess et al. is assignment into the chronological feed, not a choice. +- B6 `alavi2026when`: "Counter-delegation creates a privacy channel" deleted. +- (iv) Ruling 4: "No entry reaches full-text depth" became "At the first pass no entry reached full-text depth; most entries named below now have full-text cards". +- B1 driving-automation group bearing: Matsuo added (the reason, not the cue, carried the result). diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_J_who_holds_the_convention.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_J_who_holds_the_convention.bib new file mode 100644 index 00000000..918082b7 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_J_who_holds_the_convention.bib @@ -0,0 +1,576 @@ +% Facet J, 2026-10-13. Agent-drafted; every entry's note gives its read status. Reused keys at the end are copied from literature/references.bib. + +@article{galanter1974why, + author = {Galanter, Marc}, + title = {{Why the “Haves” Come Out Ahead: Speculations on the Limits of Legal Change}}, + journal = {Law \& Society Review}, + volume = {9}, + number = {1}, + pages = {95-160}, + year = {1974}, + doi = {10.2307/3053023}, + note = {Crossref verified. Read in part (card: library/galanter1974why.md; main text pp. 95-112, 113-125, 129-137, 141-151) from the open-access Wisconsin Law repository PDF / Cambridge Core PDF, 2026-10-13.} +} + +@article{tyler1988what, + author = {Tyler, Tom R.}, + title = {{What is Procedural Justice?: Criteria Used by Citizens to Assess the Fairness of Legal Procedures}}, + journal = {Law \& Society Review}, + volume = {22}, + number = {1}, + pages = {103-135}, + year = {1988}, + doi = {10.2307/3053563}, + note = {Crossref verified; read in full (card: library/tyler1988what.md; Cambridge Core PDF, journal pagination).} +} + +@book{lind1988social, + author = {Lind, E. Allan and Tyler, Tom R.}, + title = {{The Social Psychology of Procedural Justice}}, + publisher = {Plenum}, + year = {1988}, + doi = {10.1007/978-1-4899-2115-4}, + note = {Crossref verified; metadata only.} +} + +@book{thibaut1975procedural, + author = {Thibaut, John and Walker, Laurens}, + title = {{Procedural Justice: A Psychological Analysis}}, + publisher = {Erlbaum}, + year = {1975}, + url = {https://doi.org/10.2307/448155}, + note = {Not accessible. Existence shown only by three book reviews indexed in Crossref (10.2307/1371953, 10.2307/1142510, 10.2307/448155; the url is the Western Political Quarterly review, not the book). Reviews not opened. Publisher and year not checked against a catalogue record.} +} + +@book{tyler1990why, + author = {Tyler, Tom R.}, + title = {{Why People Obey the Law}}, + publisher = {Princeton University Press}, + year = {2006}, + doi = {10.1515/9781400828609}, + note = {Not accessible. DOI is the 2006 Princeton reprint (Crossref); the 1990 Yale University Press edition is reviewed at 10.1017/s000819730009958x (not opened).} +} + +@article{colquitt2001justice, + author = {Colquitt, Jason A. and Conlon, Donald E. and Wesson, Michael J. and Porter, Christopher O. L. H. and Ng, K. Yee}, + title = {{Justice at the millennium: A meta-analytic review of 25 years of organizational justice research.}}, + journal = {Journal of Applied Psychology}, + volume = {86}, + number = {3}, + pages = {425-445}, + year = {2001}, + doi = {10.1037/0021-9010.86.3.425}, + note = {Crossref verified (five authors); abstract only (OpenAlex).} +} + +@article{lind1993individual, + author = {Lind, E. Allan and Kulik, Carol T. and Ambrose, Maureen and de Vera Park, Maria V.}, + title = {{Individual and Corporate Dispute Resolution: Using Procedural Fairness as a Decision Heuristic}}, + journal = {Administrative Science Quarterly}, + volume = {38}, + number = {2}, + pages = {224}, + year = {1993}, + doi = {10.2307/2393412}, + note = {Crossref verified; metadata only.} +} + +@article{fuller1978forms, + author = {Fuller, Lon L. and Winston, Kenneth I.}, + title = {{The Forms and Limits of Adjudication}}, + journal = {Harvard Law Review}, + volume = {92}, + number = {2}, + pages = {353}, + year = {1978}, + doi = {10.2307/1340368}, + note = {Crossref verified (authors Fuller and Winston); metadata only.} +} + +@article{colvin2011empirical, + author = {Colvin, Alexander J. S.}, + title = {{An Empirical Study of Employment Arbitration: Case Outcomes and Processes}}, + journal = {Journal of Empirical Legal Studies}, + volume = {8}, + number = {1}, + pages = {1-23}, + year = {2011}, + doi = {10.1111/j.1740-1461.2010.01200.x}, + note = {Crossref verified; read in full (card: library/colvin2011empirical.md; Cornell author manuscript, manuscript pagination, reference list not read in full).} +} + +@article{greif1989reputation, + author = {Greif, Avner}, + title = {{Reputation and Coalitions in Medieval Trade: Evidence on the Maghribi Traders}}, + journal = {The Journal of Economic History}, + volume = {49}, + number = {4}, + pages = {857-882}, + year = {1989}, + doi = {10.1017/S0022050700009475}, + note = {Crossref verified; abstract only (OpenAlex).} +} + +@article{milgrom1990role, + author = {Milgrom, Paul R. and North, Douglass C. and Weingast, Barry R.}, + title = {{THE ROLE OF INSTITUTIONS IN THE REVIVAL OF TRADE: THE LAW MERCHANT, PRIVATE JUDGES, AND THE CHAMPAGNE FAIRS}}, + journal = {Economics \& Politics}, + volume = {2}, + number = {1}, + pages = {1-23}, + year = {1990}, + doi = {10.1111/j.1468-0343.1990.tb00020.x}, + note = {Crossref verified (Crossref family name for Weingast carries a trailing asterisk, removed here); abstract only (OpenAlex).} +} + +@article{bernstein1992opting, + author = {Bernstein, Lisa}, + title = {{Opting out of the Legal System: Extralegal Contractual Relations in the Diamond Industry}}, + journal = {The Journal of Legal Studies}, + volume = {21}, + number = {1}, + pages = {115-157}, + year = {1992}, + doi = {10.1086/467902}, + note = {Crossref verified; metadata only (no abstract).} +} + +@article{bernstein2001private, + author = {Bernstein, Lisa}, + title = {{Private Commercial Law in the Cotton Industry: Creating Cooperation through Rules, Norms, and Institutions}}, + journal = {Michigan Law Review}, + volume = {99}, + number = {7}, + pages = {1724}, + year = {2001}, + doi = {10.2307/1290478}, + note = {Crossref verified (start page 1724; end page not shown); abstract only (OpenAlex). Publisher PDF bot-walled; SSRN version 10.2139/ssrn.281437 not opened.} +} + +@article{roth2002economist, + author = {Roth, Alvin E.}, + title = {{The Economist as Engineer: Game Theory, Experimentation, and Computation as Tools for Design Economics}}, + journal = {Econometrica}, + volume = {70}, + number = {4}, + pages = {1341-1378}, + year = {2002}, + doi = {10.1111/1468-0262.00335}, + note = {Crossref verified; abstract only (OpenAlex). Harvard DASH copy returned an error.} +} + +@article{oomen2016last, + author = {Oomen, Roel}, + title = {{Last look}}, + journal = {Quantitative Finance}, + volume = {17}, + number = {7}, + pages = {1057-1070}, + year = {2017}, + doi = {10.1080/14697688.2016.1262545}, + note = {Online 22 Dec 2016. Crossref verified (online 2016, vol. 17 issue 7, 2017). Read in full (card: library/oomen2016last.md) from the LSE Research Online accepted manuscript (eprints.lse.ac.uk/68811), manuscript pagination, appendix proofs skimmed; may differ from the published version. Author is employed by Deutsche Bank.} +} + +@article{evans2018forex, + author = {Evans, Martin D.D.}, + title = {{Forex trading and the WMR Fix}}, + journal = {Journal of Banking \& Finance}, + volume = {87}, + pages = {233-247}, + year = {2018}, + doi = {10.1016/j.jbankfin.2017.09.017}, + note = {Crossref verified; read in part (card: library/evans2018forex.md; MPRA Paper No. 81583 accepted manuscript, manuscript pagination; model derivation and appendix tables skimmed).} +} + +@article{suzor2018digital, + author = {Suzor, Nicolas}, + title = {{Digital Constitutionalism: Using the Rule of Law to Evaluate the Legitimacy of Governance by Platforms}}, + journal = {Social Media + Society}, + volume = {4}, + number = {3}, + year = {2018}, + doi = {10.1177/2056305118787812}, + note = {Crossref verified; read in full (card: library/suzor2018digital.md; open-access PDF, journal pagination 1-11). 14 platforms, data from June 2015.} +} + +@article{kellogg2020algorithms, + author = {Kellogg, Katherine C. and Valentine, Melissa A. and Christin, Angéle}, + title = {{Algorithms at Work: The New Contested Terrain of Control}}, + journal = {Academy of Management Annals}, + volume = {14}, + number = {1}, + pages = {366-410}, + year = {2020}, + doi = {10.5465/annals.2018.0174}, + note = {Crossref verified; abstract only (OpenAlex).} +} + +@article{cameron2022expanding, + author = {Cameron, Lindsey D. and Rahman, Hatim}, + title = {{Expanding the Locus of Resistance: Understanding the Co-constitution of Control and Resistance in the Gig Economy}}, + journal = {Organization Science}, + volume = {33}, + number = {1}, + pages = {38-58}, + year = {2022}, + doi = {10.1287/orsc.2021.1557}, + note = {Crossref verified; abstract only (OpenAlex).} +} + +@article{jhaver2019transparency, + author = {Jhaver, Shagun and Bruckman, Amy and Gilbert, Eric}, + title = {{Does Transparency in Moderation Really Matter?}}, + journal = {Proceedings of the ACM on Human-Computer Interaction}, + volume = {3}, + number = {CSCW}, + pages = {1-27}, + year = {2019}, + doi = {10.1145/3359252}, + note = {Crossref verified; abstract only (Consensus).} +} + +@article{jhaver2019suspect, + author = {Jhaver, Shagun and Appling, Darren Scott and Gilbert, Eric and Bruckman, Amy}, + title = {{"Did You Suspect the Post Would be Removed?"}}, + journal = {Proceedings of the ACM on Human-Computer Interaction}, + volume = {3}, + number = {CSCW}, + pages = {1-33}, + year = {2019}, + doi = {10.1145/3359294}, + note = {Crossref verified; read in full (card: library/jhaver2019suspect.md; journal pagination 192:1-192:33).} +} + +@article{tyler2019social, + author = {Tyler, Tom and Katsaros, Matt and Meares, Tracey and Venkatesh, Sudhir}, + title = {{Social media governance: can social media companies motivate voluntary rule following behavior among their users?}}, + journal = {Journal of Experimental Criminology}, + volume = {17}, + number = {1}, + pages = {109-127}, + year = {2019}, + doi = {10.1007/s11292-019-09392-z}, + note = {Crossref verified (online 27 Dec 2019; volume 17 issue 1, 2021 per Crossref pages 109-127). Print year 2021. Abstract only (Consensus); abstract not re-checked by auditor.} +} + +@article{myerswest2018censored, + author = {Myers West, Sarah}, + title = {{Censored, suspended, shadowbanned: User interpretations of content moderation on social media platforms}}, + journal = {New Media \& Society}, + volume = {20}, + number = {11}, + pages = {4366-4383}, + year = {2018}, + doi = {10.1177/1461444818773059}, + note = {Crossref verified; abstract only (Consensus).} +} + +@article{wong2022meta, + author = {Wong, David and Floridi, Luciano}, + title = {{Meta’s Oversight Board: A Review and Critical Assessment}}, + journal = {Minds and Machines}, + volume = {33}, + number = {2}, + pages = {261-284}, + year = {2022}, + doi = {10.1007/s11023-022-09613-x}, + note = {Crossref verified (online 24 Oct 2022; vol. 33 issue 2); read in full (card: library/wong2022meta.md). Print year 2023.} +} + +@article{vanderdonk2022ex, + author = {Van Der Donk, Berdien}, + title = {{Ex-ante and Ex-post Access Restrictions in Marketplaces’ Terms of Service}}, + journal = {European Review of Private Law}, + volume = {30}, + pages = {909-928}, + year = {2022}, + doi = {10.54648/erpl2022044}, + note = {Crossref verified (family name given by Crossref as Van Der Donk); abstract only (Consensus).} +} + +@article{palka2025foreign, + author = {Pałka, P. and Porębski, A. and Pałosz, R. and Wiśniewska, K.}, + title = {{Do Foreign Service Providers Simply Ignore the EU Consumer Law? An Empirical Study of 100 Online Terms of Service}}, + journal = {Journal of Consumer Policy}, + volume = {48}, + number = {3}, + pages = {233-268}, + year = {2025}, + doi = {10.1007/s10603-024-09580-z}, + note = {Crossref verified (given names: initials only in Crossref); read in full (card: library/palka2025foreign.md; journal pagination 233-268).} +} + +@article{maffie2025enshittification, + author = {Maffie, Michael David and Hurtado, Hector}, + title = {{The Enshittification of Work: Platform Decay and Labour Conditions in the Gig Economy}}, + journal = {British Journal of Industrial Relations}, + volume = {64}, + number = {1}, + pages = {5-20}, + year = {2025}, + doi = {10.1111/bjir.70004}, + note = {Crossref verified (online 5 Aug 2025; vol. 64 issue 1); abstract only (Consensus).} +} + +@article{nagarajrao2026fareshare, + author = {Nagaraj Rao, Varun and Dalal, Samantha and Schwartz, Andrew and Liaqat, Amna and Calacci, Dana and Monroy-Hernández, Andrés}, + title = {{FareShare: A Tool for Labor Organizers to Estimate Lost Wages and Contest Arbitrary AI and Algorithmic Deactivations}}, + journal = {Proceedings of the ACM on Human-Computer Interaction}, + volume = {10}, + number = {2}, + pages = {1-32}, + year = {2026}, + doi = {10.1145/3788052}, + note = {Crossref verified (six authors); read in full (card: library/nagarajrao2026fareshare.md; arXiv 2505.08904 v1 preprint, preprint pagination).} +} + +@book{gillespie2019custodians, + author = {Gillespie, Tarleton}, + title = {{Custodians of the Internet: Platforms, Content Moderation, and the Hidden Decisions That Shape Social Media}}, + publisher = {Yale University Press}, + year = {2018}, + doi = {10.12987/9780300235029}, + note = {Crossref issued 2019 (online date); the book was published by Yale University Press in 2018; metadata only. Subtitle and print year not checked against the Crossref record.} +} + +@misc{douek2023meta, + author = {Douek, Evelyn}, + title = {{The Meta Oversight Board and the Empty Promise of Legitimacy}}, + year = {2023}, + doi = {10.2139/ssrn.4565180}, + note = {Crossref verified (SSRN record). Read in full (card: library/douek2023meta.md) in the 2024 Harvard JOLT print, vol. 37 no. 2, pp. 373-445; the SSRN posting is 2023, the journal print 2024.} +} + +@inproceedings{suzor2016responsibilities, + author = {Suzor, Nicolas}, + title = {{The Responsibilities of Platforms: A New Constitutionalism to Promote the Legitimacy of Decentralized Governance}}, + booktitle = {Selected Papers of AoIR 2016: The 17th Annual Conference of the Association of Internet Researchers}, + address = {Berlin}, + year = {2016}, + url = {https://spir.aoir.org/ojs/index.php/spir/article/view/9055}, + note = {No DOI found. Downloaded 4-page PDF 2026-10-13. Read in full (card: library/suzor2016responsibilities.md); cited by paragraph order.} +} + +@article{klonick2018new, + author = {Klonick, Kate}, + title = {{The New Governors: The People, Rules, and Processes Governing Online Speech}}, + journal = {Harvard Law Review}, + volume = {131}, + pages = {1598--1670}, + year = {2018}, + url = {https://harvardlawreview.org/wp-content/uploads/2018/04/1598-1670_Online.pdf}, + note = {No DOI or Crossref record found. Verified by the PDF running head (Vol. 131:1598) and table of contents. Read in full (card: library/klonick2018new.md): body of all four Parts and the Conclusion; footnotes read selectively. Pin cites: appeals p. 1648; forces p. 1664; lack of an appeals system p. 1665; exit pp. 1629-30, 1668.} +} + +@techreport{oversight2022q1, + author = {{Oversight Board}}, + title = {{Q1 2022 Transparency Report}}, + year = {2022}, + url = {https://oversightboard.com/attachment/731010438013851}, + note = {Downloaded 64-page PDF 2026-10-13. Read in part (card: library/oversight2022q1.md): pp. 1-25 and prose; Annex II scanned only. The submitted-cases count (479,649) is the Board own preliminary manual estimate.} +} + +@misc{apple2026guidelines, + author = {{Apple Inc.}}, + title = {{App Review Guidelines}}, + year = {2026}, + url = {https://developer.apple.com/app-store/review/guidelines/}, + note = {Fetched 2026-10-13; page states Last Updated June 8, 2026. A living page; quote only with the retrieval date. Read in part (card: library/apple2026guidelines.md).} +} + +@misc{eu2019p2b, + author = {{European Parliament and Council of the European Union}}, + title = {{Regulation (EU) 2019/1150 on promoting fairness and transparency for business users of online intermediation services}}, + year = {2019}, + url = {https://eur-lex.europa.eu/eli/reg/2019/1150/oj}, + note = {OJ L 186, 11.7.2019, p. 57. Quotations checked against OJ text 2026-10-13; applies from 12 July 2020. Arts. 3, 4, 5, 11, 12 first read by the facet from the UK-retained version; the card (library/eu2019p2b.md) read Arts. 1-19 (except 14(5)-(6)) in the Official Journal text, which contains Art. 5(6)-(7).} +} + +@misc{eu2022dsa, + author = {{European Parliament and Council of the European Union}}, + title = {{Regulation (EU) 2022/2065 on a Single Market For Digital Services (Digital Services Act)}}, + year = {2022}, + url = {https://eur-lex.europa.eu/eli/reg/2022/2065/oj}, + note = {OJ L 277, 27.10.2022, p. 1. Applies from 17 February 2024. Facet first read Arts. 14, 17, 20, 21 from an unofficial mirror; the card (library/eu2022dsa.md) read the operative text, Arts. 1-93, in the Official Journal text (read in part: recitals not read).} +} + +@misc{seattle2024deactivation, + author = {{Seattle Office of Labor Standards}}, + title = {{Model Notice of Deactivation (based on SMC 8.40.070), revised 11/2024}}, + year = {2024}, + url = {https://www.seattle.gov/laborstandards/ordinances/app-based-worker-deactivation-rights-ordinance}, + note = {Agency model form, read in full (card: library/seattle2024deactivation.md; https://www.seattle.gov/documents/Departments/LaborStandards/Notice_of_Deactivation_Regular_form.pdf). The ordinance text (SMC 8.40) was not read. Per the OLS page: passed August 2023, effective 1 January 2025.} +} + +@misc{globalfxc2024code, + author = {{Global Foreign Exchange Committee}}, + title = {{FX Global Code: A set of global principles of good practice in the foreign exchange market}}, + year = {2024}, + url = {https://www.globalfxc.org/docs/fx_global.pdf}, + note = {Fetched 2026-10-13, 84 pages. PDF footer reads Updated December 2024; key renamed from globalfxc2024code. Read in full (card: library/globalfxc2024code.md; printed pages 1-82). Principles 7, 8, 9, 17, 36, 39 and 40 bear on notice, reasons and concerns.} +} + +@techreport{fsb2014foreign, + author = {{Financial Stability Board}}, + title = {{Foreign Exchange Benchmarks: Final Report}}, + year = {2014}, + month = sep, + url = {https://www.fsb.org/wp-content/uploads/r_140930.pdf}, + note = {133 pages. Read in part (card: library/fsb2014foreign.md): pp. 1-34 and the Appendix 3 summary in full; about 85 pages skimmed. Working group members were central banks and regulators (Appendix 2).} +} + +@techreport{fsb2015progress, + author = {{Financial Stability Board}}, + title = {{Foreign Exchange Benchmarks: Report on progress in implementing the September 2014 recommendations}}, + year = {2015}, + month = oct, + url = {https://fsb.org/uploads/FX-Benchmarks-progress-report.pdf}, + note = {Read in full (card: library/fsb2015progress.md; body pp. 1-13, Appendix B charts not read).} +} + +@misc{fca2014fines, + author = {{Financial Conduct Authority}}, + title = {{FCA fines five banks {\pounds}1.1 billion for FX failings and announces industry-wide remediation programme}}, + year = {2014}, + month = nov, + url = {https://www.fca.org.uk/news/press-releases/fca-fines-five-banks-for-fx-failings}, + note = {Press release, first published 12 Nov 2014. Read in full (card: library/fca2014fines.md). The five Final Notices (linked from it) were not opened.} +} + +@misc{nydfs2015barclays, + author = {{New York State Department of Financial Services}}, + title = {{In the Matter of Barclays Bank PLC, Barclays Bank PLC, New York Branch: Consent Order under New York Banking Law section 44}}, + year = {2015}, + month = nov, + url = {https://www.dfs.ny.gov/system/files/documents/2020/04/ea151117_barclays.pdf}, + note = {Signed November 2015 (day illegible in the scan); announced 18 Nov 2015. Read in full (card: library/nydfs2015barclays.md). The WHEREAS statements are recitals in a negotiated consent order, not findings; Barclays states its own purpose for last look in the recitals.} +} + +@misc{nydfs2015barclayspr, + author = {{New York State Department of Financial Services}}, + title = {{NYDFS Announces Barclays to Pay Additional \$150 Million Penalty, Terminate Employee for Automated, Electronic Foreign Exchange Trading Misconduct}}, + year = {2015}, + month = nov, + url = {https://www.dfs.ny.gov/reports_and_publications/press_releases/pr1511181}, + note = {Press release dated 18 Nov 2015. Read in full (card: library/nydfs2015barclayspr.md).} +} + +@misc{cftc2017fxcm, + author = {{U.S. Commodity Futures Trading Commission}}, + title = {{CFTC Orders Forex Capital Markets, LLC (FXCM), Its Parent Company, FXCM Holdings, LLC and FXCM's Founding Partners, Dror Niv and William Ahdout, to Pay a \$7 Million Penalty for FXCM's Defrauding of Retail Forex Customers}}, + year = {2017}, + month = feb, + url = {https://www.cftc.gov/PressRoom/PressReleases/pr7528-17}, + note = {Release 7528-17, dated 6 Feb 2017; title from the page HTML title element. Read in full (card: library/cftc2017fxcm.md). The order itself was not opened.} +} + +@misc{bis2019fx, + author = {{Bank for International Settlements}}, + title = {{FX settlement risk remains significant}}, + year = {2019}, + month = dec, + howpublished = {BIS Quarterly Review, 8 December 2019}, + url = {https://www.bis.org/publ/qtrpdf/r_qt1912x.htm}, + note = {Quotations checked against the page by auditor, 2026-10-13 (the page fetches with a browser user-agent). Authors not stated on the page.} +} + +@misc{us2072usc, + author = {{United States}}, + title = {{28 U.S.C. section 2072: Rules of procedure and evidence; power to prescribe}}, + url = {https://www.law.cornell.edu/uscode/text/28/2072}, + note = {Cornell LII; read in full 2026-10-13 (card: library/us2072usc.md). Statute, so no year given.} +} + +@misc{us2073usc, + author = {{United States}}, + title = {{28 U.S.C. section 2073: Rules of procedure and evidence; method of prescribing}}, + url = {https://www.law.cornell.edu/uscode/text/28/2073}, + note = {Cornell LII; read in full 2026-10-13 (card: library/us2073usc.md). Added Pub. L. 100-702, title IV, section 401(a), 19 Nov 1988, 102 Stat. 4648. Subsection (e): failure to comply does not invalidate a rule.} +} + +@misc{us455usc, + author = {{United States}}, + title = {{28 U.S.C. section 455: Disqualification of justice, judge, or magistrate judge}}, + url = {https://www.law.cornell.edu/uscode/text/28/455}, + note = {Cornell LII; read subsections (a) and (b) 2026-10-13.} +} + +@misc{us9usc10, + author = {{United States}}, + title = {{9 U.S.C. section 10: Same; vacation; grounds; rehearing}}, + url = {https://www.law.cornell.edu/uscode/text/9/10}, + note = {Cornell LII; read subsections (a) and (b) 2026-10-13.} +} + +@misc{mathews1976eldridge, + author = {{Supreme Court of the United States}}, + title = {{Mathews v. Eldridge, 424 U.S. 319 (1976)}}, + year = {1976}, + url = {https://www.law.cornell.edu/supremecourt/text/424/319}, + note = {Cornell LII; read in part (card: library/mathews1976eldridge.md; syllabus, opinion of the Court, dissent). The syllabus puts the three factors at pp. 332-335.} +} + +@misc{tumey1927ohio, + author = {{Supreme Court of the United States}}, + title = {{Tumey v. Ohio, 273 U.S. 510 (1927)}}, + year = {1927}, + url = {https://www.law.cornell.edu/supremecourt/text/273/510}, + note = {Cornell LII; read in full (card: library/tumey1927ohio.md). Locators are paragraph numbers from the card count; LII prints no U.S. Reports pages.} +} + +@misc{caperton2009massey, + author = {{Supreme Court of the United States}}, + title = {{Caperton v. A.T. Massey Coal Co., 556 U.S. 868 (2009)}}, + year = {2009}, + url = {https://www.law.cornell.edu/supct/html/08-22.ZS.html}, + note = {Cornell LII; read in full (card: library/caperton2009massey.md; syllabus, opinion of the Court, both dissents). Cited by Part; no U.S. Reports pagination in the LII text.} +} + +% --- reused from references.bib --- + +@article{rosenblat2016algorithmic, + author = {Rosenblat, Alex and Stark, Luke}, + title = {Algorithmic Labor and Information Asymmetries: A Case Study of {Uber}'s Drivers}, + journal = {International Journal of Communication}, + year = {2016}, + volume = {10}, + pages = {3758--3784}, + url = {https://ijoc.org/index.php/ijoc/article/view/4892}, + note = {Read in part: abstract, introduction to p. 3763 and the surge-pricing section, from the open-access PDF. Crossref returned only the SSRN working-paper record (10.2139/ssrn.2686227), so the journal article is verified through the PDF's own running head, not a DOI.} +} + +@article{rahman2021invisible, + author = {Rahman, Hatim A.}, + title = {{The Invisible Cage: Workers’ Reactivity to Opaque Algorithmic Evaluations}}, + journal = {Administrative Science Quarterly}, + volume = {66}, + number = {4}, + pages = {945--988}, + year = {2021}, + doi = {10.1177/00018392211010118}, + note = {Crossref verified; read in full (card: library/rosenblat2016algorithmic.md; IJoC pp. 3758-3784).} +} + +@article{vaccaro2020end, + author = {Vaccaro, Kristen and Sandvig, Christian and Karahalios, Karrie}, + title = {{"At the End of the Day Facebook Does What It Wants"}}, + journal = {Proceedings of the ACM on Human-Computer Interaction}, + volume = {4}, + number = {CSCW2}, + pages = {1--22}, + year = {2020}, + doi = {10.1145/3415238}, + note = {Crossref verified (3 authors); abstract only. Large online experiment; no appeal design improved perceptions of fairness, accountability, trustworthiness or control against a no-appeal baseline. Participant number not retrieved.} +} + +@article{vaccaro2021contestability, + author = {Vaccaro, Kristen and Xiao, Ziang and Hamilton, Kevin and Karahalios, Karrie}, + title = {{Contestability For Content Moderation}}, + journal = {Proceedings of the ACM on Human-Computer Interaction}, + volume = {5}, + number = {CSCW2}, + pages = {1--28}, + year = {2021}, + doi = {10.1145/3476059}, + note = {Crossref verified (4 authors); abstract only. Participatory design workshops.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_J_who_holds_the_convention.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_J_who_holds_the_convention.md new file mode 100644 index 00000000..86621e1f --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_J_who_holds_the_convention.md @@ -0,0 +1,411 @@ +# J — Who holds the convention: courts, foreign-exchange markets and platforms + +*Audited 2026-10-13 (AUDIT_05.md); WRONG, OVERSTATED and STITCHED rows corrected in place.* + +*Facet J, 2026-10-13. Agent-drafted. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** The author's ruling D8 (2026-10-13): "we just dont know yet. we can look at how courts and forex markets work, compare that platforms (which i think take a highly draconian position), and look for where the convention bends towards". I compare three settings in which a third party stands between two parties: (1) courts and arbitration, (2) foreign-exchange (FX) dealer markets and their settlement and benchmark infrastructure, (3) digital platforms (social media, app stores, marketplaces, ride-hail). For each I ask who writes the procedural conventions, what the losing party can do, whether the third party is interested and what holds it, what the evidence says about effects, and in whose direction the conventions have moved. The target is premise P10 of `ARGUMENT_MAP.md` (the scribes shared their readers' interest; a platform owner does not) and `T_adjudication.md` §5.1. + +**The question the synthesis answers.** Do courts and FX markets hold their conventions in a way platforms do not, and does the record support "a highly draconian position" for platforms? Short answer, argued in (c): partly. Platform terms of service give the owner near-total discretion (Suzor), and until 2019 no statute I could open limited that. But the FX comparison does not flatter FX: the FX Global Code is voluntary, its last-look rule leaves "sole discretion, based upon the validity and price check processes" with the dealer, and five banks were fined by the FCA for control failures that allowed traders to attempt to manipulate fix rates (the Code was written later by central banks and private-sector participants, and the sources do not say these were the same firms). By one measure, the EU Platform-to-Business Regulation and the Digital Services Act now require platforms by law to give notice, reasons and a complaint route; the Code asks related things of FX dealers by voluntary principle: that the basis for rejecting requests be disclosed (Principle 9), that rejection reasons be recorded and information be available to clients on request (Principles 36, 39), that substantial changes to terms be agreed before any transaction (Principle 40, which is stronger than DSA Art. 14(2), a duty only to inform), and that concerns can be raised through a channel (Principle 7). The Code names no complaint body or sanction. I compared texts, not practice. + +**Legend for read status.** *Read in full* = I opened the whole text this session. *Read in part* = I opened the text and read the named sections or passages. *Abstract only* = a database abstract (Crossref, OpenAlex, Consensus) and nothing else; all content comes from that abstract. *Metadata only* = Crossref confirmed the record; no abstract or text. *Not accessible* = no text, record only. A source is quotable only at full-text depth (RESEARCH_PLAN §3); quotations below are from text I opened, and each needs the version of record before print where I say so. Downloaded copies are in `~/iit-playground/dxd_foundation_local/pdfs/J/` (outside the repo). + +**Count by read status (one bib key each; 56 keys, of which four are reused from `references.bib`):** read in full 6 (`us2072usc`, `us2073usc`, `us455usc`, `us9usc10`, `suzor2016responsibilities`, `seattle2024deactivation`); read in part 19; abstract only 22; metadata only 7; not accessible 2. Entries that share a heading in (b) have separate keys in the `.bib`. These counts are as at drafting; the 2026-10-13 full-text cards raised the read status of most entries (see each entry and "Changes after full-text cards"). + +--- + +## (a) Search log + +All searches ran on 2026-10-13. "Fetch" = curl with a browser user-agent; "WebFetch" = the summarising fetch tool (its extractions are paraphrase-risk, so I quote only text I also held as a file, except where I say so). + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | Repo reads | `RESEARCH_PLAN.md`, `SOURCE.md`, `ARGUMENT_MAP.md`, `T_adjudication.md` §5, `REUSE.md`, `H_candidate_moves.md` (head), `references.bib` greps | n/a | Reusable keys found: `rosenblat2016algorithmic`, `rahman2021invisible`, `vaccaro2020end`, `vaccaro2021contestability`. Nothing on courts, FX, Galanter, Tyler or the law merchant was in the repo. | +| 2 | Crossref `/works/{doi}` | 36 DOIs | 35 resolved to the intended title | Wrong guesses caught and re-searched: `10.2307/1290356` (is West, "Private Ordering at the World's First Futures Exchange"; correct Bernstein 2001 is `10.2307/1290478`); `10.1111/j.1740-1461.2010.01201.x` (is Braslow and Cheit; correct Colvin is `...01200.x`); `10.1080/14697688.2017.1303910` (404; correct Oomen is `10.1080/14697688.2016.1262545`); `10.1016/j.jbankfin.2017.10.006` and `10.1007/s11579-018-0224-5` (wrong papers; correct Evans is `10.1016/j.jbankfin.2017.09.017`). | +| 3 | Crossref `query.bibliographic` | 20 title queries | 3 per query | Greif 1993 (AER) **not found**; Van Loo 2016 **not found**; Klonick 2018 **not found** (found in file, see row 9); Cartea, Jaimungal and Walton (last look) and Greif "Political foundations of the law merchant" returned HTTP 429 and wrong hits, not retried. Thibaut and Walker 1975 exists only as book-review records (Crossref `10.2307/1371953`, `10.2307/1142510`, `10.2307/448155`). | +| 4 | OpenAlex `/works/doi:` | 15 DOIs | 10 abstracts | Used for abstracts and open-access links. No abstract: Fuller 1978, Bernstein 1992, Evans 2018, Douek 2023, Lind and Tyler 1988. | +| 5 | Unpaywall | 14 DOIs | HTTP 422 on all | I had no registered address (the `oa_acquire.py` file held none). Replaced by OpenAlex. | +| 6 | Fetch | FX Global Code, `globalfxc.org/docs/fx_global.pdf` | 84 pages | Read the Foreword and Principle 17. The text does not print an edition date; The PDF footer reads "Updated December 2024"; I call it "the current GFXC PDF (December 2024)". | +| 7 | Fetch | FSB, *Foreign Exchange Benchmarks: Final Report*, 30 Sept 2014 (`fsb.org/wp-content/uploads/r_140930.pdf`) | 133 pages | Read summary and recommendations. | +| 8 | Fetch | FSB progress report, 1 Oct 2015 (`fsb.org/uploads/FX-Benchmarks-progress-report.pdf`) | PDF | Found via WebSearch. | +| 9 | Fetch | Klonick, *The New Governors*, `harvardlawreview.org/wp-content/uploads/2018/04/1598-1670_Online.pdf` | 73 pages | Read §III.B.4 and the Conclusion. | +| 10 | Fetch | NYDFS press release `pr1511181` and consent order `ea151117_barclays.pdf` | order 3,468 words | First URL (2019/12 path) returned an HTML bot page; the 2020/04 path from the press release worked. | +| 11 | WebSearch + Fetch | FCA press release 12 Nov 2014 | HTML | The five Final Notices were linked but not opened. The direct PDF guess (`ubs-ag-2014.pdf`) returned 404. | +| 12 | WebSearch + Fetch | CFTC press release 7528-17 | HTML | | +| 13 | Fetch | EUR-Lex (P2B, DSA, Platform Work Directive), five URL forms, plus the Publications Office resource URL | HTTP 202 empty or error | **EUR-Lex is bot-walled.** P2B text taken from `legislation.gov.uk/eur/2019/1150` (the retained-EU-law version, with small UK substitutions marked). DSA text taken from `eu-digital-services-act.com` (an unofficial mirror). Platform Work Directive 2024/2831 **not opened**. | +| 14 | Fetch | Cornell LII: 28 U.S.C. §§2072, 2073, 455; 9 U.S.C. §10; *Mathews v. Eldridge*; *Tumey v. Ohio*; *Caperton* | HTML | Full statutory text and opinions. | +| 15 | Fetch | Galanter 1974, via University of Wisconsin Law repository (`repository.law.wisc.edu/s/uwlaw/item/31285`) | 66-page PDF | Open-access repository copy. | +| 16 | Fetch | Oomen, *Last look*, LSE Research Online accepted manuscript | 28 pages | | +| 17 | Fetch | Michigan Law Review (Bernstein 2001), Chicago Unbound (Bernstein 1992), SSRN (Douek), BIS Galati 2002, Roth (Harvard DASH) | HTML bot walls or 404 | **Not opened.** Abstract from OpenAlex only. | +| 18 | Consensus | `gig platform worker deactivation appeals process fairness drivers account suspension empirical study` | 20 | FareShare, Maffie and Hurtado, Wiener et al., Putri et al. | +| 19 | Consensus | `effect of explanations or appeals on users after content removal or account suspension on platform, behavior outcomes` | 20 | Jhaver (two), Tyler et al. 2019, Myers West, Gonçalves et al. | +| 20 | Consensus | `online platform terms of service unilateral modification study of marketplace seller termination` | 19 | van der Donk, Pałka et al. | +| 21 | WebSearch | Suzor 2018 open copy | 9 | Found the 2016 AoIR conference paper (4 pp.), which I read in full. The 2018 article is abstract only. | +| 22 | WebFetch | BIS Quarterly Review Dec 2019, `r_qt1912x.htm` | extraction | BIS direct fetch returned its own 404 page for every URL I tried, including Galati 2002. This note is read through the extraction only. | +| 23 | WebSearch | Seattle deactivation ordinance; Oversight Board annual numbers; Epic v. Apple; WM fix window change | 10 each | Seattle: model notice downloaded from `seattle.gov`; ordinance text not found. Oversight Board: Q1 2022 and Q4 2021 transparency reports downloaded. Epic v. Apple: news reports only. | +| 24 | Scholar Gateway | not tried | 0 | Failed in every earlier run. | + +Not done: no search for studies of Amazon seller suspension appeals, App Store rejection rates, Uber/Lyft deactivation counts, or ESMA retail-CFD interventions; no FX studies of retail customers' last-look experience; no EU Platform Work Directive text; no Cartea et al. on last look; no court-delay statistics. + +--- + +## (b) Entries + +Format: **key** — reference. *Verification.* *Read status.* *What it says* with locators. *Bearing.* + +### B1. Courts and arbitration + +**galanter1974why** — Galanter, M. (1974). Why the "haves" come out ahead: Speculations on the limits of legal change. *Law & Society Review, 9*(1), 95-160. https://doi.org/10.2307/3053023. *Verification:* Crossref title match; open-access repository PDF (Wisconsin Law). *Read status:* read in part (card: library/galanter1974why.md; main text pp. 95-112, 113-125, 129-137 and 141-151 read closely, footnotes selectively; pp. 126-128, 138-141, 145-148 in part). +- Defines the actors: "We might divide our actors into those claimants who have only occasional recourse to the courts (one-shotters or OS) and repeat players (RP) who are engaged in many similar litigations over time." (p. 97) +- On rules: "RPs can play for rules as well as immediate gains. First, it pays an RP to expend resources in influencing the making of the relevant rules by such methods as lobbying." (p. 100, item 6). +- On private forums: "Where they rely on third parties in dispute-resolution, it is likely to take a form (such as arbitration or a domestic tribunal) detached from official sanctions and applying domestic rather than official rules." (p. 111; the card reads this as describing relations among repeat players of comparable size in a dense, continuing relationship, not the platform case) +- He lists nine RP advantages (items 1-9, pp. 98-103), among them "It is the RP who writes the form contract" (p. 98). Note 4 (p. 97) says welfare clients and taxpayers "are not pure OSs, since there is next year's tax bill and next month's welfare check", and pp. 129-133 predict private ordering where relationships are dense and enduring. +- Footnote 3 (near p. 97) says the analysis "might be applied to the regulatory and rule-making phases of legal process." +- The summary table lists "cost and delay barriers" and "passivity" under "institutional facilities" among the advantages of the "haves". +- *Bearing:* the foundational source for Q4/Q5 on courts. Supports the point that the party who plays repeatedly shapes the rules, so a court is not a neutral holder of conventions. Complicates the idea that a court-like convention is protection for the weaker party. It also bears on platforms, with a qualification: the platform is the extreme repeat player, but the driver or seller is in a continuing, unequal relationship and so is not a pure one-shotter (n. 4, p. 97; pp. 129-133) (my application, not Galanter's). + +**tyler1988what** — Tyler, T. R. (1988). What is procedural justice? Criteria used by citizens to assess the fairness of legal procedures. *Law & Society Review, 22*(1), 103-135. https://doi.org/10.2307/3053563. *Verification:* Crossref; OpenAlex abstract. *Read status:* read in full (card: library/tyler1988what.md; journal pagination; reference list skimmed). +- Telephone survey (Chicago, spring 1984) of 652 analysed citizens with recent experience of police and courts (1,575 interviewed; callers to the police 47%, people stopped by the police 31%, court users 23%; p. 110). "procedural justice has a major influence on both satisfaction and evaluation." Seven issues matter, including "(4) the extent to which opportunities for representation were provided" and "(6) the opportunities for error correction". "the meaning of procedural justice varied according to the nature of the situation, not the characteristics of the people involved." +- *Bearing:* evidence that voice (representation), correctability and other criteria affect how parties judge a procedure and the authority, apart from the outcome. It does not support the Q2 list as one yardstick for all three settings. Its main finding is that the weights on criteria vary with the situation (16 of 48 situation comparisons significant, against 5 of 64 trait comparisons, p. 125), and it concludes there are "no universally fair procedures" (p. 131), so the grid needs weights per setting. "Notice" is not among the criteria Tyler measured; correctability was measured only as whether the respondent knew of an agency to which they could have complained, and 33% did (p. 113). Self-report in police and court encounters; it says nothing about FX dealers or platforms. + +**lind1988social** — Lind, E. A., & Tyler, T. R. (1988). *The Social Psychology of Procedural Justice.* Plenum. https://doi.org/10.1007/978-1-4899-2115-4. *Verification:* Crossref. *Read status:* metadata only. *Bearing:* the standard synthesis the author named; not opened. A human with library access should read it before the paper cites "procedural justice". + +**thibaut1975procedural** — Thibaut, J., & Walker, L. (1975). *Procedural Justice: A Psychological Analysis.* Erlbaum. *Verification:* the book exists as the subject of three contemporaneous reviews indexed in Crossref (`10.2307/1371953`, `10.2307/1142510`, `10.2307/448155`); I did not open them. No DOI for the book. *Read status:* not accessible. *Bearing:* none yet; named by the author as a lead. + +**tyler1990why** — Tyler, T. R. (1990). *Why People Obey the Law.* Yale University Press (Princeton University Press reprint 2006, https://doi.org/10.1515/9781400828609). *Verification:* Crossref (the 2006 reprint record); the 1990 edition is also reviewed at `10.1017/s000819730009958x`. *Read status:* not accessible. *Bearing:* none yet. + +**colquitt2001justice** — Colquitt, J. A., Conlon, D. E., Wesson, M. J., Porter, C. O. L. H., & Ng, K. Y. (2001). Justice at the millennium: A meta-analytic review of 25 years of organizational justice research. *Journal of Applied Psychology, 86*(3), 425-445. https://doi.org/10.1037/0021-9010.86.3.425. *Verification:* Crossref (five authors). *Read status:* abstract only. +- 183 justice studies. Distributive, procedural, interpersonal and informational justice "contribute incremental variance explained in fairness perceptions", with distinct relations to "job satisfaction, organizational commitment, evaluation of authority, ... withdrawal, performance." +- *Bearing:* extends procedural-justice effects from courts to employer-employee authority, the nearest evidence to platform work. Abstract only; I did not check effect sizes. + +**lind1993individual** — Lind, E. A., Kulik, C. T., Ambrose, M., & de Vera Park, M. V. (1993). Individual and corporate dispute resolution: Using procedural fairness as a decision heuristic. *Administrative Science Quarterly, 38*(2), 224-251. https://doi.org/10.2307/2393412. *Verification:* Crossref. *Read status:* metadata only. *Bearing:* a lead on how parties judge a third party's procedure when they cannot judge the outcome; not read. + +**fuller1978forms** — Fuller, L. L., & Winston, K. I. (1978). The forms and limits of adjudication. *Harvard Law Review, 92*(2), 353-409. https://doi.org/10.2307/1340368. *Verification:* Crossref (authors Lon L. Fuller and Kenneth I. Winston; start page 353). *Read status:* metadata only. *Bearing:* the author named it; not opened. Lead for "what makes a third party's decision adjudication rather than management". + +**colvin2011empirical** — Colvin, A. J. S. (2011). An empirical study of employment arbitration: Case outcomes and processes. *Journal of Empirical Legal Studies, 8*(1), 1-23. https://doi.org/10.1111/j.1740-1461.2010.01200.x. *Verification:* Crossref (volume 8, issue 1, pages 1-23). *Read status:* read in full (card: library/colvin2011empirical.md; Cornell author manuscript, manuscript pagination; reference list not read in full). +- 3,945 American Arbitration Association employment cases, 2003-2007, from employer-promulgated procedures. "(1) the employee win rate among the cases was 21.4 percent, which is lower than employee win rates reported in employment litigation trials"; "(3) mean time to disposition in arbitration was 284.4 days for cases that settled and 361.5 days for cases decided after a hearing, which is substantially shorter than times to disposition in litigation"; "(4) ... in 97 percent of these cases the employer paid 100 percent of the arbitration fees beyond a small filing fee, pursuant to AAA procedures". +- Body (card): employees won 16.9% against repeat employers and 31.6% against one-shot employers (p < .01; p. 19); mean awards were $16,134 against $40,546 in the prose (pp. 19-20; Table 3's columns look transposed); a repeat employer-arbitrator pairing went with a 12.0% win rate against 23.4% without one (p. 21). 59.1% of cases settled and the arbitration-litigation comparison is not matched ("apples and oranges", p. 7); the repeat-employer gap is not proven to be bias, since explanations of size and expertise predict it too (p. 19). +- *Bearing:* a private tribunal chosen by the stronger party: faster, cheaper for the employee, with a lower win rate and lower awards than litigation. Complicates both "courts are costly" and "private ordering favours the weaker party". Selection into arbitration is not controlled in the abstract. + +**us2072usc** / **us2073usc** — 28 U.S.C. §§2072, 2073. https://www.law.cornell.edu/uscode/text/28/2072 and /2073. *Verification:* Cornell LII, full text. *Read status:* read in full. +- §2072(a): "The Supreme Court shall have the power to prescribe general rules of practice and procedure and rules of evidence for cases in the United States district courts ... and courts of appeals." §2072(b): "Such rules shall not abridge, enlarge or modify any substantive right." +- §2073(a)(2): committees "shall consist of members of the bench and the professional bar, and trial and appellate judges." §2073(c)(1): committee meetings "shall be open to the public" unless the committee, in open session and with a majority present, closes all or part of a meeting and states its reason; meetings need "sufficient notice" (§2073(c)(2)). The Judicial Conference, not the Supreme Court, appoints the committees (§2073(a)(1)-(2), (b)). §2073(d): the recommending body "shall provide a proposed rule, an explanatory note on the rule, and a written report explaining the body's action, including any minority or other separate views." §2073(e): "Failure to comply with this section does not invalidate a rule prescribed under section 2072 or 2075 of this title." +- *Bearing:* Q1 for courts. Procedure is written by judges and lawyers under a statute, with open meetings and written reports, and may not change substantive rights. There is a notice-of-meeting duty (§2073(c)(2)) but no statutory comment period, and under (e) the process duties are not enforceable by invalidating a rule; the role of Congress (§2074) is not opened. + +**us455usc** / **us9usc10** — 28 U.S.C. §455; 9 U.S.C. §10. https://www.law.cornell.edu/uscode/text/28/455 ; /9/10. *Read status:* read in full. +- §455(a): "Any justice, judge, or magistrate judge of the United States shall disqualify himself in any proceeding in which his impartiality might reasonably be questioned." +- 9 U.S.C. §10(a)(2): a court may vacate an arbitral award "where there was evident partiality or corruption in the arbitrators". +- *Bearing:* Q3. The legal order holds the third party to disinterest by recusal and vacatur. Both are available after the fact and only to the party who can bring them. + +**mathews1976eldridge** — *Mathews v. Eldridge*, 424 U.S. 319 (1976). https://www.law.cornell.edu/supremecourt/text/424/319. *Read status:* read in part (the due-process passage). "The fundamental requirement of due process is the opportunity to be heard 'at a meaningful time and in a meaningful manner.'" The Court weighs "the private interest that will be affected by the official action; second, the risk of an erroneous deprivation ... and the probable value ... of additional or substitute procedural safeguards; and finally, the Government's interest, including the function involved and the fiscal and administrative burdens". The syllabus puts the three factors at pp. 332-335 (card); no other pin cites. *Bearing:* Q2. A named test that trades procedural protection against administrative cost. Binds the state, not private platforms (Suzor makes this point). + +**tumey1927ohio** / **caperton2009massey** — *Tumey v. Ohio*, 273 U.S. 510 (1927), LII full text; *Caperton v. A.T. Massey Coal Co.*, 556 U.S. 868 (2009), LII (one passage only). *Read status:* *Tumey* read in full (card: library/tumey1927ohio.md); *Caperton* read in full (card: library/caperton2009massey.md; syllabus, opinion of the Court and both dissents). +- *Tumey*: it "violates the Fourteenth Amendment and deprives a defendant in a criminal case of due process of law to subject his liberty or property to the judgment of a court, the judge of which has a direct, personal, substantial pecuniary interest in reaching a conclusion against him in his case." The mayor-judge got $12 in costs on conviction and none on acquittal. +- *Caperton*: recusal is required "where 'the probability of actual bias on the part of the judge or decisionmaker is too high to be constitutionally tolerable'" (quoting *Withrow*). +- *Tumey* details (card; the ¶ numbers are the card's count of the LII text, which prints no U.S. Reports pages): fines were split between the State and the village, half the village's share going to a "secret service fund" (¶¶14, 19-22); over May-December 1923 the mayor took $696.35 on top of his salary (¶36); pay-only-on-conviction was not embedded in custom (¶64) and about $100 a month in costs was not de minimis (¶65); the test is structural, not about the particular man (¶66); the mayor's dual role as chief executive for village finances is a second ground (¶67); "kinship, personal bias, state policy, remoteness of interest" are left to legislative discretion (¶38). The rule binds the State only and the remedy comes after the fact, at the accused's initiative. +- *Bearing:* the most direct legal precedent for the author's question. Courts, unlike platforms, are barred from deciding disputes in which the decider has a pecuniary stake in one outcome, and the bar is constitutional, not contractual. It complicates P10: the legal order has already ruled that an interested third party cannot hold the convention, and it enforces that through an outside body. Pin cites for Tumey are the card's paragraph numbers; Caperton is cited by Part. + +### B2. Law merchant and private ordering + +**greif1989reputation** — Greif, A. (1989). Reputation and coalitions in medieval trade: Evidence on the Maghribi traders. *Journal of Economic History, 49*(4), 857-882. https://doi.org/10.1017/S0022050700009475. *Verification:* Crossref. *Read status:* abstract only. "an economic institution based upon a reputation mechanism utilized by Mediterranean traders to confront the organizational problem associated with the exchange relations between merchants and their overseas agents" in a setting of "limited legal contract enforceability". *Bearing:* the enforcer is the community of repeat traders. Greif's 1993 AER paper (named by the author) was not found. + +**milgrom1990role** — Milgrom, P. R., North, D. C., & Weingast, B. R. (1990). The role of institutions in the revival of trade: The law merchant, private judges, and the Champagne fairs. *Economics & Politics, 2*(1), 1-23. https://doi.org/10.1111/j.1468-0343.1990.tb00020.x. *Verification:* Crossref. *Read status:* abstract only. "The system of judges used to enforce commercial law before the rise of the state was such an institution, and it successfully encouraged merchants (1) to behave honestly, (2) to impose sanctions on violators, (3) to become adequately informed about how others had behaved, (4) to provide evidence against violators of the code, and (5) to pay any judgments assessed against them". *Bearing:* the best case that a private convention can bind a third party (the judge), through reputation across repeated interaction. It is the model the FX dealer market invokes, and the FX record below shows where it failed. + +**bernstein1992opting** — Bernstein, L. (1992). Opting out of the legal system: Extralegal contractual relations in the diamond industry. *Journal of Legal Studies, 21*(1), 115-157. https://doi.org/10.1086/467902. *Verification:* Crossref. *Read status:* metadata only. + +**bernstein2001private** — Bernstein, L. (2001). Private commercial law in the cotton industry: Creating cooperation through rules, norms, and institutions. *Michigan Law Review, 99*(7), 1724 ff. https://doi.org/10.2307/1290478. *Verification:* Crossref (start page 1724; end page not shown, so the reference gives 1724 ff.). *Read status:* abstract only (publisher PDF bot-walled; SSRN `10.2139/ssrn.281437`, not opened). +- "The cotton industry has almost entirely opted out of the public legal system, replacing it with one of the oldest and most complex systems of private commercial law." Most international cotton sales are governed "by the rules of the Liverpool Cotton Association." The system relies on "reputation-based nonlegal sanctions". +- *Bearing:* a modern case of conventions held by the trade association, in a setting where the parties are repeat dealers on both sides. It supports the claim that conventions can be held by non-state bodies. It does not cover a setting where one side is a single firm. + +**roth2002economist** — Roth, A. E. (2002). The economist as engineer: Game theory, experimentation, and computation as tools for design economics. *Econometrica, 70*(4), 1341-1378. https://doi.org/10.1111/1468-0262.00335. *Verification:* Crossref. *Read status:* abstract only (Harvard DASH link returned an error). "Market design involves a responsibility for detail, a need to deal with all of a market's complications, not just its principle features." Cases: the National Resident Matching Program and FCC spectrum auctions. *Bearing:* the market-design view of who writes the rules: a designer. Facet H already holds the matching evidence (P8); I add nothing beyond the abstract. + +### B3. Foreign-exchange and settlement + +**globalfxc2024code** — Global Foreign Exchange Committee (footer "Updated December 2024"). *FX Global Code: A set of global principles of good practice in the foreign exchange market.* https://www.globalfxc.org/docs/fx_global.pdf. *Verification:* fetched 2026-10-13, 84 pages; the PDF footer reads "Updated December 2024" (key renamed from `globalfxc2024code`). *Read status:* read in full (card: library/globalfxc2024code.md; December 2024 edition, printed page numbers; earlier editions not read). +- Status: "The Global Code does not impose legal or regulatory obligations on Market Participants nor does it substitute for regulation, but rather it is intended to serve as a supplement to any and all local laws, rules, and regulation" (Foreword I). +- Who holds it: "The Global Code is maintained by the Global Foreign Exchange Committee (GFXC). The GFXC was established in May 2017 as a forum bringing together central banks and private sector participants". It "assesses regularly whether particular foreign exchange market developments warrant specific revisions" (Foreword I). +- Adoption: "The Statement, like the Code, is voluntary" (Foreword II). "the decision of what steps should be undertaken, and in what manner, resides with each Market Participant" (Foreword II). +- Principle 17: "Market Participants employing last look should be transparent regarding its use and provide appropriate disclosures to Clients." "In the context of last look, the Market Participant has sole discretion, based upon the validity and price check processes, over whether the Client's trade request is accepted or not, leaving the Client with potential market risk in the event the trade request is not accepted." "Last look should not be used for purposes of information gathering with no intention to accept the Client's request to trade." The disclosure the Code asks for is general (whether price changes may affect the decision, the typical decision time, the purpose); I found no requirement to give a client a reason for an individual rejection. +- Further principles (card; all voluntary): Principle 8 (pp. 12-13): a Principal has "no obligation to execute the order until both parties are in agreement", must use acceptance discretion reasonably and fairly, and a changed capacity should be "agreed by both parties"; Principle 9 (p. 14): tell clients "the basis on which trade requests and/or orders might be rejected"; Principle 36 (p. 40): record the reasons behind electronic rejections, give clients transparency "where appropriate", and make information "available to Clients upon request"; Principle 39 (p. 42): "At the request of a Client, Market Participants should be able to provide information regarding the actions taken in handling a specific transaction"; Principle 40 (p. 42): "Where changes are substantial, these should be agreed before any transaction"; Principle 7 (p. 11): confidential channels for personnel or external parties to raise concerns. The Code covers wholesale participants, not the retail public (p. 5), and I found no complaint body, sanction or register of breaches in its 82 pages. +- *Bearing:* Q1-Q3 for FX. The convention is written by a committee of central banks and dealers and is voluntary; the dealer keeps "sole discretion, based upon the validity and price check processes". This **qualifies** the claim that FX shows a convention that holds the third party, and it parallels the platform position (the holder can decline). On change of terms the Code (Principle 40) asks more than DSA Art. 14(2), though voluntarily and for wholesale clients. I read only the current (December 2024) text, so this entry does not show that the Code bent toward clients' interest on last look compared with earlier editions. + +**oomen2016last** — Oomen, R. (2017). Last look. *Quantitative Finance, 17*(7), 1057-1070 (online 2016). https://doi.org/10.1080/14697688.2016.1262545. *Verification:* Crossref; LSE Research Online accepted manuscript (`eprints.lse.ac.uk/68811`). *Read status:* read in full (card: library/oomen2016last.md; accepted manuscript, manuscript pagination; appendix proofs skimmed); the accepted manuscript may differ from the published version. +- "the liquidity provider owns finality of the deal." "This contradicts the notion commonly quoted that asymmetric last look is 'unfair' and symmetric last look is somehow 'fair'. I argue that the choice of last look design is not an issue of fairness, but one of preference over liquidity access, trading style, and tolerance to execution risk." (§1) +- Author affiliation footnote: "global co-head of electronic FX spot trading at Deutsche Bank A.G." +- *Bearing:* the dealer-side argument that last look is not unfair. It **qualifies** a reading in which last look is simply a dealer privilege, more narrowly than the abstract suggests: in the model, equal effective spreads can be reached by any combination of design and protocol, but in the benign-flow case (Figure 8, §4) the designs "can lead to materially different execution risk profiles despite attaining the same effective spread", so the trader's risk is where the rules differ; the equivalence also depends on competing LPs and uninformed flow. It comes from a dealer. It is a model, not evidence of customers' outcomes. Use as the best statement of the opposing view. The published year is 2017 (vol. 17); the key's 2016 is the online date. + +**fsb2014foreign** — Financial Stability Board. (2014). *Foreign Exchange Benchmarks: Final Report*, 30 September 2014. https://www.fsb.org/wp-content/uploads/r_140930.pdf. *Verification:* fetched, 133 pages. *Read status:* read in part (card: library/fsb2014foreign.md; pp. 1-34 and the Appendix 3 summary read in full; about 85 pages of the principle-by-principle analysis skimmed). +- "it is the incentive and opportunity for improper trading behaviour of market participants around the fix, more than the methodology for computing the fix (although the two interact), which could lead to potential adverse outcomes for clients." (p. 2) The report also says the structure can create "an opportunity and an incentive for dealers to try to influence the exchange rate – allegedly including by collusion or otherwise inappropriate sharing of information". +- Recommendation 1: "The group recommends the fixing window be widened from its current width of one minute." Recommendation 3: "WM should expand their consultation activities to include a named user group". Recommendation 13: "stronger demonstration by market participants of compliance with the codes of the various foreign exchange committees". (pp. 3-4) +- "it is possible that the authorities will ultimately conclude that regulatory change is needed" (p. 4). +- *Bearing:* Q1, Q6. The benchmark owner (WM) wrote the method; the change was proposed by a working group chaired by Debelle (Reserve Bank of Australia) and Fisher (Bank of England) whose members were central banks, regulators, the ECB, IOSCO and the FSB secretariat (Appendix 2, pp. 33-34), with market participants consulted (p. 1), after misconduct allegations. The group says its work "was completely independent of the various conduct investigations" (p. 1). Records allegations (the report says "allegedly") that dealers used the structure to influence the rate; it does not say that the dealers broke a convention. + +**fsb2015progress** — Financial Stability Board. (2015). *Foreign Exchange Benchmarks: Report on progress in implementing the September 2014 recommendations*, 1 October 2015. https://fsb.org/uploads/FX-Benchmarks-progress-report.pdf. *Read status:* read in full (card: library/fsb2015progress.md; 21-page PDF, body pp. 1-13; Appendix B charts not read). "On 15 February 2015, WM widened the window used to calculate benchmark rates from 1 minute for 'traded' currencies and 2 minutes for 'non-traded' currencies, to 5 minutes for all currency pairs." *Bearing:* Q6. Dated change in a convention five months after the FSB report. + +**fca2014fines** — Financial Conduct Authority. (2014, 12 November). *FCA fines five banks £1.1 billion for FX failings and announces industry-wide remediation programme* [press release]. https://www.fca.org.uk/news/press-releases/fca-fines-five-banks-for-fx-failings. *Read status:* read in full (card: library/fca2014fines.md; the press release as HTML; Final Notices not opened). Fines "totalling £1,114,918,000"; "ineffective controls at the Banks allowed G10 spot FX traders to put their Banks' interests ahead of those of their clients, other market participants and the wider UK financial system." The FCA also launched "an industry-wide remediation programme" requiring senior management to "attest that this work has been completed." *Bearing:* Q3, Q6. The party that held the conventions (the dealer banks) breached them for 2008-2013 (per the FCA); a regulator enforced after the fact. + +**nydfs2015barclays** — New York State Department of Financial Services. (2015, November, signed, day illegible in the scan; announced 18 November). *In the Matter of Barclays Bank PLC, Barclays Bank PLC, New York Branch: Consent Order under New York Banking Law §44.* https://www.dfs.ny.gov/system/files/documents/2020/04/ea151117_barclays.pdf. And the press release `pr1511181` ("NYDFS Announces Barclays to Pay Additional $150 Million Penalty, Terminate Employee for Automated, Electronic Foreign Exchange Trading Misconduct"). *Read status:* read in full (card: library/nydfs2015barclays.md; the 14-page order and the press release). +- Recitals: "Barclays employed its Last Look order holding and rejection protocols broadly and indiscriminately, without first verifying if latency or arbitrage was involved"; "there was a lack of transparency both internally and with customers regarding Last Look". +- ¶13: trades were treated as toxic "[w]henever prices within this holding period moved against Barclays and in favor of the customer beyond a certain undisclosed loss threshold". ¶24: "the client would receive a simple rejection message: 'NACK,' which stood for 'Not Acknowledged.'" ¶25: when clients asked, Barclays "failed to disclose the reason that trades were rejected, instead citing technical issues or providing vague responses". The press release quotes an internal email: "Our Team generally does not share this info with the client, and just say it was a business reject." +- *Bearing:* Q2, Q3, Q5. The clearest dated instance of a dealer using an unreviewable discretion against its customers. It also shows the redress: a regulator, not the client. The order says Barclays "states" last look was meant as a defence (recital); that is Barclays' account. The statements that Barclays used last look "broadly and indiscriminately" and that there was "a lack of transparency" are recitals (WHEREAS clauses) in a negotiated consent order, not findings of the Department; the order finds only that the Bank conducted business "in an unsafe and unsound manner" (¶29). + +**cftc2017fxcm** — U.S. Commodity Futures Trading Commission. (2017, 6 February). Release 7528-17. https://www.cftc.gov/PressRoom/PressReleases/pr7528-17. *Read status:* read in full (card: library/cftc2017fxcm.md; the release; the underlying Order was not opened). FXCM "engaged in false and misleading solicitations of FXCM's retail foreign exchange (forex) customers by concealing its relationship with its most important market maker and by misrepresenting that its 'No Dealing Desk' platform had no conflicts of interest with its customers"; a $7 million penalty; the respondents "agree to withdraw from CFTC registration". *Bearing:* Q5. Retail FX customers' protection came through regulator action on disclosure. The release adds that the undisclosed market maker rebated "approximately 70 percent of its revenue" to FXCM, about $77 million from 2010 through 2014. The press release alone does not say how many customers were affected. + +**bis2019fx** — Bank for International Settlements. (2019, 8 December). FX settlement risk remains significant. *BIS Quarterly Review.* https://www.bis.org/publ/qtrpdf/r_qt1912x.htm. *Verification:* WebFetch only; authors not stated on the page; my own fetches of BIS returned its 404 page (the page fetches with a browser user-agent). *Read status:* read in part; the quotations below were checked against the page by the auditor, 2026-10-13. "The bankruptcy of Bankhaus Herstatt in 1974 demonstrated how FX settlement risk can undermine financial stability." CLS is "a specialist institution that settles FX transactions on a payment-versus-payment (PvP) basis". "The proportion of trades with PvP protection appears to have fallen from 50% in 2013 to 40% in 2019." "an estimated $8.9 trillion worth of FX payments at risk on any given day". *Bearing:* Q3, Q6. Settlement risk was addressed by a market-built institution (CLS, from 2002) but about 60% of trades by this count remain outside PvP. That **qualifies** the use of CLS as a clean success. The extraction says the page does not describe who governs CLS. + +**evans2018forex** — Evans, M. D. D. (2018). Forex trading and the WMR fix. *Journal of Banking & Finance, 87*, 233-247. https://doi.org/10.1016/j.jbankfin.2017.09.017. *Verification:* Crossref. *Read status:* read in part (card: library/evans2018forex.md; MPRA accepted manuscript, 82 pages, manuscript pagination; model derivation and about 40 pages of appendix tables skimmed). Evans finds that price changes above the bootstrap's 95th percentile occur four to twelve times as often as expected just before the month-end Fix (p. 23), that pre- and post-Fix changes correlate negatively (p. 28), and that the pattern is "pervasive across all currency pairs and throughout the decade" (p. 3), though Table 3 has pair-years without it. He cannot test collusion, because banks' trading data are proprietary (pp. 35-36), and WM announced a methodology change in October 2014 (p. 7, n. 9). *Bearing:* the academic analysis of the fix; it does not show that dealers broke a convention. His appendix lists the FSB among the authorities "investigating the forex trading activities" of banks, but the FSB says its work "was completely independent of the various conduct investigations" (`fsb2014foreign`, p. 1); and Evans finds extraordinary pre-Fix volatility where the FSB (seven currencies, April-September 2013) finds the rise in average volatility at the Fix "not especially large" (p. 18). The measures differ (tail frequency against mean), so they do not strictly conflict. + +### B4. Platforms + +**suzor2016responsibilities** — Suzor, N. (2016). The responsibilities of platforms: A new constitutionalism to promote the legitimacy of decentralized governance. *Selected Papers of AoIR 2016*, Berlin, 5-8 October. https://spir.aoir.org/ojs/index.php/spir/article/view/9055 (4 pp.). *Read status:* read in full. +- "In legal terms, the discretion of the platform owner is practically absolute. The language of constitutional rights – freedom of speech and association, requirements of due process and natural justice, rights to participate in the democratic process – has almost no application in the 'private' sphere". +- "I examine the legal terms and conditions of fifteen of the largest English-language social media platforms. ... In all cases examined, the terms of service provided broad, unfettered discretion to platform owners." "Like constitutional documents, terms of service grant powers; but unlike constitutions, they rarely limit those powers or regulate the ways they are exercised." +- *Bearing:* the main source for "draconian" in the ToS sense. Note the count: the 2016 paper says fifteen platforms; the 2018 article names fourteen in its note 3 (p. 9) and says it excluded the Russian-language platform VK from the top 15 by traffic, with data collected in June 2015, so the two counts reflect a change of sample, not an error (this paper does not say why it has fifteen). + +**suzor2018digital** — Suzor, N. (2018). Digital constitutionalism: Using the rule of law to evaluate the legitimacy of governance by platforms. *Social Media + Society, 4*(3). https://doi.org/10.1177/2056305118787812. *Verification:* Crossref; OpenAlex abstract. *Read status:* read in full (card: library/suzor2018digital.md; open-access PDF, journal pagination 1-11). Values proposed: governance that is "consensual, transparent, equally applied and relatively stable, and fairly enforced" (abstract, p. 1); analysis of the Terms of Service of 14 platforms (VK excluded from the top 15; note 3, p. 9; data from June 2015; U.S. and common-law centred, pp. 4-5) and "how poorly platforms perform on these criteria". The body repeats the 2016 sentence: "In all cases examined, the Terms of Service provided broad, unfettered discretion to platform owners" (p. 6). Facebook alone limits termination to violations of its terms; the other 13 reserve termination "for any reason (or no reason)" (pp. 5-6). All terms can be changed by the platform alone; "approximately half of the platforms committed to some responsibility to inform users" of changes, and some give no notice (p. 7). No terms set up an internal dispute process (pp. 7-8). The method reads formal documents only, so it shows "the outer legal bounds only" (pp. 5, 8-9). *Bearing:* as above; the note of "relatively stable" speaks directly to unilateral change. + +**klonick2018new** — Klonick, K. (2018). The new governors: The people, rules, and processes governing online speech. *Harvard Law Review, 131*(6), 1598-1670. https://harvardlawreview.org/wp-content/uploads/2018/04/1598-1670_Online.pdf. *Verification:* fetched PDF, 73 pages; no Crossref match found. *Read status:* read in full (card: library/klonick2018new.md; body of all four Parts and the Conclusion; the 486 footnotes read selectively). +- Appeals vary: "Facebook allows an appeal of the removal of only a profile or page — not individual posts or content"; at YouTube "account suspensions, 'strikes' on an account, and content removal are all appealable"; "At Twitter, any form of action related to the Twitter Rules can be appealed." (§III.B.4, p. 1648; the section begins at 1647; as of the article's 2017 data) +- Conclusion (p. 1665): "the lack of an appeals system for individual users and the open acknowledgment of different treatment and rule sets for powerful users over others reveal that a fair opportunity to participate is not currently a prioritized part of platform moderation systems." +- Platforms respond to "a diverse pluralistic group of forces that informs updates of their content-moderation policies and procedures" (Part IV, p. 1664; the forces she names are government requests, media coverage, civil-society groups and the process of moderation itself, pp. 1649-58). She treats user exit separately and sceptically: Post's "market for rules" misses that users stay and demand change, and platforms are "complementary, not substitute, goods" (pp. 1629-30); on p. 1668 exit sits alongside lobbying as the user's only recourse. +- She also calls the system "impressively intricate and responsive" (p. 1664). Her complaint is lack of accountability ("users are simply dependent on the whims of these corporations", p. 1668) and unequal treatment of powerful users (p. 1665), not severity. "Draconian" is the author's word (D8), not hers. +- *Bearing:* Q1, Q2, Q6. Records 2017-18 appeals practice and names the forces that move platform rules (users, governments, advertisers). It is a legal-descriptive study from interviews and policy documents, not an outcomes study. + +**douek2023meta** — Douek, E. (2023 SSRN posting; journal print 2024). The Meta Oversight Board and the empty promise of legitimacy. *Harvard Journal of Law & Technology, 37*(2), 373-445 (SSRN `10.2139/ssrn.4565180`). *Read status:* read in full (card: library/douek2023meta.md; the 2024 journal print; pp. 423-424 only in part, footnotes skimmed). The Board's decisions "have force only insofar and for as long as Meta ... agrees they do" (p. 375); its funding sits in an irrevocable trust and independent trustees alone can remove members (pp. 383-384, 397); it "does not have the power to review decisions Meta makes about accounts, advertising, events, groups, amplification, or the platforms' features and affordances" (p. 387); it issued 45 decisions in three years (p. 394); Meta judged its own original decision wrong in 32 of 50 shortlisted cases in 2022 (n. 139, p. 396). *Bearing:* a critical reading of the Board that supports P10 (Meta's support is "contingent on these returns being greater than the costs of complying", p. 439) while crediting real independence inside a narrow remit. + +**wong2022meta** — Wong, D., & Floridi, L. (2022). Meta's Oversight Board: A review and critical assessment. *Minds and Machines, 33*(2), 261-284. https://doi.org/10.1007/s11023-022-09613-x. *Read status:* read in full (card: library/wong2022meta.md; open-access PDF, reference list skimmed). Strengths: "its ability to enhance the transparency of content moderation decisions", "its assertiveness in interpreting its jurisdiction and overruling Meta". Weaknesses: "its limited jurisdiction, limited impact, Meta's control over the OB's precedent, and its lack of diversity." *Bearing:* the Board is a court-like body whose jurisdiction and precedent Meta controls. Case rulings bind Meta (unless implementing them "could violate the law") and policy recommendations do not, so "controls" holds most fully for precedent and scope (pp. 263-264). The text gives two 2021 counts, 27 decisions (p. 268) and 20 (p. 269), each attributed to the Board's annual report; cite neither without checking the report. + +**oversight2022q1** — Oversight Board. (2022). *Q1 2022 transparency report.* https://oversightboard.com/attachment/731010438013851. *Read status:* read in part (card: library/oversight2022q1.md; pp. 1-25 and the prose of the whole report; Annex II, pp. 26-64, scanned only). "In Q1 2022, an estimated 479,649 cases were submitted by users" (counted manually, "an estimate and preliminary"). "In this period, the Oversight Board decided two cases, overturning Meta's decision in both." "The Board's decisions are binding, and Meta must implement them within seven days of publication." 94% of submissions were appeals to restore content. *Bearing:* Q2, Q5. The appeal exists, binds on the content it decides, and reaches a tiny fraction of those who file. The ratio is a quarter's snapshot, and the two decisions came from cases selected in earlier quarters, so it compares different cohorts. One of the two decisions (2021-016-FB-FBR) was a Meta referral, so only one comes from the user pool counted in the denominator (p. 12). Meta judged its own original decision wrong in 14 of the 20 shortlisted user cases (70%, against 39% from Q4 2020 to Q4 2021) and reversed itself in all 14 before the Board ruled (p. 9). Policy recommendations bind nobody (pp. 18-21). + +**apple2026guidelines** — Apple Inc. *App Review Guidelines* (page shows "Last Updated: June 8, 2026"). https://developer.apple.com/app-store/review/guidelines/. *Fetched 2026-10-13.* *Read status:* read in part (card: library/apple2026guidelines.md; introduction, Rejections and Appeals paragraphs, 5.6, 3.1, 1.1-1.2, 4.7; sections 1.3-1.4, 4.4-4.6 and 5.1.4-5.5 skimmed). +- "This is a living document; new apps presenting new questions may result in new rules at any time." +- "Rejections: Our goal is to apply these guidelines fairly and consistently, but nobody's perfect." "Appeals: If you disagree with the outcome of your review, please submit an appeal. ... You may also suggest changes to the guidelines themselves". +- 5.6: "Your Developer Program account will be terminated if you engage in activities or actions that are not in accordance with the Developer Code of Conduct. To restore your account, you may provide a written statement detailing the improvements you plan to make. If your plan is approved by Apple and we confirm the changes have been made, your account may be restored." +- *Bearing:* Q1, Q2. The owner writes the rule, changes it "at any time", hears the appeal and restores accounts at its own discretion. It contains an appeal and a suggestion route, which **contradicts** a reading of "no appeal". It gives no notice period for changes. The card adds that the page hands upkeep of the convention to the developer (filtering, reporting and blocking for user-generated content, 1.2; the same for chatbots and mini-apps, 4.7.1) and that the rules differ by place (3.1.1(a) US storefront; 2.5.6 browser-engine entitlements for the EU and Japan; alternative marketplaces in some markets), without naming a law as the reason. + +**eu2019p2b** — Regulation (EU) 2019/1150 of the European Parliament and of the Council of 20 June 2019 on promoting fairness and transparency for business users of online intermediation services. Official text: https://eur-lex.europa.eu/eli/reg/2019/1150/oj (OJ L 186, 11.7.2019, p. 57; EUR-Lex bot wall; the auditor checked the quotations against the Official Journal text, 2026-10-13). Text first read: `legislation.gov.uk/eur/2019/1150` (UK-retained version; edits are only to jurisdictional references, e.g. Art. 4(4)(b) "the law", Art. 5(6)-(7) omitted, Art. 12 "United Kingdom"). The card read the Official Journal text: Art. 5(6)-(7) are in it, and 5(6) is the no-algorithm-disclosure clause below (recital 27 repeats it). *Read status:* read in part (card: library/eu2019p2b.md; Arts. 1-19 read in the official text except Art. 14(5)-(6); recitals read selectively); the pinpoints below (15 days, Art. 3(2); 30 days before termination, Art. 4(2); Arts. 11 and 12) match the official text. +- Art. 3(2): changes to terms must be notified "on a durable medium"; "shall not be implemented before the expiry of a notice period which is reasonable and proportionate ... That notice period shall be at least 15 days". The business user "shall have the right to terminate the contract ... before the expiry of the notice period." Art. 3(3): changes "implemented ... contrary to the provisions of paragraph 2 shall be null and void." Exceptions: legal obligation, or "an unforeseen and imminent danger" from "fraud, malware, spam, data breaches or other cybersecurity risks" (Art. 3(4)). +- Art. 4: restriction or suspension requires "a statement of reasons ... prior to or at the time" it takes effect; termination of the whole service requires it "at least 30 days prior"; reasons must refer "to the specific facts or circumstances" and to the ground in the terms (Art. 4(5)); exceptions where the user "repeatedly infringed" the terms (Art. 4(4)(c), 4(5)). +- Art. 5(1): the terms must "set out ... the main parameters determining ranking and the reasons for the relative importance of those main parameters". Art. 5(6): no duty to disclose "algorithms or any information that, with reasonable certainty, would result in ... deception of consumers or consumer harm". +- Art. 11: an internal complaint system that is "easily accessible and free of charge", with "transparency and equal treatment", and outcome communicated "in an individualised manner"; the duty does not apply to providers that are small enterprises (Art. 11(5)). Art. 12: platforms must name "two or more mediators", must "bear a reasonable proportion of the total costs", and mediation does not affect the right "to initiate judicial proceedings" (Art. 12(1), (4), (5)). +- *Bearing:* Q1, Q2, Q6. A statute that gives business users court-like procedural rights (notice, reasons, complaint, mediation) and voids non-conforming unilateral change. It is the strongest source against "the owner can change the convention alone": in the EU a change breaching Art. 3(2) is void, unless the business user waives the period or an Art. 3(4) exception applies; it limits how the owner changes terms, not whether it decides alone. Limits: applies to business users, not drivers or consumers; the complaint and mediation duties do not apply to providers that are small enterprises; mediators are named by the platform. + +**eu2022dsa** — Regulation (EU) 2022/2065 (Digital Services Act). Official text: https://eur-lex.europa.eu/eli/reg/2022/2065/oj (not opened; bot wall). Text read: `eu-digital-services-act.com/Digital_Services_Act_Article_{14,17,20,21}.html` (an unofficial mirror that labels itself "the final text"; the auditor later checked the quotations against the Official Journal text, OJ L 277, 27.10.2022, p. 1). *Read status:* read in part (card: library/eu2022dsa.md; operative text, Arts. 1-93, read in the Official Journal English text; about 150 of 156 recitals not read). The facet's own first reading of Arts. 14(2), 14(4), 17, 20, 21 came from the mirror. +- Art. 14(2): "Providers of intermediary services shall inform the recipients of the service of any significant change to the terms and conditions." (No notice period.) Art. 14(4): "diligent, objective and proportionate manner ... with due regard to the rights and legitimate interests of all parties involved". +- Art. 17(1): "a clear and specific statement of reasons" for restrictions on visibility, payments, service and accounts. Art. 17(3): must include "the facts and circumstances relied on", "information on the use made of automated means", the legal or contractual ground "and explanations as to why", and "information on the possibilities for redress". Art. 17(2): does not apply to "deceptive high-volume commercial content" and applies only "where the relevant electronic contact details are known". +- Art. 20: "access to an effective internal complaint-handling system" for "at least six months" (20(1)); complaints handled "in a timely, non-discriminatory, diligent and non-arbitrary manner" (20(4)); "reasoned decision" (20(5)); decisions "taken under the supervision of appropriately qualified staff, and not solely on the basis of automated means" (20(6)). +- Art. 21(1): the recipient may "select any out-of-court dispute settlement body that has been certified"; the body "shall not have the power to impose a binding settlement" (21(2)/(3)); the right to go to court is untouched. +- Enforcement and algorithms (card): users may complain to their Digital Services Coordinator (Art. 53) and seek compensation for damage from an infringement (Art. 54); fines reach 6% of worldwide annual turnover (Art. 52(3); for very large platforms Art. 74(1)), with periodic penalties of 5% of average daily turnover (Art. 76). For platforms with 45 million or more monthly EU recipients the Regulation reaches the algorithm: systemic-risk assessment of recommender systems (Art. 34(2)(a)), annual independent audit (Art. 37(1)), and an explanation of "the design, the logic, the functioning and the testing of their algorithmic systems" to a regulator on request (Art. 40(3)). +- *Bearing:* Q1, Q2, Q6. Extends P2B-type rights to consumers and content creators on online platforms. Limits: the out-of-court body cannot bind; the statement-of-reasons duty has exceptions; there is no notice period for term changes; Art. 19 exempts micro and small online platforms from Art. 20 unless they are very large platforms. + +**seattle2024deactivation** — Seattle Office of Labor Standards. *Model Notice of Deactivation* (revised 11/2024), "based on requirements of SMC 8.40.070". https://www.seattle.gov/documents/Departments/LaborStandards/Notice_of_Deactivation_Regular_form.pdf. *Read status:* read in full (a form). "Network companies are required to provide app-based workers with notice of deactivation 14 days before the date of deactivation and again on the date of deactivation." The notice must give "Reasons for deactivation (must include portion of Company's policy that was violated and specific incident(s) that violated the policy)" and the records relied on; "You will have at least until ___ (90 days from date of notice) to challenge the deactivation" and the company "has up to 14 days to respond". The company certifies "under penalty of perjury" that the records are true copies. *Bearing:* Q2, Q6. A local law imposes court-like notice, reasons, records and appeal on the platform, but the company has "up to 14 days to respond with a decision on the challenge" (form, p. 2): the first-instance reviewer is still the company that made the decision, and nothing on the form requires an independent reviewer. The Seattle OLS page gives the enactment (passed August 2023, effective 1 January 2025); the SMC 8.40 text itself was not read; the form is an agency's model, not the law. + +**nagarajrao2026fareshare** — Nagaraj Rao, V., Dalal, S., Schwartz, A., Liaqat, A., Calacci, D., & Monroy-Hernández, A. (2026). FareShare: A tool for labor organizers to estimate lost wages and contest arbitrary AI and algorithmic deactivations. *PACM HCI, 10*(2), 1-32. https://doi.org/10.1145/3788052. *Verification:* Crossref (six authors; Consensus lists a different order and count, so I follow Crossref). *Read status:* read in full (card: library/nagarajrao2026fareshare.md; arXiv preprint v1, 30 PDF pages, preprint page numbers; the version of record may differ). Says "Recent U.S. state policies now mandate appeals processes and recovering compensation during periods of wrongful deactivation" (abstract); a 3-month deployment with the State of Washington's largest rideshare union gave 122 user sign-ups linking 178 accounts, of which 143 synced and 35 did not (pp. 16-17). *Bearing:* Q6. Evidence that the new appeal rights come with a bottleneck, assembling earnings evidence, which the tool automates. It does not show that the rights need worker-side tools to be used: the union had already won over 1,400 reactivations and about $3.5 million before the tool (pp. 7, 10). No outcome data; the 80% overturn rate it cites comes from another study of Seattle drivers (p. 7). + +**rosenblat2016algorithmic** (reused) — Rosenblat and Stark (2016), *IJoC* 10, 3758-3784. See `X_counter_case.md`. *Read status:* read in full (card: library/rosenblat2016algorithmic.md; IJoC pp. 3758-3784). The interface is the instrument of the interested party's "soft control" (p. 3761). *Bearing:* the driver side of Q5. + +**rahman2021invisible** (reused) — Rahman (2021), *ASQ* 66(4), 945-988. *Read status:* abstract only (repo entry). "the criteria for success and changes to those criteria are unpredictable" (per facet H). Qualitative; shows what unnotified change does to workers. + +**kellogg2020algorithms** — Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. *Academy of Management Annals, 14*(1), 366-410. https://doi.org/10.5465/annals.2018.0174. *Read status:* abstract only. Algorithmic control works through "restricting and recommending ... recording and rating ... replacing and rewarding"; workers resist through "algoactivism". *Bearing:* the review the author named; supports that platform conventions serve the employer; not read beyond the abstract. + +**cameron2022expanding** — Cameron, L. D., & Rahman, H. (2022). Expanding the locus of resistance: Understanding the co-constitution of control and resistance in the gig economy. *Organization Science, 33*(1), 38-58. https://doi.org/10.1287/orsc.2021.1557. *Read status:* abstract only. Comparative ethnography of two platforms; workers' resistance "decreases in each subsequent stage of the labor process". *Bearing:* Q6 (collective and individual resistance); no outcome on rule change. + +**vaccaro2020end** / **vaccaro2021contestability** (reused) — see `H_candidate_moves.md`. *Read status:* abstract only. The 2020 experiment (per the repo note) found no appeal design improved perceived fairness over a no-appeal baseline. *Bearing:* **complicates** the claim that an appeal channel is what restores fairness; the abstract is all I have and participant numbers are not retrieved. + +**jhaver2019transparency** — Jhaver, S., Bruckman, A., & Gilbert, E. (2019). Does transparency in moderation really matter? User behavior after content removal explanations on Reddit. *PACM HCI, 3*(CSCW), 1-27. https://doi.org/10.1145/3359252. *Read status:* abstract only. 32 million Reddit posts; "offering explanations for content moderation reduces the odds of future post removals"; "explanations provided by human moderators did not have a significant advantage over explanations provided by bots". *Bearing:* Q4. Observational. Reasons change what users do next. + +**jhaver2019suspect** — Jhaver, S., Appling, D. S., Gilbert, E., & Bruckman, A. (2019). "Did you suspect the post would be removed?" *PACM HCI, 3*(CSCW), 1-33. https://doi.org/10.1145/3359294. *Read status:* read in full (card: library/jhaver2019suspect.md; journal pagination 192:1-192:33; Appendix A items 15-24 only). Survey of Reddit users whose posts had just been removed (907 respondents). The abstract says 37% "did not understand why their post was removed"; the body's 36.9% is of 849 coded open-ended answers (Table 7, 192:19). The abstract says "users who read community guidelines or receive explanations for removal are more likely to perceive the removal as fair and post again"; the body shows the opposite for the rules: readers of the rules rated removals less fair (β = -.314, Table 5; 192:16-17), which the authors call "surprising", while receiving an explanation went with a fairer rating (β = .384; 192:17). *Bearing:* Q4. + +**tyler2019social** — Tyler, T., Katsaros, M., Meares, T., & Venkatesh, S. (2019 online; vol. 17, 2021). Social media governance: Can social media companies motivate voluntary rule following behavior among their users? *Journal of Experimental Criminology, 17*(1), 109-127. https://doi.org/10.1007/s11292-019-09392-z. *Verification:* Crossref (vol. 17 issue 1; online 27 Dec 2019). *Read status:* abstract only. In two studies, "users who indicate that the procedures involved in the removal of their post have been fair are less likely to post new violating content"; in study 2 users randomly sent "a message emphasizing the procedural justice of platform procedures" were "less likely to post new violating content" than controls. *Bearing:* Q4. The only source I found that carries Tyler's procedural-justice programme onto a platform with random assignment. The sample and platform are not in the abstract; the message was about procedures, not a change in them. + +**myerswest2018censored** — Myers West, S. (2018). Censored, suspended, shadowbanned: User interpretations of content moderation on social media platforms. *New Media & Society, 20*(11), 4366-4383. https://doi.org/10.1177/1461444818773059. *Read status:* abstract only. Survey (n = 519) of users who had experienced moderation, on folk theories and "seeking redress". *Bearing:* Q2, Q4. + +**vanderdonk2022ex** — van der Donk, B. (2022). Ex-ante and ex-post access restrictions in marketplaces' terms of service. *European Review of Private Law, 30*(6), 909-928. https://doi.org/10.54648/erpl2022044. *Read status:* abstract only. Case study of the terms of the five most visited online marketplaces in the EU "confirms the existence of a fragmented approach towards ex-post access restrictions"; "a 'right to access a platform' does not exist"; current practice does "not ... live up to the safeguards in the new Digital Services Act". *Bearing:* Q5. The compliance gap against the DSA, per the abstract. + +**palka2025foreign** — Pałka, P., Porębski, A., Pałosz, R., & Wiśniewska, K. (2025). Do foreign service providers simply ignore the EU consumer law? An empirical study of 100 online terms of service. *Journal of Consumer Policy, 48*(3), 233-268. https://doi.org/10.1007/s10603-024-09580-z. *Read status:* read in full (card: library/palka2025foreign.md; journal pagination; appendix regression tables skimmed). 100 terms; those from providers headquartered outside the EU are "significantly more likely to contain unfair terms". *Bearing:* Q5. Terms still contain unfair clauses after the EU consumer-law regime (all 100 contained clauses that could be deemed unfair, and 99% let the provider alter the economic relationship unilaterally; p. 234). The comparison is within the EU market, so it says nothing about platform terms elsewhere, and EU-headquartered firms were affected too (23 non-Polish EU firms plus 15 Polish). + +**maffie2025enshittification** — Maffie, M. D., & Hurtado, H. (2025). The enshittification of work: Platform decay and labour conditions in the gig economy. *British Journal of Industrial Relations, 64*(1), 5-20. https://doi.org/10.1111/bjir.70004. *Verification:* Crossref (two authors; volume 64 issue 1; online 5 Aug 2025). *Read status:* abstract only. 30 interviews; gig companies "claw back value from workers over time" through "burden shifting", "feature addition and alteration" and "market manipulation". *Bearing:* Q6. Qualitative evidence that platform conventions move toward the owner as market power grows. + +**gillespie2019custodians** — Gillespie, T. (2019, Crossref date). *Custodians of the Internet.* Yale University Press. https://doi.org/10.12987/9780300235029. *Read status:* metadata only. Named by the author; not opened. + +### B5. Items I could not place in an entry +Van Loo, "The corporation as courthouse" (*Yale J. Reg.* 33, 2016): no Crossref record, not opened. Greif 1993 (AER): not found. Cartea, Jaimungal and Walton on last look: not retrieved. See (d). + +--- + +## (c) Synthesis + +### Q1. Who writes the conventions, who can change them, with what notice? + +**Courts.** Written by judges and lawyers under statute. In the federal system the Supreme Court prescribes procedural and evidence rules, which "shall not abridge, enlarge or modify any substantive right" (28 U.S.C. §2072); committees of "members of the bench and the professional bar", appointed under the Judicial Conference, recommend; they meet in public unless the committee, in open session and with a majority present, closes all or part of a meeting and states its reason (§2073(a)(2), (c)(1)); meetings need "sufficient notice" (§2073(c)(2)); and they must supply an explanatory note and a written report with minority views (§2073(d)). I found no statutory comment period in the text I opened (there is a duty of notice of meetings, §2073(c)(2)); failure to comply "does not invalidate a rule" (§2073(e)), so the process duties are not enforceable by striking a rule down; Congress's role (§2074) is unread. Arbitration procedures come from the administering body (the AAA in Colvin's data, in cases arising from "employer-promulgated arbitration procedures") or the trade association (the Liverpool Cotton Association in Bernstein 2001). + +**FX.** The FX Global Code is written and maintained by the GFXC, "a forum bringing together central banks and private sector participants", established May 2017; it "does not impose legal or regulatory obligations", and adopting it, or a Statement of Commitment, is voluntary (Foreword I-II). The WM/Reuters fix is calculated by the WM Company, which widened its window to five minutes for all currency pairs on 15 February 2015 (from one minute for "traded" and two for "non-traded" currencies), after the FSB group recommended widening it and left the width to WM (FSB 2014, rec. 1; FSB 2015, §2.1). Last-look practice is for each dealer to disclose; the Code asks it to explain at a minimum whether price changes in either direction may affect the decision, the typical decision time and the purpose (Principle 17). The Code does not say how the terms are agreed. Notice of change to clients: the Code asks, voluntarily, that substantial changes to terms be "agreed before any transaction" (Principle 40, p. 42) and that a changed capacity be "agreed by both parties" (Principle 8, p. 12). + +**Platforms.** Written and changed by the owner. Apple: "new apps presenting new questions may result in new rules at any time." Suzor (2016) found "broad, unfettered discretion to platform owners" in the terms of all fifteen platforms he examined; his 2018 article repeats the sentence for fourteen platforms (June 2015 data; p. 6) and finds that thirteen of them reserve termination "for any reason (or no reason)" (pp. 5-6). Statute now sets a floor for some users: P2B Art. 3(2) requires a notice period of at least 15 days for business users and voids changes that skip it (Art. 3(3)); DSA Art. 14(2) requires only that recipients be informed of "any significant change", with no period. + +### Q2. What can a party do when the third party decides against them? + +| | Notice | Reasons | Hearing / complaint | Appeal | Exit | +|---|---|---|---|---|---| +| Court | Not read | Not read | Opportunity to be heard "at a meaningful time and in a meaningful manner" (*Mathews*, quoting *Armstrong v. Manzo*); recusal (§455) | Not read (§2072(c) lets the rules define finality for appeals under 28 U.S.C. §1291; §1291 not opened) | Not read | +| Arbitration | By contract | Not read | Hearing; vacatur only for corruption, partiality, misconduct, excess of powers (9 U.S.C. §10) | Very limited | Employer-promulgated procedures (Colvin; whether a pre-dispute clause binds the employee is not in the abstract) | +| FX dealer (last look) | Disclosure of practice in general (Pr. 17); substantial changes to terms "agreed before any transaction" (Pr. 40), voluntary | Basis disclosed (Pr. 9); reason recorded, shared "where appropriate" and available on request (Pr. 36, 39); none required per trade unasked; Barclays sent "NACK" | No complaint body in the Code; confidential channel to raise concerns (Pr. 7) | None in the Code; regulator after the fact | Yes in wholesale: trade with another liquidity provider; client keeps "market risk" (inference, no source read) [unverified: see AUDIT_05] | +| Retail FX | Not read | Not read | Not read | None read; regulator enforcement after the fact (CFTC FXCM order, 6 February 2017) | Not read | +| Platform (ToS) | At the owner's option; EU: 15 days (P2B), "inform" (DSA) | Statement of reasons: EU (P2B Art. 4; DSA Art. 17), Seattle | Internal complaint: EU (P2B Art. 11; DSA Art. 20), Seattle 90 days (the company decides, within 14 days) | Apple: appeal; Meta: Oversight Board (binding on content; in Q1 2022 users submitted an estimated 479,649 cases (the Board's preliminary count) and the Board decided two cases (one of them a Meta referral), selected earlier, overturning Meta in both; the ratio compares different cohorts, and Meta had already reversed itself in 14 of the 20 shortlisted user cases (p. 9)) | Yes in principle; cost is lock-in (not measured here) | + +### Q3. Is the third party interested, and what holds it? + +- **Courts.** The judge is barred by the Constitution from deciding where he has a "direct, personal, substantial pecuniary interest" (*Tumey*), and by statute where "impartiality might reasonably be questioned" (§455(a)). Arbitrators can be challenged for "evident partiality" (9 U.S.C. §10(a)(2)). What holds the court: appeal, recusal, precedent, and an outside body that can review. Galanter's point is that the repeat player still shapes rules through lobbying and litigation selection (pp. 97, 100). +- **FX.** The dealer is the counterparty. In last look "the liquidity provider owns finality of the deal" (Oomen §1), and the Code leaves the dealer "sole discretion, based upon the validity and price check processes", over acceptance, and bars use of last look for information gathering (Principle 17). What holds it: reputation among repeat dealers (Greif, Milgrom-North-Weingast describe the mechanism for medieval traders), competition among liquidity providers, and regulators. The 2008-2013 record (FCA) and Barclays (NYDFS, whose statements are recitals in a negotiated consent order, not findings) show reputation and competition did not prevent breaches; regulators acted afterwards. The FSB wrote that the structure "can create an opportunity and an incentive for dealers to try to influence the exchange rate – allegedly including by collusion". +- **Platforms.** The owner is an interested party. What holds it: regulation (P2B; the DSA, enforced by regulators with fines up to 6% of worldwide turnover, a complaint route and a damages claim, Arts. 52-54, 74; Seattle), litigation (I saw only news reports of *Epic v. Apple*), reputation and user exit (Klonick names government requests, media coverage and civil-society groups as the forces that move platform rules and treats exit sceptically, pp. 1629-30, 1668), and a self-built body (the Oversight Board, with Meta controlling "the OB's precedent" per Wong and Floridi). +- **Result.** In none of the three settings is the third party uninterested except the court. FX and platforms both rely on outside regulators to hold an interested party. The difference is that FX also has a voluntary code and the platform owner has none beyond its own terms. + +### Q4. What does the evidence say about effects? + +- **Procedure affects judgments.** In Tyler's 652 interviews, procedural justice had "a major influence on both satisfaction and evaluation" of police and courts (Tyler 1988; the weights of the criteria vary by situation, p. 131). The organisational-justice meta-analysis (Colquitt et al. 2001, 183 studies) extends this to employment. +- **It transfers to platforms.** Tyler et al. (2019) report random assignment of a procedural-justice message reducing later violations [unverified: see AUDIT_05]; Jhaver et al. (2019) find removal explanations are associated with fewer later removals, with bot explanations showing no significant disadvantage against human ones in the removal-odds models. Counter: Vaccaro et al. (2020, per repo note) found no appeal design improved perceived fairness over no appeal. All abstract only. +- **Repeat players win more, and the private forum is not a refuge.** Galanter (1974) states the argument; Colvin (2011) finds a "repeat employer effect": employee win rates and awards are "significantly lower where the employer is involved in multiple arbitration cases" (abstract); in the body, 16.9% against 31.6% (repeat against one-shot employers, p. 19) and 12.0% against 23.4% where the same employer and arbitrator recur (p. 21). Overall the employee win rate was 21.4% in AAA employment arbitration, lower than rates reported for litigation trials (the comparison is not matched, p. 7), and in 97% of cases the employer paid all fees beyond a small filing fee (abstract). Speed in arbitration was higher. +- **Law merchant works under conditions.** Milgrom-North-Weingast and Bernstein describe private judges and trade-association tribunals that work among repeat dealers; the FX record shows the same structure failing where dealers allegedly colluded (FSB 2014). +- **FX last look.** The only quantitative statement I opened is Oomen's model, which argues last look "is not an issue of fairness" and that the trader's effective transaction cost "need not be affected by" the choice of design and protocol, though execution risk differs (§4). The NYDFS order shows misuse in one bank. I found no study of customers' realised costs. +- **Platform workers.** Qualitative: unpredictable criteria (Rahman 2021), soft control (Rosenblat and Stark 2016), platform decay (Maffie and Hurtado 2025). No study I opened measures what a statute-mandated notice or appeal changed for drivers. + +### Q5. Test "platforms take a highly draconian position" + +Where it holds. (i) On the face of the contract: Suzor (2016) found "broad, unfettered discretion to platform owners" in the terms of all fifteen platforms he examined; his 2018 article repeats the sentence for fourteen platforms (June 2015 data; p. 6), finds that thirteen of them reserve termination "for any reason (or no reason)" (pp. 5-6) and that none sets up an internal dispute process (pp. 7-8). (ii) Rule change: Apple may add rules "at any time"; Rahman records workers facing "unpredictable" criteria. (iii) Appeals: in 2017-18 Facebook allowed an appeal only of the removal of a profile or page (Klonick, p. 1648), and Klonick concluded that "the lack of an appeals system for individual users" and unequal treatment of powerful users showed that fair participation was not a prioritized part of platform moderation (p. 1665); at YouTube and Twitter, broader actions were appealable (p. 1648). She also calls the system "impressively intricate and responsive" (p. 1664), so her complaint is lack of accountability and unequal treatment, not severity. (iv) Scale: in Q1 2022 users submitted an estimated 479,649 cases to the Oversight Board (the Board's preliminary count) and the Board decided two cases (one of them a Meta referral), selected earlier, overturning Meta in both; the ratio compares different cohorts, and Meta had already reversed itself in 14 of the 20 shortlisted user cases (p. 9). + +Where it fails or needs qualification. +1. **FX is not a gentler holder.** The Code is voluntary; last look leaves "sole discretion, based upon the validity and price check processes" with the dealer; the Code requires disclosure of the basis on which requests may be rejected (Principle 9) and a record of the reason for each electronic rejection, shared with clients "where appropriate" and made available "upon request" (Principle 36), with handling information on request (Principle 39); I found no requirement to give a client, unasked, the reason for an individual rejection; Barclays sent clients "NACK", and an internal email said the team "generally does not share this info with the client". Five of the dealers active in the market were fined by the FCA for control failures that allowed traders to attempt to manipulate fix rates; the Code was written later by central banks and private-sector participants (Foreword I), and the sources do not say these were the same firms. Retail customers appear in the record only through regulators (CFTC/FXCM). Oomen, a dealer, argues last look is not unfair. +2. **Courts are costly and slow, and favour repeat players.** Galanter lists "cost and delay barriers" among the haves' advantages. Colvin finds arbitration faster than litigation, with a lower employee win rate and lower awards; this qualifies the case for a private forum, not for court procedure. +3. **Platforms are not uniform and are now regulated.** Apple has an appeal route and restores accounts at its discretion. In the EU, a change to terms that skips the 15-day notice period is void as against a business user (P2B Art. 3(2)-(3)), unless the business user waives the period or an Art. 3(4) exception applies. For online platforms that are not micro or small enterprises (Art. 19), the DSA requires a statement of reasons for restrictions imposed for illegal content or incompatibility with terms (Art. 17), an internal complaint system (Art. 20) and decisions taken "under the supervision of appropriately qualified staff, and not solely on the basis of automated means" (Art. 20(6)). Seattle's ordinance (SMC 8.40) requires notice 14 days before most deactivations, with reasons and records, and gives the worker 90 days to challenge through the company, which decides the challenge itself within 14 days (nothing on the model form requires an independent reviewer); the OLS cannot investigate whether the reason was permissible before 1 June 2027. On notice, reasons and complaint, the EU instruments bind platforms by law; the Code asks related things of FX dealers by voluntary principle (reasons and handling information on request, Principles 9, 17, 36, 39; substantial changes to terms agreed before any transaction, Principle 40, which is stronger than DSA Art. 14(2); a channel for concerns, Principle 7). I compared texts, not practice, and did not read the rules that bind dealers by law. +4. **The statutory rights are partial.** Out-of-court DSA bodies cannot bind (Art. 21); reasons duties have exceptions: P2B Art. 4(5) (repeat infringement leading to termination of the whole service; a legal bar on disclosure) and DSA Art. 17(2) ("deceptive high-volume commercial content"; contact details unknown); P2B's internal-complaint and mediation duties do not apply to providers that are small enterprises (Arts. 11(5), 12(7)); DSA Art. 19 exempts micro and small online platforms from Arts. 20-21 unless they are very large platforms; van der Donk finds five marketplaces' terms short of the safeguards the DSA would require (2022, before it applied); Pałka et al. find unfair clauses under EU consumer law in 100 terms of service (abstracts). +5. **Evidence on effect is thin.** Appeal designs did not improve perceived fairness in Vaccaro et al.; explanations were associated with lower odds of later removal in Jhaver et al.; no study I opened measures outcomes under the new EU or Seattle rules. + +### Q6. Where does the convention bend, and what moved it? + +| Setting | Date | Change | Mover | +|---|---|---|---| +| Court | 1927 | A judge with a pecuniary stake cannot decide (*Tumey*) | Litigation, Supreme Court | +| Court | 1976 | Due-process balancing test (*Mathews*) | Litigation | +| Court rules | 19 Nov 1988 (Pub. L. 100-702, §401(a)) | Rules via open committees, written reports (28 U.S.C. §2073; non-compliance does not invalidate a rule, §2073(e)) | Congress | +| FX | 2002 | CLS begins PvP settlement after Herstatt (1974) (Galati not opened; BIS 2019 extraction) | Market participants with G10 central banks | +| FX | 12 Nov 2014 | FCA fines five banks £1,114,918,000 for G10 spot FX conduct | Regulator, after misconduct | +| FX | 30 Sept 2014 / 15 Feb 2015 | FSB recommends a wider window (30 Sept 2014); WM adopts five minutes (15 Feb 2015) | Regulators and WM | +| FX | Nov 2015 (signed, day illegible in the scan); announced 18 Nov 2015 | NYDFS Barclays consent order on last look | Regulator | +| FX | 6 Feb 2017 | CFTC FXCM order (retail) | Regulator | +| FX | May 2017 onward | GFXC formed May 2017; Code later updated (current text: December 2024) | Central banks and industry | +| Platform | 2016-2018 | Appeal rights differ across Facebook, YouTube and Twitter (Klonick, as of 2017); scholars name the ToS gap (Suzor) | None dated; description, not a change | +| Platform | 11 Jul 2019 published (dated 20 Jun 2019); applies from 12 Jul 2020 | P2B Regulation | EU legislature | +| Platform | by Q1 2022 | Oversight Board issuing binding content decisions (Q1 2022 transparency report; the Q4 2021 report was not read) | Meta; the reason for its creation is not in the sources I read | +| Platform | 27 Oct 2022 published (dated 19 Oct 2022); applies from 17 Feb 2024 | DSA | EU legislature | +| Platform | Seattle ordinance passed Aug 2023, effective 1 Jan 2025 (OLS); model notice revised Nov 2024 | Seattle deactivation rights (SMC 8.40.070; the company decides the challenge); "recent U.S. state policies" mandate appeals (FareShare abstract) | Local and state law | +| Platform | 2025-26 | *Epic v. Apple* contempt order, per news reports only [unverified: see AUDIT_05] | Litigation | + +In courts the convention has bent toward the weaker party through litigation about the decider's interest. In FX it bent toward clients after scandal and regulators, and less through clients themselves. In platforms it is bending toward users through statute and local law, with platform self-regulation (Board, appeals) a partial response; the statutory rights stop short of independent first-instance review (in Seattle the company decides the challenge within 14 days; DSA out-of-court bodies cannot bind), and the Board reaches few cases while Meta controls its precedent. In FX the direction is shown only as regulator action after scandal; the current Code already asks, voluntarily, more than the DSA on change of terms (Principle 40), but earlier editions are unread, so no direction is shown for the Code itself. In all three, the movers were outside the third party. I found one setting (merchant tribunals among repeat dealers) where the parties themselves write the convention, and it works because the dealers are on both sides. + +### Comparison table + +| | Courts and arbitration | FX dealer markets | Platforms | +|---|---|---|---| +| Who writes | Judges and lawyers under statute; arbitration bodies and trade associations | GFXC (central banks and industry); WM for the fix; each dealer for last look | The owner; EU and local law as a floor | +| Who can change | Supreme Court after committee process; open meetings, written report | GFXC for the Code; WM for the fix; dealer alone for its terms | The owner alone; EU business users: not without 15 days' notice, unless waived or an Art. 3(4) exception applies | +| Notice | Not read; rules published | Practice disclosed generally (Pr. 17); substantial changes to terms to be agreed before any transaction (Pr. 40, voluntary) | Owner's discretion; P2B 15 days; DSA "inform"; Seattle 14 days | +| Reasons | Not read | Basis disclosed (Pr. 9); reason recorded, shared "where appropriate" and available on request (Pr. 36, 39); none required per rejection unasked | EU: statement of reasons; Seattle: reasons and records | +| Appeal | Not read; arbitration very limited | No complaint body or sanction in the Code (a channel for concerns, Pr. 7); regulator ex post | Internal complaint (EU; Seattle: decided by the company within 14 days); Apple appeal; Board for a few cases | +| Exit | Little | Wholesale: switch dealer; retail: not read | Possible; lock-in unmeasured | +| What binds the third party | Appeal, recusal, precedent, outside review | Reputation, competition, regulators ex post; Code voluntary | Statute where it exists; litigation; reputation; self-built body | +| Direction of change | Toward the weaker party's procedural protection, by litigation | Toward regulator action after scandal; a Code change toward client disclosure is not shown (earlier editions unread), though the current Code already asks more than DSA Art. 14(2) on change of terms | Toward user rights, by statute and local law; first-instance review stays with the owner (Seattle) and DSA out-of-court bodies cannot bind | + +Tyler (1988) found that the weights of fairness criteria vary with the situation and that no procedure is universally fair (p. 131), so the columns above are a descriptive grid, not one yardstick of fairness; "notice" is the facet's column, not one of Tyler's measured criteria. + +### Claims the sources contradict or qualify (author's wording first) + +1. **"platforms ... take a highly draconian position"** (D8). Qualified. True of terms-of-service discretion (Suzor 2016 and 2018, "broad, unfettered discretion") and, in the sense of thin appeals and lack of accountability, of 2017 practice (Klonick, pp. 1648, 1665). "Draconian" in the sense of severity is not supported by the sources I opened: neither Suzor nor Klonick uses the word, Suzor's finding is about discretion, and Klonick calls the system "impressively intricate and responsive" (p. 1664). Not true as a blanket statement: Apple's guidelines include an appeal route; the EU P2B and DSA require notice, reasons and complaint; Seattle requires notice and a 90-day challenge. Say "platform terms give the owner near-total discretion unless a statute limits it". +2. **Implication that FX markets show a convention that holds the third party** (D8, "how courts and forex markets work"). Qualified. The FX Global Code "does not impose legal or regulatory obligations" and is voluntary; last look is at the dealer's "sole discretion, based upon the validity and price check processes"; the fix was a target of attempted manipulation by traders at five fined banks (FCA 2014), and the FSB reported allegations of collusion (FSB 2014). FX shows a convention held by dealers and corrected by regulators. +3. **Implication that courts are the model for the weaker party.** Qualified. Galanter (1974) puts "cost and delay barriers" and repeat-player advantage in the courts; Colvin (2011) finds arbitration faster than litigation, with a lower employee win rate and lower awards. +4. **P10 / T §5.1: a convention the owner can withdraw alone is no convention.** Qualified by two sources. The FX Code can also be declined by any participant (Foreword II), although it asks, voluntarily, that substantial changes to terms be agreed before any transaction (Principle 40). Conversely, in the EU a platform that changes terms must give business users at least 15 days' notice (with exceptions) and the business user may terminate; the owner still decides the change alone. This weakens, but does not refute, "a convention the owner can withdraw alone is no convention". The Apple guidelines ("at any time") are the case where the owner can withdraw alone. +5. **Any claim that platforms lack court-like bodies.** Qualified. The Oversight Board decides cases and its decisions are "binding" (Q1 2022 report), but in Q1 2022 users submitted an estimated 479,649 cases (the Board's preliminary count) and the Board decided two cases (one a Meta referral), selected earlier, so the ratio compares different cohorts, and Meta had already reversed itself in 14 of 20 shortlisted cases before the Board ruled (p. 9); and Meta controls its precedent (Wong and Floridi, abstract). +6. **Any claim that the interested decider is unremarkable.** Contradicted by *Tumey* (1927): a judge with a "direct, personal, substantial pecuniary interest" cannot decide. The legal order already treats interest in the outcome as disqualifying. + +### Inferences that are mine and have no source + +1. Reading Galanter's repeat player as the platform and the driver or seller as a one-shotter. Galanter's examples are insurers, prosecutors and finance companies, and his note 4 (p. 97) says clients in continuing relationships "are not pure OSs", so the driver or seller fits only with that qualification. +2. That the court's disqualification of interested deciders (*Tumey*, §455) is the right model for the platform case. No source in this facet makes that comparison. +3. That EU platforms are bound by law to give notice, reasons and a complaint route that FX dealers are asked to give only by voluntary principle. The comparison is not uniform: on change of terms, Code Principle 40 (agreed before any transaction) asks more than DSA Art. 14(2) (inform). I compared texts of P2B, DSA and the Code (Principles 7, 8, 9, 17, 36, 39, 40); I did not compare practice or the MiFID II and national rules that bind dealers (not opened). +4. That the movers in all three settings were "outside the third party". Each row of the Q6 table is sourced; the generalisation is mine. +5. That a merchant tribunal works because "the dealers are on both sides". Milgrom-North-Weingast and Bernstein abstracts describe repeat dealing; the "both sides" reading is mine. +6. That the current (December 2024) Code's last-look text bends toward clients compared to 2015. I did not read earlier editions. +7. That the answer for the paper is to name no holder (D8). This facet does not choose. + +--- + +## (d) Gaps and unverified leads + +**Opened only through a bot wall or not at all (a human with access should open them).** +- EUR-Lex official texts: the cards read P2B and the DSA in the Official Journal English text (the facet had used the UK-retained version and an unofficial mirror). The Platform Work Directive (EU) 2024/2831 was not opened, and it is the instrument most relevant to drivers (human review of automated decisions). +- Michigan Law Review (Bernstein 2001), SSRN (Bernstein SSRN; Douek was later read in the 2024 journal print), Roth 2002 (Harvard DASH 404), BIS Galati (2002) "Settlement risk in foreign exchange markets and CLS Bank" (BIS Quarterly Review, Dec 2002, pp. 55-65; every BIS URL returned the BIS 404 page to curl). +- Books: Thibaut and Walker (1975), Lind and Tyler (1988), Tyler (1990), Gillespie (2019), Fuller (1978, *Harvard Law Review*), Greif (1993, *American Economic Review*), Bernstein (1992, JLS): metadata or nothing. +- Van Loo, "The corporation as courthouse" (*Yale J. Reg.* 33, 2016); no record found. +- FCA Final Notices (Citibank, HSBC, JPMorgan, RBS, UBS; linked from the press release at `fca.org.uk/your-fca/documents/final-notices/2014/`), CFTC and DOJ FX orders, and the Fair and Effective Markets Review (2015) are unread. +- Cartea, Jaimungal and Walton on last look; studies of retail FX customers; ESMA retail CFD measures: not searched. +- Epic v. Apple: only news reports; read the district court's 30 April 2025 order and the Ninth Circuit's opinion before citing. +- Amazon, Uber and Lyft deactivation counts or appeal outcomes; platform deactivation studies beyond the abstracts above; any study of whether the P2B notice rule changed platform behaviour (a review of the Commission's evaluation of P2B would be a start). +- SMC 8.40 text (only the model notice and the OLS page were read; the OLS page gives passage in August 2023 and effect on 1 January 2025). +- Oversight Board annual reports for 2023 onward (I read only Q1 2022; the 2023 figures of 398,597 appeals and 53 decisions came from a search summary, not a file, and are **not** used above). + +**Problems in the record I could not resolve.** +- Suzor counts 15 platforms (2016 paper) and 14 (2018 article); the card resolves this: the 2018 article excluded VK from the top 15 by traffic (note 3, p. 9), and both papers report "unfettered discretion" (2016, para. 6; 2018, p. 6). +- The BIS 2019 figures come from an extraction. +- Oomen's accepted manuscript may differ from the published article; the author is a dealer employee. + +## Changes after audit (2026-10-13) + +- Q1, courts: the committee meeting rule now reads "with a majority present" (a quorum), not "majority vote", and adds the Judicial Conference and the notice requirement; the earlier wording misread §2073(c)(1). +- Q1, FX: the WM window change is no longer "after an FSB group recommended it" as a five-minute decision; the FSB recommended widening and left the width to WM. +- Q1 and Q5, last look: "set by each dealer bilaterally" removed (the Code does not say who sets the terms); "sole discretion" now carries its bound ("based upon the validity and price check processes") and the ban on information gathering. +- Q1 and Q5, Suzor: "unfettered discretion in every document (15 in 2016, 14 in 2018)" split into the 2016 finding (fifteen platforms) and the 2018 abstract (fourteen platforms "perform poorly"). Q5 (i) still holds for the 2016 sample only. +- Q2, courts: the Notice and Appeal cells now read "Not read"; the "meaningful time" quotation moved to Hearing, where Mathews places it. The comparison table's courts Notice, Reasons and Appeal cells follow. +- Q2, arbitration: "pre-dispute clause binds employee" removed; Colvin's abstract says only that the procedures were employer-promulgated. +- Q2 and Q5 item 1, FX reasons: "the Code requires disclosure of practice, not reasons" replaced; Principles 9 and 36 require disclosure of the basis for rejections and a record of reasons. The verdict "FX is not a gentler holder" now rests on the voluntary Code, the bounded "sole discretion" and the Barclays record, not on an absence of reasons duties. +- Q2, FX exit: marked as inference with no source read. Retail FX appeal: "none read". +- Q3 and Q5 item 1, FSB and FCA: "create[d]" restored to "can create ... allegedly"; the statement that "FX dealers who wrote the Code also broke the conventions at the fix" removed, because the sources do not say the same firms wrote the Code. Q4's "dealers colluded" now says "allegedly". Claim 2 and the opening summary are changed to match. +- Q4 and Q5 item 5, Jhaver: "as good as" and "did reduce" replaced by "no significant disadvantage" and "associated with lower odds". +- Q4, Colvin: the "repeat players win more" bullet now cites the repeat employer effect; Q5 item 2 and claim 3: Colvin shows a private forum doing worse on win rate and awards, so the inference that "court-like is not automatically better for the weaker party" is dropped. +- Q5 (iii), Klonick: pages corrected to 1648 and 1665 and the 1665 sentence is her general conclusion; YouTube and Twitter appeals added. +- Q5 (iv) and claim 5, Oversight Board: the 479,649 to 2 ratio compares different cohorts, so it is a weak measure of scale; stated as such. +- Q5 item 3 and claim 4, EU law: the P2B voiding rule is subject to waiver and Art. 3(4) exceptions, and DSA duties to the Art. 19 scope. Claim 4's "false for EU business users" replaced: the EU rule weakens but does not refute "no convention if the owner can withdraw alone". The "more written duties than an FX dealer" comparison is restated as law versus voluntary principle. +- Q5 item 4: "small enterprises" exemption now names the exempt party (providers); van der Donk and Pałka restated (a gap against a coming law; unfair clauses under consumer law). +- Q6: Seattle dates, P2B and DSA publication and application dates, NYDFS signing date and the 1988 enactment of §2073 added; the FSB row split into recommendation and adoption. The row "Code revised, Principle 17 adds last-look disclosure" removed because no earlier edition was read; the comparison table's FX direction-of-change cell and inference 6 are narrowed to match, and the (b) bearing that the Code "does bend toward clients" is withdrawn. +- Unverifiable items marked: Tyler et al. (2019) abstract, FX exit by switching dealer, Epic v. Apple contempt order. + +## Changes after full-text cards (2026-10-13) + +Source: `audit/CARD_FACET_CHECKS.md`, Facet J section (and notes in other sections that name Facet J); the card, including its S2 section, is the authority. Read status in (b) updated for every source below. + +- `tyler1988what`: weights on criteria vary by situation, "no universally fair procedures" (p. 131); "notice" was not measured; the Q2 list is not one yardstick for all three settings (the Tyler bearing, Q4 bullet and a note under the comparison table). Read in full (card). +- `galanter1974why`: nine repeat-player advantages (pp. 98-103); clients in continuing relationships "are not pure OSs" (n. 4, p. 97), so the driver/seller mapping carries that qualification; the p. 111 arbitration quotation concerns repeat players of comparable size. Read in part (card). +- `globalfxc2024code`: Principle 40 (substantial changes agreed before any transaction, p. 42), Principle 8 (p. 12), Principle 7 (channel for concerns, p. 11), Principles 36 and 39 (information to clients on request, pp. 40, 42); comparison-table cells for FX notice, reasons and complaint/appeal corrected, "voluntary" kept; "no notice of change" removed from Q1; header, Q5 item 3 and inference 3 reworded. Read in full (card). +- `klonick2018new`: "impressively intricate and responsive" (p. 1664); exit treated sceptically (pp. 1629-30, 1668), so it is no longer listed among her "pluralistic" forces (those are on p. 1664, Part IV); "draconian" is not her word. Read in full (card). +- `oversight2022q1`: one of the two decisions was a Meta referral; Meta reversed itself in 14 of 20 shortlisted cases before the Board ruled (p. 9); "by Q4 2021" in Q6 changed to Q1 2022 (the Q4 2021 report was not read). Read in part (card). +- `eu2022dsa`: Arts. 53-54 (regulator complaint, damages) and the 6% fine ceiling (Arts. 52(3), 74) added to the entry, Q3 and Q5 item 3. Read in part (card; official text). +- `eu2019p2b`: official EU text contains Art. 5(6)-(7); 5(6) is the no-algorithm-disclosure clause; pinpoints for Arts. 3, 4, 11, 12 confirmed. Read in part (card). +- `seattle2024deactivation`: the company decides the challenge itself within 14 days; no independent reviewer (entry, Q2 table, Q5 item 3, Q6 table, comparison table). Read in full (card). +- `us2073usc`: subsection (e), failure to comply does not invalidate a rule; notice of meetings (c)(2) but no comment period (entry, Q1, Q6 row). Read in full (card). +- `tumey1927ohio`: secret-service fund, $696.35, de minimis limit, dual role, structural test, matters left to legislatures (¶¶ per the card's count). Read in full (card). `caperton2009massey`: read in full (card). +- `mathews1976eldridge`: pin cite pp. 332-335 for the three factors (syllabus). Read in part (card). +- `suzor2016responsibilities`, `suzor2018digital`: the 15 against 14 platforms is a change of sample (VK excluded, note 3, p. 9; data from June 2015); the 2018 body also reports "broad, unfettered discretion" (p. 6), so the (d) note that only the 2016 paper does is removed; "perform poorly" was not a quotation. Both read in full (card). +- `jhaver2019suspect`: readers of the rules rated removals less fair (β = -.314, Table 5), against the abstract; 36.9% of 849 coded answers, not of 907 respondents. Read in full (card). +- `nagarajrao2026fareshare`: 122 users and 178 accounts (143 synced); the union had already won over 1,400 reactivations and about $3.5 million before the tool, so "need worker-side tools to be used" is withdrawn. Read in full (card, arXiv preprint). +- `fsb2014foreign`: working group membership was central banks and regulators, with market participants consulted; the FSB work was independent of the conduct investigations (p. 1). Read in part (card). +- `oomen2016last`: "contradicts" replaced by "qualifies": equal effective spreads, but "materially different execution risk profiles" (Figure 8, §4); effective cost "need not be affected". Read in full (card). +- `evans2018forex`: read in part; content added; the FSB and Evans volatility measures differ and do not strictly conflict. `douek2023meta`: read in full; journal print is 2024, 37(2), 373-445. +- `nydfs2015barclays`: the "indiscriminately" and "lack of transparency" statements are recitals in a negotiated consent order, not findings; signing day illegible (Q6 row, entry, Q3). Read in full (card). +- `colvin2011empirical`: repeat-player numbers added (16.9% against 31.6%; pairing 12.0% against 23.4%), the arbitration-litigation comparison is not matched, and "97% of fees" corrected to employers paying all fees beyond a small filing fee in 97% of cases. Read in full (card). +- Read-status updates with no change to content: `apple2026guidelines` (regional exceptions and delegation of duties added), `wong2022meta`, `palka2025foreign` (EU-only limits), `rosenblat2016algorithmic`, `cftc2017fxcm` (70% rebate, about $77 million), `fca2014fines`, `fsb2015progress`, `nydfs2015barclayspr`. + +Synthesis verdicts changed: + +- Q5 "draconian" verdict (claim 1 and Q5 (i), (iii)): old, "Qualified. True of terms-of-service discretion (Suzor) and of 2017 appeals (Klonick)". New: still qualified, but "draconian" in the sense of severity is not supported by the sources I opened; Suzor supports near-total discretion (both papers), Klonick supports thin appeals and lack of accountability and calls the system "impressively intricate and responsive". "Draconian" is the author's word (D8), not either source's. +- Q6 "direction of change" verdict (closing paragraph and the comparison-table cell): old, "In platforms it is bending toward users through statute and local law". New: same direction, with limits stated: the statutory rights stop short of independent first-instance review (Seattle: the company decides the challenge within 14 days; DSA out-of-court bodies cannot bind), the Board reaches few cases and Meta controls its precedent; for FX, direction is shown only as regulator action after scandal, while the current Code already asks more than DSA Art. 14(2) on change of terms (Principle 40), with earlier editions unread so no direction is shown for the Code. +- Q2 comparison grid: old, the Q2 list as a yardstick for all three settings; new, a descriptive grid only (Tyler). +- Inference 3 and the opening summary ("the Code asks the same of FX dealers"): old, EU platforms are bound by more written notice and reasons duties than FX dealers; new, bound by law where the Code asks only by voluntary principle, but not uniformly: Principle 40 asks more than DSA Art. 14(2) on change of terms. +- Claim 4 (P10) and Q1 FX: "Notice of change to clients: none required by the Code" withdrawn. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_K_training_and_cause.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_K_training_and_cause.bib new file mode 100644 index 00000000..d63616d0 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_K_training_and_cause.bib @@ -0,0 +1,404 @@ +% Facet K, 2026-10-13. Agent-drafted; not reviewed by the author. +% Keys marked "(reused)" already exist in ../references.bib. Archive.org items are lending or restricted scans read as snippets only. + +@book{quintilian1920institutio, + author = {Quintilian}, + title = {The Orator's Education ({Institutio} oratoria), Book 1}, + editor = {Butler, H. E.}, + publisher = {Loeb Classical Library, vol. 1 (London: Heinemann; New York: Putnam)}, + address = {London}, + year = {1920}, + url = {https://penelope.uchicago.edu/Thayer/E/Roman/Texts/Quintilian/Institutio_Oratoria/1A*.html}, + note = {Read in part (card: library/quintilian1920institutio.md): 1.1, 1.7, 1.5.6, 1.8.1--5 read closely; Loeb pp. 35, 37, 139, 145, 147; Butler translation (Penelope transcription). Latin not checked.} +} + +@book{dionysius1910composition, + author = {Dionysius of Halicarnassus}, + title = {On Literary Composition, being the Greek text of the {De compositione verborum}}, + editor = {Roberts, W. Rhys}, + publisher = {Macmillan}, + address = {London}, + year = {1910}, + url = {https://archive.org/details/cu31924026465165}, + note = {Read in part (card: library/dionysius1910composition.md): ch. XXV in translation, pp. 265--270, with Roberts note; chapter runs pp. 251--270, reading passage p. 269, tablet story pp. 265--267. Greek not checked. Usher Loeb supersedes for quotation.} +} + +@article{thrax1874grammar, + author = {Dionysius Thrax (attributed)}, + title = {The Grammar of {Dionysios Thrax}}, + journal = {Journal of Speculative Philosophy}, + volume = {8}, + year = {1874}, + note = {Translated by Thomas Davidson. Archive.org item grammarofdionysi00dionuoft. Read in part (card: library/thrax1874grammar.md): headnote and sections 1--14 and the closing section; sections 15--24 skimmed. Authorship of the Techne is disputed; that literature not checked.} +} + +@book{cribiore1996writing, + author = {Cribiore, Raffaella}, + title = {Writing, Teachers, and Students in {Graeco-Roman} {Egypt}}, + series = {American Studies in Papyrology}, + number = {36}, + publisher = {Scholars Press}, + address = {Atlanta}, + year = {1996}, + doi = {10.3998/mpub.9749698}, + note = {Crossref title "Writing, Teachers and Students in Graeco-Roman Egypt", dated 1995 online; closed access (Unpaywall). Archive.org writingteacherss0000crib. Read status: snippet only; printed pages not available.} +} + +@book{cribiore2001gymnastics, + author = {Cribiore, Raffaella}, + title = {Gymnastics of the Mind: {Greek} Education in {Hellenistic} and {Roman} {Egypt}}, + publisher = {Princeton University Press}, + address = {Princeton}, + year = {2001}, + doi = {10.1515/9781400844418}, + note = {Crossref title "Gymnastics of the Mind". Archive.org gymnasticsofmind0000crib. Read status: snippet only; printed pages not available.} +} + +@book{marrou1956history, + author = {Marrou, Henri-Ir{\'e}n{\'e}e}, + title = {A History of Education in Antiquity}, + publisher = {Sheed and Ward}, + address = {New York}, + year = {1956}, + url = {https://archive.org/details/historyofeducati0000marr}, + note = {Tr. G. Lamb. Scan is the 1956 Sheed and Ward edition (archive.org metadata). Read status: snippet only; printed page not available.} +} + +@book{bonner1977education, + author = {Bonner, Stanley F.}, + title = {Education in Ancient {Rome}: From the Elder {Cato} to the Younger {Pliny}}, + publisher = {Methuen}, + address = {London}, + year = {1977}, + url = {https://archive.org/details/educationinancie0000bonn}, + note = {Read status: snippet only; printed page not available.} +} + +@book{morgan1998literate, + author = {Morgan, Teresa}, + title = {Literate Education in the {Hellenistic} and {Roman} Worlds}, + publisher = {Cambridge University Press}, + address = {Cambridge}, + year = {1998}, + url = {https://archive.org/details/literateeducatio0000morg}, + note = {No Crossref hit. Read status: snippet only; printed page not available.} +} + +@book{johnson2010readers, + author = {Johnson, William A.}, + title = {Readers and Reading Culture in the {High} {Roman} {Empire}: A Study of Elite Communities}, + publisher = {Oxford University Press}, + address = {Oxford}, + year = {2010}, + doi = {10.1093/acprof:oso/9780195176407.001.0001}, + note = {(reused) Crossref title matches. This facet: snippet only (archive.org readersreadingcu0000john); printed pages not available.} +} + +@book{reynolds1996medieval, + author = {Reynolds, Suzanne}, + title = {Medieval Reading: Grammar, Rhetoric and the Classical Text}, + series = {Cambridge Studies in Medieval Literature}, + publisher = {Cambridge University Press}, + address = {Cambridge}, + year = {1996}, + doi = {10.1017/cbo9780511470356}, + note = {Crossref title "Medieval Reading". Archive.org medievalreadingg0000reyn. Read status: snippet; whole-book zero hits for word separation, scriptio continua, Saenger, silent reading.} +} + +@book{orme2001medieval, + author = {Orme, Nicholas}, + title = {Medieval Children}, + publisher = {Yale University Press}, + address = {New Haven}, + year = {2001}, + url = {https://archive.org/details/medievalchildren00orme}, + note = {Read status: snippet; whole-book zero hits for word separation, scriptio continua, Saenger.} +} + +@book{orme2006medievalschools, + author = {Orme, Nicholas}, + title = {Medieval Schools: From {Roman} {Britain} to {Renaissance} {England}}, + publisher = {Yale University Press}, + address = {New Haven}, + year = {2006}, + url = {https://archive.org/details/medievalschoolsf0000orme}, + note = {Read status: snippet; whole-book zero hits for word separation, scriptio continua, Saenger.} +} + +@book{black2001humanism, + author = {Black, Robert}, + title = {Humanism and Education in Medieval and {Renaissance} {Italy}: Tradition and Innovation in {Latin} Schools from the Twelfth to the Fifteenth Century}, + publisher = {Cambridge University Press}, + address = {Cambridge}, + year = {2001}, + url = {https://archive.org/details/humanismeducatio0000blac}, + note = {Read status: snippet; whole-book zero hits for word separation, scriptio continua, Saenger.} +} + +@book{grendler1989schooling, + author = {Grendler, Paul F.}, + title = {Schooling in {Renaissance} {Italy}: Literacy and Learning, 1300--1600}, + publisher = {Johns Hopkins University Press}, + address = {Baltimore}, + year = {1989}, + url = {https://archive.org/details/schoolinginrenai0000gren}, + note = {Read status: snippet; whole-book zero hits for word separation, scriptio continua, Saenger.} +} + +@book{riche1976education, + author = {Rich{\'e}, Pierre}, + title = {Education and Culture in the Barbarian {West}, Sixth through Eighth Centuries}, + publisher = {University of South Carolina Press}, + address = {Columbia, SC}, + year = {1976}, + url = {https://archive.org/details/educationculture0000rich}, + note = {Tr. J. J. Contreni (publisher confirmed by archive.org metadata: University of South Carolina Press, Columbia, 1976; translator from memory, verify). Read status: snippet; zero hits for word separation, scriptio continua.} +} + +@book{saenger1997space, + author = {Saenger, Paul}, + title = {Space Between Words: The Origins of Silent Reading}, + publisher = {Stanford University Press}, + address = {Stanford}, + year = {1997}, + doi = {10.1515/9781503619081}, + note = {(reused) This facet: snippets only (archive.org spacebetweenword0000saen); pages from facets A and C where cited. DOI is the 2022 De Gruyter reissue of the Stanford 1997 book; page numbers checked against the 1997 print via Google Books by the auditor.} +} + +@article{saenger1982silent, + author = {Saenger, Paul}, + title = {Silent Reading: Its Impact on Late Medieval Script and Society}, + journal = {Viator}, + volume = {13}, + pages = {367--414}, + year = {1982}, + doi = {10.1484/J.VIATOR.2.301476}, + note = {(reused) This facet: snippets from the restricted scan viator_1982_13 (whole volume). Per-snippet pages not available.} +} + +@incollection{saenger1999history, + author = {Saenger, Paul}, + title = {The History of Reading}, + booktitle = {Literacy: An International Handbook}, + editor = {Wagner, Daniel A. and Venezky, Richard L. and Street, Brian V.}, + publisher = {Westview Press}, + address = {Boulder, CO}, + year = {1999}, + url = {https://archive.org/details/literacyinternat0000unse}, + note = {Ch. 2; start page 11 from an OCR contents snippet, end page not seen. Publisher confirmed (archive.org): Westview Press, Boulder, CO, 1999. Chapter opening runs pp. 11--12 (page break 13 follows the scope sentences); end page not seen. Read status: snippet only.} +} + +@book{parkes1992pause, + author = {Parkes, M. B.}, + title = {Pause and Effect: An Introduction to the History of Punctuation in the {West}}, + publisher = {Scolar Press}, + address = {Aldershot}, + year = {1992}, + url = {https://archive.org/details/pauseeffectintro0000park}, + note = {(reused) This facet: snippets from the 1993 University of California Press scan. Zero hits for Saenger.} +} + +@book{cavallo1999history, + editor = {Cavallo, Guglielmo and Chartier, Roger}, + title = {A History of Reading in the {West}}, + publisher = {University of Massachusetts Press}, + address = {Amherst}, + year = {1999}, + url = {https://archive.org/details/historyofreading0000unse}, + note = {(reused key; facet A) Translated by Lydia G. Cochrane. Snippets from the editors' opening survey; which editor wrote the line-and-page passage and its page: to confirm. Quotations from chapter 1 must cite the chapter author (Svenbro, svenbro1999archaic).} +} + +@incollection{svenbro1999archaic, + author = {Svenbro, Jesper}, + title = {Archaic and Classical {Greece}: The Invention of Silent Reading}, + booktitle = {A History of Reading in the {West}}, + editor = {Cavallo, Guglielmo and Chartier, Roger}, + publisher = {University of Massachusetts Press}, + address = {Amherst}, + year = {1999}, + url = {https://archive.org/details/historyofreading0000unse}, + note = {Ch. 1. First running head seen is p. 38; the quoted passages run pp. 51--52; end page not verified. Added after AUDIT_06.} +} + +@incollection{rouse1982statim, + author = {Rouse, Richard H. and Rouse, Mary A.}, + title = {{Statim invenire}: Schools, Preachers, and New Attitudes to the Page}, + booktitle = {Renaissance and Renewal in the Twelfth Century}, + editor = {Benson, Robert L. and Constable, Giles}, + publisher = {Harvard University Press}, + address = {Cambridge, MA}, + year = {1982}, + pages = {201--225}, + url = {https://archive.org/details/renaissancerenew0000unse}, + note = {(reused key; the earlier note said not accessible) Title and start page 201 confirmed from the volume's contents in the scan; end page from running heads. Snippets only. Whole-volume zero hits for word separation, Saenger, silent reading.} +} + +@book{illich1993vineyard, + author = {Illich, Ivan}, + title = {In the Vineyard of the Text: A Commentary to {Hugh's} {Didascalicon}}, + publisher = {University of Chicago Press}, + address = {Chicago}, + year = {1993}, + note = {(reused) Not accessible. Known through Saenger's review.} +} + +@article{saenger1995illich, + author = {Saenger, Paul}, + title = {{In the Vineyard of the Text: A Commentary to Hugh's ``Didascalicon''}. {Ivan Illich} (review)}, + journal = {The Library Quarterly}, + volume = {65}, + number = {2}, + pages = {244--246}, + year = {1995}, + doi = {10.1086/602782}, + note = {(reused) Crossref title and pages match. This facet: snippets from scan sim_library-quarterly_1995-04_65_2.} +} + +@book{mak2011page, + author = {Mak, Bonnie}, + title = {How the Page Matters}, + publisher = {University of Toronto Press}, + year = {2011}, + note = {(reused) Not accessible this facet.} +} + +@book{chartier1994order, + author = {Chartier, Roger}, + title = {The Order of Books: Readers, Authors, and Libraries in {Europe} between the Fourteenth and Eighteenth Centuries}, + publisher = {Stanford University Press}, + year = {1994}, + url = {https://archive.org/details/orderofbooksread0000char}, + note = {Snippet search only; zero hits for word separation; cites Saenger in notes.} +} + +@book{carruthers2008book, + author = {Carruthers, Mary}, + title = {The Book of Memory: A Study of Memory in Medieval Culture}, + edition = {2}, + publisher = {Cambridge University Press}, + year = {2008}, + url = {https://archive.org/details/bookofmemorystud0000carr}, + note = {Snippet search only; zero hits for word separation, scriptio continua; one note cites Saenger and Parkes on word spacing.} +} + +@article{hanna1999saenger, + author = {Hanna, Ralph}, + title = {Review of {Paul Saenger}, Space between Words}, + journal = {Speculum}, + volume = {74}, + number = {3}, + pages = {830--831}, + year = {1999}, + doi = {10.2307/2886848}, + note = {(reused) Crossref title and pages match. This facet: snippets from scan sim_speculum_1999-07_74_3.} +} + +@article{richter2001saenger, + author = {Richter, Michael}, + title = {Review of {Paul Saenger}, Space Between Words}, + journal = {The American Historical Review}, + volume = {106}, + number = {2}, + pages = {627--628}, + year = {2001}, + doi = {10.1086/ahr/106.2.627}, + note = {(reused) Crossref match. This facet: snippets from scan sim_american-historical-review_2001-04_106_2.} +} + +@article{mulchahey2001saenger, + author = {Mulchahey, M. Mich{\`e}le}, + title = {Review of {Paul Saenger}, Space Between Words}, + journal = {Canadian Journal of History}, + volume = {36}, + number = {2}, + pages = {323--325}, + year = {2001}, + doi = {10.3138/cjh.36.2.323}, + note = {(reused) Crossref match. This facet: snippets from scan sim_canadian-journal-of-history_2001-08_36_2.} +} + +@article{walker2000saenger, + author = {Walker, Thomas D.}, + title = {Review of {Paul Saenger}, Space between Words}, + journal = {The Library Quarterly}, + volume = {70}, + number = {1}, + pages = {166--167}, + year = {2000}, + doi = {10.1086/603182}, + note = {(reused) Crossref match. This facet: snippets from scan sim_library-quarterly_2000-01_70_1.} +} + +@article{aronoff2002saenger, + author = {Aronoff, Mark}, + title = {Review of {Paul Saenger}, Space between words}, + journal = {Language in Society}, + volume = {31}, + number = {4}, + pages = {624--628}, + year = {2002}, + doi = {10.1017/s0047404502254056}, + note = {(reused) Facet B had the abstract only; this facet: snippets of the review from scan sim_language-in-society_2002-09_31_4.} +} + +@article{bauml2002saenger, + author = {B{\"a}uml, Franz H.}, + title = {Review of Saenger (1997): Space between words}, + journal = {Written Language \& Literacy}, + volume = {5}, + number = {1}, + pages = {131--134}, + year = {2002}, + doi = {10.1075/wll.5.1.08bau}, + note = {(reused) Crossref match. Not accessible.} +} + +@article{raritan2000review, + author = {Anonymous (reviewer not identified)}, + title = {Review essay on {Saenger}, {Manguel} and {Johns}}, + journal = {Raritan}, + volume = {20}, + number = {1}, + year = {2000}, + note = {Scan sim_raritan_summer-2000_20_1; author, title and pages not found. Snippets only. Do not cite until the author is identified.} +} + +@article{nq2000saenger, + author = {Anonymous (reviewer not identified)}, + title = {Review of {Paul Saenger}, Space Between Words}, + journal = {Notes and Queries}, + volume = {47}, + number = {1}, + year = {2000}, + note = {Scan sim_notes-and-queries_2000-03_47_1; reviewer and pages not found. Snippets only.} +} + +@book{blake2002literacy, + author = {Blake, Brett Elizabeth}, + title = {Literacy and Learning: A Reference Handbook}, + publisher = {ABC-CLIO}, + address = {Santa Barbara}, + year = {2002}, + url = {https://archive.org/details/literacylearning0000blak}, + note = {Snippets only. Secondary handbook for teachers; repeats Saenger.} +} + +@book{baron2000alphabet, + author = {Baron, Naomi S.}, + title = {Alphabet to Email: How Written {English} Evolved and Where It's Heading}, + publisher = {Routledge}, + address = {London}, + year = {2000}, + url = {https://archive.org/details/alphabettoemailh00baro_0}, + note = {Snippets only.} +} + +@book{lyons2010history, + author = {Lyons, Martyn}, + title = {A History of Reading and Writing in the Western World}, + publisher = {Palgrave Macmillan}, + address = {Basingstoke}, + year = {2010}, + url = {https://archive.org/details/historyofreading0000lyon}, + note = {Snippets only; secondary summary of Saenger.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_K_training_and_cause.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_K_training_and_cause.md new file mode 100644 index 00000000..2ebdcee8 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_K_training_and_cause.md @@ -0,0 +1,418 @@ +# K — Training before and after word separation, and every causal claim made from separation + +*Audited 2026-10-13 (AUDIT_06.md); WRONG, OVERSTATED and STITCHED rows corrected in place.* + +*Facet K, 2026-10-13. Agent-drafted. The author has not reviewed it, and none of the text is the author's.* + +**What this facet covers.** (1) How readers were taught before word separation (Quintilian, Dionysius of Halicarnassus, the grammarians, school papyri; Cribiore, Marrou, Bonner, Morgan, Johnson). (2) How they were taught after it (Reynolds, Orme, Black, Grendler, Riché, Saenger on novices). (3) Whether any scholar claims, tests or denies that separation shortened, eased or removed the training. (4) A catalogue of every causal claim made from separation to anything else, with its reception. (5) Reverse causation: readers who produced separation. Facets A, B and C hold Saenger's record, the dispute over silent reading, and the historians of literacy; I do not repeat them and cite them by name where I build on them. + +**The question the synthesis answers.** Is the absence of a "separation shortened training" claim real or an artefact of where facets B and C looked, and what do scholars actually claim from separation? That tells the author what the sentence at `draft.md:263` can rest on (rulings D1, D2). + +**Legend for read status.** +- *Read in full*: whole text read from an open copy (Quintilian 1.1 and 1.8.1–2 in Butler's Loeb translation via Penelope; Roberts's 1910 Dionysius; Davidson's 1874 Thrax, §§1–2). Updated 2026-10-13 from the full-text cards: Quintilian is now *read in part (card)*: 1.1, 1.7, 1.5.6 and 1.8.1–5 read closely; Dionysius is *read in part (card)*: ch. XXV English, pp. 265–270, read; the Demosthenes analysis, pp. 252–264, skimmed; Thrax is *read in part (card)*: §§1–14 and the closing section read, §§15–24 skimmed. +- *Snippet*: read as OCR snippets of about one line (roughly 20 words) returned by the archive.org full-text search (`be-api.us.archive.org/fts/v1/search`) over the borrowable scan named in the entry. I quote only text a snippet shows; where I join two overlapping snippets I say so. The API gives no printed page, so I write "page not available" and give the chapter or section where the snippet shows one. I did not borrow any book. Quotations are normalised for OCR hyphenation and spacing only. +- *Abstract only*, *not accessible*: as in facets A and B. +- *Zero hits (whole-book)*: the API reports a total of 0 for a phrase across the whole scanned book. This is a reliable null for the OCR text (a missed OCR word is possible), so I report it as "no hit for X in the lending scan of Y". + +**A note on tool use.** Google Books search-inside and the Google Books API returned zero results or a daily-quota error throughout, so every page locator for a modern book below that is not given comes from facets A–C, or is marked "page not available". I used the Unpaywall API once with an e-mail address that is not the one named in the plan (it is the address in my session context); I did not use it again. + +--- + +## (a) Search log + +| # | Source | Exact query | Hits | Notes | +| --- | --- | --- | --- | --- | +| 1 | Penelope (LacusCurtius), Butler's Quintilian | `Thayer/E/Roman/Texts/Quintilian/Institutio_Oratoria/1A*.html`, `1B*`, `1C*` | 3 pages fetched | Read 1.1.15–37 and 1.8.1–2. Perseus's own text URL returned a blank record; Penelope worked | +| 2 | archive.org item `cu31924026465165` | djvu text of Roberts 1910 | full text | Ch. XXV read in the translation; Greek not checked | +| 3 | archive.org item `grammarofdionysi00dionuoft` | djvu text | full text | Davidson 1874, §§1–2 | +| 4 | archive.org FTS, item `gymnasticsofmind0000crib` (Cribiore 2001) | `"word division"`; `"divided into syllables"`; `scriptio continua`; `"word separation"`; `syllabic`; `"bars or dots"`; `"uncertain readers"`; `"lectional signs"`; `Quintilian reading`; `Dionysius Halicarnassus` | 0; 1 each (multi-snippet); `"word division"` 0 | Item is access-restricted to guests, FTS still returns snippets | +| 5 | same, Cribiore 1996 `writingteacherss0000crib` | `"word division"`; `"syllabic division"`; `"divided into syllables"`; `scriptio continua`; `"separation of words"`; ten stitching queries | all > 0 | Section "Syllabic Division, Word Separation, and Verse Distinction" located | +| 6 | same, Marrou `historyofeducati0000marr` | `"word division"`; `"scriptio continua"`; `"not only was there no punctuation"`; `"separated as in our modern primers"` | 0; 1; 1; 1 | Three overlapping snippets stitched | +| 7 | same, Bonner `educationinancie0000bonn` | `"word division"`; `"scriptio continua"`; `praelectio`; `"silent reading"`; `"learning to read"` | 1; 0; 1; 0; 1 | Only a plate caption mentions syllabic word-division | +| 8 | same, Morgan `literateeducatio0000morg` | `"word division"`; `"scriptio continua"`; `"divided into syllables"`; `syllables` | 0; 0; 1; 1 | | +| 9 | same, Johnson 2010 `readersreadingcu0000john` | `"word division"`; `"scriptio continua"`; `Saenger`; `"years of training"`; `Quintilian`; `"emerging readers"` | 1; 1; 1; 0; 1; 1 | Located the "emerging readers" sentence | +| 10 | same, Reynolds 1996 `medievalreadingg0000reyn` | `"word division"`; `"word separation"`; `"scriptio continua"`; `"separation of words"`; `"spacing"`; `Saenger`; `"silent reading"`; `"space between words"` | all 0 | Whole-book null on every separation term | +| 11 | same, Orme 2001 `medievalchildren00orme` and Orme 2006 `medievalschoolsf0000orme` | the same eight queries | all 0 in both | Whole-book null in both books | +| 12 | same, Black 2001 `humanismeducatio0000blac` | same eight | `"word separation"`, `"word division"`, `"scriptio continua"`, `Saenger`: 0; `"silent reading"`: 1; `"spacing"`: 1 (line spacing) | | +| 13 | same, Grendler 1989 `schoolinginrenai0000gren` | same eight | all 0 except `"spacing"` (1, about typography) | | +| 14 | same, Riché 1976 `educationculture0000rich` | `"word separation"`; `"scriptio continua"`; `"spaces between"`; `Saenger` | all 0 | | +| 15 | same, Saenger 1997 `spacebetweenword0000saen` | about 35 phrase queries (novice, syllabary, shortcuts, heresy, erotic, autograph, dictation, scholastic, private devotion, literacy diffusion) | each 1 (multi-snippet) | Snippets only; pages from facets A and C where they gave one | +| 16 | same, Saenger 1982 in `viator_1982_13` (whole volume) | `"word separation" literacy`; `"word separation" beginners`; `"word separation" "foreign language"` | 1; 0; 0 | | +| 17 | same, Parkes 1993 `pauseeffectintro0000park` | 10 queries (foreign language, grammar of legibility, silent reader, ordinatio, learners, Saenger) | `"learners"` 0, `Saenger` 0 | The book does not cite Saenger | +| 18 | same, Cavallo and Chartier `historyofreading0000unse` | `"word separation"`; `"silent reading"`; `"spaces between words"`; `"easier to read"` + 12 stitching queries | all > 0 | Attribution of the passages to the Introduction is mine; see K13 | +| 19 | same, Rouse and Rouse in `renaissancerenew0000unse`; Rouse 1979 `preachersflorile0000rous` | `"word separation"`; `"silent reading"`; `Saenger`; `"scriptio continua"`; `"spacing"` | all 0 | | +| 20 | same, Chartier `orderofbooksread0000char`; Carruthers `bookofmemorystud0000carr`; Stock `listeningfortext00stoc` | `"word separation"`; `Saenger`; `"silent reading"` | 0 for `"word separation"` in all three | Each cites Saenger in notes only | +| 21 | same, reviews in `sim_` periodical scans | Speculum 74(3); AHR 106(2); Canadian J. of History 36(2); Library Quarterly 70(1) and 65(2); Language in Society 31(4); Raritan 20(1); Notes and Queries 47(1) | all returned the review | Snippets only | +| 22 | same, Wagner, Venezky and Street, `literacyinternat0000unse` | `"The History of Reading" Saenger`; `"word separation"`; literacy-diffusion phrases | found Saenger's own chapter | | +| 23 | same, Blake 2002 `literacylearning0000blak`; Baron 2000 `alphabettoemailh00baro_0` | `"word separation"`; `Saenger`; `"spaces between words"` | > 0 | Secondary uses only | +| 24 | archive.org FTS, global (unscoped) | `"word separation" "learning to read"`; `"word separation" "easier to learn to read"`; `"word separation" "spread of literacy"`; `"word separation" "wider literacy"`; `"separation of words" "mass literacy"`; `"Illich" "word separation" Hugh`; `"word separation" "reference reading"`; `"word separation" "private devotion" Saenger` | 623; 1; 175; 10; 91; 70; 54; 59 | Read the top 12 hits of each by title and snippet. None argues that separation shortened teaching. Hits were Saenger, Cavallo and Chartier, Lyons 2010, Blake 2002, Baron 2000, handbooks and school-teaching texts | +| 25 | archive.org `advancedsearch` | titles for about 30 books | identifiers found for all except Mak 2011, Illich 1993 and Bäuml | Illich and Mak: no FTS route | +| 26 | Crossref | 8 DOIs (Hanna, Richter, Walker, Mulchahey, Bäuml, Aronoff, Saenger's Illich review, Gilliard) | 8/8 resolved, titles match | | +| 27 | Crossref bibliographic | Cribiore 1996 and 2001, Reynolds, Johnson | DOIs found: 10.3998/mpub.9749698 (Cribiore 1996, closed on Unpaywall); 10.1515/9781400844418 (Cribiore 2001); 10.1017/cbo9780511470356 (Reynolds); 10.1093/acprof:oso/9780195176407.001.0001 (Johnson) | Morgan 1998: no Crossref hit | +| 28 | WebSearch (2) | Irish scribes and foreign-language learners; Saenger reviews and critique of the causal link | 20 | Popular pages and blogs only; nothing used | +| 29 | Google Books API; `books.google.com/books?jscmd=SearchWithinVolume2` | Saenger `w3vZaFoaa3EC`: `syllabary`, `novice` | 429 quota; `number_of_results: 0` | Not usable this session | +| 30 | Scholar Gateway | not tried | n/a | Failed in every earlier run, per the plan | +| 31 | Unaccessed | JSTOR, Brepols, Cambridge Core for Hanna, Bäuml, Gilliard, Parkes 1976, Rouse and Rouse offprint | not opened | Section (d) | + +--- + +## (b) Entries + +### Part 1. Ancient training + +#### K1. Quintilian, *Institutio oratoria* 1.1 and 1.8 + +**quintilian1920institutio** — Quintilian. *The Orator's Education* [*Institutio oratoria*], Book 1. Tr. H. E. Butler (Loeb), Loeb Classical Library, vol. 1 (London: Heinemann; New York: Putnam), 1920. Penelope/LacusCurtius transcription, `penelope.uchicago.edu/Thayer/E/Roman/Texts/Quintilian/Institutio_Oratoria/1A*.html` and `1C*.html`. The Loeb page numbers below are the ones the transcription marks (odd pages carry the English). +- *Verification:* the transcription carries section numbers; I did not check the Latin. +- *Read status:* **read in part (card: library/quintilian1920institutio.md)**: Preface, 1.1 in full (1.1.15–37), the opening of 1.2, 1.5.6, 1.7 in full and 1.8.1–5 read closely; the rest of Book 1 not read. +- *What it says (Butler's words):* + - 1.1.15: "Some hold that boys should not be taught to read till they are seven years old, that being the earliest age at which they can derive profit from instruction and endure the strain of learning." Quintilian argues against waiting. + - 1.1.24–26 on method: he disapproves of teaching "the names and order of the letters before their shapes"; "The method, however, to which we have objected in teaching the alphabet, is unobjectionable when applied to syllables." + - 1.1.30, c. Loeb p. 35: "As regards syllables, no short cut is possible: they must all be learnt, and there is no good in putting off learning the most difficult; this is the general practice, but the sole result is bad spelling." + - 1.1.32–33, c. p. 37: "You will hardly believe how much reading is delayed by undue haste." "Reading must therefore first be sure, then connected, while it must be kept slow for a considerable time, until practice brings speed unaccompanied by error." + - 1.1.34, p. 37 (the sentence Heilmann reads as about learners): "For to look to the right, which is regularly taught, and to look ahead depends not so much on precept as on practice; since it is necessary to keep the eyes on what follows while reading out what precedes, with the resulting difficulty that the attention of the mind must be divided, the eyes and voice being differently engaged." + - 1.8.1–2, Loeb p. 147: "Reading remains for consideration. In this connexion there is much that can only be taught in actual practice, as for instance when the boy should take breath, at what point he should introduce a pause into a line, where the sense ends or begins, when the voice should be raised or lowered, and when he should increase or slacken speed… In this portion of my work I will give but one golden rule: to do all these things, he must understand what he reads." + - 1.7.17, p. 139: Quintilian rejects an old convention of adding an *e* before *i* in endings because "such a practice will be an actual hindrance to those who are learning to read"; so he judges a spelling convention by whether it hinders learners. His test, "within the limits prescribed by usage," is that "words should be spelt as they are pronounced" (1.7.30–31, p. 145). Chapter 1.7 is about letters, doubling, diphthongs and abbreviations, not the spacing of words. +- *What it does not say:* it contains no word about word division, spacing or the look of the text (1.7 included). The difficulty it names is divided attention between eye and voice, and the reader's job of finding where to breathe and pause. It places the child's slowness in the alphabet-syllable-word sequence, not in the layout. Quintilian gives no length of training in years. +- *Bearing:* **complicates** H-cost. The teaching of reading in Quintilian is a slow, staged drill in syllables, and Quintilian says phrasing and pause "can only be taught in actual practice" (1.8.1; the word *praelectio* is Bonner's, K7, not Quintilian's 1.8.1-2). That is training the reader must do whatever the page looks like. It also shows the aid that exists later (marks for pauses) was a teacher's job in his world, not the page's. + +#### K2. Dionysius of Halicarnassus, *De compositione verborum*, ch. 25 + +**dionysius1910composition** — Dionysius of Halicarnassus. *On Literary Composition, being the Greek text of the De compositione verborum*, ed. and tr. W. Rhys Roberts. London: Macmillan, 1910. Item `cu31924026465165`, archive.org. Chapter XXV, "How prose can resemble verse". Chapter XXV runs from p. 251 to p. 270 in Roberts (card); the Plato tablet story is on pp. 265–267 and the reading passage is on p. 269. +- *Verification:* archive.org metadata; title and editor on the title page of the djvu text. +- *Read status:* **read in part (card: library/dionysius1910composition.md)**: the English of ch. XXV, pp. 265–270, read through to the end of the reading passage and its notes, and the first page of the chapter; pp. 252–264 skimmed. I did not read the Greek. +- *The passage (Roberts's translation, ch. XXV, p. 269):* "When we are taught to read, first we learn off the names of the letters, then their forms and their values, then in due course syllables and their modifications, and finally words and their properties, viz. lengthenings and shortenings, accents, and the like. After acquiring the knowledge of these things, we begin to write and read, syllable by syllable and slowly at first. And when the lapse of a considerable time has implanted the forms of words firmly in our minds, then we deal with them without the least difficulty, and whenever any book is placed in our hands we go through it without stumbling, and with incredible facility and speed." +- *Roberts's note on the passage:* he cites Wilkins's remark that "Some have supposed that Dionysius here describes the method of acquiring the power of reading, not by learning the names of the letters first, but by learning their powers", then rejects it ("hardly consistent with his language") and cites Athenaeus's chant of syllables, and an Attic terracotta school plate "doubtless intended for use in schools, contains a number of syllables". +- *What it does and does not say:* Dionysius uses reading acquisition as an analogy for the trained writer's ease with word-order; he says nothing about word division. What it gives the paper: (i) the Greek sequence letters, forms, values, syllables, words; (ii) "a considerable time" is the cost, and what it buys is fluent, effortless reading of whatever book is handed over. That last point is the same one the physiological critics of Saenger make (facet B, Heilmann): trained readers of unspaced text read fluently. +- *Bearing:* **complicates** the author's sentence at one remove: the ancient sources locate the burden in a long drill that ends in fluency, not in a defect of the page. Not evidence for or against separation lowering the drill. + +#### K3. Dionysius Thrax (attributed), *Technē grammatikē*, §§1–2 + +**thrax1874grammar** — *The Grammar of Dionysios Thrax*, tr. Thomas Davidson, *Journal of Speculative Philosophy* 8 (1874). Item `grammarofdionysi00dionuoft`. The authorship of the *Technē* as we have it is disputed in the scholarship; I did not check that literature and call the work "attributed". +- *Read status:* **read in part (card: library/thrax1874grammar.md)**: headnote and §§1–14 (pp. 3–11) and the closing section read closely; §§15–24 skimmed. Davidson's headnote says the work's authenticity "is now generally admitted" (p. 3); I did not check this against current scholarship, so I still call the work "attributed". +- *Quotation (§2, "On Reading"):* "Reading is the rendering of poetic or prose productions without stumbling or hesitancy. It must be done with due regard to expression, prosody, and pauses… from the pauses, the meaning intended to be conveyed." §1 lists "Trained reading with due regard to Prosody" as the first of the six parts of grammar. Davidson's footnote: "Prosody… includes everything designated by diacritical marks — aspiration, accentuation, quantity, and sometimes pauses." +- *Bearing:* the first part of the ancient grammarian's work is *trained reading* with attention to pauses, so the reader's task of phrasing was taught, as an art, by the grammarian. **Complicates** H-cost; says nothing about separation. + +#### K4. Cribiore, *Writing, Teachers, and Students in Graeco-Roman Egypt* (1996) + +**cribiore1996writing** — Cribiore, R. (1996). *Writing, Teachers, and Students in Graeco-Roman Egypt*. American Studies in Papyrology 36. Atlanta: Scholars Press. DOI 10.3998/mpub.9749698 (Crossref title: "Writing, Teachers and Students in Graeco-Roman Egypt"; year given as 1995 online; closed on Unpaywall). Lending scan `writingteacherss0000crib`. +- *Verification:* Crossref record; archive.org metadata. +- *Read status:* **snippet**. Section "Syllabic Division, Word Separation, and Verse Distinction" located within ch. 3, "The School Exercises"; printed pages not available. +- *What the snippets say (joined from overlapping windows):* + - "Besides the syllabaries, many exercises show how the division of words into syllables was [practised]…" + - "…the division of words into syllables or the separation of words from one another to facilitate reading, will elucidate how students worked and learned…" + - "Once a student had practiced enough with words divided into syllables, in theory he was ready for more serious and challenging reading. The continuous blocks of writing,…" + - "…individual words were not separated by spaces—scriptio continua. This feature of ancient writing must…" (cut), and "…distinguish words written in continuous blocks (scriptio continua). According to Quintilian, people usually advised the beginner to keep looking to the right when starting to read…" + - On Latin: "Although in Latin word division appears regularly until the end of the first century AD,…"; and, of wax tablets, "a transitional stage: a few tablets show interpunct and word division by space". +- *Bearing:* **supports** the point that teachers knew and used the aid for novices: school texts divide words into syllables, or separate words, and teachers move the student on to continuous script ("ready for more serious and challenging reading"). This is the clearest contemporary evidence in this facet that the aid existed in antiquity, was reserved for beginners, and was removed at the next stage of training. I could not see her wording about *why* the continuous text followed, or any claim that it shortened the course. + +#### K5. Cribiore, *Gymnastics of the Mind* (2001) + +**cribiore2001gymnastics** — Cribiore, R. (2001). *Gymnastics of the Mind: Greek Education in Hellenistic and Roman Egypt*. Princeton: Princeton University Press. DOI 10.1515/9781400844418. Scan `gymnasticsofmind0000crib`. +- *Verification:* Crossref (title matches; Crossref type "edited-book"); archive.org. +- *Read status:* **snippet**; printed pages not available. +- *Snippets (verbatim, normalised):* "Sometimes books addressed the needs of uncertain readers through consistent word separation. Part of a hexameter poem about Hero…" (cut); "exercises at this level show words divided into syllables by means of bars or dots or included in sets according to number of syllables, that is, monosyllables…"; "Since books for beginning readers were not produced on a regular basis, models prepared by teachers fulfilled…" (cut); "teachers' models presented longer excerpts in scriptio continua that addressed relatively advanced students"; "…of letters without word separation (scriptio continua) and without pauses between major sections"; "initial link — the letters of the alphabet — that led to syllables, words, and sentences"; "the effectiveness of building on syllables to teach reading and writing properly was not questioned throughout…" (cut). On Dionysius: "the first century b.c.e., the writer Dionysius of Halicarnassus, for instance, says that it was…" (cut). +- *Bearing:* **supports**. Books for the uncertain reader carried "consistent word separation", exercises for beginners divided words into syllables, and the advanced student's models were in *scriptio continua*. Ancient teachers matched the page to the learner's stage. This is the strongest source here for a design precedent in antiquity itself, and it cuts against "ignited": the aid was known and was deliberately withdrawn. + +#### K6. Marrou, *A History of Education in Antiquity* + +**marrou1956history** — Marrou, H.-I. (1956). *A History of Education in Antiquity*. Tr. G. Lamb. New York: Sheed and Ward (many reprints; the scan I searched is `historyofeducati0000marr`, a later reprint whose year and pagination I did not capture). +- *Read status:* **snippet**; printed page not available. Locator: chapter on the Hellenistic school, section "The Reading Lesson — The Alphabet — The Syllables — Words — Texts and Anthologies" (contents snippet). +- *Snippets, joined from three overlapping windows (the join is mine):* "It was not so easy for the child to read these as it is for our modern child today, because the Greeks used a scriptio continua in which not only was there no punctuation but even the words themselves were not separated from each other." On syllables: "the syllables separated as in our modern primers". Elsewhere: "manuscripts we possess, were always in scriptio continua." +- *Bearing:* **supports** a one-line claim that scriptio continua made the child's reading harder, with no evidence offered in the snippets. This is the nearest I found, in a standard history of education, to the author's intuition; it is an assertion about *difficulty for the child*, not about training length or the number of readers. I could not tell what "these" refers to. + +#### K7. Bonner, *Education in Ancient Rome* + +**bonner1977education** — Bonner, S. F. (1977). *Education in Ancient Rome: From the Elder Cato to the Younger Pliny*. London: Methuen; Berkeley: University of California Press. Scan `educationinancie0000bonn`. +- *Read status:* **snippet**; printed page not available. +- *Snippets:* a plate caption, "Section of teaching manual, showing syllabic word-division (third-century B.C. papyrus, MS 65445, Cairo Museum)"; "orderly progression from letters to syllables, and from syllables to words, and from words to sentences"; "a praelectio, or 'preliminary reading'" with "the most essential [part]…" cut. On Roman children learning Greek: "This must have presented at first a formidable obstacle, and even children learning to read in their own language needed some initial…" (the "obstacle" is Greek as a second language, not script layout). +- *Bearing:* **supports** the existence of the aid in a third-century BC school manual. No hit for `scriptio continua`, `silent reading` or `"separate the words"`. + +#### K8. Morgan, *Literate Education in the Hellenistic and Roman Worlds* + +**morgan1998literate** — Morgan, T. (1998). *Literate Education in the Hellenistic and Roman Worlds*. Cambridge: Cambridge University Press. Scan `literateeducatio0000morg`. +- *Read status:* **snippet**; printed page not available. +- *Snippets:* "four-syllable names beginning with 'E', divided into syllables to make them easier to read" (a school text); "…with learning letters, syllables and words, though there is no direct evidence" (cut); Quintilian quoted: "There is no short-cut to syllables: they must all be learned thor[oughly]". +- *Bearing:* **supports** K4 and K5 on one point: syllable division was a school device "to make them easier to read". No hit for `"word division"` or `scriptio continua`. + +#### K9. Johnson, *Readers and Reading Culture in the High Roman Empire* (2010) + +**johnson2010readers** (key reused; facet B) — Johnson, W. A. (2010). *Readers and Reading Culture in the High Roman Empire: A Study of Elite Communities*. Oxford: Oxford University Press. DOI 10.1093/acprof:oso/9780195176407.001.0001 (Crossref title matches). Scan `readersreadingcu0000john`. +- *Read status:* **snippet**; chapter 1 area, printed page not available. Facet B read only reviews of this book. +- *Snippets (joined from overlapping windows; the join is mine):* "Greeks and Romans knew perfectly well, for instance, the utility of word division—the Greek school texts on papyri bear eloquent testimony to the need for emerging readers to practice syllable and word division." Also: "…word division was discarded by the Romans in deference to Greek aesthetic and cultural traditions"; "…necessitating years of advanced study to fully apprehend." (the closing mark in the OCR is probably note 39; the words are Johnson's own, per the auditor); "The upshot of all this is that the bookroll culture in the high empire was one designedly reserved for elite of a certain stripe, able and willing…" (cut); "Strict functionality, clearly, is not a priority in bookroll design." +- *Bearing:* **supports** the "known and reserved for novices" point from a classicist who rejects Saenger's physiology. It adds the other half: for the elite, the training was the point. Johnson reads the cost of unspaced script as a social filter, not a defect. **Complicates** the author's sentence: if the training was designed to exclude, removing it is a change in who may enter, a claim about the elite's choice, not an accident of interface. + +--- + +### Part 2. Training after separation + +#### K10. Reynolds, *Medieval Reading* (1996) + +**reynolds1996medieval** — Reynolds, S. (1996). *Medieval Reading: Grammar, Rhetoric and the Classical Text*. Cambridge Studies in Medieval Literature 27. Cambridge: Cambridge University Press. DOI 10.1017/cbo9780511470356 (Crossref title "Medieval Reading"). Scan `medievalreadingg0000reyn`. +- *Read status:* **snippet**; chapter 1 ("Learning to read: the classics and the curriculum"), printed pages not available. +- *Snippets:* "In other words, even at its very earliest stages, learning to read 'means learning to read a foreign language. This has the most profound effect on the reading practice I analyse in this book" (the OCR shows the closing quotation mark misplaced); "crucial to realise that the alphabet the puer learns is not the alphabet of the mother-tongue"; "Sometimes the Psalter itself was used to teach the syllables"; "the most common way to learn the alphabet, the syllables and the paternoster". On the glossator: "Glosses like this take the place of punctuation… It is the lack of punctuation in the manuscript text that causes the glossator [to…]" (cut). +- *Zero hits (whole-book):* `"word separation"`, `"word division"`, `"scriptio continua"`, `"separation of words"`, `"spacing"`, `Saenger`, `"silent reading"`, `"space between words"`. +- *Bearing:* **complicates** the author's sentence strongly. In the standard book on how medieval readers learned, the training is an alphabet, syllables, psalter, Donatus course in a foreign language, and the book's account of medieval reading never mentions word separation. For the Insular case it agrees with Parkes: the demand came first (learners of a foreign tongue). + +#### K11. Orme, *Medieval Children* (2001) and *Medieval Schools* (2006) + +**orme2001medieval** — Orme, N. (2001). *Medieval Children*. New Haven: Yale University Press. Scan `medievalchildren00orme`. **orme2006medievalschools** — Orme, N. (2006). *Medieval Schools: From Roman Britain to Renaissance England*. New Haven: Yale University Press. Scan `medievalschoolsf0000orme`. +- *Read status:* **snippet**; printed pages not available (ch. 7 "Learning to Read" in the first; the Anglo-Saxon chapter of the second). +- *Snippets:* Orme 2001: "children learning to read were made to decipher the Paternoster as their first piece of…" (cut); "Grammars had to be written that taught Latin as a foreign language, unlike Roman grammars which assumed an understanding…"; a chapter section "Learning Syllables". Orme 2006: "For many, learning to read must have involved learning Latin as a foreign language, making it difficult to pick up such knowledge informally"; "The alphabet of letters learnt in Anglo-Saxon England was the Latin alphabet"; "Once pupils could read words and sentences, they could learn Latin texts". +- *Zero hits (whole-book), both books:* `"word separation"`, `"word division"`, `"scriptio continua"`, `"separation of words"`, `"spacing"`, `Saenger`, `"silent reading"`, `"space between words"`. +- *Bearing:* **complicates**. Same pattern as Reynolds: the burden is Latin, and the sequence is alphabet, prayers, syllables. The authority on medieval children's schooling never mentions separation. + +#### K12. Black, Grendler, Riché + +**black2001humanism** — Black, R. (2001). *Humanism and Education in Medieval and Renaissance Italy*. Cambridge: Cambridge University Press. Scan `humanismeducatio0000blac`. **grendler1989schooling** — Grendler, P. F. (1989). *Schooling in Renaissance Italy: Literacy and Learning, 1300–1600*. Baltimore: Johns Hopkins University Press. Scan `schoolinginrenai0000gren`. **riche1976education** — Riché, P. (1976). *Education and Culture in the Barbarian West, Sixth through Eighth Centuries*. Tr. J. J. Contreni. Columbia: University of South Carolina Press. Scan `educationculture0000rich`. +- *Read status:* **snippet**; printed pages not available. +- *Black snippets:* "The introduction to reading by learning the alphabet and then syllables was a technique going back [to antiquity]" (cut); "earlier middle ages, too, children began with the alphabet and proceeded to syllables: as Remigius of Auxerre…" (cut); "standard reading syllabus of carta/tabula, psalter and Donatus"; "silent reading of pupil-owned copies" in a late medieval Italian classroom (from Bonvesin). +- *Grendler snippets:* "pages of syllables, the good scholar was ready to read words composed of two two-letter syllables"; "the alphabet… the Pater Noster, Ave Maria, Symbolum"; "First, the pupils read aloud and learn the alphabet. Next, they begin to join…". +- *Riché snippets:* "First, letters and syllables, then the 'eight parts of speech'"; "old he learned the notae litterarum—that is, the alphabet—at home". +- *Zero hits (whole-book):* for `"word separation"`, `"scriptio continua"` in all three; Black and Grendler also have none for `Saenger`. Riché also none for `Saenger` (the book predates him). +- *Bearing:* **complicates** and **refutes** the "no longer required the training" reading for the period after separation: Italian schools from the fourteenth to the sixteenth centuries, long after canonical separation, still taught the alphabet, syllable pages, psalter and Donatus in sequence, and Black says that method goes back to antiquity. The instruction did not change shape when the page changed. + +#### K13. Saenger on novices (1997; 1999 chapter) + +**saenger1997space** (key reused; facets A–C) — Saenger, P. (1997). *Space Between Words: The Origins of Silent Reading*. Stanford: Stanford University Press. Scan `spacebetweenword0000saen`. +- *Read status:* **snippet** (this facet); facets A and C hold earlier snippets and the page checks. +- *On novices and the syllabary method (ch. 1; Blake 2002 cites the passage as "Saenger 1997, 4, 5, 6"; the auditor found this sentence is about Japanese kana, p. 3, not Greek or Latin teaching; on p. 6 Saenger says young children "tend initially to write in scriptura continua"):* "The exclusive use of a syllabary method of presenting text to novice readers, necessitating the oral reconstruction of units of meaning from phonetic components…"; Also, in the same area, a fragment on "conditions for rapid lexical access" that "allows Chinese children to develop silent reading at an earlier age". I could see only these fragments and cannot state his full argument for the chapter. +- *The nearest sentence to the author's:* "Word separation, the diastole, the hyphen, and eventually the trait d'union in Insular books provided shortcuts for achieving the reading skills that an elite among the ancients had mastered only through a prolonged and arduous grammatical apprenticeship." (p. 85, confirmed by the auditor). It is a list: word separation is one of four devices, the other three being marks. It names *Insular books* and the reading skills of an elite; Saenger's learner argument is on p. 91. It is Saenger's own claim that the aid replaced the elite's apprenticeship with a shortcut. He gives it without a measure of time or numbers. A second Anglo-Saxon grammarian (Bede) follows in the next sentence. +- *Also (p. 13 per facet A, wording confirmed here):* "As a consequence, even readers of modest intellectual capacity could read more swiftly…" +- *Explicit disclaimers:* the book "is not a general history of literacy" (facet A, p. x; wording confirmed in my snippet: "which, I would stress, is not a general history of literacy. Instead, I have attempted to explain the physiological…"). On p. 11 (facet A) he lists "the greater diffusion of literacy throughout all social strata of the population" among the benefits modern readers see in ease of reading; my snippet confirms the wording. +- *Bearing:* **supports** H-cost in its narrowest form (Insular books, an elite's reading skills, a prolonged apprenticeship shortened) and **does not support** H-cause for literacy. Note it is the one passage in the corpus that says the apprenticeship was shortened. It is a claim without a test. + +**saenger1999history** — Saenger, P. (1999). The history of reading. In D. A. Wagner, R. L. Venezky and B. V. Street (Eds.), *Literacy: An International Handbook* (Boulder, CO: Westview Press), ch. 2, starting p. 11 (contents snippet, OCR garbled; end page not seen). Scan `literacyinternat0000unse`. Not in the repo bib. +- *Verification:* the contents list in the scan shows "2 The History of Reading, Paul Saenger"; the chapter's opening matches. +- *Read status:* **snippet**. +- *Snippets (the key ones):* "The history of written communication in the West may be divided [into…]" and "first comprises the discipline of the history of reading, the second constitutes the history of literacy. The history of reading deals with the qualitative question of how people…" (cut); "The history of literacy is concerned with quantitative questions relating to the diffusion of these various reading skills to different strata of society." And: "…after 1200 constitute events more important for the history of literacy than for the history of reading and are thus marginal [to this chapter]"; "questions of social class and gender, central to the history of literacy, are of only marginal import to [it]" (the closing words are cut). On the twelfth century: "the far more legible and visually perfected written Latin language of the twelfth century was a second language to readers…" (cut). Also: "Totally silent reading, without movement of the tongue and…". +- *Bearing:* **the single most useful sentence for Q3.** In 1999 Saenger himself separates the history of reading (how people read) from the history of literacy (the diffusion of reading skills to strata of society), says his chapter is about the first, and treats class and gender as marginal to it. He therefore does not claim that separation spread literacy. **Refutes H-cause as a claim of Saenger's.** + +#### K14. Parkes, *Pause and Effect* (1992/1993) + +**parkes1992pause** (key reused; facet A). Lending scan `pauseeffectintro0000park` (the University of California Press edition, 1993). +- *Read status:* **snippet**; facet A holds earlier snippets and the chapter pages (ch. 2, pp. 20–29; ch. 3, pp. 30–34). +- *This session, verbatim:* "Because many had to learn Latin as a foreign language, there was a need for conventions which made it easier to read. Over the centuries these…" (cut); "New conventions, such as word separation, features of layout and punctuation, were developed to make it easier for readers to extract the information conveyed in the written medium"; "The Anglo-Saxons learned the practices of word-separation, layout and punctuation from Irish teachers"; "…words as pattern-units, in order to speed the eye of the reader from one letter to the next within a word"; "Texts were mostly read aloud." Of the sixth century on: "attitudes… Isidore of Seville (c. 560–636) could state a preference for silent reading". +- *Zero hits:* `Saenger`, `"learners"`. +- *Bearing:* Parkes gives a **cause running from readers to script** (many had to learn Latin as a foreign language, hence the need for conventions) and the effect as ease of reading. He names no literacy outcome and never cites Saenger. **Supports** H-cost; **opposes** the author's direction (page → reader) only in the sense that the readers came first. + +--- + +### Part 3. Causal claims from separation, and their reception + +#### K15. Saenger 1982 (*Viator* 13) and 1997: the claims + +**saenger1982silent** (key reused) — the whole of *Viator* 13 (1982) is in the scan `viator_1982_13`; it is access-restricted but FTS returns snippets. Facet A could not open the article; I read snippets only. +- *Read status:* **snippet**; article pp. 367–414; per-snippet pages not available. +- *Snippets:* "…the reduced paravoleal vision which resulted from lack of word separation. The introduction of word separation in the seventh century permitted fewer and longer saccades" (physiological claim; the part Vatri and Heilmann reject; facet B); "…ruling class, literate in a script without word separation, had no clear images of words correctly spelled…"; "…writing, and revising Roman scripts without word separation made autograph composition slow and tedious". + +The catalogue of claims in 1997 (each quote from my snippets of `spacebetweenword0000saen`; "page not available" unless a page is from facet A): + +| Claim from separation to… | Wording | Evidence offered | Reception | +| --- | --- | --- | --- | +| Silent, rapid reading | Reviewers' form, not Saenger's: see K16. Saenger: "even readers of modest intellectual capacity could read more swiftly" (p. 13, facet A) | The physiology of reading, manuscripts | Contested: Knox (silent reading attested long before any spacing; he never mentions word separation or the training of readers, so he bears on the silent-reading half only), Vatri, Heilmann (facet B); O'Donnell (2000, ¶3): "some will dissent from … his claim that the textual practices he observes were the necessary and sufficient condition of the rise of silent reading", and the stretch from the physiology chapter to the theory is "the most venturesome and vulnerable part of the book" (¶4); Cavallo and Chartier (K17) | +| Reference reading and the scholastic book | "thoughts, which without these tools would not have been possible" (cf. facet A, p. 122); "Separated script was an indispensable medium for this new learning" (page not available); "use, implying the scanning of text, would not have been possible in aerated script" | Manuscript layout and page apparatus, twelfth to thirteenth centuries | Claimed, not tested by any reviewer I could read; Rouse and Rouse explain the apparatus without separation (K18) | +| Private devotion | "…prayers in the first-person singular intended for intense private devotion; no equivalent can be found in unseparated [books]" (page not available) | Insular prayer books | No reviewer I could read comments on this link | +| Heresy and censorship | Section heading "Private Study and Heresy" (contents, p. 264); "Already in the eleventh century, heresy was linked to solitary intellectual curiosity and speculation"; "The transition to silent reading and composition, by providing a new dimension of privacy…" (cut); "In England, the mere possession of Lollard writings was legal grounds for formal charges of heresy" | Cases (Berengar of Tours, Lollards) | Not tested by any reviewer I could read; the Raritan reviewer (K16) objects to the "relentless teleology" of the whole | +| Authorship and autograph composition | "authors had generally composed by dictation to a scribe, in large measure because of the difficulty of composing while writing in [scriptura continua]"; Guibert of Nogent wrote on parchment "because of the lack of privacy when composition depended upon dictation to a secretary" | Authors' own manuscripts (Richer of Reims's autograph corrections) | Not tested | +| Erotic and private reading | "Private reading stimulated a revival of the antique genre of erotic art" | Decoration in books, Guibert and Odo of Orléans | Not tested | +| Vernacular reading | Section "Lay Society and Vernacular Texts" (contents, p. 265); "…the same practice of silent reading and composition for vernacular literary texts that…" (cut) | Vernacular manuscripts | Hanna notes the vernacular lag (K16); Richter (K16) | +| Literacy diffusion | None. p. 11 lists it as a *modern* advantage; 1999 chapter assigns it to a different history (K13) | n/a | n/a | + +#### K16. Reviews of Saenger 1997 (reachable at snippet depth) + +Facet B opened O'Donnell (TMR 2000) and Garrett (1998) in full and Aronoff's abstract. Below, the reviews I could reach through scans of the journal issues. Printed review pages come from Crossref. + +**hanna1999saenger** — Hanna, R. (1999). Review. *Speculum* 74(3), 830–831. DOI 10.2307/2886848 (Crossref title matches). Scan `sim_speculum_1999-07_74_3`. **Snippet.** +- "While there is much in *Space between Words* that will be of great interest to students of late Caroline books and their…" (cut). "Saenger argues that the distance between Latin and the Celtic and Germanic vernaculars…" (cut). "Saenger argues that, not only is word division native-tongue-determined, it may also reflect other features of…" (cut). "While Saenger carefully describes a revolution that would underlie such a Derridean…" (cut). "Chapters 8–9, for example, examine the proliferation of word separation as linked with specific authors (Gerbert, Abbo, Heriger), their texts, and their…" (cut). +- *Bearing:* Hanna reports Saenger's Insular-learner explanation as an argument. I could not see whether she accepts it, or what her objection is. + +**richter2001saenger** — Richter, M. (2001). Review. *American Historical Review* 106(2), 627–628. DOI 10.1086/ahr/106.2.627. Scan `sim_american-historical-review_2001-04_106_2`. **Snippet.** +- "…Saenger, generally sober and rather understated, calls this the 'great divide' (p. 12) and 'a revolution' (p. 48). The practice began in those countries that had received Latin as a foreign language. Pride of place goes here to Ireland, the first Western country to adopt…" +- "But it took more than half a millennium before space between words found acceptance in the Romania rather than the traditional scriptura continua. Saenger suggests convincingly that this was due to the relatively greater familiarity there with the Latin language…" (cut). +- *Bearing:* **a reviewer who endorses the reverse cause** ("convincingly"): separation took hold where Latin was foreign and lagged where it was familiar. He notes the half-millennium. + +**mulchahey2001saenger** — Mulchahey, M. M. (2001). Review. *Canadian Journal of History* 36(2), 323–325. DOI 10.3138/cjh.36.2.323. Scan `sim_canadian-journal-of-history_2001-08_36_2`. **Snippet.** +- "Saenger is convinced, however, that the critical juncture in the history of reading in the Latin West came earlier, when medieval scribes began to make concessions to the reader in the form of visual cues such as punctuation and word separation which greatly aided the mental recognition of words and units of meaning." "Thus, as a minor premise, the emergence of proper word separation over the course of the medieval period was likely the single most important factor enhancing readers' ability to comprehend written text rapidly and silently." "Saenger's book is impressive for the sheer weight of archival research that lies behind it." Her close: "…alike a very great service, in teaching us how to read the *Space Between Words*." +- *Bearing:* a summary in her own words ("as a minor premise… likely the single most important factor"); she does not dissent. Reception: positive, not tested. + +**walker2000saenger** — Walker, T. D. (2000). Review. *The Library Quarterly* 70(1), 166–167. DOI 10.1086/603182. Scan `sim_library-quarterly_2000-01_70_1`. **Snippet.** "[The] thrust of his work is that the introduction of word separation led to the development of silent reading during the Middle Ages." Descriptive; "a textbook for this specialized field" of the chapters on the nomenclature of word separation. No evaluation of the causal claim. + +**aronoff2002saenger** — Aronoff, M. (2002). Review. *Language in Society* 31(4), 624–628. DOI 10.1017/s0047404502254056. Scan `sim_language-in-society_2002-09_31_4`. **Snippet** (facet B had the abstract only). +- "…more plausible) reason that S gives for why word separation might have begun in the British Isles is that Latin was a foreign language for its Irish and British learners." (the opening of the sentence is cut) +- "Word separation is the simplest device for ensuring quick recognition of words in succession, while…"; "…easily lend themselves more readily to rapid silent comprehension"; "The level of detail in this core half of the book is truly remarkable." Abstract (facet B): "impressive, fascinating, and exasperating". +- *Bearing:* a linguist calls the learner explanation "plausible" and, in the part visible, accepts the speed claim as ordinary psycholinguistics. I did not see the "exasperating" passage. + +**raritan2000review** — Review essay on Saenger, Manguel and Johns (author not identified), *Raritan* 20(1) (Summer 2000), review of Saenger at p. 105 per the issue contents. Scan `sim_raritan_summer-2000_20_1`. **Snippet.** "Nonetheless, I am put off by the relentless teleology of *Space Between Words*. Like Eric Havelock, who similarly argued long ago for the progressive development of writing and its attendant [effects]…" (cut); "If Saenger's claims are right, then (much like Johns) he has produced an argument about the origins of…" (cut). Also: "exclusively a story of Latin text production Saenger tells— made possible the quicker, silent, and, in turn, more private apprehension of the written word." +- *Bearing:* the clearest reviewer objection I found to the *form* of the causal claim (teleology). Author unnamed; citable only after a person identifies him or her. + +**nq2000saenger** — Review in *Notes and Queries* 47(1) (March 2000), reviewer unnamed in my snippets. Scan `sim_notes-and-queries_2000-03_47_1`. **Snippet.** Descriptive: "Saenger shows that word separation was pioneered by non-romance cultures"; "appear to have been won over by the advantages of word separation until much later". No dissent in the snippets. + +**Not reached:** Bäuml 2002 (*Written Language & Literacy* 5(1), 131–134, DOI 10.1075/wll.5.1.08bau, verified at Crossref), Lyons 2001, the *Choice* note, any review by Ganz or Brown (I found none under those names). Section (d). + +#### K17. Cavallo and Chartier (eds.), *A History of Reading in the West* (1999) + +**cavallo1999history** (key reused; facet A). Scans `historyofreading0000unse` and `historyofreading0000cava`. +- *Read status:* **snippet**; the sentences below are from the volume's opening survey; I could not confirm from snippets which editor wrote them or the page, so I attribute them to "the editors' Introduction (to be confirmed)". The two passages on antiquity quoted below ("absence of breaks" and "sonorization") are Jesper Svenbro's, ch. 1, pp. 51-52 (AUDIT_06). +- *Verbatim (joined from overlapping windows):* "The first is the line, where in the early Middle Ages, the appearance of word separation made silent reading possible. The next is the page, which underwent two transformations: in the fourteenth and fifteenth centuries…" (cut). On antiquity (Svenbro, not the editors): "The absence of breaks in the written line was indeed a hindrance to silent reading, and it remained so. It was not an insurmountable obstacle, however, as we can surmise…" (cut); Svenbro's next sentence cites Saenger: "According to Paul Saenger, in the Middle Ages word division was a necessary condition for the dissemination of the practice of silent reading". And (Svenbro): "…sonorization of writing was negatively programmed by the absence of spaces between words, and since sonorization was valued as a good in itself, why should anyone feel the need to abandon the scriptio continua that was a technical impediment to the…" (cut). +- *Bearing:* book historians state the claim *separation made silent reading possible* flatly, then Svenbro (ch. 1) qualifies it for antiquity ("not an insurmountable obstacle"), while citing Saenger for word division as "a necessary condition for the dissemination of the practice of silent reading", and gives a social reason for not abandoning *scriptio continua*. The line, then the page: separation is the first of a chain of layout changes, which supports problem 4 (a bundle). No literacy claim in the snippets. + +#### K18. The apparatus of the scholastic book: Rouse and Rouse, Parkes 1976, Illich, Chartier, Carruthers, Mak + +**rouse1982statim** (key reused; marked "not accessible" in the repo bib because of an earlier run; I now have snippets) — Rouse, R. H. and Rouse, M. A. (1982). *Statim invenire*: Schools, preachers, and new attitudes to the page. In R. L. Benson and G. Constable (Eds.), *Renaissance and Renewal in the Twelfth Century* (Cambridge, MA: Harvard University Press), pp. 201–225 (the contents snippet gives p. 201; the end page is from the running heads). Scan of the volume `renaissancerenew0000unse`. +- *Read status:* **snippet**. +- *Snippets:* the title page: "Statim invenire: Schools, Preachers, and New Attitudes to the Page"; "a reader facing a book of eight hundred pages might well be grateful"; "…'STATIM INVENIRE' direct ancestor of all later alphabetical…"; a note citing Parkes's "The influence of the concepts of *ordinatio* and *compilatio* on the development of the book". +- *Zero hits (whole volume):* `"word separation"`, `Saenger`, `"silent reading"`, `"scriptio continua"`, `"spacing"`. One hit for `"separation of words"`, about hyperbaton. +- *Bearing:* **the standard account of the twelfth-century reference apparatus does not use word separation as a cause.** The title gives the cause as schools, preachers and new attitudes to the page. So the Rouses and Saenger stand side by side: Saenger claims separation was necessary for the apparatus (K15), the Rouses explain the apparatus by the readers' need to find things quickly and do not mention separation. No scholar I read adjudicates between them. + +**Parkes 1976 (*ordinatio* and *compilatio*)** — cited in the Rouses' note and in *Pause and Effect* (hits in K14). Not opened. In *Pause and Effect* the snippet ties scholastic readers' needs to punctuation: "In these circumstances scholastic readers needed to identify the boundaries of…" (cut); "The habits of scholastic authors also influenced the use of punctuation". **Not accessible** as an article: *Medieval Learning and Literature: Essays Presented to Richard William Hunt* (Oxford: Clarendon Press, 1976), pp. 115–141; page range from memory, unverified. + +**illich1993vineyard** (key reused) — Illich, I. (1993). *In the Vineyard of the Text: A Commentary to Hugh's* Didascalicon. Chicago: University of Chicago Press. **Not accessible** (no FTS route). What I know is Saenger's review: **saenger1995illich** — Saenger, P. (1995). Review of Illich. *The Library Quarterly* 65(2), 244–246. DOI 10.1086/602782 (Crossref title matches). Scan `sim_library-quarterly_1995-04_65_2`. **Snippet.** "…his thesis is clearly that Hugh of Saint Victor represented the crucial change from the oral reader mentality that had originated…"; "However, his assertion that the change to visual silent reading was occurring precisely during the period when Hugh composed and revised his *Didascalicon*…" (cut); "…to skim the written page, a practice only possible due to the word separation that was standard in twelfth-century Parisian…"; "In fact the canonical word separation necessary for silent reading (superbly illustrated by Illich's dust jacket reproduced…" (cut). +- *Bearing:* Saenger, reviewing Illich in 1995, supplies the missing mechanism himself ("a practice only possible due to the word separation"). Illich's claim as the reviewer reports it is about the date of the shift to visual reading; the separation link is Saenger's gloss. I cannot say what Illich wrote on separation. + +**chartier1994order**, **carruthers2008book**: scans `orderofbooksread0000char` and `bookofmemorystud0000carr`. **Snippet.** Chartier, *The Order of Books* (1994): no hit for `"word separation"` or `scriptio continua`; notes cite Saenger on silent reading in "Communities of Readers". Carruthers, *The Book of Memory* (2nd ed., 2008): no hit for `"word separation"` or `scriptio continua`; one note reads "for earliest medieval uses of word spacing see especially Saenger, *Space Between Words*, and Parkes, *Pause and Effect*". Neither builds an argument on separation. **Mak 2011** (**mak2011page**): **not accessible**; not on archive.org, Google Books quota-blocked. Stock 1983 and 1990: `listeningfortext00stoc` returns no hit for `"word separation"` (facet C had Stock 1983 at part depth). + +#### K19. Later uses + +**blake2002literacy** — Blake, B. E. (2002). *Literacy and Learning: A Reference Handbook*. Santa Barbara: ABC-CLIO. Scan `literacylearning0000blak`. **Snippet.** A school-literacy handbook with a chapter "How We First Read: Texts with No Word Separation". "It is hard to imagine that the…" and "Why did it take so long for writers to make reading easier by leaving spaces between words?" Cites Saenger 1997, pp. 2, 4–6 on how children learn to read in Chinese, Burmese, and with recitation as "an important part of elementary reading instruction". **Bearing:** the one place in the corpus that moves Saenger's ease-of-reading argument into a book about teaching reading. It is a reference handbook for teachers. It repeats Saenger and adds no evidence. + +**baron2000alphabet** — Baron, N. S. (2000). *Alphabet to Email: How Written English Evolved and Where It's Heading*. London: Routledge. Scan `alphabettoemailh00baro_0`. **Snippet.** Two separate snippets: "spaces between words when copying original Latin texts, making it visually easier to make out both the lexicon and grammar", and "copying texts in Irish (that is, their native tongue), they didn't introduce word separation, but grouped words…" (cut). The auditor found the two snippets are adjacent sentences: Insular (Irish) scribes introduced spaces when copying Latin; copying texts in Irish, their native tongue, they grouped words by stress or phrasing instead. Also "The evolutionary process was slow." **Supports** the learner-first direction from a linguist reading Saenger. + +**lyons2010history**: Lyons, M. (2010). *A History of Reading and Writing in the Western World*. Basingstoke: Palgrave Macmillan. Scan `historyofreading0000lyon`. **Snippet.** Section "Word separation and silent reading": "from the seventh century onwards, according to Paul Saenger, the separation of words significantly assisted…" (cut); "But by the twelfth century, word separation was universal in Europe". Secondary summary; no new evidence. + +--- + +## (c) Synthesis + +### Q1. Training before separation + +The sources agree on the sequence. Dionysius: letters, forms, values, syllables, words, "syllable by syllable and slowly at first" (ch. XXV). Quintilian 1.1.30–33 says syllables cannot be shortcut and reading "must be kept slow for a considerable time" before speed arrives. Bonner and Marrou both describe a syllable stage. None gives a length for learning to read alone; Cribiore (2001, p. 171) speaks of "one or two years of elementary instruction" for many students. Quintilian argues for starting before seven (1.1.15-20). + +Two points matter for the author's sentence. + +1. **The aid existed and was kept for novices.** Cribiore (1996) describes school exercises in which words are divided into syllables, "or the separation of words from one another" is used "to facilitate reading"; for higher-level texts she lists "separation of words by spaces or bars" among the features that mark a text as meant for students (p. 72). Cribiore (2001) describes syllable division "by means of bars or dots". Cribiore (1996, p. 48) writes that once a student had practiced with words divided into syllables, "in theory he was ready for more serious and challenging reading", but that "the continuous blocks of writing, however, made word divisions necessary as a further aid for the pupil" until accents could take over. Cribiore (2001) reports "consistent word separation" for "uncertain readers" and says teachers copied short excerpts of Isocrates "with words and syllables separated for students who were learning to read", while "some teachers' models presented longer excerpts in scriptio continua that addressed relatively advanced students". Bonner's plate caption shows "syllabic word-division" in a third-century BC manual. Johnson (2010, p. 20) writes that "the Greek school texts on papyri bear eloquent testimony to the need for emerging readers to practice syllable and word division." Teachers divided syllables, and sometimes words, for novices; syllable division gave way to ordinary texts, and Cribiore (1996, p. 48) has word division continuing as a further aid until accents replaced it. +2. **The unspaced text was the end of training, not its obstacle (my inference).** Dionysius, using reading as an analogy for composition, says that after "a considerable time" we go through "any book" "with incredible facility and speed"; he says nothing about word division. Johnson describes a literature "itself challenging, in many prominent cases necessitating years of advanced study to fully apprehend", and concludes that "the bookroll culture in the high empire was one designedly reserved for elite of a certain stripe" (2010, p. 31). Marrou is the plainest source here on the child's difficulty with *scriptio continua*: "It was not so easy for the child to read these as it is for our modern child today", because of *scriptio continua* ("these" are the ordinary texts that followed passages split into syllables); Cribiore (1996, p. 48) says the continuous blocks made word division "necessary as a further aid for the pupil". + +Thrax's "trained reading" shows that part of the burden was teaching phrasing and pause, which Quintilian says "can only be taught in actual practice" (1.8.1; Thrax §2); spacing does not do this. Bonner (K7) names the *praelectio*. + +### Q2. Training after separation + +The same sequence continued, and the sources do not describe a change when separation arrived. Reynolds says that "even at its very earliest stages, learning to read means learning to read a foreign language" [unverified: see AUDIT_06] (inner quotation marks dropped; confirm in print), with an alphabet that "is not the alphabet of the mother-tongue". Orme (2006) says that "for many, learning to read must have involved learning Latin as a foreign language". Black reports that the sequence of alphabet then syllables is "a technique going back" before the Middle Ages, and that the Italian reading syllabus remained *carta*, psalter, Donatus. Grendler records syllable pages, then words, in the Renaissance primers. Riché describes letters and syllables first, then the eight parts of speech. In each of Reynolds, both Orme books, Black (for separation terms), Grendler and Riché, a whole-book search finds no hit for "word separation", "scriptio continua" or "Saenger" (log rows 10–14). The standard accounts of how medieval and Renaissance children learned to read do not mention separation as a feature of the teaching. + +Saenger's account of novices in the Introduction concerns Japanese kana (p. 3); on p. 6 he adds that young children "tend initially to write in scriptura continua" because it represents uninterrupted speech. The other nearby text is the "shortcuts" sentence (K13). + +### Q3. Does any scholar claim, test or deny that separation shortened, eased or removed the training? + +I found no work that tests or denies it, and none that claims it for populations, other than Saenger's p. 11 and p. 13 remarks (below), in these sources: the archive.org lending scans of Cribiore (1996, 2001), Marrou, Bonner, Morgan, Johnson, Reynolds, Orme (two books), Black, Grendler, Riché, Parkes, Saenger (1982, 1997, 1999), Cavallo and Chartier, Rouse and Rouse, Chartier, Carruthers, Stock; the reviews of Saenger in the issue scans of *Speculum*, *AHR*, *CJH*, *Library Quarterly*, *Language in Society*, *Raritan*, *Notes and Queries*; and eight unscoped full-text queries over the whole archive.org text corpus (log rows 1–24). The facet C "silence" is therefore real, but it is narrower than "nobody says anything": three kinds of statement come close. + +- **For the narrow form.** Saenger 1997 (p. 85): "Word separation, the diastole, the hyphen, and eventually the trait d'union in Insular books provided shortcuts for achieving the reading skills that an elite among the ancients had mastered only through a prolonged and arduous grammatical apprenticeship." Parkes (p. 1 per facet A; my snippet; Parkes introduces the point for Latin in the West generally): "Because many had to learn Latin as a foreign language, there was a need for conventions which made it easier to read." Aronoff and Richter report the same Insular explanation. Marrou says the child found *scriptio continua* harder. Blake (2002) repeats the ease argument for a teaching audience. Svenbro (in Cavallo and Chartier 1999, pp. 51-52) allows that the absence of breaks "was not an insurmountable obstacle" to silent reading, but reports Saenger's view that in the Middle Ages "word division was a necessary condition for the dissemination of the practice of silent reading". +- **Against the wide form.** Saenger (1999) separates the history of reading from the history of literacy, which "is concerned with quantitative questions relating to the diffusion of these various reading skills to different strata of society", and holds that class and gender are "of only marginal import to the history of reading". Saenger 1997: "not a general history of literacy". +- **Silent where it should speak.** Reynolds, Orme, Black, Grendler and Riché contain no hit for "word separation" or "scriptio continua" in the OCR of the scans I searched; Riché (1976) and Grendler (1989) were written before Saenger's book, and Black and Riché do discuss silent reading. + +The author's sentence therefore has one nearby statement, Saenger's "shortcuts" sentence, and it is about Insular books and the reading skills of an elite (Saenger's argument that Insular readers were learners of Latin appears on p. 91), bundles separation with three marks, and brings no measure. + +### Q4. The catalogue of causal claims + +See the table in K15, and K16 to K19 for reception. In short: the claims run from separation to **silent and rapid reading** (Saenger; repeated by Cavallo and Chartier, Mulchahey, Walker, Lyons, Garrett); to **reference reading and scholasticism** (Saenger; absent from the Rouses' own explanation); to **private devotion, heresy, authorship, erotic and vernacular reading** (Saenger alone, with Illich's date for visual reading glossed by Saenger). Reviewers repeat the claims and do not test the individual links. The two negative reactions I could read are of kind: O'Donnell on "necessary and sufficient" (facet B) and the Raritan reviewer on "relentless teleology". Facet B holds the critics who reject the physiology. + +### Q5. Reverse causation + +| Scholar | Direction | Wording | +| --- | --- | --- | +| Parkes | readers to script | "Because many had to learn Latin as a foreign language, there was a need for conventions which made it easier to read." | +| Saenger | both ways | Irish scribes separated for learners; the Continent adopted for scholars (facet A, pp. 83–97, 120); in 1997 he also writes of the scribe's "unconscious quest" (p. 33, facet A) | +| Richter (reviewer) | readers to script | "The practice began in those countries that had received Latin as a foreign language." Romance regions lagged because of "the relatively greater familiarity there with the Latin language" | +| Aronoff (reviewer) | readers to script | Saenger's learner reason, "much more plausible" than the Syriac model, for why word separation "might have begun" in the British Isles: "Latin was a foreign language for its Irish and British learners" | +| Baron | readers to script | Insular (Irish) scribes introduced spaces when copying Latin; copying texts in Irish, their native tongue, they grouped words by stress or phrasing instead | +| Svenbro (in Cavallo and Chartier 1999, ch. 1) | both | "since sonorization was valued as a good in itself, why should anyone feel the need to abandon the scriptio continua that was a technical impediment to the development of silent reading?" | +| Johnson | social, elite choice | Word division in Latin literary texts "was discarded by the Romans in deference to Greek aesthetic and cultural traditions"; the bookroll "designedly reserved" for an elite | + +Two reviewers single out the learner explanation or the Romance lag for approval (Aronoff: "much more plausible" than the Syriac model; Richter: "convincingly" for the Romance lag); the others I could read describe the silent-reading claim without testing it. + +### Q6. Verdict + +| Claim | Class | Basis | +| --- | --- | --- | +| Insular separation arose to ease reading for learners of Latin as a foreign language | (b) claimed, endorsed by two reviewers as plausible, untested | Saenger (p. 91, 83-89), Parkes; Richter and Aronoff report it, Aronoff as "much more plausible"; Baron. Reynolds and Orme show only that Latin was a foreign language | +| Ancient teachers divided syllables, and sometimes words, for novices; syllable division gave way to ordinary texts for advanced students | (a) supported | Cribiore 1996, 2001 ("some"), Johnson, Bonner, Morgan; Marrou | +| Separation shortened Insular learners' apprenticeship (the "shortcuts" sentence) | (b) claimed, untested | Saenger 1997 p. 85; I found no reviewer comment on it in the snippets I could read (Hanna and Baeuml unread) [unverified: see AUDIT_06]; no measure | +| Separation made silent reading possible, or rapid reading | (b) claimed and contested | Saenger; Cavallo and Chartier; Mulchahey; against: Knox (silent reading attested before spacing; no mention of word separation), Vatri, Heilmann, O'Donnell 2000 ¶3 on "necessary and sufficient" | +| Separation enabled reference reading and the scholastic apparatus | (b) claimed; Rouse and Rouse give an explanation that does not mention it | Saenger; Rouse and Rouse (1982) explain the apparatus by readers' need to find things and do not mention separation; the paper predates Saenger's book, so it is not a reply to him | +| Separation → private devotion, heresy, autograph composition, erotic reading, vernacular reading | (b) claimed by one author, not tested | Saenger 1997 and 1982 | +| Separation → greater literacy, or a larger body of readers | (c) no source I read says separation raised literacy | Saenger lists "the greater diffusion of literacy throughout all social strata" among the advantages of ease of reading that moderns perceive (p. 11) and says "even readers of modest intellectual capacity could read more swiftly" (p. 13) | +| Separation → the populace no longer needed training | (c) claimed by nobody found | Reynolds, Orme, Black, Grendler show the training (alphabet, syllables, psalter, Latin) unchanged | + +**What the sentence at `draft.md:263` can be made to say that scholars support.** I do not rewrite it. The supported material, with the quotations, is: +1. Teachers in antiquity already used word or syllable division as a novice aid (Cribiore 1996, p. 48: "in theory he was ready for more serious and challenging reading", with word division kept "as a further aid"; Johnson, p. 20: "the need for emerging readers to practice syllable and word division"). +2. In the Insular case, separation and other marks were said to give a "shortcut" to what an elite had learned by "a prolonged and arduous grammatical apprenticeship" (Saenger 1997, p. 85) for readers who "had to learn Latin as a foreign language" (Parkes, speaking of Latin in the West generally). +3. The training in the later Middle Ages and Renaissance was an alphabet, syllable and psalter course in a foreign language (Reynolds; Orme; Black; Grendler), with no source giving separation a place in it. +4. "Populace": no source credits separation with reaching the populace; Saenger's own 1999 chapter assigns diffusion to the history of literacy, though his 1997 book lists "the greater diffusion of literacy throughout all social strata" among the advantages of ease of reading that moderns perceive (p. 11). + +### Claims the sources contradict or qualify + +1. **"It changed the interface so that reading no longer required the training" (`draft.md:263`).** Qualified. Saenger's "shortcuts" for Insular learners is the nearest statement and it shortens an apprenticeship; it does not remove one (Saenger 1997 p. 85). After separation was standard, Italian and English schools still taught alphabet, syllables, psalter and Latin (Black; Grendler; Orme 2006: "for many, learning to read must have involved learning Latin as a foreign language"). +2. **"The populace."** No source supports it. Saenger (1999) separates the history of reading from the history of literacy, which "is concerned with quantitative questions relating to the diffusion of these various reading skills to different strata of society", and holds that class and gender are "of only marginal import to the history of reading"; the readers served by separation were Latin learners, then clerics and scholars (facets A and C). +3. **That the difficulty of unspaced script was the main training burden.** Reynolds: "even at its very earliest stages, learning to read means learning to read a foreign language" [unverified: see AUDIT_06]. The burden named is Latin. Dionysius and Johnson both treat the end state as fluent. +4. **That the unspaced page had no aid for the novice.** Cribiore 1996 and 2001, Bonner and Johnson show ancient teachers separating syllables and words for beginners. +5. **That antiquity could not have seen the benefit.** Johnson (2010, p. 20): "Greeks and Romans knew perfectly well, for instance, the utility of word division". Their choice not to use it for adult texts is a social one. +6. **Facet A's gloss of the "shortcuts" sentence as about separation alone.** The sentence bundles "Word separation, the diastole, the hyphen, and eventually the trait d'union". +7. **Silent reading as the product.** Svenbro (in Cavallo and Chartier 1999, pp. 51-52) allows that the absence of breaks "was not an insurmountable obstacle" to silent reading, but reports Saenger's view that in the Middle Ages "word division was a necessary condition for the dissemination of the practice of silent reading". +8. **Reference reading as caused by separation.** Rouse and Rouse (1982) explain the apparatus by readers' need to find things and do not mention separation; the paper predates Saenger's book, so it is not a reply to him. + +--- + +## (d) Gaps and unverified leads + +**Could not open; where a human with library access finds them.** +- Cribiore 1996 (closed; Unpaywall) and 2001: read the sections on syllabic division and word separation, and the Hero passage, for her stated reasons and the length of each stage. +- Marrou, *History of Education in Antiquity*, the Hellenistic reading-lesson chapter: what "these" refers to in "It was not so easy for the child to read these". +- Saenger 1997 p. 3 (Japanese kana) and p. 6 (novices), p. 85 ("shortcuts"; page confirmed by the auditor), p. 264 ("Private Study and Heresy"), p. 265: page numbers for every K15 claim marked "page not available". +- Saenger 1982 (*Viator* 13, 367–414): Brepols bot wall; the open "bronze" copy at `brepolsonline.net/doi/epdf/10.1484/J.VIATOR.2.301476`. +- Saenger 1999 in Wagner, Venezky and Street (1999), ch. 2: confirm start and end pages; check the sentences on diffusion in print. +- Cavallo and Chartier 1999, Introduction: which editor wrote the "line and page" passage, and its page. +- Hanna (*Speculum* 74(3), 830–831, DOI 10.2307/2886848): her evaluation, not just her summary. Bäuml (*Written Language & Literacy* 5(1), 131–134, DOI 10.1075/wll.5.1.08bau). Gilliard 1993. Lyons 2001 review. +- Illich 1993 and Mak 2011: no route to the text. Driver's *Speculum* review of Mak (DOI 10.1017/s0038713413001322). +- Parkes 1976, on *ordinatio* and *compilatio*, and Rouse and Rouse 1982 in print: for whether either mentions separation in the body. +- The *Raritan* reviewer and the *Notes and Queries* reviewer: names. +- Stutzmann, any dated-manuscript test of Saenger's chronology (facet A could not find one). +- Dionysius of Halicarnassus: the Greek; Usher's Loeb translation, which supersedes Roberts for any quotation in the paper. +- Dionysius Thrax: the authenticity literature, before the paper cites the *Technē* as ancient evidence. + +**Unverified or weak.** The joins of snippets in Marrou, Johnson, Cribiore 1996, Cavallo and Chartier and Baron are mine; a print copy should confirm each sentence. Nothing in this facet is quotable in the paper at "full text" depth, by the repo's rule (`../algorithmacy_ot/README.md:29-41`), except Quintilian, Roberts's Dionysius and Davidson's Thrax. + +**Not found despite search.** Any study measuring the length or cost of learning to read spaced against unspaced Latin for medieval learners; any modern experiment on Latin or Greek novices. The closest modern evidence is in facet D (children and second-language readers). + +## Changes after audit (2026-10-13) + +- Q1, length of training: "None gives a length of training in years" narrowed to "a length for learning to read alone"; Cribiore (2001, p. 171) gives "one or two years of elementary instruction" for many students. +- Q1 point 1, "teachers took the aid away": now says syllable division gave way to ordinary texts, while word division continued "as a further aid" until accents (Cribiore 1996, p. 48). The Cribiore 1996 pieces (syllable exercises; "spaces or bars" at p. 72; 2001 "bars or dots") are no longer run together, and "advanced students in scriptio continua" is limited to "some" models. +- Q1 point 2: Dionysius is an analogy for composition and says nothing about word division; the "end of training, not its obstacle" heading is marked as my inference. Johnson's "years of advanced study" describes the literature, not the script. "Marrou is the only source" changed to "the plainest source", since Cribiore (1996, p. 48) also reports the difficulty. +- Q1, praelectio: not in Quintilian 1.8.1-2; the point now rests on Quintilian's "can only be taught in actual practice", and the praelectio is credited to Bonner. +- Q2, Saenger on novices: the "syllabary" sentence is about Japanese kana (p. 3), not pre-separation Greek or Latin teaching; removed as evidence about Latin teaching and replaced with the p. 6 statement. +- Q2, Orme and Reynolds: "for many" restored in Orme (2006); Reynolds' inner quotation marks dropped and flagged unverified. +- Q3, Cavallo and Chartier: the two quotations are Svenbro's. Removed from "Against the wide form" (they say nothing on literacy diffusion) and moved to "For the narrow form", because Svenbro's next sentence cites Saenger for word division as "a necessary condition for the dissemination of the practice of silent reading". The Q5 table row and claim 7 changed to match. +- Q3, Saenger 1999: "marginal to his chapter" replaced by his actual statement that class and gender are "of only marginal import to the history of reading"; the scope disclaimer does not settle whether Saenger links ease of reading to literacy, and "I found no work that ... claims it for populations" now carries an exception for Saenger 1997 pp. 11 and 13. +- Q3, "Silent where it should speak": now states the OCR zero hit and that Riché and Grendler predate Saenger's book. +- Q3 closing sentence: the "shortcuts" sentence is about Insular books and an elite's reading skills; the learner argument is on p. 91. +- Q5, Baron: the scribes are Insular (Irish), not Anglo-Saxon, and the Irish-language copying result is the reverse of the earlier reading. +- Q5 closing sentence: "the only causal claim a reviewer endorses as 'convincing' and 'plausible'" replaced; "convincingly" (Richter) is about the Romance lag and "much more plausible" (Aronoff) is comparative and hedged. +- Q6 row 1 (Insular separation to ease reading for Latin learners): (a) supported changed to (b) claimed, endorsed by two reviewers as plausible, untested. Reynolds and Orme show only that Latin was foreign, and no source tests it. +- Q6 row 2 (ancient teachers): restated as syllables, and sometimes words, divided for novices; syllable division gave way to ordinary texts. Class (a) kept. +- Q6 row 3 (shortcuts): "no reviewer comments" changed to "none in the snippets I could read (Hanna and Baeuml unread)", marked unverified. +- Q6 row 5 and claim 8 (reference reading): "rival explanation" withdrawn; the Rouses' 1982 paper predates Saenger's book, so it is not a reply to him. +- Q6 row 7 and the supported-sentence list item 4 (literacy): "claimed by nobody found" changed to "no source I read says separation raised literacy", with Saenger's p. 11 and p. 13 remarks stated. +- Claims 1 and 3: Orme's "for many" restored; the Reynolds marks flagged. + + +## Changes after full-text cards (2026-10-13) + +Source: `audit/CARD_FACET_CHECKS.md`, Facet K section and the B-section notes that name Facet K; the card, including its S2 section, is the authority. + +- `quintilian1920institutio`: 1.1.34 is on p. 37 (not pp. 37–39); 1.8.1–2 on p. 147 (not p. 148); 1.7.17 (p. 139) added: he judges a spelling convention by whether it hinders learners, and 1.7 says nothing about spacing; read status corrected from "read in full" to read in part (card), since the rest of Book 1 was not read. +- `dionysius1910composition`: chapter XXV runs pp. 251–270, not "between running heads 265 and 269"; the reading passage is on p. 269 and the tablet story on pp. 265–267; "says nothing about word division" confirmed. Read in part (card). +- `thrax1874grammar`: quotations confirmed; read status corrected from "read in full for §§1–2" to read in part (card: §§1–14 and the closing section; §§15–24 skimmed); "attributed" kept. +- `knox1968silent` (K rows on silent reading): Knox never mentions word separation or the training of readers; the two table rows now say so. Read in full (card). +- `odonnell2000saenger` (same rows): quotation now exact (¶3 "some will dissent from … necessary and sufficient"; ¶4 "most venturesome and vulnerable part"); he calls the chronicle sound and the causal claim the likelier to fail. Read in full (card). `odonnell1994parkes` is not cited in this facet; the card's correction (the *Viator* article is cited in the 2000 review, not the 1994 one) concerns facet B, so no change here. +- `ker2012johnson`: not cited in this facet; the card's note (Ker does not discuss training of novices, so no purchase on Johnson's "emerging readers" sentence) needs no change. + +Synthesis verdicts changed: none. Q1 still reads that the ancient sources place the burden in a long drill, not in the page; 1.7.17 adds that Quintilian weighs a convention by its cost to learners, but he never applies the test to spacing. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_L_disclosure_and_choice.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_L_disclosure_and_choice.md new file mode 100644 index 00000000..be2e7ec8 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_L_disclosure_and_choice.md @@ -0,0 +1,333 @@ +# L — Disclosure and choice: the literatures on disclosure the project has not yet covered + +*Facet L, 2026-10-13. Agent-drafted. The author has not reviewed it, and none of the text is the author's.* + +*Audited 2026-10-13 (AUDIT_07.md); WRONG and OVERSTATED rows corrected in place.* + +**What this facet covers.** Section 4 of the author's outline ("the failure of disclosure"), and the ruling behind it (SOURCE.md, D3): "disclosure of the algorithm doesnt affect our ability to navigate it. disclosure is a literacy based intervention, not an algorithmacy based one." Five literatures: (Q1) the law-and-economics and behavioural critique of mandated disclosure, with the targeted-transparency work that says when disclosure works; (Q2) privacy notices and consent, the largest tested case of disclosure; (Q3) nudging and boosting, the nearest rival to "design, don't inform"; (Q4) meta-analyses of explainable-AI and human-AI decision studies; (Q5) any study of disclosure or transparency in a two-sided, mediated or strategic setting with a behavioural outcome. + +**The question the synthesis answers.** What do these literatures say about whether disclosure changes what people do, by what mechanism it fails or succeeds, whether the nudge/boost distinction already draws the paper's line, and whether anyone has tested disclosure where two humans coordinate through an algorithm. Where they cut against the paper, I say so. + +**Legend for read status.** +- *Read in full*: I opened the whole text this session and read it through, or searched it end to end for the passages I cite. +- *Read in part*: I opened the text and read named sections; the entry says which. +- *Abstract only*: a database abstract (Crossref, Consensus, a publisher page) and nothing else. Every number comes from that abstract, and the entry says so. +- *Not accessible*: no text. The source appears only as a lead. +- *Held*: already in `references.bib` and `library/`; I cite it by key and reread nothing. + +Page locators come from the page markers in the copy I read. Where the copy is an author or working-paper version, the entry says so; every quotation needs a check against the version of record before print. All PDFs sit in `~/iit-playground/dxd_foundation_local/pdfs/L/`. + +--- + +## (a) Search log + +| # | Source | Exact query or request | Hits | Notes | +|---|---|---|---|---| +| 1 | Repo | Read RESEARCH_PLAN.md, SOURCE.md, T_adjudication.md §9, library/INDEX.md (251 cards), H_candidate_moves.md header and §1 table, X_counter_case.md index lines; grep of `references.bib` and facets for Ben-Shahar, Loewenstein, Fung, McDonald, nudge, boost, Schemmer, Vaccaro, Thaler, Hertwig, Lai | n/a | Held: Kelley 2009, Seira 2017, Kling 2012, Mejia 2021, Li 2026, Jhaver 2019 (survey), Poursabzi 2021, Vasconcelos 2023, Bucinca 2021, Cabrera 2023, Dietvorst 2018. **Not held:** every Q1 and Q3 source, McDonald, Obar, any meta-analysis. `references.bib` has no Hertwig on boosting (only Hertwig and Erev 2009). **Lai et al. 2023 is not in `references.bib`, `library/` or any facet `.bib`, although the brief says it is held.** I read it (entry 29). | +| 2 | OpenAlex | 11 title searches for the core sources | 0 | Every call returned "Rate limit exceeded ... daily budget used up". Not retried (resets midnight UTC). | +| 3 | Crossref `query.bibliographic` | 28 title and author queries across Q1–Q5 (Ben-Shahar and Schneider; Loewenstein, Sunstein and Golman; Willis; Weil et al.; McDonald and Cranor; Obar and Oeldorf-Hirsch; Johnson et al.; Hertwig and Grüne-Yanoff; Vaccaro et al.; Fernandes et al.; Bakos et al.; Acquisti et al.; Utz et al.; Mertens et al.; Maier et al.; DellaVigna and Linos; Yeung; Lorenz-Spreen et al.; Kozyreva et al.; Bar-Gill; Thaler and Sunstein) | 2–3 each | Gave the DOIs. Several calls returned HTTP 429 and I reran them after a pause. **Not found in Crossref:** McDonald and Cranor (I/S, no DOI), Willis 2008 (Iowa Law Review, no DOI), Thaler and Sunstein 2008 (only book reviews). | +| 4 | Crossref `/works/{doi}` | 38 DOIs, one by one | 37 resolved at first pass; Vaccaro 429 then resolved on rerun | Titles compared with the entries below. Two corrections: Normann and Sternberg's SSRN title differs from the title Consensus shows; Bauer et al. is dated 2024 in Crossref (vol. 35). | +| 5 | WebSearch | "Ben-Shahar Schneider The Failure of Mandated Disclosure University of Pennsylvania Law Review 159 pdf" | 10 | Found Chicago Unbound, Michigan working paper, an IFS-hosted offprint. | +| 6 | Direct download | Michigan repository, Chicago Unbound (bepress) | HTML, not PDF | bepress returned an HTML wall. Not retried or worked around. | +| 7 | Direct download | `ifs.org/wp-content/uploads/2012/11/Ben-Shahar-2011-Failure.pdf` | 103-page PDF | The offprint's text layer is a substituted-glyph font, unreadable. I OCRed it with tesseract at 150 dpi and read the OCR. Quotations below come from the OCR; I checked the p. 743 passage against the page image and the rest only against the OCR text, so verify each against the journal. | +| 8 | WebSearch | "Loewenstein Sunstein Golman Disclosure: Psychology Changes Everything pdf" | 10 | CMU author copy opened. | +| 9 | WebSearch | "Weil Fung Graham Fagotto The effectiveness of regulatory disclosure policies" and "Fung Graham Weil Full Disclosure ... targeted transparency" | 9 and 9 | Journal article paywalled (Wiley); full book PDF is on the Harvard Ash Center site (304 pp.). | +| 10 | Direct download | `transparencypolicy.net/...pdf` | 4-page Taubman policy brief only | Not the journal article. I did not use it. | +| 11 | WebSearch and download | Hertwig and Grüne-Yanoff 2017; Grüne-Yanoff and Hertwig 2016 | author-hosted PDFs on a UC San Diego faculty page | Springer returned a script challenge page and WebFetch hit a redirect to an identity provider; I stopped there. The UCSD faculty copies are author-hosted. | +| 12 | WebSearch | "Johnson ... Beyond nudges: Tools of a choice architecture Marketing Letters 2012 pdf" | 10 | Only repository records, no open PDF found (St Gallen, Erasmus, Princeton records opened; no file link). Publisher PDF not tried (Springer wall). | +| 13 | Europe PMC REST | DOI lookups for Mertens et al. and Maier et al. | PMC8740589, PMC9351501 | Read from PMC full-text XML. | +| 14 | arXiv API | titles for Schemmer, Lai, Utz, Nouwens, Cen, Erlei, Jhaver, Bauer, Hagenbach | found Schemmer 2205.05126, Lai 2112.11471, Utz 1909.02638, Nouwens 2001.02479, Cen 2405.05596 | Erlei, Jhaver, Bauer, Hagenbach not on arXiv by that search. | +| 15 | Direct download | AAAI OJS (Erlei), ACM DL (Jhaver), SSRN (Bakos; Obar) | 404, HTML wall, 403 | Blocked; recorded. WebFetch of SSRN returned HTTP 403. | +| 16 | Consensus | "effect of disclosing ranking algorithm criteria to sellers on seller behavior field experiment marketplace" | 20 | Q5. Found Klein et al., Hui et al., Dash et al., Nosko and Tadelis (held). Abstracts only. | +| 17 | Consensus | "algorithmic management transparency gig workers experiment effect on worker behavior earnings" | 20 | Q5. All but one hit is survey or interview work (Li Y-N et al. 2025; Gong 2025; Wiener et al. 2021; Cram et al. 2022). Rahman and Bucher et al. are qualitative. None randomises transparency. | +| 18 | Consensus | "explanation of content moderation decisions randomized experiment subsequent user rule-following behavior" | 20 | Q5. Found Jhaver et al. 2019, Tyler et al. 2019, Jhaver et al. 2024, Srinivasan et al. 2019. | +| 19 | Consensus | "disclosing the algorithm to strategic users: field or lab experiment on gaming and behavior change when scoring rule is revealed (credit, ranking, ratings)" | 20 | Q5. Found Cen et al. 2024, Bauer et al. 2024, Hagenbach et al. 2025, Cofone and Strandburg (not carded). Most hits are strategic-classification theory. | +| 20 | Consensus | "ride-hailing drivers pay transparency or dispatch algorithm information randomized experiment driver behavior" | 20 | Q5. Found Mejia and Parker, Li et al. 2026 (both held). Others are models or surveys. Angrist et al. 2021 randomises a pay contract, not information. | +| 21 | Consensus | "experiment human negotiation or market through algorithmic intermediary with disclosed versus undisclosed rules two human participants" | 20 | Q5. Found Erlei et al. 2020, Casella et al. 2025, Normann and Sternberg 2021, Ozkes et al. 2024. | +| 22 | Scholar Gateway | one natural-language query on disclosure to the parties of a mediated transaction | `ACCESS_DENIED` | Failed, as in earlier runs. | +| 23 | WebSearch | `"decision experience design" OR "decision-experience design" choice architecture nudge algorithmic` | 10 | No page uses the exact phrase. Nearest: Mills and Sætra, "The autonomous choice architect" (AI & Society, 2022; abstract in a search snippet only), and digital-nudging papers (Weinmann et al., Münster). | +| 24 | Crossref | `"decision experience design"`; `"decision architecture" algorithmic intermediary`; `digital nudging choice architecture algorithmic intermediary two-sided` | 4 each | No title uses "decision experience design". "Decision architecture" returns an AI-governance paper (MDPI AI 2026) and an intermediary-liability chapter; neither opened. | + +**Searches that found nothing for Q5.** (i) A randomised or quasi-experimental study in which the rule of an algorithmic intermediary is disclosed to two human parties who coordinate through it, with a behavioural outcome for both: none in searches 16–21. (ii) A field test of ranking-criterion disclosure to sellers (Amazon, Booking, Airbnb, Etsy): none; hits were models or interview studies. (iii) A randomised test of dispatch or pay-algorithm disclosure to ride-hail or delivery workers: none; Li et al. 2026 (held) is a natural experiment in which transparency came bundled with a pay guarantee. Search results are not proof of absence; the exact queries are above. + +--- + +## (b) Entries + +Entries follow the questions. For each: reference, key, verification, read status, what it argues, quotations, bearing. "Bearing" says how the entry bears on D3 ("disclosure is a literacy-based intervention and does not affect our ability to navigate the algorithm"), on H-cost and H-cause where relevant, and on the four problems. + +### Q1. The critique of mandated disclosure + +**1. Ben-Shahar, O., & Schneider, C. E. (2011). The failure of mandated disclosure. *University of Pennsylvania Law Review, 159*, 647 ff. (final page and issue number not verified).** Key `benshahar2011failure`. +- *Verified:* Crossref resolves the 2010 SSRN working-paper record (10.2139/ssrn.1567284), which carries the same title and authors. The 2011 article has no DOI that Crossref returns; the Chicago Unbound record gives the volume and page (search 5). +- *Read status:* **read in full**, from a 103-page offprint hosted at ifs.org (a text layer I could not use; I read an OCR of all pages). Page numbers are the law review's own, taken from the running heads. The offprint is not the publisher's page, so check each quotation against the journal. +- *What it argues:* Mandated disclosure is a regulatory technique that "chronically fails." The evidence is that disclosers do not always disclose, disclosees do not read, understand or use, and decisions do not improve; the cause is false assumptions about people, about the decisions, and about the long chain of things that must go right. +- *Quotations.* + - "Although mandated disclosure addresses a real problem and rests on a plausible assumption, it chronically fails to accomplish its purpose." (p. 651) + - "We are not asking what information people need to make good decisions. We are asking whether a regulatory technique—mandated disclosure—works." (p. 651) + - "disclosees often do not read disclosed information, do not understand it when they read it, and do not use it even if they understand it." (p. 665) + - "Financial literacy education is an important component of the disclosure and informed consent paradigm, and much evidence now shows that this effort has largely failed." (p. 667; the footnote cites Willis 2008, pp. 208–09) + - On simplification: "for decades experts have labored intelligently and earnestly to present complex information accessibly, and it is now clear that only modest progress is possible." (p. 713) And: "Unfortunately, complexity cannot be explained simply." (p. 713) + - "Our task is not to propose an alternative." (p. 651) In the conclusion: "We might have had a fifth purpose: to prescribe a regulatory alternative." (p. 742) + - **The concession.** "brief, simple, easy disclosures are at least preferable ... perhaps to the point of using symbols instead of sentences." Los Angeles grade cards "seem to have influenced consumers—and, in turn, led to cleanlier restaurants. This is information that anyone can understand, and the single datum is enough to affect customers' choice of restaurants." (p. 743) +- *Bearing.* **Supports** D3: the mechanism they give is that disclosure hands a text to a reader and the reader cannot or will not read it, and their treatment of literacy and financial-literacy education as the weak complement is the very pairing the paper draws. **Qualifies** it in two ways. First, they say disclosure "regularly fails", not that it has no effect; their target is a regulatory technique, and they decline to say what replaces it. Second, their positive examples are the Los Angeles grade cards (offered with a hedge: "There is little real evidence, but we find some reason to think", p. 743) and Medicare satisfaction ratings (p. 743), and on pp. 744 and 748 they add that grade cards also work by making disclosers "try to get their grades up", so that "disclosees need not rely on the disclosed data". *My inference, not theirs:* that a grade is a re-representation that removes the reading task, the author's type (i) move. Their own stated reason is that the datum is brief and simple ("information that anyone can understand", p. 743). A paper that quotes them should say they support "the reader cannot be made the remedy" and that their positive examples are brief, symbolic signals whose effect may run partly through the discloser. + +**2. Loewenstein, G., Sunstein, C. R., & Golman, R. (2014). Disclosure: Psychology changes everything. *Annual Review of Economics, 6*, 391–419.** Key `loewenstein2014disclosure`. DOI 10.1146/annurev-economics-080213-041341 (Crossref confirms title, authors, volume, pages). +- *Read status:* **read in part** (author copy from CMU, 32 pp.): abstract, §3 opening, §3.8, §§4.1–4.8, §6, and the passages on privacy notices. I did not read §§3.2–3.7 line by line. +- *What it argues:* Disclosure rules work through two channels, the recipients' choices and the disclosers' reactions. Psychology weakens the first (limited and motivated attention, biased probabilities) and often helps the second. +- *Quotations.* + - "In many cases, disclosure does not much affect the recipients of the information but does significantly affect the behavior of the providers, sometimes for the better and sometimes for the worse." (abstract, p. 391) + - "A pervasive theme is that disclosure may have little effect on recipients but large effects on providers." (p. 398) + - "In the classic account of how disclosure works (e.g., Fung et al. 2007), mandated disclosure leads to changes in the behavior of disclosees, which in turn cause disclosers to clean up their act." (p. 403; the telltale-heart effect, §3.8) The next sentences name the Los Angeles hygiene grades as "a case often cited as a paradigm for successful disclosure" and add: "But in many situations (to some extent including the case of restaurant hygiene ratings), an industry response can be found amid little evidence of a consumer response" (p. 403); footnote 8 says Jin and Leslie found the grade cards "also led inspectors to distort their ratings". + - Kling et al. (held): "Plan switching was 28% in the intervention group, but only 17% in the comparison group ... this intervention combined a number of different features (comparative and personalized information), so we cannot isolate a single mechanism that explains its effectiveness" (p. 407). + - Choi et al. 2010 and Beshears et al. 2010 on mutual funds: "These findings suggest that making disclosure simple and accessible may provide insufficient help to consumers who lack the tools to process that information." (p. 408) + - On the privacy notices that "no one reads on Internet sites, the volume of information is overwhelming and not worth the investment one must make to read it." (p. 411) + - On intermediaries: "Instead of attempting to provide information directly to consumers, disclosure requirements could make information available in standardized formats so that intermediaries can arise to process it, make sense of it, and (perhaps for a fee) provide it in a form that is usable to its end users." (p. 411, §4.6) + - "policy makers need to be vigilant against the risk that mandatory information disclosure policies will be implemented as a substitute for other, often more effective, regulatory interventions" (p. 412, §4.8) +- *Bearing.* **Supports** D3 on the recipient side, and gives the paper a second effect it can use: disclosure acts on the discloser. Applied to Nadella (P15, P19), the discloser is the model's operator, which is the party who holds the log. **Qualifies** D3 by reporting that "simplified disclosure, standardized disclosure, vivid disclosure, and social comparison information can all be used to enhance the effectiveness of disclosure policies" (abstract, p. 391) and that an intermediary can do the reading for the person. The intermediary route is a form of counter-delegation (type iii), so the review gives the paper a bridge from "disclosure" to its own remedies, and also a warning: once an intermediary reads the disclosure on the person's behalf, the paper's triad has acquired a fourth party. "Doesn't affect" is stronger than their claim, as T §9.2 already noted. + +**3. Fung, A., Graham, M., & Weil, D. (2007). *Full disclosure: The perils and promise of transparency*. Cambridge University Press.** Key `fung2007full`. DOI 10.1017/9780521699617 (Crossref confirms). +- *Read status:* **read in part** (full 304-page PDF from the Harvard Ash Center site; read ch. 1, ch. 4 in the sections on user and discloser embeddedness, and Tables 4.2–4.3). Page numbers are the book's printed pages, which fall 20 behind the PDF pages. +- *What it argues:* A "targeted transparency" policy has to put the information inside the user's existing decision routine and the discloser's routine ("double embeddedness"). Whether users act depends on perceived value, compatibility with the routine, and comprehension. +- *Quotations.* + - "Transparency systems alter decisions only when they provide pertinent information that enables people to substantially improve their choices without imposing significant additional costs." (p. 55; the next sentence contrasts window-front grades in Los Angeles with Internet databases: "the added cost of obtaining new information ... is very low") + - "For transparency systems to be effective, it is necessary but not sufficient that information become embedded in existing decision-making processes." (p. 55) + - "Both systems convert data and expert interpretation into simple normative signals such as stars or letter grades." (p. 57; on the Los Angeles hygiene grades and rollover ratings; the text adds that "underlying details can be accessed by those who want to study them") + - "Information is comprehensible when users have the capacity to relate it to the decisions they face." (p. 59). Workers "were able to grasp only about 60 percent of the information" in hazard data sheets (p. 59). + - "Highly effective transparency policies, then, are doubly embedded." (p. 65) +- *Bearing.* **Qualifies** the paper's line. Fung et al. give the success condition as a design property of the disclosure (compatibility with the routine, a simple signal at the point of choice), not as reader skill. That is the paper's position about moving work from the reader to the artefact, reached from the transparency side. Their lead success, the Los Angeles grades, is the case Ben-Shahar and Schneider use. Fung et al. themselves say a rating system is "one way to simplify the format of complex information" (p. 57) and that effectiveness needs double embeddedness, so the discloser's response counts as much as the user's (p. 65). *My inference, not theirs:* that the best disclosures therefore work "by re-representation, not by reading". The sources support the first half (simple, low-cost signals); neither states the second, and a letter grade is still read. + +**4. Weil, D., Fung, A., Graham, M., & Fagotto, E. (2006). The effectiveness of regulatory disclosure policies. *Journal of Policy Analysis and Management, 25*(1), 155–181.** Key `weil2006effectiveness`. DOI 10.1002/pam.20160 (Crossref confirms, dated 2005 online). +- *Read status:* **not accessible** (Wiley wall). The four-page Taubman brief of the same authors (2005) was opened and not used. I rely on the book (entry 3), which covers the same cases. + +**5. Fernandes, D., Lynch, J. G., Jr., & Netemeyer, R. G. (2014). Financial literacy, financial education, and downstream financial behaviors. *Management Science, 60*(8), 1861–1883.** Key `fernandes2014financial`. DOI 10.1287/mnsc.2013.1849 (Crossref confirms). +- *Read status:* **read in part**, and only the **June 2013 working-paper version** (103 pp., NEFE report; page numbers are its own). I read the abstract, the decay analysis and the policy discussion. I did not read the three empirical studies. +- *What it argues:* A meta-analysis of financial-literacy interventions finds a small effect that decays. +- *Quotations.* + - "We find that interventions to improve financial literacy explain only 0.1% of the variance in financial behavior." (abstract, p. 2) + - "at delays of 20 months or greater, there is no significant effect of even 24 hours of instruction" (p. 14) + - "Others have advocated defaults, 'nudges', and 'choice architecture' ... as less costly and more effective alternatives to financial education ... But defaults work best when almost all consumers have similar needs. When consumers' needs are heterogeneous, one needs to know something to decide for oneself. Here, 'just-in-time' financial education may have promise, alone and embedded in decision support systems that help select tailored options." (p. 22) +- *Bearing.* The best quantified evidence that "teach the person" is a weak remedy (**supports** the paper), and the authors' own recommendation, education delivered at the moment of decision and embedded in decision support, is a case for a hybrid. They also name the limit of defaults: defaults fail when needs differ. *My inference, not theirs:* the paper's triad is such a case, because each party's intent differs. Treat this as a counter-case to "design, don't educate" for heterogeneous needs. + +**6. Willis, L. E. (2008). Against financial-literacy education. *Iowa Law Review, 94*, 197.** Key `willis2008against`. +- *Read status:* **not accessible**. I know it only from Ben-Shahar and Schneider's footnote (p. 667, n. cited at "94 IOWA L. REV. 197, 208-09 (2008)", "surveying studies finding that financial education programs have no effect or small paradoxical results"). Crossref has no record. A human with a law-library account should fetch it. + +**7. Bar-Gill, O. (2012). *Seduction by contract: Law, economics, and psychology in consumer markets*. Oxford University Press; and Bar-Gill, O. (2019). Algorithmic price discrimination: When demand is a function of both preferences and (mis)perceptions. *University of Chicago Law Review, 86*, 217.** Keys `bargill2012seduction`, `bargill2019algorithmic`. +- *Read status:* **not accessible** (book DOI 10.1093/acprof:oso/9780199663361.001.0001 resolves at Crossref; the 2019 article's abstract came from a search result and not from a database record I opened). The 2019 abstract, as returned by search, says the normative verdict depends on whether discrimination targets preferences or "(mis)perceptions". Do not quote. It is the nearest law-and-economics treatment of an algorithm that exploits misperception, so it belongs in the reading list for §4 and §7. + +### Q2. Privacy notices and consent + +**8. McDonald, A. M., & Cranor, L. F. (2008). The cost of reading privacy policies. *I/S: A Journal of Law and Policy for the Information Society, 4*(3), 543–568.** Key `mcdonald2008cost`. No DOI (Crossref has none). AUDIT_07 confirmed volume, issue and pages from the journal-paginated copy at the Stanford repository. +- *Read status:* **read in part** (the 22-page author's pre-press version; abstract, method and the cost estimate). The numbers are the authors' own modelled estimates: reading time from the word counts of 75 popular sites at 250 words per minute (the 244-hour point estimate), skimming time from a 212-person study (the 154-hour estimate). The 244 hours is what a person would spend *if* they read every policy word for word at every new site (the authors' own framing: "if website users were to read the privacy policy for each site they visit just once a year"); it is not an observed behaviour. +- *Quotations.* "We calculated the average time to read privacy policies in two ways." "...the national opportunity cost for just the time to read policies is on the order of $781 billion." (abstract, p. 1) "...using the point estimate of 244 hours per year to read privacy policies per person means an average of 40 minutes a day." (p. 18 of the pre-press; p. 563 of the journal) +- *Bearing.* **Supports** the cost side of D3 with a modelled figure for the reader's task. It models time, not behaviour or welfare; it is a cost estimate, not an effect. + +**9. Obar, J. A., & Oeldorf-Hirsch, A. (2018/2020). The biggest lie on the Internet: Ignoring the privacy policies and terms of service policies of social networking services. *Information, Communication & Society, 23*(1), 128–147.** Key `obar2018biggest`. DOI 10.1080/1369118x.2018.1486870 (Crossref confirms; the year on the article is 2018 online, 2020 in print). +- *Read status:* **abstract only, and not even that.** The publisher page and SSRN were walled (searches 15). Crossref holds no abstract. I saw a search-engine summary reporting 543 students, a fictitious network, and 74% choosing "quick join"; I did not verify these figures. **Do not cite numbers from this entry.** A human with library access should read the paper. + +**10. Bakos, Y., Marotta-Wurgler, F., & Trossen, D. R. (2014). Does anyone read the fine print? Consumer attention to standard-form contracts. *Journal of Legal Studies, 43*(1), 1–35.** Key `bakos2014does`. DOI 10.1086/674424 (Crossref confirms). +- *Read status:* **not accessible** (SSRN delivery returned HTML). Title and DOI only. + +**11. Utz, C., Degeling, M., Fahl, S., Schaub, F., & Holz, T. (2019). (Un)informed consent: Studying GDPR consent notices in the field. *Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security*, 973–990.** Key `utz2019informed`. DOI 10.1145/3319535.3354212; arXiv 1909.02638 (Crossref confirms). +- *Read status:* **read in part** (18-page arXiv v2: abstract, §1 summary, experiment descriptions and figure data; the discussion not read closely). +- *What it argues:* Three field experiments with "more than 80,000 unique users" on one German website vary where the consent notice sits, what choices it offers, and how the choices are framed. +- *Quotations.* "...seemingly small implementation decisions can substantially impact whether and how people interact with consent notices." (abstract) "...nudging visitors to accept privacy-invasive defaults leads more visitors to accept cookies, whereas in a privacy-by-default (opt-in) setting, less than 0.1 % of visitors allow cookies to be set for all purposes." (§1) +- *Bearing.* The consent decision moves with the layout of the notice, not with what the notice says. **Supports** the paper's contrast: the lever is the interface. It also shows the cost of that lever: the same design power produces the dark patterns of entry 12. Design is a tool that either party to the notice can aim in either direction. + +**12. Nouwens, M., Liccardi, I., Veale, M., Karger, D., & Kagal, L. (2020). Dark patterns after the GDPR: Scraping consent pop-ups and demonstrating their influence. *Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems*, 1–13.** Key `nouwens2020dark`. DOI 10.1145/3313831.3376321; arXiv 2001.02479 (Crossref confirms). +- *Read status:* **read in part** (abstract, introduction, related-work summary of Utz et al.). +- *Quotations.* "We found that dark patterns and implied consent are ubiquitous; only 11.8% meet the minimal requirements that we set based on European law." "...notification style (banner or barrier) has no effect; removing the opt-out button from the first page increases consent by 22–23 percentage points; and providing more granular controls on the first page decreases consent by 8–20 percentage points." (abstract). The field experiment had 40 participants. +- *Bearing.* As entry 11, with a smaller sample (40 participants; the 22–23 points are the effect of removing the "reject all" button from the first page, 95% CI 14.9 to 29.0 points). The choice set differs between conditions, so the information offered is not strictly constant. + +**13. Acquisti, A., Brandimarte, L., & Loewenstein, G. (2015). Privacy and human behavior in the age of information. *Science, 347*(6221), 509–514.** Key `acquisti2015privacy`. DOI 10.1126/science.aaa1465 (Crossref confirms; abstract from Crossref). +- *Read status:* **abstract only.** Crossref abstract: privacy behaviour turns on "people's uncertainty about the consequences", "the context-dependence of people's concern", and whether concerns are "malleable—manipulable by commercial and governmental interests". Nothing more used. + +### Q3. Nudging, choice architecture and boosting + +**14. Hertwig, R., & Grüne-Yanoff, T. (2017). Nudging and boosting: Steering or empowering good decisions. *Perspectives on Psychological Science, 12*(6), 973–986.** Key `hertwig2017nudging`. DOI 10.1177/1745691617702496 (Crossref confirms). +- *Read status:* **read in full** (author-hosted PDF, 14 pp.). Page numbers follow the page markers in that copy and are probably right to the page; check them against the journal. +- *What it argues:* Behavioural policy has two families. Nudges "steer people in a particular direction while preserving their freedom of choice"; boosts "foster people's competence to make their own choices". The two differ on seven dimensions: immediate target, research programme, causal pathway, assumed cognitive architecture, reversibility, programmatic ambition, normative implications. +- *Quotations.* + - "we define boosts as interventions that target competences rather than immediate behavior ... A boost may enlist human cognition (e.g., decision strategies ...), the environment (e.g., information representation or physical environment), or both." (p. 977) + - "In sum, the SMT nudge does not aim to foster people's competences. Instead, it skillfully designs an external choice architecture ... the nudge approach steers behavior without taking the detour of honing new competences, whereas the boost approach invests in building on and developing people's competences." (p. 978) + - Reversibility test: "If, ceteris paribus, the policy maker eliminates an efficacious ... behavioral intervention and behavior reverts to its preintervention state, then the policy is likely to be a nudge. If ... behavior persists when an intervention is eliminated, then the policy is more likely to be a boost." (p. 981, the rule begins at the foot of p. 980) + - Qualification the authors attach to that test: "nudges that affect behavior repeatedly may produce behavioral routines through learning that 'survive' the removal of the nudge in the choice architecture. In such cases, our empirical criterion indicates that the nudge intervention has a boosting 'side effect'" and "The nudge has thus turned into a boost and had lasting effects." (p. 981) + - On disclosure: "Sunstein (2016) introduced the notion of educative nudges, citing reminders, warnings, and information such as nutrition labels as examples. In our view, educative nudges and short-term boosts largely overlap. Both ... require—in contrast to classic nudges, such as defaults—a modicum of motivation and cognitive skill." And: "The mere provision of information is often not enough. Health statistics or nutritional information, for instance, bring no benefits if they are intransparent (e.g., reliant on conditional probabilities), overwhelming (like software license agreements), or misleading." (p. 977) + - "competences are often best fostered by redesigning aspects of individuals' external environment or by teaching them how to redesign them." (p. 980) + - On the 2008 book: "words such as 'competence,' 'knowledge,' 'skills,' and empowerment do not even feature as entries in the book's index." (p. 981) +- *Bearing on Q3.* **Does the distinction draw the paper's line? Partly, and the match is not clean.** + 1. *Where it matches.* The nudge side, "an external choice architecture" that works with the person's processes untouched, is the paper's "change the interface so that reading no longer requires the training". The reversibility test gives the paper a check. *My inference, not theirs:* spacing persists in the page whether or not the reader is trained, and (from the held reading-science results) the benefit goes if the page loses its spaces, which is the nudge signature. The authors' own qualification (p. 981) cuts the other way for H-cause: a nudge met repeatedly can build routines that survive its removal, which is a boosting side effect, so the test cannot settle H-cost against H-cause by itself. + 2. *Where it does not.* Hertwig and Grüne-Yanoff report Sunstein's (2016) category of "educative nudges" (reminders, warnings and nutrition labels; Sunstein adds disclosure requirements, p. 981) and say these "largely overlap" with *short-term* boosts: both require "a modicum of motivation and cognitive skill" (p. 977). They then set long-term boosts apart, as aimed at situations "that will not be reached by mandated disclosure requirements, warnings, and labels" (p. 981). The overlap is close to the paper's own D3 claim (disclosure is literacy-based) reached from the behavioural-policy side, in the authors' own terms, so the paper can cite it. But their *boost* is a competence-building programme (risk literacy, financial rules of thumb, representation training) that the paper would also class as literacy. So the boost/nudge split puts "literacy" on the boost side and "interface" on the nudge side only if one forgets that boosts may "enlist ... the environment (e.g., information representation)" (p. 976). The natural-frequency format of entry 15 is a re-representation that is also a boost. In their taxonomy the scribes' space would be a representation that matches the reader's cognitive algorithms and, if it raised reading skill that persisted, a boost; if it only lowered the cost for as long as the page stayed spaced, a nudge. The distinction does not decide the question the author asks, whether the work moves from person to artefact. The relevant axis in this literature is *reversibility*, not *person versus environment*. + 3. *The triad.* Both families assume one chooser and a choice architect whose objective is the chooser's good, or the policymaker's. Neither says what to do when the architect is an adaptive third party with its own objective that reads two choosers. See entry 20. + +**15. Grüne-Yanoff, T., & Hertwig, R. (2016). Nudge versus boost: How coherent are policy and theory? *Minds and Machines, 26*(1–2), 149–183.** Key `gruneyanoff2016nudge`. DOI 10.1007/s11023-015-9367-9 (Crossref confirms; online 2015, volume year 2016). +- *Read status:* **read in part** (author-hosted PDF, UCSD faculty page; abstract, §1, and the statistical-representation section; the heuristics literature in the middle not read). The copy carries odd-page markers only, so page ranges are approximate. +- *Quotations.* "Some authors have suggested the term 'educate' for these kinds of policies ... In our view, boosting goes beyond education and the provision of information. For example, in order to boost decision makers' skills, policy designers need to identify information representations that match the cognitive algorithms of the human mind, thus using the environment (e.g., external representations) as an ally to foster insight and decision-making skills." (pp. 151–152) On natural frequencies: "Changing the representation of statistical information in health brochures ... from probabilities to natural frequencies ... can be understood as an act of framing information ... The key difference to the nudge approach, however, is that statistical information is framed in such a way that people understand it" (pp. 157–158). "...difficulties understanding statistical information are seen not as an incorrigible mental deficiency of, say, doctors or patients, but as largely attributable to poor or intentionally misleading information." (p. 159) +- *Bearing.* The last quotation is the nearest sentence in this literature to the author's "the remedy was never to reserve reading for those who could bear the load but to change the interface until the load fell" (`draft.md:337`). The boost side blames the *format*, not the person, and fixes the format; but the authors add that "The boost approach does not stop at improving the format of statistical information. It also has an explicit educational goal" (p. 159 ff.), so the education is part of a boost, not optional. This is a **parallel** the paper can cite, and the authors' own wording ("poor or intentionally misleading information") fits an adaptive third party that writes the format. It also complicates the paper: the same authors use "boost" for a re-representation, which means the paper's "literacy versus algorithmacy" label would not map onto their "boost versus nudge" label. + +**16. Johnson, E. J., Shu, S. B., Dellaert, B. G. C., Fox, C., Goldstein, D. G., Häubl, G., Larrick, R. P., Payne, J. W., Peters, E., Schkade, D., Wansink, B., & Weber, E. U. (2012). Beyond nudges: Tools of a choice architecture. *Marketing Letters, 23*(2), 487–504.** Key `johnson2012beyond`. DOI 10.1007/s11002-012-9186-1 (Crossref confirms; 12 authors). +- *Read status:* **not accessible** (Springer wall; no open copy found). Abstract from the search results: the tools split into those "structuring the choice task" and those "describing the choice options". Do not quote. A human should read the section on describing options: it treats the format of an option description as a design variable, which is the nearest design-literature home for "re-representation". + +**17. Thaler, R. H., & Sunstein, C. R. (2008). *Nudge: Improving decisions about health, wealth, and happiness*. Yale University Press.** Key `thaler2008nudge`. +- *Read status:* **not accessible**; cited only as quoted in entry 14 ("choice architecture is inevitable, and hence certain influences on choices are also inevitable", p. 21, as quoted there). No DOI; Crossref returned only reviews. Hertwig and Grüne-Yanoff say it puts "educative" information among nudges (via Sunstein 2016, not opened). + +**18. Mertens, S., Herberz, M., Hahnel, U. J. J., & Brosch, T. (2022). The effectiveness of nudging: A meta-analysis of choice architecture interventions across behavioral domains. *PNAS, 119*(1), e2107346118.** Key `mertens2022effectiveness`. DOI 10.1073/pnas.2107346118 (Crossref confirms). +- *Read status:* **read in part** (Europe PMC full text of the *uncorrected* article: abstract, the intervention-category results, the discussion of the categories). **A published correction (PNAS 119(19), e2204059119, 4 May 2022; the authors cite four observations from a retracted paper and coding errors) changes the numbers below.** The corrected values, from the correction notice, are in square brackets. +- *Quotations.* "Drawing on more than 200 studies reporting over 450 effect sizes (n = 2,149,683) ... choice architecture interventions overall promote behavior change with a small to medium effect size of Cohen's d = 0.45 (95% CI [0.39, 0.52]) [corrected: d = 0.43 (95% CI [0.38, 0.48]); k = 447 effect sizes from 212 publications] ... Across behavioral domains, interventions that target the organization and structure of choice alternatives (decision structure) consistently outperform interventions that focus on the description of alternatives (decision information) or the reinforcement of behavioral intentions (decision assistance)." (abstract) Category effects ran "from d = 0.31 to 0.55" [corrected: 0.28 to 0.54; information 0.34, structure 0.54, assistance 0.28]; structure beat information by b = 0.17 (95% CI [0.03, 0.31], P = 0.02) [corrected: b = 0.19, 95% CI [0.08, 0.31], P = 0.001]. In Mertens et al.'s own taxonomy, the paper's re-representation sits on the *information* side: decision information includes interventions that "translate existing descriptions of choice options into more comprehensible or relevant information", and the "translation" technique averaged d = 0.31 [corrected 0.28], among the lowest of nine techniques. Mertens et al. also report their own bias sensitivity analysis (d = 0.31 under moderate and d = 0.08 under severe publication bias) and, in their reply (PNAS 119(31), e2202928119), agree that publication bias is "problematic". *Decision information* interventions include "measures that increase the visibility of otherwise covert information (e.g., feedback devices and nutrition labels)" and ones that "translate existing descriptions of choice options into more comprehensible or relevant information". +- *Bearing.* **Supports** the paper's ordering on measured effects: restructuring the choice beats informing, in the largest synthesis. **Complicates** it in one respect the paper needs to face: re-representation (simplifying, reframing, translating) is classed as *information*, the weaker category. Cautions in entry 19. + +**19. Maier, M., Bartoš, F., Stanley, T. D., Shanks, D. R., Harris, A. J. L., & Wagenmakers, E.-J. (2022). No evidence for nudging after adjusting for publication bias. *PNAS, 119*(31), e2200300119.** Key `maier2022no`. DOI 10.1073/pnas.2200300119 (Crossref confirms; 6 authors). +- *Read status:* **read in full** (Europe PMC text of this short letter). +- *Quotation.* "Strikingly, there is an absence of evidence for an overall effect and evidence against an effect in the 'information' and 'assistance' intervention categories, whereas the evidence is undecided for 'structure' interventions." (main text, on Table 1). The letter re-analyses the Mertens et al. data with bias-correcting models (title: "No evidence for nudging after adjusting for publication bias"). +- *Bearing.* The most useful measured sentence for §4 in this facet, and it cuts both ways. Maier et al. also note that "all intervention categories and domains apart from 'finance' show evidence for heterogeneity, which implies that some nudges might be effective, even when there is evidence against the mean effect." After correction, informing interventions (the category that includes disclosure-like measures) show evidence *against* an effect, which supports D3. The structure category is only "undecided", so the evidence that restructuring works is thin too. A paper that says "design beats disclosure" on the strength of Mertens must also report Maier. **Also read** DellaVigna and Linos (entry 20a): in two nudge units' 126 RCTs the average effect was 1.4 percentage points, "still sizable and highly statistically significant", an 8.1% increase in the July 2020 working paper (8.0% in the Econometrica abstract), against 8.7 points (33.5%; 33.4% in Econometrica) in academic papers. The working paper says publication bias "can account for the full difference in effect sizes"; the version of record says selective publication "explains about 70 percent of the difference". The study does not compare structure with information interventions. + +**20a. DellaVigna, S., & Linos, E. (2022). RCTs to scale: Comprehensive evidence from two nudge units. *Econometrica, 90*(1), 81–116.** Key `dellavigna2022rcts`. DOI 10.3982/ecta18709 (Crossref confirms). *Read in part*: the July 2020 NBER working paper 27594 (abstract and §1 only). Quotation as above (working-paper abstract); the version-of-record wording differs, see entry 19. + +**20b. Yeung, K. (2017). 'Hypernudge': Big Data as a mode of regulation by design. *Information, Communication & Society, 20*(1), 118–136.** Key `yeung2017hypernudge`. DOI 10.1080/1369118x.2016.1186713 (Crossref confirms). *Not accessible* (paywall, no open copy). Title and DOI only. It is the standard source for the claim that algorithmic choice architecture is dynamic and personalised, which is the nudge literature's closest approach to an adaptive third party. A human should read it for §5 and §7 as well. + +**20c. Mills, S., & Sætra, H. S. (2022). The autonomous choice architect. *AI & Society, 39*(2), 583–595.** Key `mills2022autonomous`. DOI 10.1007/s00146-022-01486-z (Crossref confirms title, authors, year, volume and pages). *Not accessible*; I saw only a search-result snippet saying that an algorithm "infused with an objective it has been programmed to maximise" becomes the choice architect. **Unverified lead**, but on topic: the nearest item I found to "choice architecture where the architect is an adaptive algorithm". Read before §4 is final. + +**20d. Lorenz-Spreen, P., Lewandowsky, S., Sunstein, C. R., & Hertwig, R. (2020). How behavioural sciences can promote truth, autonomy and democratic discourse online. *Nature Human Behaviour, 4*(11), 1102–1109.** Key `lorenzspreen2020how`. DOI 10.1038/s41562-020-0889-7. And **Kozyreva, A., Lewandowsky, S., & Hertwig, R. (2020). Citizens versus the internet. *Psychological Science in the Public Interest, 21*(3), 103–156.** Key `kozyreva2020citizens`. DOI 10.1177/1529100620946707. Both Crossref-confirmed; *not read* (I downloaded the Nature page but did not read it; the SAGE copy was walled). They apply nudges and boosts to digital environments and are the closest literature to a nudge/boost account of algorithmic platforms. Leads only. + +**On the term.** I found no use of "decision experience design" or "decision-experience design" (searches 23, 24; earlier Z found only the author's post). Nearest phrases: "choice architecture", "digital nudging" (Weinmann, Schneider and vom Brocke, Münster; title seen in a search result, not opened), "autonomous choice architect" (entry 20c), "decision design" (Morii 2026, held under Z). + +### Q4. Meta-analyses and reviews of explainable-AI and human-AI decision studies + +**21. Schemmer, M., Hemmer, P., Nitsche, M., Kühl, N., & Vössing, M. (2022). A meta-analysis of the utility of explainable artificial intelligence in human-AI decision-making. *Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society (AIES '22)*.** Key `schemmer2022meta`. arXiv 2205.05126 v2. The arXiv record resolves; I did not look up an AIES DOI. +- *Read status:* **read in full** (10-page arXiv v2). Page numbers not given; I cite the section. +- *What it argues:* Meta-analysis of studies with between-group accuracy comparisons on classification tasks, effect size as Hedges' g. +- *Quotations and numbers.* "We observe a statistically positive impact of XAI on users' performance ... However, we find no effect of explanations on users' performance compared to sole AI predictions." (abstract) XAI-assisted against humans alone: 33 studies, 5,083 participants, SMD 0.59 (95% CI 0.39 to 0.79), I² = 85.4%, prediction interval −0.54 to 1.71. AI-assisted against XAI-assisted: 11 studies, 999 observations, SMD 0.07 (95% CI −0.15 to 0.30; z = 0.63, p = 0.53), prediction interval −0.59 to 0.74. (Results, "AI Assistance vs. XAI Assistance") Limitations: "the current existing sample of XAI studies contains just online studies" and "the meta-analysis is limited with regard to drawing causal conclusions." (Limitations) +- *Bearing.* **Supports** D3 in the narrow case: adding an explanation to an AI prediction does not move accuracy on average, and the true effect in a new study could be either sign. The gain over humans alone comes from the AI's prediction and not from the explanation. **Limits:** one person and one machine, accuracy on a classification task, online samples. Nothing here is a coordination between two parties. + +**22. Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. *Nature Human Behaviour, 8*(12), 2293–2303.** Key `vaccaro2024when`. DOI 10.1038/s41562-024-02024-1 (Crossref confirms on rerun). +- *Read status:* **read in part** (journal PDF, 29 pp.: abstract, introduction, results on moderators, discussion of moderators and limitations). +- *Quotations.* "...a preregistered systematic review and meta-analysis of 106 experimental studies reporting 370 effect sizes ... human–AI combinations performed significantly worse than the best of humans or AI alone (Hedges' g = −0.23; 95% confidence interval, −0.39 to −0.07) ... we found performance losses in tasks that involved making decisions and significantly greater gains in tasks that involved creating content." (abstract, p. 2293) "...much of the recent research in human–AI collaboration has focused on using AI systems to help humans make decisions by providing not only suggested decisions but also confidence levels or explanations. But we found that neither of these factors significantly affected the performance of human–AI systems." (p. 2294) The discussion puts it as "on average across our 300+ effect sizes" (p. 2297); the coding table splits 198 effect-size rows without and 180 with an explanation. The explanation test is a between-study moderator, not a within-experiment contrast. "...we think researchers may wish to de-emphasize this line of inquiry" (p. 2297). Against humans alone, the combination did better (g = 0.64; 95% CI 0.53 to 0.74; p. 2295). +- *Bearing.* **Supports** D3, with a stronger statement than Schemmer's: across 300-odd effect sizes, whether an explanation was shown did not moderate performance. **Qualifies** it: the losses are specific to decision tasks, and gains come when humans outperform AI. The authors flag high unexplained heterogeneity and too few studies to detect interactions such as explanation by AI type (p. 2297). + +**23. Lai, V., Chen, C., Smith-Renner, A., Liao, Q. V., & Tan, C. (2023). Towards a science of human-AI decision making: An overview of design space in empirical human-subject studies. *Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency*.** Key `lai2023towards`. DOI 10.1145/3593013.3594087 (Crossref confirms the published title). +- *Read status:* **read in part** (arXiv 2112.11471 v1, 36 pp., under the earlier title "Towards a Science of Human-AI Decision Making: A Survey of Empirical Studies"; abstract, the metrics section on understanding and reliance). Not the FAccT version, and the author list differs between versions; confirm. +- *What it does:* Surveys "over 100 papers" and summarises design choices in tasks, AI assistance elements and metrics. It does not pool effects. +- *Quotation.* "...objective and subjective understanding do not always align due to the phenomena of illusory confidence with which one believes they understood the model more than they actually did." (metrics section, "Understanding") +- *Bearing.* It is a map of what is measured and not a finding about whether explanations help. Its use for §4 is the observation that "understanding" has both subjective and objective measures and they come apart. Note: this source was described as held; I found it in no bib or card. It needs a card. + +### Q5. Disclosure in two-sided, mediated or strategic settings with a behavioural outcome + +I found **no study that discloses an intermediary's rule to two human parties and measures what both do** (searches 16–21). The nearest studies divide into three groups: disclosure in a person–platform dyad, disclosure of ratings or information in a two-sided market, and laboratory experiments with two humans and a computer in a mediating or counterpart role. All of the following are abstract only unless stated. + +**24. Cen, S. H., Ilyas, A., Allen, J., Li, H., & Madry, A. (2024). Measuring strategization in recommendation: Users adapt their behavior to shape future content. *Proceedings of the 25th ACM Conference on Economics and Computation*, 203–204.** Key `cen2024measuring`. DOI 10.1145/3670865.3673634; arXiv 2405.05596 (Crossref confirms the record, 5 authors; the 2-page EC record is an extended abstract of the arXiv paper). *Read in part* (arXiv, 51 pp.: abstract, design, hypotheses; not the analysis). +- *Design:* A lab music player; participants randomly told that the algorithm would attend more to likes, or to dwell time, or told nothing. +- *Quotation.* "participants who are told that the algorithm mainly pays attention to 'likes' and 'dislikes' use those functions 1.9x more than participants told that the algorithm mainly pays attention to dwell time." (abstract) +- *Bearing.* **Cuts against "disclosure doesn't affect".** A one-sentence statement of the algorithm's rule changed behaviour by a large factor. It does not say the people did better. And it is a person and a platform. The effect is on *acting on* the algorithm, which is the engagement the paper wants, so it can be read as support for the paper's section 5 ("people understand through acting") and as a qualification of D3: disclosure did change what people did. D3's phrasing "our ability to navigate it" leaves the question whether the changed behaviour was navigation or gaming. + +**25. Erlei, A., Nekdem, F., Meub, L., Anand, A., & Gadiraju, U. (2020). Impact of algorithmic decision making on human behavior: Evidence from ultimatum bargaining. *Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, 8*, 43–52.** Key `erlei2020impact`. DOI 10.1609/hcomp.v8i1.7462 (Crossref confirms; abstract from Crossref). *Abstract only* (AAAI link returned 404). +- *Closest to the triad in this list.* Ultimatum bargaining, 1,178 participants; one side may use a decision-support system. Abstract: "Explanations increase the number of unique system inquiries, but users appear less willing to follow the system's recommendation." "People who negotiate with a user who has a decision support system, but cannot use one themselves, react to its introduction by demanding a better deal for themselves, thereby decreasing overall cooperation levels." "Interpretability mitigates perceptions of unfairness." "...higher levels of transparency might inadvertently hurt cooperation through changes in fairness perceptions." +- *Bearing.* A **two-human strategic setting with a behavioural outcome and a transparency manipulation**, which neither round found. The algorithm assists one party; it does not read both. Explanation lowered following, and transparency changed the *other* party's behaviour via fairness. Both effects are on third parties and not on the reader of the explanation, which is the kind of effect the paper says disclosure cannot be assumed to have. Obtain the full text before using; abstract-level. + +**26. Jhaver et al. (2019), "Did you suspect the post would be removed?", is held as `jhaver2019suspect` (survey; see T §9.2). Jhaver, S., Bruckman, A., & Gilbert, E. (2019). Does transparency in moderation really matter? *Proceedings of the ACM on Human-Computer Interaction, 3*(CSCW), 1–27.** Key `jhaver2019does`. DOI 10.1145/3359252 (Crossref confirms). *Abstract only* (Crossref). +- Abstract: "Most importantly, we show that offering explanations for content moderation reduces the odds of future post removals. Additionally, explanations provided by human moderators did not have a significant advantage over explanations provided by bots". The sample is 32 million Reddit posts; the design is regression on observational data. +- *Bearing.* **Cuts against D3.** A platform's reason for an action, shown to the affected user, lowered later removals. Observational, not randomised; one platform; a user and a moderation system, not two parties. The held survey (β = −.314 for those who had read the rules, T §9.2) points the other way on perceived fairness, so the two Jhaver papers disagree and the paper should cite both. Its mechanism is what the authors call education about norms. That is literacy-based by the paper's definition, and it worked. + +**27. Tyler, T., Katsaros, M., Meares, T., & Venkatesh, S. (2019). Social media governance: Can social media companies motivate voluntary rule following behavior among their users? *Journal of Experimental Criminology, 17*(1), 109–127.** Key `tyler2019social`. DOI 10.1007/s11292-019-09392-z (Crossref confirms title, authors, volume and pages). *Abstract only* (Consensus). A randomised message "emphasizing the procedural justice of platform procedures" lowered subsequent violating posts against no-message controls (study 2). *Bearing:* a short disclosure about *how* the platform decides changed behaviour. It is again a person-to-platform dyad and the content is a statement of fairness, not the algorithm's rule. + +**28. Klein, T. J., Lambertz, C., & Stahl, K. O. (2016). Market transparency, adverse selection, and moral hazard. *Journal of Political Economy, 124*(6), 1677–1713.** Key `klein2016market`. DOI 10.1086/688875 (Crossref confirms; abstract from Consensus). *Abstract only.* +- Abstract: "The improvement was achieved by reducing strategic bias in buyer ratings. It led to a significant increase in buyer satisfaction with seller performance, but not to an increase in seller exit." Sellers "preferred" improving behaviour to exiting. +- *Bearing.* A two-sided market (buyers, sellers, eBay) where a change in what the *mediator publishes* changed seller behaviour. The abstract says the change removed strategic bias in ratings: a redesign of the rating channel, not a disclosure of the platform's algorithm. It is a **type (i) example with a behavioural outcome on both sides**, though a rating system and not an algorithmic intermediary. I did not read the mechanism. + +**29. Further abstract-only hits, grouped.** +- *Bauer, K., & Gill, A. (2024). Mirror, mirror on the wall: Algorithmic assessments, transparency, and self-fulfilling prophecies. Information Systems Research, 35*(1), 226–248.* Key `bauer2024mirror`. DOI 10.1287/isre.2023.1217 (Crossref confirms). Abstract: "the disclosure of fundamentally erroneous algorithmic scores evokes self-fulfilling prophecies that endogenously steer the behavior of scored individuals toward their assessment". *Bearing:* disclosure changed behaviour, and not usefully. Person and algorithm only. +- *Hagenbach, J., & Salas, A. (2025). Strategic information disclosure to classification algorithms: An experiment. Experimental Economics, 28*(4), 952–973.* Key `hagenbach2025strategic`. DOI 10.1017/eec.2025.10030 (Crossref confirms). Abstract: "Information about the algorithm functioning makes subjects identify correlations they otherwise do not see but also overthink. Overall, this information decreases the frequency of optimal disclosure strategies." *Bearing:* telling people how the algorithm works made their strategy worse. One person against an algorithm. +- *Casella, A., Friedman, E., & Perez Archila, M. (2025). On the fragility of mediation: Theory and experimental evidence. Journal of the European Economic Association, 24*(3), 801–850.* Key `casella2025fragility`. DOI 10.1093/jeea/jvaf049 (Crossref confirms). Abstract: two subjects negotiate with and without an algorithmic mediator; "While messages to the mediator are more sincere, we find that peaceful resolution is not more frequent." *Bearing:* a genuine **two-humans-plus-algorithmic-mediator experiment**. It does not manipulate disclosure of the mediator's rule per the abstract. Pass to facet P. +- *Hui, X., Jin, G. Z., & Liu, M. (2024). Designing quality certificates: Insights from eBay. Journal of Marketing Research, 62*(1), 40–60.* Key `hui2024designing`. DOI 10.1177/00222437241270222 (Crossref confirms). Abstract: after eBay's certification criteria shifted from consumer reports to administrative data, "this change motivates seller effort in dimensions highlighted by the new criteria, as well as enabling sellers to more precisely target their effort at the threshold." *Bearing:* published criteria in a two-sided market change seller behaviour, partly by letting sellers aim at the threshold. A case where disclosure of the rule changes what the regulated party does. +- *Dash, A., Chakraborty, A., Ghosh, S., Mukherjee, A., & Gummadi, K. P. (2024). Investigating nudges toward related sellers on e-commerce marketplaces: A case study on Amazon. PACM HCI, 8*(CSCW2), 1–31.* Key `dash2024investigating`. DOI 10.1145/3686994 (Crossref confirms). Abstract: "when customers are shown the rectified metrics for the different sellers, their preference toward Related Sellers is almost halved." *Bearing:* showing corrected seller metrics changed buyers' choices. The buyer-side effect of an information repair; design of the experiment not known from the abstract. +- *Jhaver, S., Rathi, H., & Saha, K. (2024). Bystanders of online moderation: Examining the effects of witnessing post-removal explanations. CHI '24, 1–9.* Key `jhaver2024bystanders`. DOI 10.1145/3613904.3642204. Abstract: bystanders who witness post-removal explanations "significantly increase their posting activity and interactivity levels as compared to their matched control set". Quasi-experiment on two subreddits. +- *Normann, H.-T., & Sternberg, M. (2021). Human-algorithm interaction: Algorithmic pricing in hybrid laboratory markets.* Key `normann2021human`. SSRN, DOI 10.2139/ssrn.3840789 (Crossref title and authors; Consensus lists the paper as "Hybrid Collusion"). Abstract: "(Un)certainty about the actual presence of an algorithm does not significantly affect collusion." Algorithm as a market player, not an intermediary. +- *Ozkes, A. I., Hanaki, N., Vanderelst, D., & Willems, J. (2024). Ultimatum bargaining: Algorithms vs. humans. Economics Letters, 244*, 111979.* Key `ozkes2024ultimatum`. DOI 10.1016/j.econlet.2024.111979. Abstract: subjects "do not differentiate ... between an explained and unexplained algorithm". Algorithm as a counterpart. +- *Wang, Q., Huang, Y., Jasin, S., & Singh, P. V. (2020). Algorithmic transparency with strategic users.* Key `wang2020algorithmic`. SSRN, DOI 10.2139/ssrn.3652656. A model, no behavioural data: "in some cases, even the predictive power of the algorithm can increase if the firm makes the algorithm transparent. By contrast, users may not always be better off". + +**Held, reused by key.** `mejia2021when` (what each party sees, when: a field experiment in a triad; the bias moved from request to cancellation), `li2026platform` (earnings transparency bundled with a pay guarantee: engagement up, and "greater transparency may induce strategic driver behavior"; natural experiment), `seira2017are`, `kling2012comparison`, `kelley2009nutrition`, `poursabzi2021manipulating`, `vasconcelos2023explanations`, `bucinca2021trust`, `cabrera2023improving`, `rahman2021invisible`, `dietvorst2018overcoming`, `eu2019p2b`, `bansal2019updates`. I reread none of them for this facet. + +--- + +## (c) Synthesis + +### Q1. What mechanisms do the mandated-disclosure critics give, and what evidence? + +*Mechanisms of failure.* (1) The disclosee does not read: Ben-Shahar and Schneider put it as "do not read ... do not understand ... do not use" (p. 665). (2) Length and complexity: "complexity cannot be explained simply" (p. 713). (3) Attention, motivation and probability errors (Loewenstein et al., §3). (4) Disclosure serves as a substitute for harder regulation (L-S-G p. 412; Ben-Shahar and Schneider pp. 651, 684). (5) The financial-education complement fails: 0.1% of variance in Fernandes et al. (87 effect sizes; r² = .0011), and no significant effect at delays of 20 months or more. *Mechanisms of success.* (1) A brief, simple signal at the point of choice (grades, stars; Fung et al. pp. 55–57; Ben-Shahar and Schneider p. 743, where the restaurant example is hedged: "There is little real evidence"). L-S-G (p. 403 and n. 8) warn that for the restaurant grades the industry response is clearer than the consumer response. (2) Effect on the discloser, not the disclosee (L-S-G p. 403). (3) An intermediary that reads for the person (L-S-G p. 411). (4) Double embeddedness (Fung et al. p. 65). *Evidence.* The critique rests on case surveys and field and lab experiments of mixed design. Kling et al. is a bundle (held). *My inference, not the sources':* the successes the critics name are almost all brief symbolic signals (letter grades) or discloser effects. None of the sources calls them re-representations or says they work "not by reading". + +What the paper can keep: the diagnosis that the remedy requires a reader. What it must change: the critics do not say disclosure never works and do not equate disclosure with literacy. Their positive cases resemble the paper's type (i) and (iii) moves (my reading, not theirs). Ben-Shahar and Schneider decline to name what replaces disclosure (pp. 651, 742). Fernandes et al. do name one: "just-in-time" education, alone or embedded in decision support, and machine-readable "Smart Disclosures" read by "infomediaries" (WP p. 22). + +### Q2. Privacy notices and consent + +Estimated: the time cost of reading (McDonald and Cranor, a modelled 244 hours a year if every policy were read in full). Measured: the dependence of consent on layout (Utz et al.: under 0.1% accept all purposes in the opt-in setting, on one German website; Nouwens et al.: 22–23 points from removing the reject-all button, 40 participants). Not read: Obar and Oeldorf-Hirsch's 543-student result [unverified: see AUDIT_07] and Bakos et al.'s clickstream study (abstract: only one or two of every 1,000 shoppers open the license agreement). Together these say that consent tracks the interface even where the notice's topic is constant; the conditions do differ in the choices offered, and Utz et al.'s wording change had a small effect (CV = .08). That fits the paper's claim. It also shows an asymmetry the paper does not mention: the interface lever has been used mostly by the discloser to produce consent. The notice problem is itself a triad (user, site, third-party trackers), and no study here asks what the user can do to the notice, which is the engagement the author wants. + +### Q3. Nudge, boost, and the paper's line + +The nudge/boost distinction does not draw the paper's line. It draws a nearby line, and the paper should say which. The parts that agree: a nudge changes the choice environment and leaves the person's competence unchanged; the reversibility test points to spacing as nudge-like (my inference; the authors' own qualification, p. 981, is that repeated nudges can turn into boosts, which is the H-cause question); Hertwig and Grüne-Yanoff say educative nudges (Sunstein's category, which includes disclosure) "largely overlap" with short-term boosts, both needing "a modicum of motivation and cognitive skill" (p. 977). The parts that do not: a boost can be a re-representation (natural frequencies), so "environment versus person" is not the axis; both literatures assume one chooser and a benevolent or at least non-adaptive architect. Measured effects add a warning: Mertens finds structure beats information (corrected category averages run from d = 0.28 to 0.54; contrast b = 0.19, 95% CI 0.08 to 0.31), and files re-representation ("translation", d = 0.28) under information; Maier shows that, after bias correction, there is evidence *against* an effect for information nudges and only "undecided" evidence for structure nudges, with heterogeneity that leaves room for some effective nudges; Mertens et al.'s reply concedes publication bias; DellaVigna and Linos find 1.4 percentage points at scale in nudge-unit trials, still significant, and attribute about 70 percent of the gap from academic papers to selective publication. "Design the decision environment" in this literature means the arrangement of options for one chooser (defaults, order, effort); the paper's version restructures what a coordinating pair sees of each other through a mediator (Mejia and Parker's field experiment on what the driver sees and when; the phrasing is facet H's). Nobody in the nudge literature I opened treats that as choice architecture, though digital-nudging work (Yeung; Mills and Sætra, unopened) gets near it. I found no use of "decision experience design". + +### Q4. Do explanations improve decisions? + +Not on average. Schemmer et al.: the explanation estimate is 0.07 SMD (95% CI −0.15 to 0.30; 11 studies), not distinguishable from zero, over the AI prediction alone. Vaccaro et al.: explanation and confidence were not significant moderators in a 106-experiment, 370-effect-size dataset (the authors say "300+ effect sizes" for the moderator result; it is a between-study contrast), and human–AI combinations did worse than the best single party in decision tasks (g = −0.27). The heterogeneity is large in both (Schemmer I² 57% for the AI-versus-XAI contrast; Vaccaro says much is unexplained), so "does not help on average" is correct and "never helps" is not. The held single studies are mixed rather than in agreement: Vasconcelos finds a highlighted-path explanation lowers overreliance only in the hard task; Bucinca finds that cognitive forcing functions (not explanations) beat simple explanations only on wrong-AI questions, and still did worse than no AI there; Poursabzi finds a clearer model left people worse at catching its mistakes in some experiments. This is the strongest quantitative support for D3 in the facet, with the limits that every study is one human and one machine, mostly online and with accuracy as the outcome. + +### Q5. Two-sided, mediated or strategic disclosure + +The two earlier rounds found none, and so do I for the specific design (an intermediary's rule disclosed to two humans, behaviour of both measured). What exists is adjacent and mostly cuts against "disclosure does not affect behaviour": a sentence about the algorithm changed how people used it by 1.9 times (Cen et al.); removal explanations lowered later removals (Jhaver et al., observational); a message stressing the platform's procedural justice lowered later rule violations against no-message controls in a randomised study (Tyler et al., abstract); introducing a decision-support system for one bargainer made the other demand a better deal, an effect the abstract attributes to perceived unfairness, which interpretability mitigated (Erlei et al.; the abstract does not say that transparency itself changed the demands). Equally, it can hurt: erroneous scores disclosed produced self-fulfilling behaviour (Bauer et al.), and information about the algorithm lowered the quality of strategy (Hagenbach et al.). The honest summary is that disclosure changes behaviour, and often in the wrong direction or on the wrong party. D3 as worded ("doesn't affect") is not supported; "does not reliably help the person coordinate" is. + +### Claims in the paper's position that these sources contradict or qualify + +1. **"Disclosure of the algorithm doesn't affect our ability to navigate it" (D3).** Contradicted as a claim about behaviour by Cen et al. 2024 (1.9x between the likes-told and dwell-told groups), Jhaver et al. 2019 (observational), Tyler et al. (randomised message, abstract), Erlei et al. (abstract; the behavioural effect on the counterpart follows availability of the system) and Loewenstein et al. (disclosure affects disclosers). Supported as a claim about *performance* by Schemmer, Vaccaro, and the held Poursabzi, Bucinca, Cabrera. The wording needs to say "behaviour" and "performance" apart. +2. **"Disclosure is a literacy-based intervention."** Supported by Hertwig and Grüne-Yanoff (educative nudges, a category that includes disclosure on Sunstein's account, "largely overlap" with short-term boosts, p. 977; long-term boosts are set apart, p. 981) and by Ben-Shahar and Schneider's pairing of disclosure and literacy education. Qualified by Fung et al. and Ben-Shahar and Schneider, whose successful disclosures (grades) are brief symbolic signals (that they "work as re-representations" is my reading, not theirs, and both also credit the discloser's response); and by Loewenstein et al., for whom disclosure acts through the discloser. +3. **"Change the interface and the training is not needed."** Qualified by Fernandes et al. (needs differ, so defaults and fixed designs fail; just-in-time education is the authors' recommendation) and by Hertwig and Grüne-Yanoff (a boost is the more durable remedy: behaviour is likely to "revert to the prenudging state" when a nudge is removed, p. 981, unless repeated exposure builds routines that survive it). +4. **"Design beats informing."** Mertens supports it (corrected: structure d = 0.54 against information 0.34), though re-representation counts there as information; Maier undermines the evidence for structure (undecided) while finding evidence against information nudges; DellaVigna and Linos show smaller effects at scale but do not compare structure with information. +5. **"Nadella's remedy is disclosure."** Not tested here. Loewenstein et al.'s point that disclosure reaches the discloser suggests the sharper reading: P15 and P19 mostly bind the organisation that holds the log. +6. **"Nobody has tested disclosure in a triad."** Still true for the exact design; Erlei et al. (abstract) is a two-human bargaining experiment with an algorithm available to one side and explanation and interpretability manipulated, and Casella et al. is a two-human experiment with an algorithmic mediator. The claim should be "I found no study that discloses an intermediary's rule to both parties". + +--- + +## What the outline should add (section 4, with echoes in 3, 5, 6 and 9) + +| # | Section | Claim | Carrying sources | Read status | +|---|---|---|---|---| +| 1 | 4 | The law-and-economics case that mandated disclosure fails names the same mechanism the paper does: it needs a reader who does not read, understand or use. | Ben-Shahar and Schneider 2011 (pp. 651, 665, 713); Loewenstein, Sunstein and Golman 2014 (pp. 398, 408, 411) | both read (Ben-Shahar and Schneider in full via OCR of an offprint; L-S-G in part) | +| 2 | 4 | The disclosures the critics call successful are brief symbolic signals at the point of choice (a grade, a star), with a hedge from Ben-Shahar and Schneider ("little real evidence") and a discloser-side route. *My reading:* these resemble the paper's type (i) move rather than being an exception to it. | Ben-Shahar and Schneider 2011 (pp. 743, 744, 748); Fung, Graham and Weil 2007 (pp. 55–59); caution: Loewenstein et al. 2014 (p. 403, n. 8). *The "re-representation" label is my inference; the sources say "brief, simple" signals and credit discloser response too.* | read | +| 3 | 4 | Disclosure often acts on the discloser, not the person, so the paper should ask which party in a triad the disclosure reaches. | Loewenstein et al. 2014 (abstract, p. 403); Fung et al. 2007 (p. 65) | read in part | +| 4 | 4 | Explanations do not measurably improve accuracy on average (Schemmer: SMD 0.07, 95% CI −0.15 to 0.30; Vaccaro: not a significant moderator, a between-study contrast) and may not be the moderator that matters. | Schemmer et al. 2022; Vaccaro et al. 2024 (pp. 2294, 2297) | read in full; read in part | +| 5 | 4 | The notice-and-consent literature is a tested case: consent changes with the layout and choice set of the notice (Utz: one German site; Nouwens: 40 participants), and reading every policy would cost an estimated 244 hours a year (McDonald and Cranor, modelled). | McDonald and Cranor 2008; Utz et al. 2019; Nouwens et al. 2020 | read in part (Obar and Oeldorf-Hirsch and Bakos et al. not accessible) | +| 6 | 4 or 9 | Where disclosure *has* changed behaviour (algorithm rule told to users; removal reasons; a fairness message) the effect is on how people use or act on the system, and it is sometimes harmful. State D3 as a claim about performance, and report these. | Cen et al. 2024 (read in part); Jhaver et al. 2019, Tyler et al. 2019, Bauer et al. 2024, Hagenbach et al. 2025 (abstracts only) | mixed | +| 7 | 4 | Name the nearest rival and say where the line falls. The nudge/boost split is not the paper's split: disclosure sits with "educative nudges", which overlap with short-term boosts; a boost may be a re-representation; and the discriminating test is reversibility, which the authors themselves qualify (repeated nudges can become boosts, p. 981). The paper's contribution is to put the architect in the triad. | Hertwig and Grüne-Yanoff 2017 (pp. 977–981); Grüne-Yanoff and Hertwig 2016 (pp. 151–159) | read in full; read in part | +| 8 | 4 | The measured case for "design over information" is weaker than it looks, and re-representation counts as information in Mertens et al.'s taxonomy. Use the corrected Mertens numbers (d = 0.43; structure 0.54 against information 0.34). | Mertens et al. 2022 (with its correction); Maier et al. 2022; DellaVigna and Linos 2022 (cite the version of record: about 70 percent of the gap) | read in part; read in full; read in part | +| 9 | 5 | A fair statement of the evidence on acting on algorithms: disclosure of a rule changes strategy at once (Cen et al.), which is the engagement the paper wants and also the gaming it fears. | Cen et al. 2024; Hagenbach et al. 2025 | read in part; abstract only | +| 10 | 6 or 9 | Where needs differ, defaults and fixed designs fail and the literature recommends just-in-time support. This is a limit on "design, don't educate". | Fernandes et al. 2014 (WP version p. 22) | read in part | +| 11 | 8 | Two nearest experiments with two humans and an algorithm: Erlei et al. (decision-support availability for one side; explanations and interpretability appear in the abstract's results; ultimatum) and Casella et al. (algorithmic mediator, negotiation; no disclosure manipulation per the abstract). Check both before the paper says no triadic disclosure experiment exists. | Erlei et al. 2020; Casella et al. 2025 | abstracts only | +| 12 | 9 | State what was not found: no study discloses an intermediary's rule to both parties of a coordination and measures both. Give the queries (search log). | searches 16–22 | n/a | + +--- + +## (d) Gaps and unverified leads + +- **Not accessible; a human with library access should fetch:** Ben-Shahar and Schneider, *More than you wanted to know* (Princeton, 2014; Crossref 10.2307/j.ctt5hhrqj; I read only the 2011 article); Weil et al. 2006 (JPAM; Wiley); Willis 2008 (Iowa Law Review; Crossref has no record); Bar-Gill 2012 (OUP) and 2019 (U Chicago Law Review 86:217); Thaler and Sunstein 2008; Johnson et al. 2012 (Springer; read the section on describing options); Yeung 2017; Acquisti et al. 2015; Bakos et al. 2014; Obar and Oeldorf-Hirsch 2018 (the figures I saw in a search summary are unverified); Mills and Sætra 2022; Lorenz-Spreen et al. 2020; Kozyreva et al. 2020. +- **Abstract only (verify before any use):** Erlei et al. 2020 (obtain AAAI full text; the OJS link I tried returned 404), Jhaver et al. 2019 (CSCW, ACM), Tyler et al. 2019, Klein et al. 2016, Bauer et al. 2024, Hagenbach et al. 2025, Casella et al. 2025, Hui et al. 2024, Dash et al. 2024, Jhaver et al. 2024, Normann and Sternberg 2021, Ozkes et al. 2024, Wang et al. 2020. Author lists for entries 24–29 were checked against Crossref. +- **Version problems.** Ben-Shahar and Schneider read as an OCRed offprint (ifs.org). Fernandes et al. read as the June 2013 working paper. McDonald and Cranor read as the author's pre-press (page numbers differ in the journal). The two Hertwig and Grüne-Yanoff PDFs are author-hosted, and their page markers are partial. Lai et al. read as arXiv v1. +- **Tools that failed.** OpenAlex (daily budget), Scholar Gateway (access denied), Springer, SAGE, Wiley, tandfonline, bepress, ACM DL and SSRN (walled or 403). None was bypassed. +- **Out of scope but adjacent, passed on:** Casella et al. 2025 and Erlei et al. 2020 to facet P; Yeung 2017, Mills and Sætra 2022 and the Chicago-Unbound/Hertwig digital-environment papers to facets M and O; Rahman 2021 and Bucher et al. 2021 (algorithmic opacity, held or qualitative) to facet M. + +--- + +## Changes after audit (2026-10-13) + +- Synthesis Q1, mechanisms of success: restaurant grade example now carries Ben-Shahar and Schneider's own hedge (p. 743) and the Loewenstein et al. warning (p. 403, n. 8); Fung locator corrected from pp. 54-57 to pp. 55-57. +- Synthesis Q1, "measured successes are re-representations": relabelled as my inference; neither source says "re-representation" or "not by reading". +- Synthesis Q1, "both decline to name what replaces": corrected; Fernandes et al. do name just-in-time education, decision support and Smart Disclosure. +- Synthesis Q2: McDonald and Cranor's 244 hours relabelled "modelled", not "measured"; "same information in every condition" withdrawn; sample sizes added. +- Synthesis Q3: Hertwig and Grüne-Yanoff pages corrected (boosts p. 977; reversibility p. 981); their qualification that repeated nudges become boosts added; disclosure-with-boosts claim narrowed to Sunstein's educative nudges and short-term boosts; Mertens numbers replaced by the corrected ones; re-representation shown to sit in the weaker "information" category; DellaVigna and Linos restated (about 70 percent, still significant). +- Synthesis Q4: Vaccaro "370" qualified ("300+", between-study moderator); Schemmer 0.07 described as not distinguishable from zero; held-study "agree" replaced by "mixed". +- Synthesis Q5: Erlei et al. claim corrected (system availability, not transparency, moved the counterpart's demands); Tyler et al. "recidivism in an RCT" reworded. +- Contradicted-claims items 1-4 reworded to match the above. +- Should-add item 2: relabelled as my reading, with the Ben-Shahar and Schneider hedge. Kept with change. +- Should-add item 4: numbers added, qualified. Kept with change. +- Should-add item 5: 244 hours relabelled modelled; sample sizes added. Kept with change. +- Should-add item 7: reversibility pages and qualification corrected. Kept with change. +- Should-add item 8: use corrected Mertens numbers and the version-of-record DellaVigna wording. Kept with change. +- Should-add item 11: Erlei et al. description narrowed. Kept with change. +- Should-add items 1, 3, 6, 9, 10, 12: unchanged. +- Entry 15 (Grüne-Yanoff and Hertwig): "the education that follows is then optional" removed; the authors say the boost approach has an explicit educational goal. +- Bib: Ben-Shahar and Schneider pages; McDonald and Cranor volume, issue, pages; Schemmer DOI and pages; Mertens correction noted; DellaVigna version noted; Tyler, Casella, Hui, Mills and Sætra, Obar print years. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_O_custody_of_conventions.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_O_custody_of_conventions.bib new file mode 100644 index 00000000..ee865bd2 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_O_custody_of_conventions.bib @@ -0,0 +1,501 @@ +% Facet O, 2026-10-13. Agent-drafted. The author has not reviewed it, and none of the text is the author's. +% Read status is in each note. Keys marked "reused" exist in literature/references.bib. + +@misc{charlemagne1900letter, + author = {Charlemagne}, + title = {{Letter to Abbot Baugulf of Fulda (De litteris colendis), c. 780--800; English translation by D. C. Munro (1900)}}, + year = {1900}, + howpublished = {Translations and Reprints from the Original Sources of European History, vol. 6, no. 5, pp. 12--14 (University of Pennsylvania Press); text in Fordham Internet Medieval Sourcebook; Latin in Boretius, Capitularia I, no. 29, p. 78}, + url = {https://sourcebooks.web.fordham.edu/source/carol-baugulf.asp}, + note = {Read in full (translation, not the Latin). The Fordham page gives the Boretius and Munro citation.} +} + +@book{boretius1883capitularia, + editor = {Boretius, Alfred}, + title = {{Capitularia regum Francorum, vol. 1}}, + series = {Monumenta Germaniae Historica, Legum sectio II}, + year = {1883}, + url = {https://archive.org/details/capitulariaregum01bore}, + note = {Read in part: no. 22, Admonitio generalis (789), c. 72, p. 60 (OCR text from archive.org, item capitulariaregum01bore). Year taken from the archive.org record.} +} + +@book{dummler1881alcuini, + editor = {D{\"u}mmler, Ernst}, + title = {{Poetae Latini aevi Carolini, vol. 1}}, + series = {Monumenta Germaniae Historica, Poetarum Latinorum medii aevi I}, + year = {1881}, + url = {https://archive.org/details/poetaelatiniaevi00dm}, + note = {Read in part: Alcuini carmina no. XCIV, p. 320 (OCR text from archive.org). Year from the archive.org record.} +} + +@phdthesis{westwell2018dissemination, + author = {Westwell, Arthur}, + title = {{The Dissemination and Reception of the Ordines Romani in the Carolingian Church, c.750--900}}, + school = {University of Cambridge}, + year = {2018}, + doi = {10.17863/cam.23255}, + url = {https://doi.org/10.17863/cam.23255}, + note = {Read in part: abstract, historiographic introduction, keyword search, conclusion. DataCite DOI (Crossref 404); doi.org resolves to the Cambridge Apollo item. About liturgical texts, not script.} +} + +@misc{piazzoni11bookhands, + author = {Piazzoni, Ambrogio M.}, + title = {{11.1 Bookhands of the Frankish regions}}, + howpublished = {Latin Paleography: Thematic Pathways on the Web, Vatican Library (Spotlight)}, + year = {n.d.}, + url = {https://spotlight.vatlib.it/latin-paleography/feature/11-1-bookhands-of-the-frankish-regions}, + note = {Read in part (sections 11.1.1 to 11.1.3). Web exhibition text under copyright; no date shown on the page. Author line from the page source.} +} + +@book{mckitterick1989carolingians, + author = {McKitterick, Rosamond}, + title = {The Carolingians and the Written Word}, + publisher = {Cambridge University Press}, + year = {1989}, + url = {https://archive.org/details/carolingianswrit0000mcki}, + note = {Reused key. Not reopened in facet O (archive.org search-inside blocked for guests). Facet C read it in part, search-inside.} +} + +@incollection{ganz1995book, + author = {Ganz, David}, + title = {{Book production in the Carolingian empire and the spread of Caroline minuscule}}, + booktitle = {The New Cambridge Medieval History, vol. 2}, + pages = {786--808}, + year = {1995}, + doi = {10.1017/chol9780521362924.032}, + note = {Abstract only (Consensus record). DOI resolved at Crossref, title and pages match.} +} + +@incollection{mostert2016pray, + author = {Mostert, Marco}, + title = {{`\ldots but they pray badly using corrected books': errors in early Carolingian copies of the Admonitio generalis}}, + booktitle = {Religious Franks}, + publisher = {Manchester University Press}, + year = {2016}, + doi = {10.7765/9781784997519.00015}, + note = {Abstract only (Consensus record; publisher PDF returned 403). DOI resolved at Crossref. Pages 112 to 127 come from a library-catalogue search snippet, not Crossref.} +} + +@incollection{parkes2016carolingian, + author = {Parkes, M. B.}, + title = {{Carolingian renovatio: augmenting old notation with new symbols}}, + booktitle = {Pause and Effect (Routledge reprint)}, + pages = {30--34}, + year = {2016}, + doi = {10.4324/9781315247243-4}, + note = {Abstract only (Consensus record). DOI resolved at Crossref (author given as Parkes, M. there). See facet A for Parkes 1992.} +} + +@article{dumville1987english, + author = {Dumville, David}, + title = {{English Square minuscule script: the background and earliest phases}}, + journal = {Anglo-Saxon England}, + volume = {16}, + pages = {147--179}, + year = {1987}, + doi = {10.1017/s0263675100003884}, + note = {Abstract only (Consensus record). DOI resolved at Crossref.} +} + +@article{steinova2015psalmos, + author = {Steinov{\'a}, E.}, + title = {{Psalmos, notas, cantus: on the meanings of nota in the Carolingian period}}, + journal = {Speculum}, + volume = {90}, + number = {2}, + pages = {424--457}, + year = {2015}, + doi = {10.1017/s0038713415000275}, + note = {Abstract only (Consensus record). DOI resolved at Crossref.} +} + +@misc{rfc2026, + author = {Bradner, S.}, + title = {{The Internet Standards Process -- Revision 3}}, + howpublished = {RFC 2026 (BCP 9), IETF}, + year = {1996}, + url = {https://www.rfc-editor.org/rfc/rfc2026}, + note = {Read in part (section 1, 1.1, participation clauses). Text from the RFC Editor.} +} + +@misc{rfc7282, + author = {Resnick, P.}, + title = {{On Consensus and Humming in the IETF}}, + howpublished = {RFC 7282, IETF}, + year = {2014}, + url = {https://www.rfc-editor.org/rfc/rfc7282}, + note = {Read in part (introduction).} +} + +@misc{rfc8890, + author = {Nottingham, M.}, + title = {{The Internet is for End Users}}, + howpublished = {RFC 8890, Internet Architecture Board (informational)}, + year = {2020}, + url = {https://www.rfc-editor.org/rfc/rfc8890}, + note = {Read in part (abstract, section 1, section 4.1).} +} + +@misc{w3c2025process, + author = {{World Wide Web Consortium}}, + title = {{W3C Process Document}}, + year = {2025}, + url = {https://www.w3.org/policies/process/20250818/}, + note = {Read in part (consensus and dissent sections; contents). Version of 18 August 2025.} +} + +@misc{w3c2024wcag22, + author = {{World Wide Web Consortium}}, + title = {{Web Content Accessibility Guidelines (WCAG) 2.2}}, + year = {2024}, + url = {https://www.w3.org/TR/2024/REC-WCAG22-20241212/}, + note = {Read in part (status block, success criteria list, conformance headings). W3C Recommendation, 12 December 2024.} +} + +@misc{doj2024ada, + author = {{U.S. Department of Justice}}, + title = {{Nondiscrimination on the Basis of Disability; Accessibility of Web Information and Services of State and Local Government Entities}}, + howpublished = {89 Fed. Reg. 31320 (final rule)}, + year = {2024}, + url = {https://www.federalregister.gov/documents/2024/04/24/2024-07758}, + note = {Abstract only (Federal Register API abstract; the full text was downloaded but not read).} +} + +@misc{doj2026extension, + author = {{U.S. Department of Justice}}, + title = {{Extension of Compliance Dates for Nondiscrimination on the Basis of Disability; Accessibility of Web Information and Services of State and Local Government Entities}}, + howpublished = {91 Fed. Reg. 20902 (interim final rule)}, + year = {2026}, + url = {https://www.federalregister.gov/documents/2026/04/20/2026-07663}, + note = {Read in part (abstract, background, cost statements). Effective 20 April 2026.} +} + +@misc{eu2019eaa, + author = {{European Parliament and Council of the European Union}}, + title = {{Directive (EU) 2019/882 on the accessibility requirements for products and services (European Accessibility Act)}}, + howpublished = {OJ L 151, 7.6.2019, p. 70}, + year = {2019}, + url = {http://publications.europa.eu/resource/celex/32019L0882}, + note = {Read in part (Arts. 2 to 4, 15, 30; recital on presumption of conformity). Official Journal text via the Publications Office (EUR-Lex itself was behind a JavaScript challenge).} +} + +@inproceedings{felt2016rethinking, + author = {Felt, Adrienne Porter and Reeder, Robert W. and Ainslie, Alex and Harris, Helen and Walker, Max and Thompson, Christopher and Acer, Mustafa Emre and Morant, Elisabeth and Consolvo, Sunny}, + title = {{Rethinking Connection Security Indicators}}, + booktitle = {Twelfth Symposium on Usable Privacy and Security (SOUPS 2016)}, + publisher = {USENIX Association}, + year = {2016}, + url = {https://www.usenix.org/conference/soups2016/technical-sessions/presentation/porter-felt}, + note = {Read in part (abstract, section 1, 5.1.5, 5.2.1 to 5.2.2, start of 6, keyword scan). No DOI. Authors are Google staff and one Berkeley student; the sample skews tech-savvy by the authors' own account.} +} + +@misc{chromium2018evolving, + author = {Schechter, Emily}, + title = {{Evolving Chrome's security indicators}}, + howpublished = {Chromium Blog, 17 May 2018}, + year = {2018}, + url = {https://blog.chromium.org/2018/05/evolving-chromes-security-indicators.html}, + note = {Read in full.} +} + +@misc{helpnet2023lock, + author = {{Help Net Security}}, + title = {{Google Chrome will lose the ``lock'' icon for HTTPS-secured sites}}, + year = {2023}, + url = {https://www.helpnetsecurity.com/2023/05/03/google-chrome-https/}, + note = {Snippet only (WebSearch summary; page not opened; original Google post not found). Do not quote.} +} + +@inproceedings{dhamija2006why, + author = {Dhamija, Rachna and Tygar, J. D. and Hearst, M.}, + title = {{Why phishing works}}, + booktitle = {Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI 2006)}, + pages = {581--590}, + year = {2006}, + doi = {10.1145/1124772.1124861}, + note = {Abstract only (Consensus record). DOI resolved at Crossref, authors and pages match.} +} + +@inproceedings{egelman2008warned, + author = {Egelman, Serge and Cranor, L. and Hong, J.}, + title = {{You've been warned: an empirical study of the effectiveness of web browser phishing warnings}}, + booktitle = {Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI 2008)}, + pages = {1065--1074}, + year = {2008}, + doi = {10.1145/1357054.1357219}, + note = {Abstract only (Consensus record). DOI resolved at Crossref, authors and pages match.} +} + +@inproceedings{akhawe2013alice, + author = {Akhawe, Devdatta and others}, + title = {{Alice in Warningland: a large-scale field study of browser security warning effectiveness}}, + booktitle = {22nd USENIX Security Symposium}, + year = {2013}, + url = {https://www.usenix.org/conference/usenixsecurity13/technical-sessions/presentation/akhawe}, + note = {Abstract only (Consensus record). USENIX page title confirmed; author list beyond the first not checked; no DOI.} +} + +@article{shangguan2019meta, + author = {Shangguan, Siyi and others}, + title = {{A meta-analysis of food labeling effects on consumer diet behaviors and industry practices}}, + journal = {American Journal of Preventive Medicine}, + volume = {56}, + number = {2}, + pages = {300--314}, + year = {2019}, + doi = {10.1016/j.amepre.2018.09.024}, + note = {Abstract only (Consensus record). DOI resolved at Crossref (eleven authors; only the first is listed).} +} + +@article{ganderats2023front, + author = {Ganderats-Fuentes, Montserrat and Morgan, Sherry}, + title = {{Front-of-package nutrition labeling and its impact on food industry practices: a systematic review of the evidence}}, + journal = {Nutrients}, + volume = {15}, + number = {11}, + pages = {2630}, + year = {2023}, + doi = {10.3390/nu15112630}, + note = {Abstract only (Consensus record). DOI resolved at Crossref; second author is Sherry Morgan (Crossref, AUDIT_10).} +} + +@article{hatton2026product, + author = {Hatton, C. and others}, + title = {{Product reformulation in response to the 2016 FDA updated Nutrition Facts label}}, + journal = {JAMA Health Forum}, + volume = {7}, + number = {5}, + pages = {e261253}, + year = {2026}, + doi = {10.1001/jamahealthforum.2026.1253}, + note = {Abstract only (Consensus record). DOI resolved at Crossref.} +} + +@article{stadelmann2018how, + author = {Stadelmann, Marcel and Schubert, Renate}, + title = {{How do different designs of energy labels influence purchases of household appliances? A field study in Switzerland}}, + journal = {Ecological Economics}, + volume = {144}, + pages = {112--123}, + year = {2018}, + doi = {10.1016/j.ecolecon.2017.07.031}, + note = {Abstract only (Consensus record). DOI resolved at Crossref} +} + +@article{waechter2015desired, + author = {Waechter, Signe and S{\"u}tterlin, Bernadette and Siegrist, Michael}, + title = {{Desired and undesired effects of energy labels: an eye-tracking study}}, + journal = {PLOS ONE}, + volume = {10}, + number = {7}, + pages = {e0134132}, + year = {2015}, + doi = {10.1371/journal.pone.0134132}, + note = {Abstract only (Consensus record). DOI resolved at Crossref.} +} + +@article{gorwa2019what, + author = {Gorwa, Robert}, + title = {{What is platform governance?}}, + journal = {Information, Communication \& Society}, + volume = {22}, + number = {6}, + pages = {854--871}, + year = {2019}, + doi = {10.1080/1369118X.2019.1573914}, + note = {Read in part (abstract, sections 1, 3, start of 4) in the OSF accepted manuscript, https://osf.io/fbu27/download; page numbers cited are the post-print's. DOI resolved at Crossref.} +} + +@book{gillespie2019custodians, + author = {Gillespie, Tarleton}, + title = {{Custodians of the Internet: Platforms, Content Moderation, and the Hidden Decisions That Shape Social Media}}, + publisher = {Yale University Press}, + year = {2018}, + doi = {10.12987/9780300235029}, + note = {Reused key (see references.bib). Metadata only; not reopened in facet O.} +} + +@article{balkin2016information, + author = {Balkin, Jack M.}, + title = {{Information fiduciaries and the First Amendment}}, + journal = {UC Davis Law Review}, + volume = {49}, + number = {4}, + pages = {1183--1234}, + year = {2016}, + url = {https://lawreview.law.ucdavis.edu/sites/g/files/dgvnsk15026/files/media/documents/49-4_Balkin.pdf}, + note = {Read in part (contents, Part II.E, Part III, the Facebook and Uber passages; Part VII by keyword). Pages 1183 to 1234 confirmed from the PDF headers. No DOI checked.} +} + +@article{khan2019skeptical, + author = {Khan, Lina M. and Pozen, David E.}, + title = {{A skeptical view of information fiduciaries}}, + journal = {Harvard Law Review}, + volume = {133}, + pages = {497}, + year = {2019}, + url = {https://scholarship.law.columbia.edu/faculty_scholarship/2451}, + note = {Abstract only (repository page via fetch; PDF returned HTTP 403). SSRN abstract 3341661. Start page from the repository citation line.} +} + +@inproceedings{raji2020closing, + author = {Raji, Inioluwa Deborah and Smart, Andrew and White, Rebecca N. and Mitchell, Margaret and Gebru, Timnit and Hutchinson, Ben and Smith-Loud, Jamila and Theron, Daniel and Barnes, Parker}, + title = {{Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing}}, + booktitle = {Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (FAT* 2020)}, + pages = {33--44}, + year = {2020}, + doi = {10.1145/3351095.3372873}, + eprint = {2001.00973}, + archiveprefix = {arXiv}, + note = {Read in part (abstract, sections 1, 2.3, 2.4, framework outline). DOI resolved at Crossref; arXiv PDF read.} +} + +@misc{eu2022dsa, + author = {{European Parliament and Council of the European Union}}, + title = {{Regulation (EU) 2022/2065 on a Single Market For Digital Services (Digital Services Act)}}, + howpublished = {OJ L 277, 27.10.2022, p. 1}, + year = {2022}, + url = {http://publications.europa.eu/resource/celex/32022R2065}, + note = {Reused key. Read in part in facet O: Arts. 21, 25, 37, 38, 52 in the Official Journal text via the Publications Office (facet J read other articles).} +} + +@misc{holznagel2025shortcomings, + author = {Holznagel, Daniel}, + title = {{Shortcomings of the first DSA audits, and how to do better}}, + howpublished = {DSA Observatory, 11 June 2025}, + year = {2025}, + url = {https://dsa-observatory.eu/2025/06/11/shortcomings-of-the-first-dsa-audits-and-how-to-do-better/}, + note = {Read in part (opening and Observation 1). Blog analysis by one author; not peer reviewed.} +} + +@misc{eu2024platform, + author = {{European Parliament and Council of the European Union}}, + title = {{Directive (EU) 2024/2831 on improving working conditions in platform work}}, + howpublished = {OJ L, 11.11.2024}, + year = {2024}, + url = {http://publications.europa.eu/resource/celex/32024L2831}, + note = {Read in part (Arts. 2, 7 to 15, 20 to 25, 29(1); recitals by keyword). Official Journal text via the Publications Office (EUR-Lex itself was behind a JavaScript challenge). Transposition by 2 December 2026 (Art. 29(1)).} +} + +@misc{boe2021rdl9, + author = {{Jefatura del Estado}}, + title = {{Real Decreto-ley 9/2021, de 11 de mayo, por el que se modifica el texto refundido de la Ley del Estatuto de los Trabajadores, para garantizar los derechos laborales de las personas dedicadas al reparto en el ámbito de plataformas digitales}}, + howpublished = {BOE-A-2021-7840}, + year = {2021}, + url = {https://www.boe.es/buscar/act.php?id=BOE-A-2021-7840}, + note = {Read in part (preamble paragraphs on algorithms; the single article; first provisions). Text from the BOE XML. Title in the XML matches.} +} + +@misc{de1972betrvg, + author = {{Bundesministerium der Justiz}}, + title = {{Betriebsverfassungsgesetz (BetrVG), sections 80, 87, 90, 95}}, + howpublished = {gesetze-im-internet.de}, + year = {n.d.}, + url = {https://www.gesetze-im-internet.de/betrvg/}, + note = {Read in full (the four sections; German text). Consolidated text as shown on the site on 2026-10-13. Year of the original Act (1972) is in the key; not checked on the page.} +} + +@phdthesis{neumann2025authority, + author = {Neumann, Denis}, + title = {{``We were an authority they couldn't bypass'': How Works Councils reshape power and control dynamics in the German platform economy}}, + school = {University of Leeds}, + year = {2025}, + url = {https://etheses.whiterose.ac.uk/id/eprint/38626/}, + note = {Abstract only (eprint page). PDF embargoed until 1 June 2029 (White Rose eTheses; checked AUDIT_10). Single-case PhD thesis, not peer reviewed.} +} + +@article{doellgast2022negotiating, + author = {Doellgast, Virginia and Wagner, Ines and O'Brady, Sean}, + title = {{Negotiating limits on algorithmic management in digitalised services: cases from Germany and Norway}}, + journal = {Transfer: European Review of Labour and Research}, + volume = {29}, + number = {1}, + pages = {105--120}, + year = {2022}, + doi = {10.1177/10242589221143044}, + note = {Abstract checked in the audit via Consensus (SAGE page returned 403); article not opened. DOI resolved at Crossref, authors and pages match.} +} + +@misc{eurofound2024hilfr, + author = {{Eurofound}}, + title = {{Collective agreement on use of AI and algorithms signed (Hilfr and 3F)}}, + howpublished = {Platform Work Repository, record 4438}, + year = {2024}, + url = {https://apps.eurofound.europa.eu/platformeconomydb/collective-agreement-on-use-of-ai-and-algorithms-signed-110187}, + note = {Read in full (the record; secondary summary; the agreement itself not opened).} +} + +@misc{wageindicator2022justeat, + author = {{WageIndicator (reproducing the ETUI Collective bargaining newsletter)}}, + title = {{Spain -- Collective agreement for food delivery platform -- January 31, 2022}}, + year = {2022}, + url = {https://wageindicator.org/labour-laws/collective-bargaining/2022/spain-collective-agreement-for-food-delivery-platform-january-31-2022}, + note = {Read in part (one paragraph). Secondary summary; the agreement not opened.} +} + +@misc{eurofound2025glovo, + author = {{Eurofound}}, + title = {{Glovo Faces 450m Spanish Social Security Bill for Misclassifying Riders}}, + howpublished = {Platform Work Repository, record 4522}, + year = {2025}, + url = {https://apps.eurofound.europa.eu/platformeconomydb/glovo-faces-450m-spanish-social-security-bill-for-misclassifying-riders-110271}, + note = {Read in full (the record; secondary summary of Spanish press).} +} + +@article{cox2010review, + author = {Cox, Michael and Arnold, Gwen and Villamayor Tom{\'a}s, Sergio}, + title = {{A review of design principles for community-based natural resource management}}, + journal = {Ecology and Society}, + volume = {15}, + number = {4}, + pages = {38}, + year = {2010}, + doi = {10.5751/es-03704-150438}, + note = {Read in part (abstract, method, Table 3, sections on principles 3, 4, 7). DOI resolved at Crossref; open PDF read.} +} + +@article{calacci2022bargaining, + author = {Calacci, Dan and Pentland, Alex}, + title = {{Bargaining with the black-box: designing and deploying worker-centric tools to audit algorithmic management}}, + journal = {Proceedings of the ACM on Human-Computer Interaction}, + volume = {6}, + number = {CSCW2}, + pages = {428:1--428:24}, + year = {2022}, + doi = {10.1145/3570601}, + note = {Read in part (abstract, introduction, results and limitations, keyword scan) in the MIT DSpace open copy. DOI resolved at Crossref.} +} + +@article{delacroix2019bottom, + author = {Delacroix, Sylvie and Lawrence, Neil D.}, + title = {{Bottom-up data Trusts: disturbing the `one size fits all' approach to data governance}}, + journal = {International Data Privacy Law}, + volume = {9}, + pages = {ipz014}, + year = {2019}, + doi = {10.1093/idpl/ipz014}, + note = {Read in part (abstract; keyword scan of the 43-page SSRN accepted manuscript). DOI resolved at Crossref. Issue and pages not checked.} +} + +@article{rosenblat2016algorithmic, + author = {Rosenblat, Alex and Stark, Luke}, + title = {Algorithmic Labor and Information Asymmetries: A Case Study of {Uber}'s Drivers}, + journal = {International Journal of Communication}, + year = {2016}, + volume = {10}, + pages = {3758--3784}, + url = {https://ijoc.org/index.php/ijoc/article/view/4892}, + note = {Reused key, copied from references.bib. Not reopened in facet O (see facets J and X).} +} + +@article{kellogg2020algorithms, + author = {Kellogg, Katherine C. and Valentine, Melissa A. and Christin, Ang{\'e}le}, + title = {{Algorithms at Work: The New Contested Terrain of Control}}, + journal = {Academy of Management Annals}, + volume = {14}, + number = {1}, + pages = {366-410}, + year = {2020}, + doi = {10.5465/annals.2018.0174}, + note = {Reused key, copied from references.bib (Crossref verified there; abstract only). Not reopened in facet O.} +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_O_custody_of_conventions.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_O_custody_of_conventions.md new file mode 100644 index 00000000..d505e066 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_O_custody_of_conventions.md @@ -0,0 +1,354 @@ +# O — Custody of conventions: who has held and enforced them, and what binds the interface owner + +*Facet O, 2026-10-13. Agent-drafted. The author has not reviewed it, and none of the text is the author's.* + +*Audited 2026-10-13 (AUDIT_10.md); WRONG and OVERSTATED rows corrected in place.* + +**What this facet covers.** The outline's section 2 says "scribes and readers shared mutual communicative interests" (the same idea is in `DRAFT.md:166`: "Environmental conventions like word separation spread because scribes and readers shared a common linguistic community"). Section 7 asks who can hold a convention against the party that owns the interface. Facets J and A hold courts, dealer foreign exchange, platform terms of service, the DSA and P2B in part, and the record of the page. I look for custody literatures they do not hold: (Q1) who held and enforced the written page in the Carolingian reform; (Q2) standards bodies, accessibility law, the browser padlock and mandated labels; (Q3) platform-governance scholarship, information fiduciaries and audit regimes; (Q4) statute and bargaining on algorithmic management (Directive (EU) 2024/2831, Spain's Royal Decree-Law 9/2021, German co-determination, collective agreements); (Q5) Ostrom's design principles and worker data collectives. + +**The question the synthesis answers.** Which custody arrangements give the weaker party something other than information, and which of them bind the interface owner so that it cannot change the convention alone? Short answer, argued in (c): very few. The Carolingian page was held by the ruler and the church, not by a scribe-reader community, so the outline's sentence has no support in the sources I opened. Standards bodies write conventions that owners adopt or drop at will; Chrome removed its "Secure" label and HTTPS scheme in 2018 on Google's own schedule (a second change, the 2023 lock icon, rests on press only [unverified: see AUDIT_10]). Among the legal instruments, only three features bind an owner against its own wish: a mandated format (labels, DSA Arts. 25 and 38), a rule that a human takes the decision and a decision that infringes the worker's rights is rectified within two weeks or compensated (Directive 2024/2831 Arts. 10(5), 11(3); the Directive has been in force since 1 December 2024, and national transposition is due on 2 December 2026), and a works council's co-determination right with a binding arbitration body (BetrVG §87(1) no. 6 and (2)). The one study I found of works councils in platform work (a single-case thesis, abstract only; the PDF is embargoed until 1 June 2029) finds co-determination "excluded from the core of the algorithmic work organisation"; the abstract does not name § 87(1) no. 6. + +**Legend for read status.** *Read in full* = I opened the whole text this session. *Read in part* = I opened the text and read the named sections or passages. *Abstract only* = a database abstract (Consensus, Crossref, OpenAlex, publisher page) and nothing else. *Snippet only* = a search-engine summary of a page I did not open. *Not accessible* = no text. A source is quotable only at full-text depth (RESEARCH_PLAN §3); every quotation below is from text I opened, and quotations from translations or OCR are marked. Downloaded copies are in `~/iit-playground/dxd_foundation_local/pdfs/O/` (outside the repo). + +**Count by read status (one bib key each, 50 keys).** Read in full 5 (`charlemagne1900letter`, `chromium2018evolving`, `eurofound2024hilfr`, `eurofound2025glovo`, `de1972betrvg`); read in part 23; abstract only 16; snippet only 2 (`helpnet2023lock`, `doellgast2022negotiating`); reused and not reopened 4 (`mckitterick1989carolingians`, `gillespie2019custodians`, `kellogg2020algorithms`, `rosenblat2016algorithmic`); not accessible 3 (no bib entry: Bischoff, Ganz 1990, Ostrom 1990). `eu2022dsa` is also a reused key, counted in "read in part" because I re-read five articles. + +--- + +## (a) Search log + +All searches ran on 2026-10-13. "Fetch" = curl with a plain browser user-agent. No bot wall, captcha or paywall was bypassed. + +| # | Source | Exact query / request | Hits | Notes | +|---|---|---|---|---| +| 1 | Repo reads | `RESEARCH_PLAN.md`, `SOURCE.md` (D8), `T_adjudication.md` §9 D8, `J_who_holds_the_convention.md`, `references.bib` greps, `DRAFT.md` grep for "scribes" | n/a | J holds courts, FX, P2B, DSA (Arts. 14, 17, 20, 53–54), Suzor, Klonick, Gillespie (metadata). A holds Parkes and Saenger. The sentence at issue is `DRAFT.md:166`. | +| 2 | Publications Office (CELLAR) | `http://publications.europa.eu/resource/celex/32024L2831` with header `Accept: application/xhtml+xml` | HTTP 200, official OJ text | EUR-Lex proper (`eur-lex.europa.eu`, three URL forms) returned HTTP 202 with an empty body (JavaScript challenge); I did not try to pass it. The Publications Office resource URL serves the same Official Journal text as XHTML. Same route for DSA (`32022R2065`) and EAA (`32019L0882`). | +| 3 | BOE | `https://www.boe.es/diario_boe/xml.php?id=BOE-A-2021-7840` | 200 | Title in the XML matches Real Decreto-ley 9/2021. | +| 4 | gesetze-im-internet.de | `/betrvg/__87.html`, `__90`, `__95`, `__80` | 200 | Official federal site; full text of the four sections. | +| 5 | Federal Register API | `documents.json?conditions[term]=Accessibility of Web Information and Services of State and Local Government Entities`; then `documents/2024-07758.json`, `2026-07663.json` and their `.txt` | 200 | 89 FR 31320 (2024) abstract; 91 FR 20902 (2026) full text. | +| 6 | RFC Editor | `rfc2026.txt`, `rfc7282.txt`, `rfc8890.txt` | 200 | | +| 7 | W3C | `/policies/process/` and `/TR/WCAG22/` | 200 | Process Document of 18 August 2025; WCAG 2.2 Recommendation of 12 December 2024. | +| 8 | USENIX | `soups2016-paper-porter-felt.pdf` | 200 | Felt et al. 2016 PDF, 12 pages. | +| 9 | Chromium blog | `blog.chromium.org/2018/05/evolving-chromes-security-indicators.html` | 200 | | +| 10 | archive.org | `capitulariaregum01bore` (MGH Capitularia I, 1883) and `poetaelatiniaevi00dm` (MGH Poetae I, 1881), `_djvu.txt` via the item's server | 200 | OCR text; I corrected OCR c/e confusions only where I say so. | +| 11 | archive.org search-inside | `fulltext/inside.php` for McKitterick 1989 (`carolingianswrit0000mcki`), query "Admonitio generalis" | "Item not available" | Guest users cannot search inside lending books. McKitterick is not reopened. Bischoff, *Latin Palaeography* (no open scan found in title and creator searches); Ganz, *Corbie in the Carolingian Renaissance* (`corbieincaroling0000ganz`, lending only). | +| 12 | Google Books API | McKitterick snippet search | HTTP 429 "Queries per day" exceeded | Not retried. | +| 13 | Fordham Medieval Sourcebook | `sourcebooks.web.fordham.edu/source/carol-baugulf.asp` | 200 | Munro's 1900 translation of Charlemagne's letter; read in full. | +| 14 | Cambridge Apollo | `doi.org/10.17863/cam.23255` → repository item → bitstream | 200, 319-page PDF | Westwell thesis; read in part. | +| 15 | Vatican Library Spotlight | `spotlight.vatlib.it/latin-paleography/feature/11-1-bookhands-of-the-frankish-regions` | 200 | Curator text by A. M. Piazzoni. | +| 16 | OpenAlex | `works?search=` (3 queries) | HTTP 429 | Retried with back-off; later `works/doi:` lookups worked (Gorwa: gold OA, OSF accepted manuscript). | +| 17 | Consensus | (i) `Carolingian reform of script and correction of texts: who benefited…`; (ii) `Caroline minuscule standardisation scriptoria Charlemagne Alcuin Tours`; (iii) `Carolingian correctio of texts and Admonitio generalis…`; (iv) `meta-analysis food labeling effects on consumer diet behaviors and industry reformulation`; (v) `energy efficiency label consumer purchase effect…`; (vi) `browser security indicators users ignore SSL lock phishing study Why phishing works` | 20 each | Source of the abstracts marked "abstract only". | +| 18 | WebSearch | Admonitio c. 72; Epistola de litteris colendis; Khan–Pozen; Gorwa; Calacci; Delacroix–Lawrence; collective agreements and algorithms; works councils; Chrome lock icon; Rider Law effects; DSA audits; Carolingian scriptoria | 9–10 each | Leads. Items I did not open are "snippet only". | +| 19 | Blocked or failed | Columbia Law repository PDF of Khan–Pozen (HTTP 403, curl and WebFetch); Taylor & Francis PDF of Gorwa (403); Manchester Hive PDF of Mostert (403); SAGE page of Doellgast et al. (403); Oxford Academic PDF of Delacroix–Lawrence (403); UAB repository (INCODING policy briefs) and Heidelberg *Francia* (Ganz review) behind an Anubis "not a bot" challenge | n/a | Recorded; not bypassed. | +| 20 | Crossref | 21 DOIs | 21 resolved to the intended title | `10.17863/cam.23255` is a DataCite DOI (Crossref 404); `doi.org` resolves it to the Cambridge repository item. Felt et al. (USENIX), the RFCs, the Chromium post and the legal texts have no DOI. | +| 21 | Scholar Gateway | not tried | 0 | Failed in every earlier run. | + +--- + +## (b) Entries + +Format: **key** — reference. *Verification.* *Read status.* *What it says* with locators. *Bearing.* + +### Q1. Who held and enforced the conventions of the written page + +**charlemagne1900letter** — Charlemagne, letter to Abbot Baugulf of Fulda ("De litteris colendis"), c. 780–800; Boretius no. 29, p. 78; English by D. C. Munro in *Translations and Reprints from the Original Sources of European History*, vol. 6, no. 5, pp. 12–14 (Philadelphia, 1900); web text in the Fordham Internet Medieval Sourcebook. *Verification:* page opened; the source line names Boretius, Munro and the series. *Read status:* read in full (translation, not the Latin). +- The ruler states the aim as correct speech: "those who desire to please God by living rightly should not neglect to please him also by speaking correctly." +- The letter judges the monasteries' own writing: "we have recognized in most of these letters both correct thoughts and uncouth expressions; because what pious devotion dictated faithfully to the mind, the tongue, uneducated on account of the neglect of study, was not able to express in the letter without error." +- The stake is Scripture: "as the skill in writing was less, so also the wisdom for understanding the Holy Scriptures might be much less than it rightly ought to be." +- Enforcement is a request backed by favour: "Do not neglect, therefore, if you wish to have our favor, to send copies of this letter to all your suffragans and fellow-bishops and to all the monasteries." +- *Bearing.* The party that fixes the standard of correctness is the king, writing to institutions that are to be "the especial servants of truth". The letter says nothing of a shared interest between the writing monk and a reader. It says the king found the product wanting. Complicates the outline's sentence. + +**boretius1883capitularia** — Boretius, A. (ed.) (1883). *Capitularia regum Francorum*, vol. 1 (MGH, Legum sectio II). No. 22, *Admonitio generalis* (23 March 789). *Verification:* archive.org item `capitulariaregum01bore`, OCR text; the heading "22. Admonitio generalis. 789. Mart. 23" and chapter 72 located in the OCR. *Read status:* read in part (c. 72 and the lines around it; OCR, with c/e confusions in the apparatus only). +- c. 72, p. 60 (chapter opens p. 59): "Et ut scolae legentium puerorum fiant. Psalmos, notas, cantus, compotum, grammaticam per singula monasteria vel episcopia et libros catholicos bene emendate; quia saepe, dum bene aliqui Deum rogare cupiunt, sed per inemendatos libros male rogant. Et pueros vestros non sinite eos vel legendo vel scribendo corrumpere; et si opus est euangelium, psalterium et missale scribere, perfectae aetatis homines scribant cum omni diligentia." +- My gloss (agent's, not a translation of record): the king orders schools and correct books in every monastery and bishopric, because people who want to pray well "pray badly through uncorrected books"; the boys must not spoil texts by reading or writing, and a gospel book, psalter or missal is to be copied by grown men with all care. +- *Bearing.* The text addresses the clergy as subordinates ("vestros") and names a defect (error in the book) and a victim (the person praying). The scribe appears as a worker to be supervised by age and care, not as a party with an interest of his own. Complicates the outline's sentence. The passage concerns correctness of text, not word separation; no source I opened ties it to spacing. + +**dummler1881alcuini** — Dümmler, E. (ed.) (1881). *Poetae Latini aevi Carolini*, vol. 1 (MGH, Poetarum Latinorum medii aevi I). *Alcuini carmina*, no. XCIV, p. 320. *Verification:* archive.org item `poetaelatiniaevi00dm`, OCR text; editor named on the title page ("Ernestus Duemmler"). *Read status:* read in part (carm. XCIV, 16 lines; OCR with c/e confusions, which I corrected silently in the quotation: "Pcr", "Nc", "ecclesia"). +- ll. 5–10: "Correctosque sibi quaerant studiose libellos, / Tramite quo recto penna volantis eat. / Per cola distinguant proprios et commata sensus, / Et punctos ponant ordine quosque suo, / Ne vel falsa legat, taceat vel forte repente / Ante pios fratres lector in ecclesia." +- ll. 13–14: "Fodere quam vites melius est scribere libros, / Ille suo ventri serviet, iste animae." +- *Bearing.* A teacher-abbot's verse instructs scribes to seek corrected exemplars and to mark sense units so that the reader in church neither reads wrongly nor falls silent. The party named as harmed is the lector, and the party giving the order is the head of the house. This is the nearest thing in the sources to "scribes working for readers", and the alignment is produced by an instruction. (My reading; the poem's use as a scriptorium inscription is not shown by what I opened.) Supports a custody story; complicates "mutual". + +**westwell2018dissemination** — Westwell, A. (2018). *The dissemination and reception of the Ordines Romani in the Carolingian Church, c. 750–900*. PhD thesis, University of Cambridge. https://doi.org/10.17863/cam.23255. *Verification:* DOI resolves at doi.org to the Apollo repository item; PDF downloaded (319 pp.). *Read status:* read in part (abstract; the historiographic introduction, printed pp. 1–6; keyword search of the whole text; the conclusion, printed pp. 288 ff.). The thesis is about liturgical texts, not script. +- Abstract: "what occurred under the Carolingians in the liturgy did not involve the imposition of the Roman rite from above. What was 'Roman' and 'correct' was decided by individuals, each in their own case, and they created and edited texts for what they needed." +- Introduction (PDF p. 16): "McKitterick pointed out that Carolingian interventions in the liturgy appear to be local initiatives, and differed in nature from one place to another. Hen revealed the total lack of evidence for any centralised imposition of…" (sentence continues on the next page). +- Conclusion (PDF p. 300): "Closeness to the centre of power, even to the shaping of the very decrees and edicts traditionally seen as demanding uniformity, has not made that supposed uniformity any more real." +- *Bearing.* A modern specialist rejects the picture of a top-down uniform reform in the liturgy. For the paper, this cuts two ways. Against "the king imposed the page": local houses chose and edited. Against "scribes and readers shared interests": the local actors were bishops and communities, "intensive networks of exchange", not scribes and lay readers. It does not speak to script or word separation. + +**piazzoni11bookhands** — Piazzoni, A. M. (curator text), "11.1 Bookhands of the Frankish regions", in *Latin Paleography: Thematic Pathways on the Web*, Vatican Library (Spotlight). https://spotlight.vatlib.it/latin-paleography/feature/11-1-bookhands-of-the-frankish-regions. *Verification:* page opened; author line "A. M. Piazzoni" in the page source. *Read status:* read in part (§§11.1.1–11.1.3). A web exhibition text, not a peer-reviewed article; the page is under copyright, so I quote briefly. +- On Corbie: "the scriptorium of the abbey constituted a real experimental laboratory of research, often linked to the figures of the various abbots that had the office of guiding the monastery." +- On the Maurdramnus Bible: "Extremely regular and calligraphic, with rounded letters and well-spaced words, this script has been considered by some paleographers as the first example of a Caroline minuscule." Also: Corbie was "closely connected to the court and therefore also to its chancery." +- *Bearing.* Script styles are tied to abbots and courts in this account. A source that holds both "well-spaced words" and an abbot's commission for the same codex, which is useful to section 2, but I read only this page. + +**mckitterick1989carolingians** (reused key) — McKitterick, R. (1989). *The Carolingians and the Written Word.* Cambridge University Press. *Read status:* not reopened this run (search-inside blocked, row 11). Facet C read it in part, found "Drivers named. The state and church: the *Admonitio generalis* of 789" and no passage on legibility or word division. The Consensus listing gives its chapters: "1. The spoken and the written word 2. law and the written word … 4. The production and possession of books: an economic dimension 5. The organization of written knowledge 6. The literacy of the laity". *Bearing.* The standard book frames the page as an instrument of law, church and economy. A library copy should be checked for chapter 4 and for what it says about who paid for and ordered scribal work. + +**ganz1995book** — Ganz, D. (1995). Book production in the Carolingian empire and the spread of Caroline minuscule. In *The New Cambridge Medieval History*, vol. 2. https://doi.org/10.1017/chol9780521362924.032. *Verification:* Crossref title match (pp. 786–808). *Read status:* abstract only. The abstract says "Book production throughout the empire on a scale required an increase in the number of scribes, and the formalisation of their training. Carolingian scriptoria developed a uniform script, Caroline minuscule" and "Carolingian scribes used enlarged letters to indicate the start of a new section, and punctuation marks to guide the reader". *Bearing.* Training and scale are institutional; the abstract does not say who ordered the uniformity. + +**mostert2016pray** — Mostert, M. (2016). "…but they pray badly using corrected books": errors in early Carolingian copies of the *Admonitio generalis*. In *Religious Franks*, Manchester UP. https://doi.org/10.7765/9781784997519.00015. *Verification:* Crossref title match; pages 112–127 are from a library-catalogue search snippet, not Crossref. *Read status:* abstract only (publisher PDF returned 403). The abstract: copies were probably made "by dictation", "no-fewer-than-seven independent strands in the tradition"; "The early copies of the Admonitio are written by experienced scribes, who wrote clear hands. These scribes have to be looked for among the elite of their monasteries' scriptoria." *Bearing.* The text that orders correct copying was itself copied in several independent strands, by the best scribes. Evidence of an institutional distribution channel (the *missi*), not of any scribe–reader bargain. + +**parkes2016carolingian** — Parkes, M. B. (2016). Carolingian renovatio: augmenting old notation with new symbols. In *Pause and Effect* (Routledge reprint), pp. 30–34. https://doi.org/10.4324/9781315247243-4. *Verification:* Crossref title match. *Read status:* abstract only. The abstract: "Books were produced on a scale without precedent… Continuous practice occasioned by the need to produce large quantities of books led to a rapid development of scribal skills" and "Scribes had developed a complex of new graphic conventions both to improve the intelligibility of texts copied in minuscule scripts, and to facilitate access to the information transmitted in the written medium." *Bearing.* The only source in this facet that gives the scribes agency in the conventions. It supports a weaker form of the outline's claim (scribes developed conventions for access), and says the production volume came from demand. See also facet A, entry on Parkes 1992. + +**dumville1987english** — Dumville, D. (1987). English Square minuscule script: the background and earliest phases. *Anglo-Saxon England* 16, 147–179. https://doi.org/10.1017/s0263675100003884. *Verification:* Crossref title match. *Read status:* abstract only. The abstract ties the English adoption of Caroline script to "a monastic revolution within the English church (supported by the monarchy but led by Benedictine ideologues…) some of whose proponents probably favoured Caroline writing as a matter of principle". Dumville's 1992 book abstract (Consensus, not separately opened) calls Caroline script "a badge of ecclesiastical reform". *Bearing.* A script as a mark of institutional allegiance, which is a custody-by-order reading. + +**steinova2015psalmos** — Steinová, E. (2015). *Psalmos, notas, cantus*: on the meanings of *nota* in the Carolingian period. *Speculum* 90(2), 424–457. https://doi.org/10.1017/s0038713415000275. *Verification:* Crossref title match. *Read status:* abstract only. The abstract: the Admonitio passage calling for schools "has caused quite a few problems for scholars on account of its seemingly corrupt grammar and ambiguous vocabulary." *Bearing.* A caution on c. 72: the passage I quote in Latin is itself disputed in its terms. Check before building on "notas". + +### Q2. Standards bodies, accessibility, the padlock, mandated labels + +**rfc2026** — Bradner, S. (1996). *The Internet Standards Process — Revision 3.* RFC 2026 (BCP 9). https://www.rfc-editor.org/rfc/rfc2026. *Verification:* RFC Editor text. *Read status:* read in part (§1, §1.1, the open-participation clauses). +- §1.1: the Internet "supports host-to-host communication through voluntary adherence to open protocols and procedures defined by Internet Standards." +- Participation: announcements "shall contain (or provide pointers to) all of the information that is necessary to support the participation of any interested individual." +- *Bearing.* The convention is written in an open process and binds nobody; adherence is voluntary. + +**rfc7282** — Resnick, P. (2014). *On Consensus and Humming in the IETF.* RFC 7282. https://www.rfc-editor.org/rfc/rfc7282. *Read status:* read in part (Introduction). "we don't let a single individual dictate decisions (a king or president), nor should decisions be made by a vote… We only require rough consensus." *Bearing.* Who can change: a working-group chair may declare rough consensus "the objection notwithstanding". + +**rfc8890** — Nottingham, M. (2020). *The Internet Is for End Users.* IAB, RFC 8890 (informational). https://www.rfc-editor.org/rfc/rfc8890. *Read status:* read in part (Abstract, §1, §4.1). +- Abstract: "when there is a conflict between the interests of end users of the Internet and other parties, IETF decisions should favor end users." +- §4.1: "The IETF community does not have any unique insight into what is 'good for end users'"; "Affected parties are rarely technical experts… Often, even when we try to engage a broader audience, their participation is minimal -- until a change affects someone in a way they don't like." +- Status line: "not an Internet Standards Track specification." +- *Bearing.* The standards body states the principle the paper wants (favour the weaker party) as aspiration, concedes that the weaker party rarely attends, and has no enforcement. Supports the claim that custody by a standards body gives voice without a hold on the owner. + +**w3c2025process** — W3C (18 August 2025). *W3C Process Document.* https://www.w3.org/policies/process/20250818/. *Read status:* read in part (§5.2 Consensus, §5.2.2 Managing Dissent; contents). "Consensus is a core value of W3C." §5.2.2: "Dissenters cannot stop a group's work simply by saying that they cannot live with a decision." Formal Objections are heard by the Advisory Board or TAG (§§5.5–5.6 in the contents). *Bearing.* Same shape as the IETF: a process for writing, an escape valve for dissent, no sanction on an implementer. + +**w3c2024wcag22** — W3C (12 December 2024). *Web Content Accessibility Guidelines (WCAG) 2.2.* https://www.w3.org/TR/2024/REC-WCAG22-20241212/. *Read status:* read in part (status block; success-criteria list; §5.2 Conformance Requirements headings). *Bearing.* A W3C Recommendation, editors from Nomensa, Oracle, the Library of Congress, W3C and Adobe. Its force comes only from being cited in law (next two entries). + +**doj2026extension** — U.S. Department of Justice. Extension of Compliance Dates for Nondiscrimination on the Basis of Disability; Accessibility of Web Information and Services of State and Local Government Entities. 91 FR 20902 (20 April 2026, interim final rule). https://www.federalregister.gov/documents/2026/04/20/2026-07663. *Verification:* Federal Register API record; full text downloaded. *Read status:* read in part (abstract, background, the statements on costs). And **doj2024ada** — 89 FR 31320 (24 April 2024), same title without "Extension": *abstract only* (API abstract; the 12,000-line text was not read). +- 2026 background: the 2024 rule "adopted the Web Content Accessibility Guidelines ('WCAG') version 2.1 Level AA… as the technical standard for web content and mobile app accessibility under title II." +- Abstract: "The compliance date for State and local government entities with a total population of 50,000 or more is extended from April 24, 2026, to April 26, 2027." +- *Bearing.* A standards-body text became binding on state and local governments by being incorporated into a regulation, and the binding date then moved a year on cost grounds. It covers public entities, not private interface owners. + +**eu2019eaa** — Directive (EU) 2019/882 (European Accessibility Act). OJ L 151, 7.6.2019, p. 70. *Verification:* Publications Office text (`celex/32019L0882`). *Read status:* read in part (Arts. 2–4, 15, 30; recital on presumption of conformity). +- Art. 4(1): "Member States shall ensure… that economic operators only place on the market products and only provide services that comply with the accessibility requirements set out in Annex I." +- Art. 15(1): "Products and services which are in conformity with harmonised standards… shall be presumed to be in conformity with the accessibility requirements of this Directive." +- Art. 30(2): penalties "shall be effective, proportionate and dissuasive." +- *Bearing.* Here the interface owner (a supplier of covered services) is bound by statute, with a voluntary technical standard as the safe harbour. The standard-writers do not enforce; the member state does. Applies "after 28 June 2025" (Art. 2). + +**felt2016rethinking** — Felt, A. P., Reeder, R. W., Ainslie, A., Harris, H., Walker, M., Thompson, C., Acer, M. E., Morant, E., & Consolvo, S. (2016). Rethinking connection security indicators. *SOUPS 2016*. https://www.usenix.org/conference/soups2016/technical-sessions/presentation/porter-felt. *Verification:* USENIX PDF. *Read status:* read in part (Abstract; §1; §5.1.5 demographics; §5.2.1–5.2.2 survey results; §6 opening; keyword scan of the rest). +- p. 1: "people often find browser security indicators confusing. Researchers have cautioned since 2002 that people don't always understand security indicators." +- §5.1.5: "our sample may be biased toward tech-savvy users" (the respondents were 90.4% male; recruited through TechCrunch, Reddit, the Chrome Web Store). +- §5.2.1 (Table 2): of 733 respondents describing the green lock, 40.1% gave a "connection" answer and 13.4% an "identity" answer; "Some respondents mentioned identity but incorrectly said that HTTPS vouched for the website's trustworthiness… (about 3% of the total)." +- §5.2.2: for the HTTP icon, "7.1% responses said they Don't know"; "9% of respondents thought the indicator was the default favicon"; "1.5% of responses thought that the HTTP indicator meant the opposite: that the page is secure." +- Abstract: "Our proposed indicators have been adopted by Google Chrome." +- *Bearing.* A measured case of an interface convention whose meaning users misread even in a tech-savvy sample; the fix was written, tested and shipped by the browser owner. The paper reports its own sample as skewed. Supports "conventions need design, not training"; complicates "conventions will never be granted voluntarily" (see c). + +**chromium2018evolving** — Schechter, E. (2018, 17 May). Evolving Chrome's security indicators. *Chromium Blog.* https://blog.chromium.org/2018/05/evolving-chromes-security-indicators.html. *Read status:* read in full (a short post). "Users should expect that the web is safe by default, and they'll be warned when there's an issue… we'll step towards removing Chrome's positive security indicators so that the default unmarked state is secure. Chrome will roll this out over time, starting by removing the 'Secure' wording and HTTPS scheme in September 2018 (Chrome 69)." *Bearing.* The owner announced and executed a change to the meaning of a convention on its own schedule. This is the plainest case of "the owner can change the convention alone". A second change (the lock icon replaced in Chrome 117, 2023) is known to me only from press summaries, see next entry. + +**helpnet2023lock** — Help Net Security (2023, 3 May). Google Chrome will lose the "lock" icon for HTTPS-secured sites. https://www.helpnetsecurity.com/2023/05/03/google-chrome-https/. *Read status:* snippet only (WebSearch summary; page not opened; the original Google post was not found). Summary: Google's team said "only 11% of study participants correctly understood the precise meaning of the lock icon" in 2021 research. Not quotable until the original is opened. + +**dhamija2006why** — Dhamija, R., Tygar, J. D., & Hearst, M. (2006). Why phishing works. *CHI 2006.* https://doi.org/10.1145/1124772.1124861. *Verification:* Crossref title and authors match (Dhamija, Tygar, Hearst; pp. 581–590). *Read status:* abstract only. "23% of the participants did not look at browser-based cues such as the address bar, status bar and the security indicators, leading to incorrect choices 40% of the time." + +**egelman2008warned** — Egelman, S., Cranor, L. F., & Hong, J. (2008). You've been warned. *CHI 2008.* https://doi.org/10.1145/1357054.1357219. *Verification:* Crossref title and authors match (Egelman, Cranor, Hong; pp. 1065–1074). *Read status:* abstract only. "when presented with the active warnings, 79% of participants heeded them, which was not the case for the passive warning that we tested—where only one participant heeded the warnings." *Bearing.* A design that interrupts (active) worked where an indicator to be read (passive) did not; compare facet L. + +**akhawe2013alice** — Akhawe, D., et al. (2013). Alice in Warningland: a large-scale field study of browser security warning effectiveness. *USENIX Security 2013.* https://www.usenix.org/conference/usenixsecurity13/technical-sessions/presentation/akhawe. *Verification:* USENIX page title matches. The author list is not checked at the publisher; I cite "Akhawe et al." in the bib. *Read status:* abstract only. "users continued through… 70.2% of Google Chrome's SSL warnings" against about a tenth of Firefox's malware warnings. "the user experience of a warning can have a significant impact on user behavior." + +**shangguan2019meta** — Shangguan, S., et al. (2019). A meta-analysis of food labeling effects on consumer diet behaviors and industry practices. *Am. J. Prev. Med.* 56(2), 300–314. https://doi.org/10.1016/j.amepre.2018.09.024. *Verification:* Crossref title match. *Read status:* abstract only. 60 studies, "labeling decreased product contents of sodium by 8.9%… and artificial trans fat by 64.3%"; "No significant heterogeneity was identified by… voluntary or legislative approaches." *Bearing.* Labels changed what the industry sold, not only what the buyer knew. Note this review found no difference between voluntary and legislative schemes, which the next entry disputes for front-of-pack labels. + +**ganderats2023front** — Ganderats-Fuentes, M., & Morgan, S. (2023). Front-of-package nutrition labeling and its impact on food industry practices: a systematic review. *Nutrients* 15(11), 2630. https://doi.org/10.3390/nu15112630. *Verification:* Crossref title and authors match (Ganderats-Fuentes, Morgan). *Read status:* abstract only. "Mandatory policies reported higher and more consistent effects on product reformulation compared to voluntary approaches. Voluntary FOP labeling resulted in low uptake and tended to be applied to healthier products… food manufacturers behave strategically by labeling healthier choices." + +**hatton2026product** — Hatton, C., et al. (2026). Product reformulation in response to the 2016 FDA updated Nutrition Facts label. *JAMA Health Forum.* https://doi.org/10.1001/jamahealthforum.2026.1253. *Verification:* Crossref title match. *Read status:* abstract only. "no evidence of substantive reformulation of packaged foods in response to label adoption." *Bearing.* A mandated label that is only a disclosure panel moved the product little. Supports the section-4 claim that disclosure alone is thin. Check the full text before print. + +**stadelmann2018how** / **waechter2015desired** — Stadelmann, M., & Schubert, R. (2018). How do different designs of energy labels influence purchases of household appliances? A field study in Switzerland. *Ecol. Econ.* 144, 112–123; Waechter, S., Sütterlin, B., & Siegrist, M. (2015). Desired and undesired effects of energy labels: an eye-tracking study. *PLOS ONE* 10(7), e0134132. https://doi.org/10.1016/j.ecolecon.2017.07.031 ; https://doi.org/10.1371/journal.pone.0134132. *Verification:* Crossref title, authors and pages match. *Read status:* abstract only. Stadelmann: in a field study with a Swiss online retailer, "the display of any of the two energy labels increases the sales proportion of energy-efficient appliances compared to the absence of any energy label." Waechter (eye-tracking): "the energy label's effect on consumers' actual product choices seems to be rather low" and the format can lead people to "disregard information about actual energy consumption." *Bearing.* A mandated EU label has a measured but modest effect, and its format can misdirect attention. Both abstracts are second-hand; I did not read the studies. + +### Q3. Platform governance, information fiduciaries, audits + +**gorwa2019what** — Gorwa, R. (2019). What is platform governance? *Information, Communication & Society* 22(6), 854–871. https://doi.org/10.1080/1369118X.2019.1573914. *Verification:* Crossref title match; OpenAlex lists the OSF accepted manuscript (https://osf.io/fbu27/download), which I read. Page numbers below are the post-print's (the file says to "consult the final version of record for page numbers"). *Read status:* read in part (abstract; §1; §3 "Self-Governance", "External Governance", "Co-Governance"; §4 opening). +- §3.1 (post-print pp. 12–14): in the self-governance mode "transparency is generally voluntary, and most platform decisions are made with minimal external oversight"; "voluntary arrangements rely on goodwill and provide limited recourse in the case of non-compliance." +- §3.3 (p. 15): co-governance proposals include press-council-style bodies and "ombudsman programmes or third-party adjudication"; the Global Network Initiative has "a system of transparency reporting, third party audits". +- p. 16: co-governance "lends itself to a philosophy that leads away from major, corporatized platforms and towards various platform cooperatives, decentralized systems, and other forms of community self-management", which Gorwa calls "admittedly unlikely in the near-term (due to scale issues, network effects, and other challenges)". +- *Bearing.* Names three models (self, external, co-governance) and says none yet has a record. Gives the paper a map for section 7. Gorwa offers no evidence that any binds a platform. + +**gillespie2019custodians** (reused key) — see `references.bib` and facet J (metadata only). Not reopened; Gillespie's "custodians" frame matches this facet's title, but I have no text to quote. + +**balkin2016information** — Balkin, J. M. (2016). Information fiduciaries and the First Amendment. *UC Davis Law Review* 49(4), 1183–1234. https://lawreview.law.ucdavis.edu/sites/g/files/dgvnsk15026/files/media/documents/49-4_Balkin.pdf. *Verification:* publisher PDF, 52 pp., page headers 1183–1234. *Read status:* read in part (contents; Part II.E on contract; Part III on fiduciaries; the passages on Facebook and Uber; Part VII only by keyword). +- p. 1199 (II.E), on the contract route: "companies like Uber and Facebook can state their privacy policies in vague terms, or change their privacy policies more or less at will, relying on terms of service or end user license agreements may offer only very limited privacy protections." +- p. 1208 (III): fiduciaries owe a duty of care and "the duty of loyalty. Fiduciaries must keep their clients' interests in mind and act in their clients' interests… a duty not to create potential or actual conflicts of interest." +- p. 1227: "'Fiduciary' does not mean 'not for profit'… It is also reasonable to expect that a social networking service is designed to facilitate social networking and not to manipulate you into voting for the candidate of its choice." +- *Bearing.* The model: a court-enforced duty of loyalty for a business that holds data, so that the owner cannot rewrite its obligations in its terms. Balkin himself says the terms route lets the owner "change… at will", the same point as J's P2B Art. 3. It is a proposal about a US constitutional problem; no case showing it binding a platform appears in what I read. + +**khan2019skeptical** — Khan, L. M., & Pozen, D. E. (2019). A skeptical view of information fiduciaries. *Harvard Law Review* 133, 497. https://scholarship.law.columbia.edu/faculty_scholarship/2451. *Verification:* the Columbia repository page and SSRN abstract 3341661 exist (WebFetch of the repository page). *Read status:* abstract only (PDF returned HTTP 403; start page from the repository citation line). The abstract's first sentence: "The concept of 'information fiduciaries' has surged to the forefront of debates on online-platform regulation." The repository summary says the authors warn that the framework could encourage complacency toward platforms' structural power and delay stronger public regulation (paraphrase of a fetch summary, not a quotation). *Bearing.* The critique that matters to section 7 is that a fiduciary label may be a substitute for binding rules. I have not read their argument; a human with library access should read Parts II–III. + +**raji2020closing** — Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., Smith-Loud, J., Theron, D., & Barnes, P. (2020). Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing. *FAT\* '20*, 33–44. https://doi.org/10.1145/3351095.3372873 (arXiv 2001.00973). *Verification:* Crossref title match; arXiv PDF (12 pp.). *Read status:* read in part (abstract; §1; §2.3–2.4; the SMACTR outline; keyword scan). +- §2.4: "External auditing, in which companies are accountable to a third party, [is] fundamentally limited by lack of access to internal processes at the audited organizations… external auditors can only access model outputs, for example by using an API. Auditors do not have access to intermediate models or training data, which are often protected as trade secrets." +- Abstract: the framework is "to be applied throughout the internal organization development lifecycle." +- *Bearing.* Authors from Google and the Partnership on AI propose that the owner audit itself. The custody sits inside the owner, which answers Q3's question on who holds the audit: the party being audited. Supports the claim that audit does not bind unless a statute or third party sets it. + +**eu2022dsa** (reused key) — Regulation (EU) 2022/2065 (Digital Services Act). Facet J read Arts. 14, 17, 20, 53–54 from a mirror, then the card read the operative text. *This run:* I read Arts. 21, 25, 37, 38, 52 in the Official Journal text from the Publications Office (`celex/32022R2065`). *Read status:* read in part. +- Art. 25(1): "Providers of online platforms shall not design, organise or operate their online interfaces in a way that deceives or manipulates the recipients of their service or in a way that otherwise materially distorts or impairs the ability of the recipients of their service to make free and informed decisions." +- Art. 38: very large platforms using recommender systems "shall provide at least one option for each of their recommender systems which is not based on profiling". +- Art. 37(1)–(2): very large platforms "shall be subject, at their own expense and at least once a year, to independent audits"; they "shall afford the organisations carrying out the audits the cooperation… including by giving them access to all relevant data and premises". +- Art. 37(6): a platform receiving a non-positive audit report "shall take due account of the operational recommendations"; "Where they do not implement the operational recommendations, they shall justify in the audit implementation report the reasons for not doing so". +- Art. 21(2): the certified out-of-court dispute body "shall not have the power to impose a binding settlement of the dispute on the parties." +- Art. 52(3): fines up to "6 % of the annual worldwide turnover". +- *Bearing.* Arts. 25 and 38 are interface rules: the owner may not design deceptively and must offer a non-profiled choice. They bind without relying on the user to read anything. Art. 37 is an audit regime with a "comply or explain" ending, and Art. 21 gives the user a body that cannot bind the platform. + +**holznagel2025shortcomings** — Holznagel, D. (2025, 11 June). Shortcomings of the first DSA audits, and how to do better. *DSA Observatory.* https://dsa-observatory.eu/2025/06/11/shortcomings-of-the-first-dsa-audits-and-how-to-do-better/. *Read status:* read in part (opening and Observation 1). "Most were produced by Big Four accounting firms, and were, in many ways, not very ambitious." For Facebook, the auditor "did not conduct auditing on Articles 14(1), 16(1), 16(5), 16(6), 17(1), 20(1), 20(3), 24(5), 25(1)…", citing the Commission's open investigations; Holznagel: "I do not see any legitimate defense for such refusal to audit." *Bearing.* The first-round evidence on the DSA audit regime: the audits ran, and covered less than the law asks. A blog analysis by one author, not a peer-reviewed study. I did not use the KPMG and Deloitte counts that appeared in search snippets (they conflict across rounds). + +### Q4. Statute and bargaining on algorithmic management + +**eu2024platform** — Directive (EU) 2024/2831 of the European Parliament and of the Council of 23 October 2024 on improving working conditions in platform work. OJ L, 11.11.2024. *Verification:* Publications Office text (`celex/32024L2831`, XHTML). *Read status:* read in part (Arts. 1–3 headings, 2, 7–15, 20–25, 29(1); recitals by keyword). + +Chapter III, "Algorithmic management" (Arts. 7–15), requires more than disclosure. I list what each article asks, quoting the text. + +| Art. | What it requires | Text | +|---|---|---| +| 7(1) | Limits on processing | Platforms "shall not, by means of automated monitoring systems or automated decision-making systems: (a) process any personal data on the emotional or psychological state of a person performing platform work; (b) process any personal data in relation to private conversations…; (c) collect any personal data… while that person is not offering or performing platform work; (d) process personal data to predict the exercise of fundamental rights…" | +| 8 | Impact assessment with worker input | The platform "shall seek the views of persons performing platform work and their representatives"; the assessment goes to workers' representatives (8(2)). | +| 9 | Information | Includes "the categories of data and the main parameters that such systems take into account and the relative importance of those main parameters" (9(1)(b)(iii)) and "the grounds for decisions to restrict, suspend or terminate the account" (9(1)(b)(iv)); portability of data (9(6)). | +| 10 | Human oversight | Evaluation of the impact of decisions "regularly and in any event every two years" with workers' representatives (10(1)); oversight staff "shall have the competence, training and authority necessary… including for overriding automated decisions" and protection from dismissal (10(2)); if a high risk of discrimination is found, "the digital labour platform shall take the steps necessary, including, if appropriate, the modification of the automated monitoring system or the automated decision-making system or the discontinuation of its use" (10(3)); "Any decision to restrict, suspend or terminate the contractual relationship or the account… shall be taken by a human being" (10(5)). | +| 11 | Human review | Right to "an oral or written explanation… without undue delay" and to a "contact person" with authority (11(1)); a written statement of reasons for suspensions, non-payment and status decisions "at the latest on the date on which it takes effect" (11(1)); a right to ask for review, answered "within two weeks" (11(2)); "Where the decision… infringes the rights of a person performing platform work, the digital labour platform shall rectify that decision without delay and in any case within two weeks… Where such rectification is not possible, the digital labour platform shall offer adequate compensation" (11(3)). | +| 12 | Safety and health | Platforms "shall not use automated monitoring systems or automated decision-making systems in a manner that puts undue pressure on platform workers" (12(3)). | +| 13 | Information and consultation | Consultation of workers' representatives "also covers decisions likely to lead to the introduction of or to substantial changes in the use of automated monitoring systems or automated decision-making systems" (13(2)), on the Directive 2002/14/EC arrangements; the representatives "may be assisted by an expert of their choice", paid by platforms with more than 250 workers in the Member State (13(3)). | +| 14–15 | No representatives; non-employee representatives | Direct information to workers (Art. 14); representatives of self-employed persons may use only Arts. 8(2), 9(1) and (4), 10(4), 11(2), "with regard to the protection of their personal data" (Art. 15). | + +- Enforcement: the GDPR supervisory authorities enforce Arts. 7–11 "as far as data-protection matters are concerned", with "the upper limit for administrative fines referred to in Article 83(5)" (Art. 24(1)); other penalties are for Member States, "effective, dissuasive and proportionate" (24(5)). Courts may order disclosure of evidence "which lies in its control" (Art. 21(1)). Art. 25: Member States "encourage the exercise of the right to collective bargaining in platform work, including… to facilitate the exercise of their rights related to algorithmic management set out in Chapter II" (the text says Chapter II; algorithmic management is Chapter III; I quote as printed and flag the slip). +- Transposition: Art. 29(1), "by 2 December 2026". Today is 13 October 2026; the Directive itself has been in force since 1 December 2024 (Art. 31). Whether any national transposing law exists I did not check [unverified: see AUDIT_10]. +- *Bearing.* What the weaker party gets beyond information: bans on certain processing (7), a human decision for the gravest outcomes (10(5)), a duty to correct within two weeks (11(3)), a duty to modify or discontinue a discriminating system (10(3)), consultation with a paid expert (13). What it does not get: a say in what the system optimises, a veto, or co-determination. Consultation under Directive 2002/14/EC is not agreement. Scope is limited to a "digital labour platform" as defined (Art. 2(1)(a)): a service "organi[sing] work performed by individuals in return for payment"; Art. 11 does not apply to persons who are business users under P2B (11(5)). + +**boe2021rdl9** — Real Decreto-ley 9/2021, de 11 de mayo, por el que se modifica el texto refundido de la Ley del Estatuto de los Trabajadores… para garantizar los derechos laborales de las personas dedicadas al reparto en el ámbito de plataformas digitales. BOE no. 113, 12 May 2021, BOE-A-2021-7840. *Verification:* BOE XML; title matches. *Read status:* read in part (preamble paragraphs on algorithms; the single article; the first three provisions). +- Art. único, point One adds Art. 64.4 d) to the Workers' Statute: the works council has the right "Ser informado por la empresa de los parámetros, reglas e instrucciones en los que se basan los algoritmos o sistemas de inteligencia artificial que afectan a la toma de decisiones que pueden incidir en las condiciones de trabajo, el acceso y mantenimiento del empleo, incluida la elaboración de perfiles." +- Point Two adds Additional Provision 23: a presumption that delivery work by persons whose employers exercise organisation, direction and control "mediante la gestión algorítmica del servicio o de las condiciones de trabajo, a través de una plataforma digital" is within the Workers' Statute. +- Preamble: algorithms bring changes "de manera ajena al esquema tradicional de participación de las personas trabajadoras en la empresa", and the information is to be respected "desde el mayor respeto a los secretos industrial y comercial de las empresas". +- *Bearing.* The Rider Law's lever is employment status. Its only algorithm-specific right is information for the works council about parameters, rules and instructions. It does not give a veto or a human-decision rule. Evidence on effects is press-level (next entries): the law was enforced on status. + +**de1972betrvg** — Betriebsverfassungsgesetz (Works Constitution Act), §§ 80, 87, 90, 95. Federal Ministry of Justice, gesetze-im-internet.de. https://www.gesetze-im-internet.de/betrvg/. *Verification:* official site, four section pages. *Read status:* read in full (each of the four sections; German text; translations are mine). +- § 87(1) no. 6: the works council "hat… mitzubestimmen" in the "Einführung und Anwendung von technischen Einrichtungen, die dazu bestimmt sind, das Verhalten oder die Leistung der Arbeitnehmer zu überwachen". The right applies only "soweit eine gesetzliche oder tarifliche Regelung nicht besteht". +- § 87(2): "Kommt eine Einigung über eine Angelegenheit nach Absatz 1 nicht zustande, so entscheidet die Einigungsstelle. Der Spruch der Einigungsstelle ersetzt die Einigung zwischen Arbeitgeber und Betriebsrat." (If no agreement is reached, the conciliation board decides, and its ruling replaces the agreement.) +- § 90(1) no. 3: the employer must inform the council in good time about the planning of "Arbeitsverfahren und Arbeitsabläufen einschließlich des Einsatzes von Künstlicher Intelligenz"; § 95(2a): the selection-guideline rules also apply "wenn bei der Aufstellung der Richtlinien… Künstliche Intelligenz zum Einsatz kommt"; § 80(3): where the council must assess the introduction or use of AI, "gilt insoweit die Hinzuziehung eines Sachverständigen als erforderlich" (an expert is deemed necessary). +- *Bearing.* This is the one custody arrangement in the sources that binds an employer against its own wish: no monitoring system without the council's agreement or a board ruling. Limits that I did not verify in a court source: how far "dazu bestimmt sind… zu überwachen" reaches decision systems that do not monitor; German case law was not opened. Applies to employees with a works council; the Rider-type freelancers are outside it. + +**neumann2025authority** — Neumann, D. (2025). "We were an authority they couldn't bypass": how works councils reshape power and control dynamics in the German platform economy. PhD thesis, University of Leeds. https://etheses.whiterose.ac.uk/id/eprint/38626/. *Verification:* eprint page opened. *Read status:* abstract only (the PDF link was not on the public page). +- The study is an "explanatory single-case study" of a food-delivery company (pseudonym BiteFleet) from "37 expert interviews, participant observation, and 39 Works Agreements, collected during a one-year field stay (2022–2023)". +- Finding three: council activity "led to noticeable improvements in working conditions, particularly in areas such as shift planning, health protection, and continued pay during illness." +- Finding four: "workplace co-determination remained excluded from the core of the algorithmic work organisation, signalling a managerial strategy based on recognising a structural incompatibility between algorithmic control and the principles of co-determination." +- *Bearing.* The strongest custody arrangement in this facet, tested on a platform, is credited, in the abstract, with improvements in shift planning, health protection and sick pay, while co-determination stayed out of the algorithm; the abstract does not say that § 87 was the lever. One case, one thesis; not peer reviewed. + +**doellgast2022negotiating** — Doellgast, V., Wagner, I., & O'Brady, S. (2022). Negotiating limits on algorithmic management in digitalised services: cases from Germany and Norway. *Transfer* 29(1), 105–120. https://doi.org/10.1177/10242589221143044. *Verification:* Crossref title and authors match. *Read status:* snippet only (SAGE returned 403; the claim "the German works council mobilised collective voice institutions to protect worker privacy and discretion" is from a search summary, not the page). Worth opening: the call-centre cases show a council limiting monitoring in a service firm that is not a platform. + +**eurofound2024hilfr** — Eurofound (2024). Collective agreement on use of AI and algorithms signed (Hilfr and 3F, Denmark). Platform Work Repository, record 4438. https://apps.eurofound.europa.eu/platformeconomydb/collective-agreement-on-use-of-ai-and-algorithms-signed-110187. *Read status:* read in full (the record; the agreement itself not opened). "the platform must disclose how exactly it came to the decision, including an explanation of the assessment, the facts, and weighting. As a result, the decision can be legally verified." "As of now, Hilfr does not make use of artificial intelligence"; "Hilfr has 38 employees"; the agreement "can be renegotiated by Hilfr and 3F until March 2025." *Bearing.* An agreement that makes the platform explain an AI-assisted termination, signed with a firm that did not yet use AI. Untested; a secondary summary. + +**wageindicator2022justeat** — ETUI Collective bargaining newsletter item, reproduced as "Spain – Collective agreement for food delivery platform – January 31, 2022", WageIndicator. https://wageindicator.org/labour-laws/collective-bargaining/2022/spain-collective-agreement-for-food-delivery-platform-january-31-2022. *Read status:* read in part (one paragraph). The agreement (Just Eat, with trade unions) "recognises the right of employees to be informed of the content of the algorithms and artificial intelligence systems the platform uses to govern how work is distributed." A search summary also mentioned an "Algorithm Commission" of union and employer representatives; I did not see that sentence on the page I read. *Bearing.* An information right again. + +**eurofound2025glovo** — Eurofound (2025). Glovo faces €450m Spanish Social Security bill for misclassifying riders. Platform Work Repository, record 4522. https://apps.eurofound.europa.eu/platformeconomydb/glovo-faces-450m-spanish-social-security-bill-for-misclassifying-riders-110271. *Read status:* read in full (record). "Labour inspections determined that Glovo's modifications to its autonomous model after the 'Rider Law' was introduced were insufficient"; "Glovo completed the hiring of 14,000 riders as salaried employees by June 2025." *Bearing.* The Rider Law bit through the status rule, years after enactment, and the platform changed its model in response to a fine; the record says nothing about the algorithm article. Secondary summary of Spanish press. + +### Q5. Commons, polycentric governance, worker data collectives + +**cox2010review** — Cox, M., Arnold, G., & Villamayor Tomás, S. (2010). A review of design principles for community-based natural resource management. *Ecology and Society* 15(4), 38. https://doi.org/10.5751/es-03704-150438. *Verification:* Crossref title match; open PDF (19 pp.). *Read status:* read in part (abstract; method; Table 3; the sections on principles 3, 4, 7). +- Abstract: "We analyzed 91 such studies to evaluate the principles empirically… We found that the principles are well supported empirically." +- Table 3: principle 3, "Collective-choice arrangements: Most individuals affected by the operational rules can participate in modifying the operational rules": 20 supportive and 6 unsupportive cases among successes against 7 and 16 among failures (ratio 2.8, p = 0.001); principle 4B, "Monitors are accountable to or are the appropriators": ratio 11.7, the largest in the table; principle 7, "The rights of appropriators to devise their own institutions are not challenged by external governmental authorities": ratio 2.2, p = 0.032. +- *Bearing.* The review tests the principles on irrigation, fisheries and forests managed by the people who use the resource. In each case the users are the appropriators and can monitor one another. A platform user does not appropriate from a shared stock, cannot see other users' treatment, and has no say over the monitor, who works for the owner. The best-supported principle (4B) is the one the platform setting lacks. I did not read Ostrom (1990) itself (see d). + +**calacci2022bargaining** — Calacci, D., & Pentland, A. (2022). Bargaining with the black-box: designing and deploying worker-centric tools to audit algorithmic management. *PACM HCI* 6(CSCW2), article 428, 1–24. https://doi.org/10.1145/3570601. *Verification:* Crossref title match; MIT DSpace open PDF (25 pp.). *Read status:* read in part (abstract; §1; §5–6 results and limitations; keyword scan). +- Abstract: the Shipt Calculator let delivery workers "track and share aggregate data about their pay"; "using data contributed by 140 workers in the Summer of 2020, finding that although the average pay per-order increased under the new payment model, almost half of workers experienced an unannounced pay cut… and many workers worked shifts that paid under their state's minimum wage." +- §6.4 limitations: "the data collected by the Calculator is not 'owned' by workers"; the 140 workers are "miniscule compared to their broader workforce"; the authors "did not perform any rigorous user studies". +- *Bearing.* A worker-built audit is the working case. It produced evidence and a bargaining resource. In the sections I read, I found no passage reporting that the platform changed its pay model in response. Information again, collected by the weaker party. + +**delacroix2019bottom** — Delacroix, S., & Lawrence, N. D. (2019). Bottom-up data trusts: disturbing the 'one size fits all' approach to data governance. *International Data Privacy Law* 9, ipz014. https://doi.org/10.1093/idpl/ipz014. *Verification:* Crossref title and authors match; the SSRN accepted manuscript (43 pp.) opened. *Read status:* read in part (abstract; keyword scan). "Bound by a fiduciary obligation of undivided loyalty, the data trustees would exercise the data rights conferred by the GDPR… on behalf of the Trust's beneficiaries… thus introducing an independent intermediary between data subjects and data collectors." *Bearing.* A proposal that the weaker parties pool their rights under a trust. The text I read cites no operating trust. + +**rosenblat2016algorithmic** and **kellogg2020algorithms** (reused keys) — see facets J and X (Rosenblat and Stark read in full; Kellogg, Valentine and Christin in `references.bib`). Not reopened here. + +--- + +## (c) Synthesis by question + +### Q1. Who held and enforced the conventions of the written page? + +In the Carolingian sources I opened, the ruler and the church held the standard and enforced it by order, correction and hierarchy. The king's letter measures the monasteries' own letters against a standard of correct language and asks the abbot, "if you wish to have our favor", to circulate the letter. The *Admonitio generalis* (c. 72) orders schools and "libros catholicos bene emendate" and says that, if a gospel book, psalter or missal has to be copied ("si opus est"), men of full age are to copy it with all care. Alcuin's poem tells scribes to find corrected books and to mark sense units so the reader in church does not read wrongly. Piazzoni ties scripts to abbots and the court. Westwell, writing on the liturgy, finds that uniformity "from above" was not achieved and that local houses chose and edited. Nothing I opened shows scribes and readers sharing interests; the interests named are the king's (correct Scripture, correct prayer), the abbot's (a regulated house) and the church's, with the lector and the person praying as the beneficiaries. Parkes's abstract gives scribes a role: they "developed a complex of new graphic conventions" to make texts intelligible. That supports a narrower statement: the people who did the work also refined the form, inside an institution that wanted correct reading. I found no historian who describes a free exchange between scribes and readers, and I could not quote McKitterick, Ganz or Bischoff, who are the authorities (d). The sources I opened concern correct language and correct books (Charlemagne's letter; *Admonitio generalis* c. 72) and, in Alcuin's poem, the marking of sense units with cola, commata and points so that the lector in church reads correctly. None of the three primary texts or Westwell mentions word separation; only Alcuin ties the matter to reading aloud. Piazzoni describes the Maurdramnus script as having "well-spaced words", as a feature of the script and not as an aim of the reform. Parkes's abstract credits the "grammar of legibility" to Irish and Anglo-Saxon scribes who taught continental pupils; that abstract (and Parkes 1992, facet A), not the court texts, is the nearest lead on word separation. + +### Q2. Standards bodies and mandated conventions + +The IETF and W3C write conventions in open consensus processes. RFC 2026 calls adherence "voluntary"; the W3C says "Dissenters cannot stop a group's work simply by saying that they cannot live with a decision" (a chair may record a decision over dissent; Formal Objections go to the Advisory Board or TAG). RFC 8890, an informational document of the Internet Architecture Board, says the IAB believes IETF decisions "should favor end users" and concedes that, even when the IETF tries to engage a broader audience, end users' "participation is minimal -- until a change affects someone in a way they don't like". Neither body can bind an implementer, and the one case I opened where an interface owner rewrote a user-facing convention, Chrome's HTTPS indicator, shows the owner acting on its own schedule (the sources do not say whether anyone outside Google was consulted). Google's 2016 paper reports how the lock was read (of 733 respondents describing the green lock, 40.1% gave a connection answer and 13.4% an identity answer, a category that includes correct answers such as "valid certificate"; about 3% of the total wrongly said HTTPS vouched for trustworthiness) in a sample that skews tech-savvy, and the Chromium post of 2018 removes the "Secure" wording on Google's schedule. Standards gain force when a statute or regulation cites them: the DOJ's 2024 rule (a regulation under title II of the ADA) adopted WCAG 2.1 AA for state and local governments, and an interim final rule of 20 April 2026 moved the compliance dates by a year (to 26 April 2027 for entities of 50,000 or more, 26 April 2028 for smaller ones); the EAA gives a presumption of conformity for products and services that meet harmonised standards (Art. 15(1); the Directive itself does not mention WCAG), exempts microenterprise service providers (Art. 4(5)) and requires effective, proportionate and dissuasive penalties (Art. 30(2)). Labels have measured effects on the seller as well as the buyer: reformulation (Shangguan, abstract: sodium −8.9%, based on 4 studies, 95% CI −17.3% to −0.6%; artificial trans fat −64.3%, 3 studies), though Shangguan found no significant difference between voluntary and legislative approaches. Ganderats-Fuentes (abstract) reports larger and more consistent reformulation under mandatory front-of-package schemes, and more under labels with intuitive than with numerical information. Hatton 2026 (abstract) found no substantive reformulation after brands adopted the revised Nutrition Facts label (a revision of an existing mandatory label, one supermarket chain, 2016-2019). Energy labels raised the sales proportion of efficient appliances in one Swiss field study (Stadelmann, abstract; no magnitude in the abstract), and the EU label's format can lead people to disregard actual energy consumption (Waechter, abstract). All of the effect sizes are abstract-level. + +### Q3. Platform-governance scholarship + +Gorwa sorts models into self-, external and co-governance and says that, under self-governance, "voluntary arrangements rely on goodwill and provide limited recourse in the case of non-compliance". Balkin proposes a duty of loyalty enforceable in court and observes that companies can "change their privacy policies more or less at will", so that relying on terms of service "may offer only very limited privacy protections". Khan and Pozen (abstract) fear that the fiduciary framework "invites an enervating complacency toward online platforms' structural power and a premature abandonment of more robust visions of public regulation". Raji and colleagues propose an audit run inside the owner, because outside auditors see only outputs. The one audit regime with statutory force, DSA Art. 37, produced in its first round audits that, on one critic's reading, skipped provisions under investigation and were "not very ambitious". Art. 37(6) lets a platform decline a recommendation if, within one month, it justifies the reasons and sets out any alternative measures. For Facebook, the auditor skipped Art. 25(1) among other provisions (Holznagel). Evidence that any of these models binds a platform: none in the sources I opened beyond the DSA's own text, whose enforcement record (fines, Art. 52) I did not search. + +### Q4. Statute and bargaining on algorithmic management + +Beyond information, Directive 2024/2831 gives workers: bans on processing emotional-state, private-conversation and off-duty data (Art. 7); a human decision for every restriction, suspension or termination (10(5)); a contact person with authority (11(1)); reasons by the date a decision takes effect (11(1)); correction within two weeks or compensation (11(3)); a duty, where oversight finds a high risk of discrimination or infringed rights, to take the necessary steps, "including, if appropriate, the modification" or discontinuation of the system (10(3)); consultation of the platform workers' representatives, who may be assisted by an expert whose proportionate expenses a platform with more than 250 workers in the Member State pays (13(2)-(3)). Arts. 10(5) and 11 cover every person performing platform work, but Art. 11 does not apply to business users under Regulation 2019/1150 (11(5)), and Art. 13 reaches representatives of platform workers who are employees (Art. 15 limits what representatives of the self-employed may use). Penalties for Arts. 7-11 run through the data-protection authorities and the GDPR fine ceiling (Art. 24(1)). It does not give a veto or a share in design. Consultation is under Directive 2002/14/EC. The Directive entered into force on the twentieth day after publication (Art. 31, 1 December 2024); the transposition deadline is 2 December 2026 (Art. 29(1)), so I found no measured effect. Whether any Member State has already transposed it I did not check [unverified: see AUDIT_10]. Spain's Royal Decree-Law gives the works council information on "parámetros, reglas e instrucciones" and a presumption of employment; the Glovo record shows the status rule enforced and says nothing of the information right. German co-determination is the strongest: § 87(1) no. 6 and (2) mean the employer cannot introduce a monitoring system without agreement or a board ruling. The one platform study I found (Neumann, a single-case PhD thesis, abstract only) says council activity led to improvements in shift planning, health protection and continued pay during illness, and that co-determination remained excluded from "the core of the algorithmic work organisation". Collective agreements I could confirm (Hilfr; Just Eat [unverified: see AUDIT_10]) give an explanation duty and an information right; the Hilfr firm had 38 employees and did not yet use AI. + +### Q5. Commons and data collectives + +Cox and colleagues find Ostrom's principles "well supported" across 91 studies, strongest for monitors accountable to the appropriators (ratio 11.7). Those studies are of local resource commons whose users monitor each other. A platform's user does not fit that description, and the review does not test interfaces. The only working case I found of a worker-side instrument is the Shipt Calculator, built by researchers with non-profit worker groups: 140 workers contributed pay data, and the authors say the data "is not 'owned' by workers". I found no passage reporting a change by the platform. Data trusts are proposals. Facet J's Seattle ordinance and the 1,400 reactivations won by a drivers' union belong in this comparison but I did not reopen them. + +### Q6. Which arrangements give the weaker party more than disclosure, and which bind the owner? + +| Arrangement | Gives the weaker party | Binds the owner against its will? | Evidence | +|---|---|---|---| +| Carolingian ruler and church | A corrected book for the reader in church | The ruler could change it; no appeal for the scribe | Primary texts; Westwell on limits of imposition | +| IETF, W3C | A seat at the table | No. Adoption voluntary; owner changes UI alone (Chrome 2018) | RFCs; Chromium post | +| WCAG via ADA rule or EAA | A legal duty for covered bodies | Yes for covered public bodies (ADA rule; dates moved a year) and for covered private suppliers (EAA; microenterprise services exempt) | 91 FR 20902; Directive 2019/882 Arts. 4, 15, 30 | +| Mandated labels | A comparable format; the seller reformulates | Yes, more under mandates | Shangguan; Ganderats-Fuentes (abstracts) | +| Information fiduciary | A court-enforced duty of loyalty (proposed) | In theory; no case found | Balkin (read in part); Khan and Pozen (abstract) | +| Internal audit (Raji) | None directly | No: the owner audits itself | Raji §2.4 | +| DSA Arts. 25, 38 | No deceptive interface (except practices covered by Directive 2005/29/EC or the GDPR, Art. 25(2); not micro or small platforms, Art. 19); a non-profiled option on very large platforms | Yes (fines to 6%, Art. 52(3); Commission fines on very large platforms, Art. 74(1)) | Official text; no enforcement record read | +| DSA Art. 37 audit | A published audit and a justification | Partly: comply or explain (37(6)) | Holznagel (blog) on first round | +| Platform Work Directive Ch. III | Human decision, rectification or compensation in two weeks where rights are infringed, bans, expert for employees' representatives | Yes once national law applies (transposition due 2 December 2026); consultation not agreement | Official text; not yet in force | +| Spanish Rider Law | Information to council; employment presumption | Status rule enforced; information right untested | BOE text; Eurofound record | +| German BetrVG § 87(1) no. 6, (2) | A veto-with-arbitration over monitoring systems | Yes | Statute; Neumann (abstract): kept out of the algorithmic core | +| Collective agreements (Hilfr, Just Eat) | Explanation duty; information right | Only by renegotiation | Secondary summaries | +| Ostrom commons | Self-rule over a shared stock | Not tested on interfaces | Cox 2010 | +| Worker-side data tool (Shipt Calculator) | Evidence for bargaining | No | Calacci 2022 (data not owned by workers) | + +The pattern (the agent's inference from the rows above; none of the sources states it, and the sample is fourteen arrangements, not a test): a convention binds the owner only when a third party with a sanction stands behind it (a regulator, a conciliation board, a court), or when the convention is a mandated format the owner cannot alter. The two halves overlap, since a mandated format also depends on a regulator's sanction. Everything the owner can renegotiate, redesign or audit itself is custody in name. + +--- + +## What the outline should add + +*Section 2 (the scribal precedent).* + +1. **Replace "scribes and readers shared mutual communicative interests" with a statement about who held the page.** Claim: in the Carolingian reform, the ruler and the church fixed and enforced the standard of correct writing, and the people named as beneficiaries were the reader in church and the person praying. Sources: Charlemagne's letter to Baugulf (read in full, translation); *Admonitio generalis* c. 72 (read in part, Latin OCR); Alcuin, carm. XCIV (read in part, OCR). Add Mostert 2016 and Ganz 1995 (abstract only) for who copied and in what numbers. The letter and c. 72 do not mention reading aloud; only Alcuin's poem does. +2. **Qualify the "reform from above" picture.** Claim: even for the liturgy, Westwell (citing McKitterick and Hen) reports local initiative and no imposition of the Roman rite from above. Source: Westwell 2018 (read in part). Note its scope is liturgical texts; it does not discuss word separation. +3. **Keep two claims apart.** (i) The sources concern correct language, correct books and (Alcuin) punctuation for the lector; no source I opened ties the reform to word separation. Piazzoni's "well-spaced words" describes a script, not a reform aim. (ii) The scribes' own practice produced new graphic conventions (Parkes 2016, abstract only, which credits Irish and Anglo-Saxon scribes with teaching the "grammar of legibility"; facet A for Parkes 1992). The outline can say scribes refined conventions; it cannot say they and readers shared a bargain. + +*Section 7 (institutional custody).* + +4. **Add the test "can the owner change it alone?" and apply it to cases.** Claims: Google's own staff designed and shipped new connection-security indicators (Felt et al. 2016, read in part) and removed the "Secure" label on its own schedule (Chromium 2018, read in full; the sources do not say whether anyone outside Google was consulted); a standard binds through statute or regulation (EAA Arts. 4, 15, read in part; 91 FR 20902, read in part); a works-council right with arbitration binds on paper (BetrVG § 87(1) no. 6 and (2), read in full), and Neumann's single case shows co-determination kept out of the algorithmic core. The sentence in `DRAFT.md:166`, "DXD conventions will never be granted voluntarily", needs this evidence: Google granted one, in its interest. +5. **State what the law gives beyond information, and date it.** Claim: Directive 2024/2831 Arts. 7, 10(3), 10(5), 11(3) and 13 go beyond disclosure; the Directive is in force, but its platform obligations apply only through national law due by 2 December 2026; Art. 10(3) requires "the steps necessary" and names modification or discontinuation only "if appropriate"; Art. 11(3) applies where a decision infringes the worker's rights; and consultation is not co-determination. Source: the Directive (read in part, official text). Contrast the Spanish Rider Law, whose algorithm right is information (BOE text, read in part). +6. **Add mandated formats and the labels evidence, with the disagreement.** Claim: labels change what sellers offer; Ganderats-Fuentes 2023 reports more reformulation under mandatory than voluntary front-of-package schemes, Shangguan 2019 found no significant difference between voluntary and legislative approaches, and Hatton 2026 found no substantive reformulation after the revised Nutrition Facts label (all abstract only; read the full texts before print). DSA Arts. 25 and 38 (read in part) are the interface-design versions of the same idea. +7. **Add the caution that the strongest custody arrangement was kept out of the algorithm.** Claim: platform works councils won rosters and sick pay and were excluded from the core algorithmic system. Source: Neumann 2025 (abstract only; single case; the thesis PDF is embargoed until 1 June 2029, so it cannot be opened before then). The abstract does not say which statutory right the councils used. +8. **Add a short typology and name Gorwa's three models.** Sources: Gorwa 2019 (read in part), Balkin 2016 (read in part), Khan and Pozen 2019 (abstract only), Raji 2020 (read in part), Holznagel 2025 (read in part). +9. **Say the commons evidence does not transfer.** Cox, Arnold and Villamayor Tomás 2010 (read in part): the best-supported principle is monitors accountable to the appropriators, which the platform user lacks. +10. **Cite the one working worker-side tool as information.** Calacci and Pentland 2022 (read in part): built by researchers with worker groups; the data are not owned by workers. + +### Claims in the paper's position that these sources contradict or qualify + +1. **"Scribes and readers shared mutual communicative interests"** (outline; `DRAFT.md:166` "shared a common linguistic community"). The sources show a ruler and abbots directing scribes to correct books; the beneficiary named is the reader in church. No source I opened shows scribe–reader mutuality. Parkes (abstract) allows scribes a role in refining conventions. Not supported as worded. +2. **"Word separation [and similar conventions] spread because scribes and readers shared a common community."** Nothing I opened tests a cause of spread; the Carolingian sources concern correction of text, not spacing. +3. **"DXD conventions will never be granted voluntarily through corporate design guidelines."** Google designed and shipped new Chrome security indicators and published its user research (Felt et al. 2016, by Google staff); in 2018 it announced removal of the "Secure" label (the Chromium post does not mention the padlock or research). No law is shown to have compelled either; whether the owner acted "voluntarily" in its own interest is the agent's inference. The claim holds where the owner's interest diverges from the user's (Hatton 2026 shows a mandated panel changed little) but "never" is too strong. +4. **"They must be institutionalized through legal frameworks or collective resistance."** The legal frameworks I opened mostly give information and consultation (Rider Law Art. 64.4 d); Directive 2024/2831 goes further but is not in force; the one arrangement that binds on paper (BetrVG § 87(1) no. 6 and (2)) sits beside a single platform case in which co-determination was excluded from the algorithmic core (the abstract does not name § 87). Collective resistance produced evidence, not a changed pay model, in the case I read. +5. **A reading of "the failure of disclosure" (section 4) that treats statute as disclosure.** Directive Arts. 10(5), 11(3) and DSA Arts. 25, 38 are not disclosure. The section-4 argument is on firmer ground against disclosure-only instruments (Nutrition Facts panel, Rider Law information right, Hilfr and Just Eat agreements). +6. **The assumption that the Carolingian reform was imposed uniformly from above.** Westwell finds local initiative in the liturgy. The reform's custodians were many, not one. +7. **That a standards body can serve as holder.** RFC 8890 (IAB) says the IETF has no unique insight into what is good for end users and that end-user participation is often minimal until a change affects them; the W3C lets a chair record a decision over dissent after the group has considered the concerns. + +--- + +## (d) Gaps and unverified leads + +- **McKitterick (1989), Ganz (1990 *Corbie in the Carolingian Renaissance*), Bischoff (*Latin Palaeography*, tr. Ó Cróinín and Ganz, Cambridge UP 1990).** The historians the brief names. All three are in copyright; search-inside returned "Item not available" and Google Books was over quota. A library copy is needed for: who ordered, paid for and supervised scribes (McKitterick ch. 4); the Corbie scriptorium and its abbots (Ganz); the origin and diffusion of Caroline minuscule (Bischoff). Until then, the claim in Q1 rests on primary texts, one thesis on the liturgy and abstracts. +- **Ostrom (1990), *Governing the Commons*, and Ostrom (2010, Nobel lecture).** Not opened; Cox et al. is a proxy. The polycentric-governance literature (Ostrom, Tiebout and Warren 1961; Aligica and Tarko) is unsearched. +- **Khan and Pozen (2019).** Abstract only (Columbia PDF 403). Read Parts II–III in the Harvard Law Review volume 133. +- **Neumann (2025) and Doellgast et al. (2022).** Abstract only for Neumann (the Leeds thesis, White Rose eTheses 38626, is embargoed until 1 June 2029; ask the author or Leeds); the Doellgast *Transfer* abstract was checked in the audit, the article was not opened. The INCODING policy briefs on Germany and Denmark (UAB repository, behind a bot challenge) would give a second source on collective bargaining and algorithms. +- **The Hilfr and Just Eat agreements themselves.** Summaries only. The "Negotiating the algorithm" manual (ETUC) was named in a search summary and not found. +- **Spanish and German case law.** No court judgement was read on the reach of § 87(1) no. 6 BetrVG to decision systems, or on Art. 64.4 d) Estatuto de los Trabajadores in use. The Rider Law's effect on the algorithm right has no evaluation in what I found. +- **DSA enforcement record.** I read the audit and penalty articles and one critique of the first audits. I did not search Commission decisions or fines under Arts. 25 and 38. +- **The 2023 lock-icon change.** Press only (Help Net Security snippet). The original Google Security blog post should be found before any figure is quoted. +- **Label and energy-label studies.** All five effect-size sources are abstracts. Read Shangguan 2019 and Ganderats-Fuentes 2023 in full before using the voluntary/mandatory contrast; the two reviews disagree on it. +- **Data collectives beyond Calacci.** Driver's Seat (named in Calacci §5), Worker Info Exchange and similar groups were not searched. Facet J's Seattle ordinance and union reactivations bear on Q5 and are not repeated here. +- **Steinová (2015) on "notas".** The Admonitio passage I quote has contested vocabulary; read her article before building anything on "notas". +- **Unchecked metadata.** Second authors and page ranges for Ganderats-Fuentes, Stadelmann and Akhawe were taken from the Consensus records; titles and DOIs were checked at Crossref or the publisher's page, but the author lists were not. They are marked in the bib. + +## Changes after audit (2026-10-13) + +- Intro and Q2: "padlock redesigned twice by Google alone" replaced by the 2018 removal of the "Secure" label on Google's schedule; the 2023 change marked unverified; "alone" dropped. +- Intro, Q4, Q6, should-add 5: Directive 2024/2831 is in force since 1 December 2024 (Art. 31); only national transposition is due on 2 December 2026. "Not yet in force" removed. +- Q4, Q6, should-add 5: Art. 10(3) modification or discontinuation only "if appropriate"; Art. 11(3) applies where rights are infringed, not to any "wrong" decision; Art. 11(5) and Art. 13/15 scope added. +- Q2: RFC 8890 "is minimal" quotation restored to "participation is minimal -- until a change affects someone in a way they don't like", attributed to the IAB. +- Q2: W3C "Dissenters cannot stop a group's work" quotation completed. +- Q2: Felt et al. 13.4% is the identity category (includes correct answers); the misreading is about 3%. +- Q2: DOJ rule called a regulation; EAA presumption covers products and services, microenterprise exemption added; WCAG not named in the EAA. +- Q2, should-add 6: Shangguan found no voluntary/legislative difference, so it no longer supports "mandatory labels change more"; Hatton concerns a revision of an existing mandatory label; Stadelmann "modestly" dropped. +- Q3: Gorwa and Balkin quotations tied to their actual subjects (voluntary arrangements; privacy policies); Khan and Pozen quoted; DSA Art. 37(6) one-month and alternative-measures text added. +- Q6 table: DSA Art. 25(2) and Art. 19 limits, Art. 74 added; WCAG/EAA row corrected; Calacci row renamed. +- Q5, should-add 10, contradicted 3 and 4: Shipt Calculator is not "worker-held" (researcher-built; data not owned by workers). +- Q6 closing synthesis ("an owner is bound only when ...") labelled the agent's inference; overlap of its two halves stated. +- Q1 and should-add 1-3: word separation: none of the primary texts or Westwell mentions it; Piazzoni "well-spaced words" and Parkes (Irish and Anglo-Saxon scribes) noted; "reading aloud" tied to Alcuin only; Admonitio "si opus est" restored. +- Should-add 4 and 7: Neumann PDF embargoed until 1 June 2029, so "open the thesis PDF" dropped; Neumann's abstract does not name § 87. +- Contradicted 3: Chrome claim rewritten (the 2018 post does not mention the padlock or research). diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_S_hmi_read_to_check.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_S_hmi_read_to_check.md new file mode 100644 index 00000000..caf8e7ff --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_S_hmi_read_to_check.md @@ -0,0 +1,136 @@ +# Facet S — The 2026 HMI trajectory and read-to-check oversight + +*Facet S, round 5, 2026-10-15. Agent-drafted (Fable). The author has not reviewed it, and none of the text is the author's.* + +## What the facet covers + +The facet serves the abstract's first three sentences (`../../ABSTRACT_v4.md`): human–machine interaction has moved to extended reality and integration; "whether on a factory headset or in an enterprise audit log ... the system still sends a person something to read so they can check the machine"; and "That approach assumes a dyad: an operator paired with a tool (Parasuraman, Sheridan, & Wickens, 2000)". Four questions from `round5_items.json`: (1) what the 2020–2026 HMI/HCI, XR, human-centred AI and oversight literature prescribes as the human's role; (2) how the supervisory-control lineage models the parties; (3) where that lineage admits several humans or teams, and whether the machine then becomes a party with its own objective; (4) evidence on how well people check machines by reading. + +Existing cards read first: `grahn2026integration`, `mourtzis2026emerging`, `parasuraman2000types`, `parasuraman2010complacency`. Also consulted from the library: `eu2024regulation`, `green2022flaws`, `sheridan1978human`, `lee2004trust`, `alonbarkat2023human`, `sterz2024quest`, `laux2024institutionalised`, `crootof2023humans`, `langer2024effective`, `hutchins1995cockpit`, `santonidesio2018meaningful`, `vaccaro2024when`, `simkute2024ironies` (headers and relevant sections, not re-carded). + +New cards (13): `bainbridge1983ironies`, `wiener1980flight`, `klein2004ten`, `onnasch2014human`, `oneill2022human`, `shneiderman2020human`, `enqvist2023human`, `enarsson2022approaching`, `koulu2020proceduralizing`, `lai2023towards`, `quandt2021systematic`, `demouracosta2022augmented`, `rayo2023cognitive`. Four of these keys (`bainbridge1983ironies`, `koulu2020proceduralizing`, `lai2023towards`, plus the long-cited `parasuraman2000types`) already stood in `../references.bib` without a card or with a partial read; they now have cards from full texts. Sections headed "Also used by facet S (2026-10-15)" were appended to `grahn2026integration`, `mourtzis2026emerging`, `parasuraman2000types`, `parasuraman2010complacency` and `eu2024regulation`. + +Full texts are saved under `dxd_foundation_local/pdfs/S/` (outside the repo). Two sources (Koulu, Enarsson et al.) were read from the publishers' open-access HTML in a browser because their PDFs refused scripted download; the Chrome tab group was shared with other agents during the session and was navigated away from under me twice, so browser reading stopped after those two and the remaining publisher-blocked items (Goddard 2012, Holzinger et al. 2024, Ariansyah et al. 2022) are listed below as abstract-only. + +## Search log + +| Source | Query (exact) | Hits | Notes | +| --- | --- | --- | --- | +| Consensus | human oversight of AI systems EU AI Act Article 14 effectiveness human role monitoring | 20 | Sterz, Enqvist, Laux (carded already), Trapani 2026, Corrêa 2025, Banks 2026, Kaltenbrunner 2026, Holzinger 2024, Langer 2025 compliance testing; Enqvist opened and carded | +| Consensus | human-autonomy teaming review empirical literature teammate versus tool | 20 | O'Neill et al. 2020/2022 (carded), Lyons 2021 (already carded), Hagemann 2026 (N=80, abstract only), Kargarnovin 2026 (104 studies, not opened), McNeese 2018, Walliser 2019 | +| Consensus | industrial augmented reality head-mounted display operator information support systematic review | 20 | Fang 2023 HMD AR review (closed), de Moura Costa 2022 (carded), Quandt 2021 (carded), Ariansyah 2022 (OA but blocked), Seeliger 2023, Egger & Masood 2020 | +| Scholar Gateway | Does the supervisory control and levels-of-automation literature model the human as a single operator ... | 0 | "ACCESS_DENIED" — connector unavailable this session | +| Semantic Scholar API (paper/search) | 16 title queries (Bainbridge 1983; Parasuraman & Riley 1997; Endsley 1995; Klein et al. 2004; Hollnagel & Woods 1983; O'Neill 2022; Endsley 2017; Endsley 2023; Shneiderman 2020; Goddard 2012; Onnasch 2014; Wiener & Curry 1980; Wickens 2015; Mackworth 1948; Koulu 2020; Enarsson 2022) | 0 returned | The search endpoint returned empty data; the batch endpoint returned 429 (rate limit) on three attempts | +| Crossref (works?query.bibliographic) | the same 16 plus Strauch 2018, Christoffersen & Woods 2002, Holzinger 2024, Ariansyah 2022, de Moura Costa 2022, Fang 2023, Lai 2023 | 25 queries, 19 resolved | DOIs used for every bib entry; 6 queries failed on rate limiting and were resolved by OpenAlex or a later works/DOI call | +| Crossref (works/DOI) | 16 DOIs of the carded and lead sources | 16 | Bib fields taken from these records (saved in scratchpad crossref_S.txt) | +| OpenAlex (works?filter=doi:…) | 25 DOIs | 25 | Open-access status and pdf locations; found DiVA (Enqvist), DepositOnce (Onnasch), arXiv (Shneiderman), PMC (O'Neill) | +| Europe PMC REST | DOI searches for 7 items; fullTextXML for PMC9284085 and PMC3240751 | O'Neill XML obtained; Goddard XML returned 500 three times | NCBI efetch for PMC3240751 says the publisher does not allow XML download | +| NASA NTRS API | Wiener Curry flight-deck automation promises problems | 2 | TM 81206 PDF downloaded | +| WebSearch | Klein et al. 2004 "Ten challenges" pdf; Hollnagel & Woods 1983 pdf (two queries); Parasuraman & Riley 1997 pdf (two); Endsley 1995 pdf (two); Onnasch 2014 depositonce; Shneiderman 2020 pdf; Christoffersen & Woods 2002 pdf; Endsley 2023 pdf; Endsley & Kiris 1995 pdf; Enarsson diva-portal; Hollnagel Risø-M-2330 | 14 queries | Located DepositOnce record (Onnasch), arXiv (Shneiderman), the Bradshaw archive (Klein, via the site's index page); none for Endsley 1995, Parasuraman & Riley, Hollnagel & Woods | +| WebFetch | archive.cps-vo.org node 62405; depositonce item page; PMC9284085; jeffreymbradshaw.net; helda.helsinki.fi handle; tandfonline full (Enarsson); sciencedirect (Holzinger); nottingham-repository preview (Ariansyah) | 8 | DepositOnce and PMC gave links; Helda (Anubis), T&F, ScienceDirect and Nottingham returned 403 or an access-denied page | +| Direct download (curl) | ckrybus.com (Bainbridge), Bradshaw Dropbox (Klein), DepositOnce (Onnasch), DiVA (Enqvist), arXiv (Shneiderman, Lai), MDPI (de Moura Costa), Frontiers (Quandt), NSF PAR (Rayo), Europe PMC XML (O'Neill) | 10 files | All verified as PDFs or JATS XML; Sage, T&F, OUP, Elsevier and PMC PDFs refused | +| Browser (shared Chrome tab) | Sage HTML (Koulu); T&F HTML (Enarsson); PMC and OUP pages (Goddard) | 2 full reads | Goddard returned the OUP abstract page only; the tab was then taken over by another agent's navigation | + +## Read-status legend + +- **read in full**: whole text read from an open copy (publisher OA, repository, author archive, arXiv, NTRS). +- **read in part**: most of the text read; the gap is stated in the card. +- **abstract only**: only the abstract or a database record was seen; no card, listed below. +- **not accessible**: no abstract beyond the title was reachable; no card. + +| Source | Status | Where read | +| --- | --- | --- | +| Bainbridge 1983 | read in full | journal PDF (mirror) | +| Wiener & Curry 1980 | read in part (NASA TM; Table 1 and guidelines 6–14 missing from the scan) | NTRS | +| Klein, Woods, Bradshaw, Hoffman & Feltovich 2004 | read in full | author archive (Bradshaw) | +| Onnasch, Wickens, Li & Manzey 2014 | read in full | TU Berlin DepositOnce | +| O'Neill, McNeese, Barron & Schelble 2022 | read in full | Europe PMC JATS XML | +| Shneiderman 2020 | read in full (arXiv v2; journal version not seen) | arXiv | +| Enqvist 2023 | read in full | DiVA (publisher PDF) | +| Enarsson, Enqvist & Naarttijärvi 2022 | read in full (HTML; footnotes 65+ skimmed) | T&F HTML in browser | +| Koulu 2020 | read in full (HTML; footnotes 9+ not shown) | Sage HTML in browser | +| Lai, Chen, Liao, Smith-Renner & Tan 2021/2023 | read in full (arXiv v1; FAccT version not seen) | arXiv | +| Quandt & Freitag 2021 | read in full | Frontiers | +| de Moura Costa, Petry & Moreira 2022 | read in full | MDPI | +| Rayo, Cooke, Flach, Hoffman, Smith & Woods 2023 | read in full (p. 183 from an image) | NSF PAR | + +## Findings per question + +### Q1. What the 2020–2026 literature prescribes as the human's role + +**Industrial XR.** The two reviews carded here and the two already carded agree that the headset's job is to present information. De Moura Costa et al. (2022): AR solves "the 'human in the loop' integration problem by augmenting information properly into the operators' field of view" (`demouracosta2022augmented`, p. 3), and "The capacity of augmented reality to display real-time information to the user field of view is essential to insert the human in the loop" (p. 45). Of 32 papers, safety features were the most common application and most were "for informational purposes only" (p. 27); 13 of the 20 safety systems were "passive" (informational), 7 "active" (p. 29); none certified (p. 29). Quandt & Freitag (2021): AR gives "cognitive relief through the location-specific and context-related superimposition of real objects with virtual information" (`quandt2021systematic`, abstract), and after 33 papers "there is no widespread adoption" (p. 14). Mourtzis & Tsakalos (2026) and Grahn et al. (2026) say the same (existing cards). Role prescribed: receive instructions, warnings and status; check the robot's intent from a preview. + +**Human-centred AI.** Shneiderman (2020) prescribes "high levels of human control AND high levels of automation" (`shneiderman2020human`, p. 1); reliability rests on "Audit trails and analysis tools to review failures & near misses" and "flight data recorders" (pp. 3–4); trust rests on independent oversight bodies (p. 4). The interface principles are "continuous visual display", "rapid, incremental, and reversible actions" and "informative feedback" (p. 13). He rejects the teammate metaphor outright (p. 5). Role prescribed: operate a controllable device; let organisations read the records. + +**Human–AI decision making (HCI).** Lai et al. survey "over 100" studies (2018–2021): the paradigm is "to have models providing predictions, then users can choose to follow or ignore" (`lai2023towards`, §4.2.4, p. 20); the dominant assistance elements are explanations of the prediction and information about the model (Table 3); the alternatives (predict first, cognitive forcing, interactive explanation, user control of inputs) are a minority (§4.2.4). Role prescribed: read a prediction and its explanation, then accept or reject. + +**Human–autonomy teaming.** O'Neill et al. review 76 studies: higher agent autonomy generally helps, reliability always helps, transparency is "mixed" and "may create a level of complacency that leads to less vigilance in overseeing or questioning the autonomous agent's work" (`oneill2022human`, Transparency and reliability). Role prescribed: team member who pushes and reads information; still "overseeing". + +**Oversight under the EU AI Act.** Enqvist (2023): the overseer works "Through human intermediaries as interpreters of the data produced by the operating algorithms"; without transparency "the human overseers would have nothing substantive to review" (`enqvist2023human`, pp. 517–518); "There is no real point in performing the oversight if the overseer is not provided with information (or system output) in such a representation that it can interpret or understand" (p. 521); the Act leaves "a great deal of room for the providers to determine the detail on what human overseers will be presented with" (p. 525). Koulu (2020): oversight is adopted because "through digital traces and metadata we can easily ascertain that the oversight requirement has been fulfilled ... to have pressed the button, or ticked the box" (`koulu2020proceduralizing`, False promise section), and risks "turning human decision-makers into rubberstamps" (Conclusion). Enarsson et al. (2022): the human in the loop is "the overseer and correcting mechanism for the algorithmic predictions" (`enarsson2022approaching`, §1.2). Article 14(4) itself, in the enacted text, asks the assigned person to understand, monitor, "correctly interpret" the output, override and stop (`eu2024regulation`). Role prescribed: interpret outputs and logs; override; stop. + +Summary for Q1: across XR, HCAI, HCI, teaming and law, the prescribed role is reading, monitoring and checking, with override and stop as the acts. The exceptions are design-side: Shneiderman's reversible actions and feedback, Lai et al.'s minority of agency-preserving workflows, and the teaming literature's directability. + +### Q2. How the supervisory lineage models the parties + +One operator and one automated system, from the start. Wiener & Curry (1980): the human "will be relegated to the post of monitor and decisionmaker, keeping watch for deviations and failures, and taking over when necessary" (`wiener1980flight`, p. 1); their Figure 1 has two dimensions, the operator monitoring the computer and the computer monitoring the operator (p. 13). Bainbridge (1983): the operator "may be expected to monitor that the automatic system is operating correctly" or take over (`bainbridge1983ironies`, p. 775); "The human monitor has been given an impossible task" (p. 776). Sheridan & Verplank (1978) and Parasuraman, Sheridan & Wickens (2000) allocate functions between "the human operator" and "automation" (existing cards). Onnasch et al. (2014) meta-analyse 18 experiments, every one an operator and an aid (`onnasch2014human`). Even the ATC example in Parasuraman et al. (controller, pilot, uplink, p. 293) keeps the operator–automation pair as the unit (existing card, S2-verified). Finding: the lineage's unit is the pair, and the automation has no objective of its own; it is an aid to the operator's task. + +### Q3. Where the lineage admits several humans or teams, and whether the machine becomes a party + +Three places, and in none does the machine acquire an objective against the people it faces. + +- **Joint cognitive systems (Hollnagel & Woods 1983; Rayo et al. 2023).** The unit is the system: "The cognition resides not only in the head of an individual, but in components of the system and their interactions" (`rayo2023cognitive`, p. 184, Cooke); a new tool "may result in adaptations that influence the work patterns of several interdependent people" (p. 183, Smith); Woods calls the levels proposals "obsolete, narrow, and/or counterproductive" (p. 185). Several humans, yes; the machine is a component of one system with one purpose. The 1983 paper itself has no open copy (not accessible). +- **Joint activity and "team player" automation (Klein et al. 2004).** An "ensemble of people" coordinating; agents must enter a Basic Compact, which "involves a commitment to some degree of goal alignment" and "one or more participants relaxing their own shorter-term goals" (`klein2004ten`, p. 91); Challenge 7 is goal negotiation (p. 93). Here the machine does have goals, but the design aim is to align them with the team's; the agent is at best "a young child or a novice" teammate (p. 94). +- **Human–autonomy teaming (O'Neill et al. 2022).** A HAT requires "a common goal as a collective"; 22 of 76 studies had two or more humans, 7 had several agents, 45 of 76 had one human and one agent; all were laboratory simulations, and the authors describe the evidence base as "dyads" (`oneill2022human`). The agent counts as a teammate only above level 5 on the Parasuraman scale, that is, with decision latitude, but never with a divergent objective. +- **Oversight law (Enqvist 2023; Enarsson et al. 2022; Koulu 2020).** The structure is a chain, provider to deployer to overseer to system, with affected persons as rights-holders without an interface (`enqvist2023human`, pp. 518–530); or decision-maker, system and subject (`enarsson2022approaching`, `koulu2020proceduralizing`). Three positions, but the act of oversight is one reader facing one system's output, and the subject does not coordinate with anyone through the system. + +Finding: the lineage admits teams and chains; it treats the machine as a tool, a component or an aligned teammate; the interested mediator between two counterparts is absent. + +### Q4. Evidence on how well people check machines by reading + +- Vigilance: Bainbridge reports from Mackworth that effective attention to a source "on which very little happens" lasts about half an hour (`bainbridge1983ironies`, p. 776); logging does not help: "people can write down numbers without noticing what they are" (p. 776). The original study (`mackworth1948breakdown`) remains inaccessible. +- Complacency and bias: `parasuraman2010complacency` (existing card): detection of automation failures 33% under constant reliability against 82% under variable reliability in multitask conditions; 75% of pilots followed a wrong automated checklist against 25% with paper. Wiener & Curry: a training manager doubled command training after automation bred "inactivity or complacency" in first officers (`wiener1980flight`, p. 9). +- Degree of automation: across 18 experiments, return-to-manual performance fell with automation (tau −.34) and collapsed when action selection was automated (tau −.90 across six studies); the authors attribute the protection of keeping the choice to the generation effect (`onnasch2014human`, pp. 483–485). +- Transparency: in teaming studies, more of the agent's reasoning to read raised workload and in some studies complacency (`oneill2022human`, Transparency and reliability; Wright et al. 2016, Sadler et al. 2016 as cited there). +- A fielded case: the Swedish Social Insurance Agency's partially automated sickness-benefit process hands administrators a machine-generated interruption document and pre-collected facts; the internal audit found investigations insufficient in 67 percent of cases, correct decisions in 76 percent of those, and only about a quarter of cases both properly investigated and correct (`enarsson2022approaching`, §4.3). +- Public-sector experiments: `alonbarkat2023human` (existing card): selective adherence to algorithmic advice in three preregistered experiments. +- Abstract-only: Goddard, Roudsari & Wyatt (2012) report, from 74 included papers, that automation bias is mediated by cognitive style, experience, trust and workload and mitigated by training, accountability, the position of advice on screen and "the provision of information versus recommendation" (OUP abstract); not carded. + +Finding: the evidence is consistent that reading to check degrades with time, reliability and the amount handed over; the one lever the dyadic literature finds is keeping the person in the choosing (Onnasch; Parasuraman & Manzey's information-analysis aids), which is the abstract's "epistemic action" in the dyad's own terms. + +## What this means for the abstract's sentence + +The facet's sentence: "Yet, whether on a factory headset or in an enterprise audit log (Nadella, 2026), the system still sends a person something to read so they can check the machine. That approach assumes a dyad: an operator paired with a tool (Parasuraman, Sheridan, & Wickens, 2000)." + +| Claim | Verdict | Grounds | +| --- | --- | --- | +| HMI has moved to XR and integration (sentence 1) | supported | `mourtzis2026emerging`, `grahn2026integration` (existing); `demouracosta2022augmented` dates the HMD turn to 2016–2018 | +| On a factory headset the system sends the person something to read to check the machine | supports a narrower form | The headset mostly sends instructions, warnings and status; checking the robot is one use (trajectory previews, digital twins) among guidance and safety, and most safety content is informational and uncertified (`demouracosta2022augmented`, pp. 27–29, 45). "Something to read" holds; "so they can check the machine" is narrower than the field's own purpose, which is to tell the worker what to do | +| In an enterprise audit log the system sends the person something to read to check the machine | supported | Audit trails as the reliability practice in `shneiderman2020human` (pp. 3–4); Article 12 logs and Article 14 interpretation duties in `enqvist2023human` (pp. 518–523) and `eu2024regulation`; Koulu on the record as the reason oversight is adopted (`koulu2020proceduralizing`); Nadella P15 (`nadella2026models`, existing) | +| That approach assumes a dyad: an operator paired with a tool | supports a narrower form | For the supervisory lineage and HCI decision-support studies the dyad holds (`wiener1980flight`, `bainbridge1983ironies`, `onnasch2014human`, `lai2023towards`; 45 of 76 teaming studies in `oneill2022human`). Joint cognitive systems, joint activity and teaming admit several humans and agents with goals (`rayo2023cognitive`, `klein2004ten`, `oneill2022human`), and oversight law is a provider–deployer–overseer chain (`enqvist2023human`). The defensible form: the approach assumes that the machine serves the person who reads it; it has no place for a mediator with an objective of its own between two counterparts | +| The attribution to Parasuraman, Sheridan & Wickens (2000) | supported, with a caveat | The model's unit is the operator–automation pair (existing card, S2-verified); its own ATC example involves controller and pilot, and Shneiderman (2020) and Woods (2023) both attack the levels model as the field's limiting frame, so the citation names the right anchor but a reviewer can point to the model's multi-person example | +| Implied: checking by reading works badly | supported, inside the dyad | `bainbridge1983ironies`, `onnasch2014human`, `parasuraman2010complacency`, `enarsson2022approaching` §4.3; no study in a triad | + +Two things cut against the abstract's direction rather than its sentence. First, the dyadic literature already contains the remedy the abstract proposes: Bainbridge's feedback "within a few seconds" (p. 777), Shneiderman's reversible actions with immediate feedback (p. 13), Onnasch et al.'s "keep the operator in decision and action selection" (p. 486), Klein et al.'s directability and mutual predictability. The abstract's "smoking gun" has to be shown to differ from these by crossing to the triad, not by being feedback or action as such. Second, Shneiderman keeps the dyad on purpose (computers "are not teammates", p. 5), so "assumes a dyad" is, for the leading HCAI statement, a chosen position, not an oversight; the paper should argue against the choice, not describe it as an assumption. + +## Abstract-only and inaccessible leads + +| Source | Status | What was seen | +| --- | --- | --- | +| Goddard, K., Roudsari, A., & Wyatt, J. C. (2012). Automation bias: a systematic review of frequency, effect mediators, and mitigators. *JAMIA, 19*(1), 121–127. doi:10.1136/amiajnl-2011-000089 | abstract only | OUP abstract page in browser: 13,821 retrieved, 74 included; mediators and mitigators listed, including "the provision of information versus recommendation". PMC3240751 exists but the PMC page, Europe PMC XML (500) and NCBI efetch (publisher forbids XML) all failed | +| Holzinger, A., Zatloukal, K., & Müller, H. (2025). Is human oversight to AI systems still possible? *New Biotechnology, 85*, 59–62. doi:10.1016/j.nbt.2024.12.003 | abstract only | Consensus record; gold OA per OpenAlex but ScienceDirect refused WebFetch and curl | +| Ariansyah, D., et al. (2022). A head mounted augmented reality design practice for maintenance assembly. *Applied Ergonomics, 98*, 103597. doi:10.1016/j.apergo.2021.103597 | abstract only | Consensus record: 3D animation 14% faster than video; eye-gaze; "overreliance" raised from a joint-cognitive-systems view. Nottingham repository preview returned 403 | +| Fang, W., et al. (2023). Head-mounted display augmented reality in manufacturing: A systematic review. *RCIM, 83*, 102567. doi:10.1016/j.rcim.2023.102567 | abstract only | Consensus record; closed | +| Hagemann, V., et al. (2026). Team member or tool? The role of framing in human-autonomy teaming. *IJHCI*. doi:10.1080/10447318.2026.2634249 | abstract only | Consensus record: N = 80, Mars habitat task, no trust or affect difference between framings | +| Kargarnovin, S., et al. (2026). From testbeds to high-stakes work: a review of human-AI teaming domains. *Frontiers in Robotics and AI*. doi:10.3389/frobt.2026.1733942 | abstract only (open, not opened for time) | Consensus record: 104 empirical studies 2015–2025 | +| Endsley, M. R. (1995). Toward a theory of situation awareness in dynamic systems. *Human Factors, 37*(1), 32–64. doi:10.1518/001872095779049543 | abstract only | Search-result abstracts; closed; no repository copy found | +| Endsley, M. R., & Kiris, E. O. (1995). The out-of-the-loop performance problem and level of control in automation. *Human Factors, 37*(2), 381–394 (`endsley1995out` in the bib) | abstract only | As before; one of the 18 studies in `onnasch2014human` | +| Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. *Human Factors, 39*(2), 230–253. doi:10.1518/001872097778543886 | abstract only | Search-result abstracts; closed | +| Endsley, M. R. (2017). From here to autonomy. *Human Factors, 59*(1), 5–27. doi:10.1177/0018720816681350 | not accessible | Closed; cited by `oneill2022human` and `shneiderman2020human` | +| Endsley, M. R. (2023). Supporting human-AI teams: Transparency, explainability, and situation awareness. *Computers in Human Behavior, 140*, 107574 | abstract only | Webinar abstract via WebSearch; closed | +| Hollnagel, E., & Woods, D. D. (1983). Cognitive systems engineering: New wine in new bottles. *IJMMS, 18*(6), 583–600. doi:10.1016/S0020-7373(83)80034-0 | not accessible | Closed; the 1982 Risø report is catalogued but not online; summarised in `rayo2023cognitive` | +| Christoffersen, K., & Woods, D. D. (2002). How to make automated systems team players. *Advances in Human Performance and Cognitive Engineering Research, 2*, 1–12. doi:10.1016/S1479-3601(02)02003-9 | abstract only | Emerald abstract via WebSearch; cited in `klein2004ten` | +| Mackworth, N. H. (1948). The breakdown of vigilance during prolonged visual search (`mackworth1948breakdown`) | not accessible | As before | +| Strauch, B. (2018). Ironies of automation: Still unresolved after all these years. *IEEE THMS, 48*(5), 419–433. doi:10.1109/THMS.2017.2732506 | not accessible | Closed | +| Wickens, C. D., Clegg, B. A., Vieane, A. Z., & Sebok, A. L. (2015). Complacency and automation bias in the use of imperfect automation. *Human Factors, 57*(5), 728–739. doi:10.1177/0018720815581940 | not accessible | Closed; the meta-analysis that would extend `onnasch2014human` | +| Lindgren, I. (2024). Ironies of automation and their implications for public service automation. *Government Information Quarterly*. doi:10.1016/j.giq.2024.101974 | not opened | Hybrid OA per OpenAlex; found via Semantic Scholar's one successful query | +| National Academies (2022). *Human-AI Teaming: State-of-the-Art and Research Needs* | not opened | Free PDF requires a login | diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_T_adjudication.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_T_adjudication.md new file mode 100644 index 00000000..1bab73ab --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_T_adjudication.md @@ -0,0 +1,435 @@ +# T — Adjudication + +*Written by Claude (lead) on 2026-10-12 from facets A–H, X and Z and audits `../audit/AUDIT_01`–`03`. +None of this is the author's text. Each verdict ends in a recommended ruling; the author rules.* + +**How to read the depth marks.** *Full* means a facet agent read the passage in the source and an +auditor confirmed it. *Part* means read from a working paper, preprint or lending scan. *Abstract* means +abstract only. *Confirmed* is used for Saenger 1997: an auditor confirmed +the wording, and for 19 of 25 items the page, by Google Books search-inside and the archive.org lending +scan (`../audit/AUDIT_01.md`); the pages still want one check against a print copy before publication. + +## 1. The history: H-cost against H-cause + +**Verdict: H-cost holds in a narrow form. H-cause has no support.** + +| Evidence | Says | Depth | Facet | +| --- | --- | --- | --- | +| Saenger 1997, pp. x, 11 | His book "is not a general history of literacy". He does list "the greater diffusion of literacy throughout all social strata of the population" among the advantages modern readers see in ease of reading (p. 11); that is a modern value the ancients did not share, not a finding about the Middle Ages | Confirmed | A | +| Saenger 1997, pp. 83, 100, 176–177 | Irish separated books before 692 and 713; "not adopted by Continental scribes until the mid-tenth century"; "standard and ubiquitous" by the mid-eleventh | Confirmed | A | +| Saenger 1997, pp. 8, 13, 122 | Ambrose read silently but "could not have read rapidly in the manner of the modern silent reader". The strong claims concern speed and reference reading among Latin readers, and a scholasticism that "without these tools would not have been possible" | Confirmed | A | +| Knox 1968, p. 435 | "Ancient books were normally read aloud"; silent reading existed for letters and documents | Full | B | +| Vatri 2012, via Heilmann 2020, p. 181 | The physiological mechanism "must be rejected" | Full (Heilmann); abstract (Vatri) | B | +| O'Donnell 2000 | Spacing as "necessary and sufficient" for silent reading is the claim most open to dissent | Full | B | +| Ten histories of literacy | None credits word separation with the spread of reading. Clanchy indexes it once, under layout | Part (full-text search of lending scans) | C | +| Scribner and Cole 1981 | The Vai script "runs across the page without division into words or phrases" and "is customarily learned in adulthood" without extended study. It shows an unspaced script can be learned informally; it is one case | Part; pages unconfirmed | C | +| Cressy 1980 | "a reasonable guess might place male illiteracy around 80%" at Elizabeth's accession in 1558, about five centuries after spacing was standard (mid-eleventh century to 1558) | Part; page unconfirmed | C | +| Bai et al. 2008; Winskel et al. 2009 | Fluent readers of Chinese read no faster with added word spaces (5,221 ms against 5,239 ms, 16 readers). Fluent Thai readers were not faster either; their 5% slowdown is unreliable (spaced sentences were 12.8% longer; item-wise p = .13) | Part | D | +| Winskel et al. 2009; Rayner et al. 1998 | Removing spaces from English slows reading by a third to a half | Part; abstract | D | +| Shen et al. 2012; Blythe et al. 2012; Li et al. 2021 | Spaces help second-language learners and beginners most; other child studies find no gain | Abstract to part | D | +| Hangul; Guido; Pacioli; Hindu-Arabic numerals | Each "one move" story dissolves into a bundle plus schools plus demand. Hangul stayed marginal for four centuries | Part | F | + +Three points follow. + +1. **Spacing lowered a cost, most for learners.** That is the whole of what the experimental and + historical record supports. Insular scribes separated words for readers who were learning Latin as a + foreign language; the Continent adopted separation three centuries later for scholars reading hard + texts. Nobody designed it for a populace. +2. **Demand may have come first.** Romans used word dividers and gave them up. Saenger and Johnson both + explain that choice socially: an elite that did not want reading made easier. The move became + attractive when the readers changed. +3. **No source ties spacing to how many people could read.** Saenger says his book is not a history of + literacy, the historians of literacy ignore spacing, and unspaced scripts have supported both fluent reading and informal learning. + +**Recommended ruling.** Keep the Part 2 ruling for this paper: the history is a design precedent. Retire +"ignited literacy". The defensible sentence is of the form "word separation lowered a cost of reading, +most for readers still learning the language; it did not by itself raise the number of readers." + +**The author's sentence at `draft.md:263`.** "…it changed the interface so that reading no longer +required the training" is stronger than the sources. Hutchins and Zhang and Norman describe a task +transformed, and the scribe took on the work the reader shed. "Populace" has no support. I flag it; the +wording is the author's to change. + +## 2. Word spacing is a legibility remedy, and the ruling rejects legibility + +**Verdict: the contrast can be drawn, with a narrower ruling and a different test.** + +- **Two senses of one word.** Parkes's "grammar of legibility" is graphic legibility, a perceptual aid. + The 30 September ruling rejects legibility as interpretability: asking a person to read an opaque + system more skilfully. Human factors treats the first as its best work (ecological interface design; + Erickson and Kellogg's sign and window). A blanket "not legibility" will read as rejecting the field. +- **"Moves work to the artefact" does not separate the paper from Nadella.** A reference monitor does + exactly that (facet G, C7). Facet E's cut is better: does the change make the person's task perceptual + at a load they can carry in real time, or leave it analytical and add volume? +- **Two further tests**, both the facet agents' inferences and mine, not findings from a source: what + task the person keeps after the artefact has done its share, and whom the move serves. Spacing left the + reader a task people do well and served the reader. "Observe everything" leaves a person reading + evidence and deciding when to halt, which Bainbridge (1983, p. 776) calls "an impossible task" when + the machine decides faster than a human can follow, and it serves the deploying organisation. +- **Absorbing a task has costs.** Risko and Gilbert name complacency and skill decay; Buçinca et al. + found the more demanding design worked best; Bainbridge's own remedy is deliberate inefficiency. + +**Recommended ruling.** Restate the first ruling for this paper as "algorithmacy is not a larger reading +load on the person", and let the paper use perceptual re-representation as its vocabulary (Zhang and +Norman's representational effect; Hutchins's task transformation, from his 1995 article "How a cockpit +remembers its speeds"). + +## 3. Nadella is not a clean opposite + +**Verdict: "exact opposite" is not supported by the text. A contrast on authority and seat is.** + +| The author's count | Verdict | Basis | +| --- | --- | --- | +| Legibility paradigm | Partly holds | "observe", "log", "auditability" and "human readable evidence" are in the text (P4, P8, P15, P18). His load-bearing controls are deterministic and need no reading (P11, P12, P17). The reader is an auditor, never a user | +| Keeps the dyad | Holds once "dyad" is defined as organisation and model | Authority rests with "Orgs" (P5, P17). "Those affected" appear once, after a failure (P20). No counterpart to a transaction appears | +| Containment versus engagement | Holds for this post | Steering rights go to organisations (P17). His other 2026 posts describe firm-level correction loops, so the paper should say "the insider post" | + +**Convergences the paper must concede in the body.** Chain-of-thought transparency "alone is not +sufficient or dependable" (P9). Controls sit outside the model (P11), which is the mechanism of bounded +outputs, though declared to the deployer and to no one else. Disclosure goes "to those affected" (P20). +The emergency brake (P19) resembles structural refusal in form only: the organisation holds it. + +**Where the human-factors evidence bites.** It undercuts real-time noticing and whole-volume review, +which is P19 and a literal reading of P15. It does not undercut reconstruction by trained auditors, and +it favours the artefact-side controls that are his backbone. Saltzer and Schroeder's "psychological +acceptability" puts a human-factors principle inside the canon he invokes and he does not use it. + +**Scope.** The post is enterprise security architecture. The author declined the "enterprise, not user" +reading, and the text supports that reading more than any other. The honest form of the contrast: the +post answers who holds authority over a model from the deployer's seat, and leaves out the people the +model acts upon. + +**Recommended ruling.** Replace "the exact opposite" with a contrast on who holds authority and whom +the design serves. Treat containment as the floor under the paper's proposal. Aim the critique at P15 +and P19. + +## 4. One move or a complex + +**Verdict: a complex of conventions.** + +- Saenger links word order and word separation (p. 17; wording confirmed, page to check); + spacing arrived with punctuation, construction notes, capitals, and later tables and running heads. +- Every parallel case in facet F is a bundle plus teaching institutions plus demand. +- In facet H the measured candidates sort by what they do with the work: + +| What the intervention does | Examples | Evidence | +| --- | --- | --- | +| Removes a task by fixing the mediator's rule | Strategy-proof matching; withholding cues; simultaneous reveal | Lab and field experiments on matching; removal is incomplete (23% still misreport, Rees-Jones and Skowronek) and depends on presentation (36% against 7% off the dominant outcome, Li 2017). The platform study of withheld cues (Mejia and Parker) says it cannot tell whether the party gained | +| Hands over a pre-computed comparison | Kling et al.: 28% against 17% switching | One randomised field experiment | +| Hands the person something to read or tune | Envelopes, seams, model cards, feed controls | Mostly satisfaction; several nulls | +| Hands the task to the person's own agent | Zhu et al. 2026 | One preprint, no mediator; the delegate's advantage does not survive correction for multiple tests (p = .033, then .10) | +| Removes exposure | Chronological feed | Two field experiments, behaviour only | + +- The strongest candidates work because the participant does **not** read or rewrite the mediator. That + collides with the engagement layer as worded ("reading and rewriting the shared mediator"). +- The removed task moves elsewhere (search and stopping in Chilean school choice; Arteaga et al.). + +**Recommended ruling.** Write "a convention" or "a complex of conventions", never "the move". Offer +mediator rules that remove a strategic task, and make the removal visible, as the best-evidenced +candidate, and present the missing experiment as the paper's research agenda (facet H, (iii) 1). + +## 5. Problems the research added + +1. **Who builds it.** Scribes shared the reader's interest; the platform that owns the interface does + not share its users' (Part 1, line 60: users "whose objectives conflict with the platform's bottom line"). "Redesign the interface" hands the remedy to the + interested party. The convention has to be one the owner cannot withdraw alone: a public standard, the + person's own agent, a regulated rule, or exit. Facet X ranks this the most damaging objection. +2. **No trained elite.** `CLAIM.md` says coordination through an algorithm is "in the condition of + unspaced script: only the trained can do it". Part 1 gives no evidence that anyone can do it reliably + (lines 66, 138, 156). The analogy needs "a task nobody has shown how to do reliably". +3. **Sensibility and artefact.** The author defines algorithmacy as a sensibility of persons and places + the remedy in artefacts. The paper has to say how the two relate. +4. **Media hold still.** External-representation research assumes a stable regularity and a designer who + knows the constraints (Hutchins; Scaife and Rogers; Vicente and Rasmussen). An adaptive, interested + mediator offers neither. This is Part 1's own caution (line 118) and the place the analogy breaks. + +## 6. What survives + +*Claude's statement of the strongest defensible thesis, for the author to accept, change or reject.* + +Scribes once changed the page so that readers no longer had to find the words. The change lowered the +cost of reading for people still learning, spread over centuries for social reasons, and came as part of +a complex of conventions. It is a clear historical case of a human-factors principle: re-represent the +task so that the artefact carries work the person would otherwise carry. + +Coordination through an algorithmic third party has no such conventions. Decision Experience Design is +the discipline that looks for them, and the evidence so far points to rules in the mediator that remove +a strategic task from the people it coordinates, made visible to them, and held in a form the mediator's +owner cannot withdraw alone. + +Nadella's insider post shares the premise that the model cannot be read and the remedy of controls +outside it. It differs in whom the controls serve. His architecture gives authority to the organisation +that deploys the model and leaves a human to read evidence and pull a brake, tasks the oversight +literature finds people do badly. It says nothing of the people the model acts upon. + +## 7. Decisions for the author + +1. History: design precedent (recommended) or cause? +2. Keep, soften or replace "no longer required the training" and "populace" (`draft.md:263`)? +3. Restate "algorithmacy is not legibility" as "not a larger reading load on the person"? +4. Replace "the exact opposite" with the contrast on authority and seat? +5. Take up the "enterprise, not user" reading you declined, now that the text favours it? +6. "A complex of conventions" in place of "a similar move"? +7. The engagement layer: keep "reading and rewriting the shared mediator", or reword so that the person + steers without a reading task? +8. Who builds the convention: standard, the person's agent, regulation, exit, or some mix? +9. Does the paper propose the missing experiment (remove a task, add a reading aid, or delegate, with a + mediator that reads both sides)? +10. The course statement rejects giving novices different support from experts. The expertise-reversal + literature expects the opposite. Defend or drop? + +## 8. Limits of this research + +- Saenger 1997 was first read from a copy of doubtful provenance; the audit then confirmed the quotations + used here by two legitimate routes, and no chapter was read whole. Parkes 1992 is known only from + snippets. Wingo 1972, + Vatri 2012, Gavrilov 1997, Burnyeat 1997, Johnson 2000, Dekker and Woods 2002, Norman 1991, and Zhang + and Norman 1995 are unread or abstract-only. The priority list is in `OUTLINE.md`. +- No book was read in full. The literacy histories were searched, not read. +- Facet H has no source at full-text depth in the strict sense; ten were read in the passages cited. +- The gap search could not reach Semantic Scholar, arXiv search or Scholar Gateway, and did not cover + X, Hacker News or podcasts. +- Several AI-safety sources in facet G come from Anthropic, and the agents and the lead are Anthropic + models. +- The Nadella post was two days old at the time of research. + +## 9. Second round (2026-10-13): after the author's rulings + +*Written by Claude (lead) on 2026-10-13. The author ruled on the ten decisions of §7 (verbatim in +`../../SOURCE.md`). Three facets followed from the rulings (I, J, K; audits `../audit/AUDIT_04`–`06`), +and 251 sources were then read in full or in large part and carded (`../library/`, +`../audit/CARD_FACET_CHECKS.md`). Sections 1–8 stand as the record of the first round; where this +section disagrees with them, this section is the later reading. A source named with a key in backticks +has a card, and the card carries the page locators.* + +### 9.1 What the full readings changed in §§1–5 + +| Earlier statement | Now | Source | +| --- | --- | --- | +| §1, point 1: spacing lowered a cost "most for learners" | Not shown in the modern evidence. Child learners of Chinese read spaced text more slowly and less accurately (word spacing F(1, 21) = 11.80, p = .002; hanzi accuracy 95.67% against 90.91%); the learner contrast in Cui et al. is not significant. What stands for learners: second-language adults in abstract-only studies, one interaction for dyslexic readers, and Saenger's Insular case | `bassetti2012effects` (submitted version), `cui2023eye`, `galliussi2020inter` | +| §1 table: fluent readers of unspaced scripts gain nothing | Mixed. Adult Thai readers fixate the target word for less time with spaces and take longer over the sentence; eight Thai readers read aloud faster with spaces (p < .08); Tibetan adults read spaced sentences faster (5,566 against 6,019 ms). Bai et al. remain a clean null for skilled Chinese adults (5,221 against 5,239 ms) | `chantavarin2025effects`, `kohsom1997adding`, `winskel2009eye`, `wang2024tibetan`, `bai2008reading` | +| §1: removing spaces slows practised readers by a third to a half | Stands, and is the robust result. Mirault's "40–70%" is his citation of Rayner and Pollatsek (1996) and Rayner et al. (1998), not his own data | `mirault2019reading`, `winskel2009eye` | +| §1, point 1: Insular scribes separated words "for readers who were learning Latin as a foreign language" | A reading held by Parkes and Saenger, not a settled finding. The reviewers' "convincingly" and "plausible" attach to narrower points; Tunbridge's thesis, supervised by Parkes, is no independent test | `../audit/AUDIT_06.md`; `tunbridge1992study` | +| §1, point 2: the Romans' teachers divided words for novices and then withdrew the aid | The aid was known and used for learners; "withdrew" is wrong. Cribiore says continuous blocks "made word divisions necessary as a further aid for the pupil" (1996, p. 48, lending scan) | `../audit/AUDIT_06.md` | +| §1: Knox supports trained readers supplying word boundaries | Knox says nothing of word boundaries. He supports "normally read aloud" only | `knox1968silent` | +| §2: avoid "cognitive offloading" for the scribes' move | The term is available. Risko and Gilbert count offloading designed into the task environment (Box 3) | `risko2016cognitive` | +| §2: Zhang and Norman, representational effect, p. 89 | p. 88 | `zhang1994representations` | +| §3: aim the critique at P15 and P19, where people perform worst | Weaker than stated. Insider-threat practice endorses retrospective review by people (CERT, practice 12, pp. 84, 89); complacency is mainly an effect of competing tasks; the alert-fatigue paper measures no analyst performance; Greenblatt's "auditor" is a model. One direct measure exists: three annotators per transcript, 103 minutes each, AUC 0.65 | `cert2022common` (partial), `parasuraman2010complacency`, `alahmadi2022false`, `greenblatt2024ai`, `kutasov2025shade` | +| §4 table: Kling et al. as "a pre-computed comparison" | The letter was a bundle: recommended default, savings figure, deadline, printout, after an interview | `kling2012comparison` (working paper) | +| §4: presentation decides who benefits (Gonczarowski et al., "with a lab experiment") | The version on file is theory only. Li (2017) and Rees-Jones and Skowronek carry the point alone | `gonczarowski2023strategyproofness`, `li2017obviously`, `reesjones2018experimental` | +| §4: bounded control beats free control (Dietvorst et al.) | Not shown. Under random assignment the capped and free conditions do not differ (p = .224). Some control raises use of the algorithm | `dietvorst2018overcoming` | +| §5 and the exposure list: the more demanding design worked best (Buçinca et al.) | Narrower. Forcing functions beat simple explanations only where the AI was wrong, gave no overall gain, and stayed below working without the AI | `bucinca2021trust` | + +### 9.2 The rulings, one by one + +**D1 and D2. Precedent, with the causal analysis reviewed.** Facet K searched the histories of +reading instruction for the training claim and for every causal claim made from word separation. + +- *Nearest support for `draft.md:263`.* Saenger writes that "Word separation, the diastole, the hyphen, + and eventually the trait d'union in Insular books provided shortcuts for achieving the reading skills + that an elite among the ancients had mastered only through a prolonged and arduous grammatical + apprenticeship" (1997, p. 85; verbatim and page confirmed, `../audit/AUDIT_06.md`). He names four + devices, gives no measure, and no reviewer read so far comments on the sentence. +- *Against "no longer required the training".* Reynolds, Black, Grendler and Orme describe the alphabet, + syllables, psalter and Latin course continuing after separation was standard (snippets). Their books + return no hit for "word separation" or for Saenger, which shows only that the phrase is absent from + the scanned text. +- *"Populace".* No source supports it. +- *Causal claims.* Saenger alone claims that separation caused silent or rapid reading, reference + reading and the scholastic apparatus, private devotion, and autograph composition. Classicists contest + the first (Knox, Vatri, Heilmann); the rest are untested in what was read. Svenbro, in Cavallo and + Chartier's volume, calls unspaced script "not an insurmountable obstacle" and then cites Saenger for + word division as "a necessary condition for the dissemination of the practice of silent reading". Nobody + found claims that separation raised the number of readers. +- *What the sentence can be made to say with support:* that separation and three other devices gave + Insular readers a shorter route to skills an ancient elite reached by long training. That is Saenger's + claim, about Insular books, and should be attributed to him. + +**D3. Disclosure is a literacy-based intervention.** The ruling settles the legibility problem of §2: +word separation disclosed nothing, since the letters were the same before and after, and the +representational effect is defined over "different isomorphic representations of a common formal +structure" (`zhang1994representations`, p. 88). The evidence on disclosure is mixed in direction, which +"doesn't affect" does not capture. + +| Finding | Direction | Card | +| --- | --- | --- | +| Reddit users who had read the rules judged their removal less fair (β = −.314); the abstract says the opposite | Disclosure did not help | `jhaver2019suspect` | +| A months-to-pay message on credit-card statements raised delinquency by 0.011 points, 7.4% of the mean, for two client groups; the interest-rate message did nothing | Disclosure harmed or did nothing | `seira2017are` | +| A clearer model made people worse at catching its mistakes in three of four experiments | Disclosure harmed | `poursabzi2021manipulating` | +| Showing where an AI errs raised accuracy mostly by raising reliance on ordinary items | Disclosure changed deference, not navigation | `cabrera2023improving` | +| Korbak et al. decline to recommend showing chain of thought to users (p. 7) | The monitoring literature itself stops short | `korbak2025chain` | +| The EU Platform-to-Business Regulation requires a description of ranking parameters and says providers are "not required to disclose algorithms" (Art. 5(6)) | The legislator stops short | `eu2019p2b` (partial) | +| A warning cue with its reason cut simulator collisions to 1 of 15, against 6 of 15 with the cue alone and 7 of 15 with none (N = 45, Fisher p = .044) | Disclosure helped | `matsuo2026educational` | +| Explanations reduced overreliance in the hard task only | Disclosure helped, conditionally | `vasconcelos2023explanations` | + +Matsuo et al. and Vasconcelos et al. are person-and-machine settings with no second party. No study +found tests disclosure in a triad with a behavioural outcome. A sentence the sources carry: disclosure +has not been shown to change how well a person navigates a mediated coordination, and in several +studies it changed nothing or made matters worse. + +**D4 and D5. The insider post.** The post names "Neuralese" (P9), as the author says. Its text does not +bear "his only move is disclosure": model diversity (P14), independent controls (P17) and containment +(P19) disclose nothing, and Chan et al. set out controls outside the agent as a design programme +(`chan2025infrastructure`). The form of the author's contrast that the text supports: what the post +gives a *person* is something to read (evidence, audit, incident disclosure) or a brake to pull, and +every other control is held by the organisation. With D5 adopted, the post's silence on the people the +model acts upon becomes the main point. The critique should not rest on the claim that people review +records badly (9.1). + +**D6. A complex, with a search for the most critical convention.** The first-round candidate, a rule in +the mediator that removes a strategic task, keeps its two experiments (36% against 7% of games off the +dominant outcome, `li2017obviously`, §5.3 of the working paper; 23% misreporting in a strategy-proof +match, `reesjones2018experimental`). D7 adds a second candidate, set out below. + +**D7. From reading and writing to deciding.** Facet I and the cards give the ruling partial support +and one correction. + +- *Support.* Weick: "people think by acting" (`weick1988enacted`). Kirsh and Maglio's epistemic + actions are "physical actions that make mental computation easier, faster, or more reliable" + (`kirshmaglio1994distinguishing`, p. 513). Ride-hail drivers with more than 2,500 trips earn 14% more + per hour than those with fewer than 100 (`cook2021gender`, working paper, observational). Heiland's + own A/B trials on a delivery app corrected riders' false theories (`heiland2023social`). +- *Condition.* Skilled intuition needs a valid environment and "feedback that is both rapid and + unequivocal" (`kahneman2009conditions`, p. 524). Delay and noise act differently: delay pushes people + toward options that pay quickly, and noise does not (`rahmandad2023delays`, pp. 13, 17). Part 1, + line 66, says an adaptive intermediary disrupts both conditions. +- *Correction.* "Pilots don't spend time reading the situation" is contradicted. Both pilots in + Hutchins's cockpit keep monitoring airspeed (`hutchins1995cockpit`, pp. 283, 285); with automation + engaged, monitoring is the pilot's task. A narrower form holds: crews caught 35.4% of deviations in + primary procedures, 14.4% in checklists and 5.6% in monitoring (`dismukes2010checklists`, partial). + People correct what they do far more often than what they watch. +- *No source supports "from reading and writing to deciding" as a historical shift.* It is the author's + proposal. +- *Inference (Claude's, unsourced).* If understanding of a mediator comes through deciding and taking in + the results, then the convention that matters most may be one that returns the result of a decision + quickly and plainly, and holds still long enough for the result to mean something. That would be a + candidate for D6 that keeps the person acting, unlike the task-removing rule. Interventions that + shorten or clean the loop have thin and mixed evidence (previews, simulation, counterfactual feedback; + `kuhl2023alien` did not replicate its own gain, p = .554). + +**D8. Who holds the convention.** Facet J, its audit and the cards give this comparison. "Draconian" is +the author's word and is tested, not adopted. + +| | Courts | Dealer foreign exchange | Platforms | +| --- | --- | --- | --- | +| Who writes the convention | Judges under statute, through committees; non-compliance with the process "does not invalidate a rule" (28 U.S.C. § 2073(e)) | The Global FX Committee of central banks and industry; each dealer for its own last look | The owner; EU law sets a floor | +| Change of terms | By rule-making or appeal | Substantial changes "should be agreed before any transaction" (Principle 40), voluntarily | By the owner alone, on 15 days' notice to EU business users (P2B Art. 3) | +| Reasons and complaint | Reasoned decisions, appeal | Basis for rejections disclosed, reasons recorded and shared "where appropriate", a channel for concerns (Principles 7, 9, 36, 39), all voluntary | Statement of reasons, internal complaint, regulator complaint and damages in the EU (DSA Arts. 17, 20, 53–54); in Seattle the company decides the worker's challenge itself | +| What binds the third party | Appeal, recusal; an interested judge is barred (*Tumey v. Ohio*, 1927) | Reputation and regulators after the fact; the 4pm fix was manipulated by dealers and five banks were fined £1.1bn on 12 Nov 2014 | Statute where it exists (fines to 6% of turnover under the DSA); otherwise the owner's terms, which gave "broad, unfettered discretion" in all 14 platforms Suzor read | +| What moved it | Litigation | Scandal, then regulators | Statute; collective action (a drivers' union won over 1,400 reactivations before any tool existed) | + +Three findings qualify the comparison. Tyler finds the weights people give to criteria of fair procedure +vary with the situation and concludes there are "no universally fair procedures" (`tyler1988what`, +p. 131), so the settings can be described on these rows and not ranked on one scale. Galanter's +one-shotter does not fit a driver or seller who deals with a platform continually (`galanter1974why`, +n. 4, p. 97). Klonick calls platform moderation "impressively intricate and responsive" while recording +the lack of an appeals system (`klonick2018new`, pp. 1664–1665), and Meta reversed itself in 14 of 20 +cases the Oversight Board shortlisted, before the Board ruled (`oversight2022q1`, p. 9). + +*Where the convention bends.* In every setting it moved toward the weaker party under outside force and +not on the third party's initiative. On paper, dealer foreign exchange promises the counterparty most +and binds the dealer least; EU platform law promises less and binds more; a platform's own terms promise +least. The author's reading holds for terms of service and for places without such statute. Premise P10 +stands: nothing found stops an owner from changing a convention alone. + +**D9. Irreducible triad.** The paper says "irreducible triad" and names Φ as the gauge of +irreducibility. Two points stay open for the author: whether the paper proposes a behavioural experiment +(no study found tests any convention in a triad with a behavioural outcome), and whether this arm +computes Φ or cites the lab's results. A Φ value classifies a triad; it does not show that a convention +helps the people in it, so the two are complements. + +**D10. The history as a guide.** This agrees with the Part 2 ruling and with what the evidence now +permits (9.1). The course statement's claim about novices and experts is not carried, on Claude's +reading of the answer. + +### 9.3 Three further matters for the author + +1. **"Algorithmacy" appears in print in 1995.** Oberschelp writes "we have to preserve logicacy and + algorithmacy by continuously training simple mental activities without artificial tools" + (`oberschelp1995developing`, p. 121), once, undefined, meaning a skill of carrying out algorithmic + steps by hand. The sense differs from the author's. Any statement of coinage should say "in this + sense" or cite him. This reaches the dissertation and the *Organization Theory* arm. +2. **A seventh flag on the author's published text.** Part 1, line 106, says forcing functions curbed + uncritical acceptance "far more effectively than explanatory text"; the paper's result is narrower + (9.1, last row). +3. **A source that changed sides.** Meyman (2026), which facet Z read as a proof that observability + cannot produce authorization, withdraws the universal claim in its current version and is a + self-published note by an author affiliated with the company it supports (`meyman2026impossibility`). + +### 9.4 What survives, restated + +*Claude's statement, for the author to accept, change or reject.* + +Readers used to spaces lose a third to a half of their speed without them. That dependence is the robust +fact; whether spaces first opened reading to learners rests on Saenger's reading of the Insular evidence +and one sentence on p. 85. Word separation re-represented the page without disclosing anything new, and +that is the feature the paper needs: it is the model of an intervention that is neither disclosure nor +training. The history is a guide to the question and no proof of the answer. + +Coordination through an algorithmic third party has two candidate conventions with some evidence. One +removes a strategic task by rule and shows that it is gone. The other, which follows from the author's +ruling that understanding now comes through deciding, returns the result of a decision fast and plainly +enough to learn from. The second is an inference with thin evidence. No convention has been tested in a +triad, and in courts, dealer markets and platforms alike the third party gave ground only under outside +force. + +Nadella's insider post answers a different question: how an organisation contains a model it cannot +read. What it offers a person is a record to read and a brake. It is silent on the people the model acts +upon. + +### 9.5 Limits of the second round + +- 251 of 576 sources are carded; 43 of the cards rest on a partial read. Saenger, Parkes, Wingo, Vatri, + Gavrilov, Burnyeat, Reynolds, Orme, Cribiore and the other monographs were read as snippets or not at + all, and facet K's page locators for them are mostly "not obtained". +- The verification pass corrected 151 of 251 cards. A card is the best record available and is not a + substitute for the source when a passage is quoted in the paper. +- Several sources were read as working papers or preprints (Li 2017, Kling 2012, Cook 2021, Bassetti and + Masterson 2012, Bolton 2013); each card names the version. +- About 40 sources have open copies behind bot walls that the agents left alone. Among them is a 2026 + post titled "wordspacing" (key `rosehill2026wordspacing`) that bears on whether the analogy has been + drawn before; the gap statement of facet Z is unchanged until someone reads it. +- The Jensen and Meckling card was written from a scan of unclear provenance and should be checked + against the journal. + +## 10. Third round (2026-10-13): literatures the outline had not covered + +*Written by Claude (lead) on 2026-10-13. The author adopted a nine-section outline (snapshot in +`../../DRAFT.md`; the governing text is his Google Doc) and asked for research on what else the +literature holds. Facets L, M, N, O and P ran, and `../audit/AUDIT_07`–`11.md` re-checked and corrected +them in place (18 wrong and 92 overstated rows across the five). Read depth is thinner than in the card +library: most sources were read in part or as abstracts, and each facet says which. Section numbers +below are the outline's.* + +| Outline section | What the round found | Sources, with the audit's caution | +| --- | --- | --- | +| 2, 7 | Ruler and church enforced the conventions of the written page. Nothing opened shows scribes and readers sharing interests as equals, and none of these texts mentions word separation | Charlemagne, letter to Baugulf (read in full, translation); *Admonitio generalis* c. 72 and Alcuin, carm. XCIV (OCR of the MGH volumes); Westwell 2018. McKitterick, Ganz and Bischoff not opened | +| 4 | The law-and-economics critique of mandated disclosure states the paper's mechanism: disclosees do not read, understand or use disclosures | Ben-Shahar and Schneider 2011, p. 665 (read in full); Loewenstein, Sunstein and Golman 2014; Fung, Graham and Weil 2007 | +| 4 | Explanations do not improve decisions on average | Schemmer et al. 2022 (SMD 0.07, 95% CI −0.15 to 0.30; read in full); Vaccaro, Almaatouq and Malone 2024 (370 effect sizes) | +| 4 | Disclosure often changes behaviour without improving outcomes. A stated property of a rule moved sincere bidding from 20.6% to 46.9% and efficiency by about one point | Masuda et al. 2022 (net effect against control p = 0.051). "Doesn't affect" is contradicted for behaviour; the claim should be about performance | +| 4 | The nearest rival, nudge against boost, sorts by whether an effect outlasts the intervention, and a boost can be a re-representation. The measured case for structure over information is modest | Hertwig and Grüne-Yanoff 2017, pp. 977, 981 (read in full); Mertens et al. 2022 as corrected; Maier et al. 2022. Mertens classes "translation" of information with informational interventions | +| 5 | Intent is sourced for withholding rules and not for noise or delay | Rahman 2019 (dissertation, pp. 128–129: anti-gaming and control); the 2021 article not opened. Chen, Mislove and Wilson 2015: surge jitter was a bug by Uber's account | +| 5 | Where the rule is hidden, people infer it from other people's outcomes; learning is social as well as enacted | Ghalme et al. 2021; Bechavod et al. 2022; Lee et al. 2015; Kellogg, Valentine and Christin 2020 | +| 6 | Convention A: the best presentation tested leaves about four in ten not playing straightforwardly; a menu description raised understanding more than play | Gonczarowski, Heffetz, Ishai and Thomas 2024 (N = 542; 59% against 56% and 48%; read in part) | +| 6 | Convention B: validity comes before speed. Frequent feedback on noisy choices lowered investment by about 25% and earnings by about 35%; more than a third of feedback interventions lower performance | Larson, List and Metcalfe 2016 (read in part); Kluger and DeNisi 1996, 1998 (abstracts); Tiefenbeck et al. 2018 for the case where real-time feedback helps | +| 7 | One statute goes beyond disclosure: a human takes termination decisions, and decisions that infringe rights are rectified or compensated | Directive (EU) 2024/2831, Arts. 7, 10(3), 10(5), 11(3), 13; in force since 1 December 2024, transposition due 2 December 2026 | +| 7 | Standards bodies bind no implementer; co-determination binds through arbitration | RFC 8890 (IAB); BetrVG §87(1) no. 6 and (2) | +| 8 | Closest precedent: two people and a computer mediator that maximises agreement, rule shown every round, no manipulation of what is shown | Casella, Friedman and Perez Archila (working paper, 2024). No study found gives the mediator its own objective. March 2021 and Chugunova and Sele 2020 not opened | +| 8 | Integrated information has been applied to human groups only as a proxy | Engel and Malone 2018 (68 groups, r = .37, no mediator; read in full). Barrett et al. 2026 (preprint): Φ "has not been computed on any real physical system" | + +**Inferences by the agents or the lead, with no source behind them.** That disclosures which work are +re-representations (the sources say "simple signals" and stress the discloser's response). That an owner +is bound only when a third party with a sanction stands behind the rule, or the convention is a mandated +format. That Convention B should read "valid and plain first, then quick". Each is usable as the paper's +argument and should not be cited as a finding. + +**What was changed in the author's document.** Fifteen word-level edits followed from this round, listed +in the commit that added this section: the two sentences on scribes and readers, the "Mediated +Distortion" bullet, the limits of both conventions, the abstract's gloss on Convention B and its verb +("argues" for "demonstrates"), citations for the literacy fallacy and the surge example, the Platform +Work Directive among the statutes, and three additions to section 8 (the measurement plan, the Casella +precedent, a discrete-time arm). + +**Still open.** Whether the paper keeps the name "radical feedback immediacy"; how it places itself +against nudge and boost; whether it treats a stated property of a rule as disclosure or as part of +Convention A; and the full texts listed in each facet's section (d). diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_W_spacing_oracy_literacy.bib b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_W_spacing_oracy_literacy.bib new file mode 100644 index 00000000..9adad500 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_W_spacing_oracy_literacy.bib @@ -0,0 +1,354 @@ +% Facet W bibliography, 2026-10-15. Agent-drafted (Fable); the author has not reviewed it. +% Every bibliographic field below was taken from the Crossref record of the DOI (or from the source page for items without a DOI) on 2026-10-15. Nothing is from memory. UTF-8 encoded. +% Reused from literature/references.bib and not repeated: saenger1997space, tunbridge1992study, garrett1998saenger, odonnell2000saenger, knox1968silent, busch2002lautes, heilmann2020reading, gavrilov1997techniques, burnyeat1997postscript, johnson2000toward, vatri2012physiology, mccutcheon2015silent, balogh1927voces (the key exists in references.bib; the entry below is kept so the card's pagination note travels with the facet), parkes1992pause, rayner1998unspaced, raynerpollatsek1996reading, epelboim1994reading, perea2009space, sainio2007role, veldre2017spelling, kasisopa2013eye, winskel2009eye, bai2008reading, blythe2012inserting, zang2013interword, seymour2003foundation, scribner1981psychology, scribner1978literacy, street1984literacy, street2003whats, ong1982orality, havelock1963preface, goody1986logic, bronner2018nonseparation, saenger2020orality, zhang1994representations, risko2016cognitive, butskhrikidze2022what, clanchy1993from. + +% ---- New cards (read in full unless noted) ---- + +@article{leinenger2014phonological, + author = {Leinenger, Mallorie}, + title = {Phonological coding during reading}, + journal = {Psychological Bulletin}, + year = {2014}, + volume = {140}, + number = {6}, + pages = {1534--1555}, + doi = {10.1037/a0037830}, + note = {Read in full (NIH author manuscript, PMC4211933; local copy dxd_foundation_local/pdfs/W/leinenger2014.xml). Crossref verified.}, +} + +@article{hendrickson1929ancient, + author = {Hendrickson, G. L.}, + title = {Ancient reading}, + journal = {The Classical Journal}, + year = {1929}, + volume = {25}, + number = {3}, + pages = {182--196}, + note = {No DOI; not in Crossref. Fields from the journal issue scan, Internet Archive item sim\_classical-journal\_1929-12\_25\_3 (December 1929). Read in full (local copy dxd_foundation_local/pdfs/W/hendrickson1929.txt).}, +} + +@article{balogh1927voces, + author = {Balogh, Josef}, + title = {„Voces paginarum“: Beiträge zur Geschichte des lauten Lesens und Schreibens}, + journal = {Philologus}, + year = {1927}, + volume = {82}, + pages = {84--109, 202--240}, + doi = {10.1524/phil.1927.82.14.84}, + note = {Crossref holds two records, 10.1524/phil.1927.82.14.84 (part IV) and 10.1524/phil.1927.82.14.204 (part IX, pages given as 204--242); the printed running heads in the Internet Archive scan (item philologuszeitschrift82) and the citations in Hendrickson 1929 and Knox 1968 give 84--109 and 202--240, used here. Read in part (pp. 84--92, 202--208, 226--228, 236--240; local copy dxd_foundation_local/pdfs/W/balogh1927.txt).}, +} + +@article{perea2015colours, + author = {Perea, Manuel and Tejero, Pilar and Winskel, Heather}, + title = {Can colours be used to segment words when reading?}, + journal = {Acta Psychologica}, + year = {2015}, + volume = {159}, + pages = {8--13}, + doi = {10.1016/j.actpsy.2015.05.005}, + note = {Read in full (author's copy, uv.es/mperea/coloursPTW.pdf, publisher layout; local copy dxd_foundation_local/pdfs/W/perea2015colours.pdf). Crossref verified.}, +} + +@article{kasisopa2016child, + author = {Kasisopa, Benjawan and Reilly, Ronan G. and Luksaneeyanawin, Sudaporn and Burnham, Denis}, + title = {Child readers' eye movements in reading Thai}, + journal = {Vision Research}, + year = {2016}, + volume = {123}, + pages = {8--19}, + doi = {10.1016/j.visres.2015.07.009}, + note = {Read in full (Maynooth University MURAL eprint 10442; local copy dxd_foundation_local/pdfs/W/kasisopa2016child.pdf). Crossref verified.}, +} + +@article{laubrock2015eyevoice, + author = {Laubrock, Jochen and Kliegl, Reinhold}, + title = {The eye-voice span during reading aloud}, + journal = {Frontiers in Psychology}, + year = {2015}, + volume = {6}, + pages = {1432}, + doi = {10.3389/fpsyg.2015.01432}, + note = {Read in full (PMC4585246, CC BY; local copy dxd_foundation_local/pdfs/W/laubrock2015.xml). Crossref verified; article number from the DOI.}, +} + +@article{li2014reading, + author = {Li, Xingshan and Bicknell, Klinton and Liu, Pingping and Wei, Wei and Rayner, Keith}, + title = {Reading is fundamentally similar across disparate writing systems: A systematic characterization of how words and characters influence eye movements in {Chinese} reading}, + journal = {Journal of Experimental Psychology: General}, + year = {2014}, + volume = {143}, + number = {2}, + pages = {895--913}, + doi = {10.1037/a0033580}, + note = {Read in full (NIH author manuscript, PMC3885613; local copy dxd_foundation_local/pdfs/W/li2014reading.xml). Crossref verified.}, +} + +@article{matzinger2021influence, + author = {Matzinger, Theresa and Ritt, Nikolaus and Fitch, W. Tecumseh}, + title = {The influence of different prosodic cues on word segmentation}, + journal = {Frontiers in Psychology}, + year = {2021}, + volume = {12}, + pages = {622042}, + doi = {10.3389/fpsyg.2021.622042}, + note = {Read in full (PMC8007974, CC BY; local copy dxd_foundation_local/pdfs/W/matzinger2021.xml). Crossref verified; article number from the DOI.}, +} + +@techreport{barton1980awareness, + author = {Barton, David and Hamilton, Mary E.}, + title = {Awareness of the segmental structure of {English} in adults of various literacy levels}, + institution = {Stanford University, Department of Linguistics}, + year = {1980}, + number = {ERIC ED222858}, + note = {Dated 20 December 1980; National Institute of Education grant NIE-G-80-0040. No DOI; fields from the ERIC record and the document. Read in full (files.eric.ed.gov/fulltext/ED222858.pdf; local copy dxd_foundation_local/pdfs/W/barton1980.pdf).}, +} + +@article{krieber2017eye, + author = {Krieber, Magdalena and Bartl-Pokorny, Katrin D. and Pokorny, Florian B. and Zhang, Dajie and Landerl, Karin and Körner, Christof and Pernkopf, Franz and Pock, Thomas and Einspieler, Christa and Marschik, Peter B.}, + title = {Eye movements during silent and oral reading in a regular orthography: Basic characteristics and correlations with childhood cognitive abilities and adolescent reading skills}, + journal = {PLOS ONE}, + year = {2017}, + volume = {12}, + number = {2}, + pages = {e0170986}, + doi = {10.1371/journal.pone.0170986}, + note = {Read in full (PMC5289712; local copy dxd_foundation_local/pdfs/W/krieber2017.xml). Crossref verified.}, +} + +@inproceedings{doyle2025assessment, + author = {Doyle, Adrian and McCrae, John P.}, + title = {An assessment of word separation practices in {Old Irish} text resources and a universal method for tokenising {Old Irish} text}, + editor = {Davis, Brian and Fransen, Theodorus and U{\'i} Dhonnchadha, Elaine and Walsh, Abigail}, + booktitle = {Proceedings of the 5th Celtic Language Technology Workshop (CLTW 5)}, + month = jan, + year = {2025}, + pages = {1--11}, + publisher = {International Committee on Computational Linguistics}, + address = {Abu Dhabi [Virtual Workshop]}, + url = {https://aclanthology.org/2025.cltw-1.1/}, + note = {No DOI in the Anthology record; fields from the Anthology BibTeX (fetched 2026-10-15) and the PDF footer. Publisher corrected from ACL to ICCL at S2 verification. Read in full (local copy dxd_foundation_local/pdfs/W/doyle2025.pdf).}, +} + +% ---- Abstract-only or record-only leads cited in the facet file (no cards) ---- + +@article{parkes1997raedan, + author = {Parkes, Malcolm}, + title = {R{\ae}dan, areccan smeagan: how the {Anglo-Saxons} read}, + journal = {Anglo-Saxon England}, + year = {1997}, + volume = {26}, + pages = {1--22}, + doi = {10.1017/s0263675100002088}, + note = {Abstract (Cambridge extract) only. Crossref verified.}, +} + +@article{moffat2011grammar, + author = {Moffat, Kaaren}, + title = {The ‘Grammar of Legibility’: Word separation in ogam inscriptions}, + journal = {Peritia}, + year = {2011}, + volume = {22--23}, + pages = {281--294}, + doi = {10.1484/j.perit.1.103290}, + note = {Abstract only. Crossref verified.}, +} + +@article{busch2015zusammenschreibung, + author = {Busch, Nathanael and Fleischer, J{\"u}rg}, + title = {Zusammenschreibung im Althochdeutschen und Altniederdeutschen bis ca. 975: Ergebnisse einer pal{\"a}ographischen Untersuchung}, + journal = {Beitr{\"a}ge zur Geschichte der deutschen Sprache und Literatur}, + year = {2015}, + volume = {137}, + number = {4}, + doi = {10.1515/bgsl-2015-0063}, + note = {Abstract only; Crossref gives no page range.}, +} + +@article{moran2025learning, + author = {Moran, P{\'a}draic}, + title = {Learning {Greek} in the ninth century: The origins of three {Irish} bilingual manuscripts}, + journal = {Sacris Erudiri}, + year = {2025}, + volume = {64}, + number = {1}, + pages = {103--140}, + doi = {10.1484/j.se.5.153405}, + note = {Abstract only. Crossref verified.}, +} + +@incollection{osullivan2017reading, + author = {O'Sullivan, Sin{\'e}ad}, + title = {Reading and the lemma in early medieval textual culture}, + booktitle = {Utrecht Studies in Medieval Literacy}, + publisher = {Brepols}, + year = {2017}, + pages = {371--396}, + doi = {10.1484/m.usml-eb.5.115028}, + note = {Abstract only; Crossref gives the series as container without a volume title. Crossref verified.}, +} + +@article{wilkinson1965concept, + author = {Wilkinson, Andrew}, + title = {The concept of oracy}, + journal = {Educational Review}, + year = {1965}, + volume = {17}, + number = {4}, + pages = {11--15}, + doi = {10.1080/0013191770170401a}, + note = {Abstract (Consensus extract) only; publisher closed. Crossref verified; the Crossref record carries no author given name beyond the initial, supplied here from the Consensus record (A. Wilkinson).}, +} + +@article{cushing2024social, + author = {Cushing, Ian}, + title = {Social in/justice and the deficit foundations of oracy}, + journal = {Oxford Review of Education}, + year = {2024}, + volume = {51}, + number = {3}, + pages = {396--413}, + doi = {10.1080/03054985.2024.2311134}, + note = {Not read: CC BY-NC-ND file in the Manchester Metropolitan University repository (figshare id 65058063) whose pre-signed download link expired before fetch; publisher PDF returned HTTP 403. Crossref verified (issued 2024-02-06; volume 51 is 2025).}, +} + +@article{olson1996towards, + author = {Olson, David R.}, + title = {Towards a psychology of literacy: on the relations between speech and writing}, + journal = {Cognition}, + year = {1996}, + volume = {60}, + number = {1}, + pages = {83--104}, + doi = {10.1016/0010-0277(96)00705-6}, + note = {Abstract only. Crossref verified.}, +} + +@article{homer1999literacy, + author = {Homer, Bruce and Olson, David R.}, + title = {Literacy and children's conception of words}, + journal = {Written Language \& Literacy}, + year = {1999}, + volume = {2}, + number = {1}, + pages = {113--140}, + doi = {10.1075/wll.2.1.07hom}, + note = {Abstract only. Crossref verified.}, +} + +@article{kolinsky1987awareness, + author = {Kolinsky, R{\'e}gine and Cary, Luz and Morais, Jos{\'e}}, + title = {Awareness of words as phonological entities: The role of literacy}, + journal = {Applied Psycholinguistics}, + year = {1987}, + volume = {8}, + number = {3}, + pages = {223--232}, + doi = {10.1017/s0142716400000278}, + note = {Abstract only. Crossref verified.}, +} + +@article{morais1986literacy, + author = {Morais, Jos{\'e} and Bertelson, Paul and Cary, Luz and Alegria, J{\'e}sus}, + title = {Literacy training and speech segmentation}, + journal = {Cognition}, + year = {1986}, + volume = {24}, + number = {1--2}, + pages = {45--64}, + doi = {10.1016/0010-0277(86)90004-1}, + note = {Abstract only. Crossref verified; author order from the Consensus record (Crossref record checked for the DOI, title, volume and pages).}, +} + +@article{morais1979does, + author = {Morais, Jos{\'e} and Cary, Luz and Alegria, J{\'e}sus and Bertelson, Paul}, + title = {Does awareness of speech as a sequence of phones arise spontaneously?}, + journal = {Cognition}, + year = {1979}, + volume = {7}, + number = {4}, + pages = {323--331}, + doi = {10.1016/0010-0277(79)90020-9}, + note = {Record only (cited by Barton \& Hamilton 1980). Crossref verified.}, +} + +@article{havron2016reading, + author = {Havron, Naomi and Arnon, Inbal}, + title = {Reading between the words: The effect of literacy on second language lexical segmentation}, + journal = {Applied Psycholinguistics}, + year = {2017}, + volume = {38}, + number = {1}, + pages = {127--153}, + doi = {10.1017/s0142716416000138}, + note = {Abstract only. Crossref verified (online 2016, volume 38 is 2017); second author from the Consensus record.}, +} + +@article{veldhuis2012offline, + author = {Veldhuis, Dorina and Kurvers, Jeanne}, + title = {Offline segmentation and online language processing units: The influence of literacy}, + journal = {Written Language \& Literacy}, + year = {2012}, + volume = {15}, + number = {2}, + pages = {165--184}, + doi = {10.1075/wll.15.2.03vel}, + note = {Abstract only. Crossref verified for title, journal, volume and pages; authors from the Consensus record.}, +} + +@article{mattys2005integration, + author = {Mattys, Sven L. and White, Laurence and Melhorn, James F.}, + title = {Integration of multiple speech segmentation cues: A hierarchical framework}, + journal = {Journal of Experimental Psychology: General}, + year = {2005}, + volume = {134}, + number = {4}, + pages = {477--500}, + doi = {10.1037/0096-3445.134.4.477}, + note = {Abstract only. Crossref verified for title, journal, volume, issue and pages; author given names from the Consensus record.}, +} + +@article{johnson2001word, + author = {Johnson, Elizabeth K. and Jusczyk, Peter W.}, + title = {Word segmentation by 8-month-olds: When speech cues count more than statistics}, + journal = {Journal of Memory and Language}, + year = {2001}, + volume = {44}, + number = {4}, + pages = {548--567}, + doi = {10.1006/jmla.2000.2755}, + note = {Abstract only. Crossref verified for title, journal, volume, issue and pages; second author from the Consensus record.}, +} + +@article{vorstius2014eye, + author = {Vorstius, Christian and Radach, Ralph and Lonigan, Christopher J.}, + title = {Eye movements in developing readers: A comparison of silent and oral sentence reading}, + journal = {Visual Cognition}, + year = {2014}, + volume = {22}, + number = {3--4}, + pages = {458--485}, + doi = {10.1080/13506285.2014.881445}, + note = {Abstract only. Crossref verified.}, +} + +@article{davidson2019writing, + author = {Davidson, Andrew}, + title = {Writing: the re-construction of language}, + journal = {Language Sciences}, + year = {2019}, + volume = {72}, + pages = {134--149}, + doi = {10.1016/j.langsci.2018.09.004}, + note = {Abstract only. Crossref verified.}, +} + +@article{kukkonen2024voices, + author = {Kukkonen, Karin and others}, + title = {Voices in reading literature}, + journal = {Language Sciences}, + year = {2024}, + volume = {106}, + pages = {101664}, + doi = {10.1016/j.langsci.2024.101664}, + note = {Abstract only; ScienceDirect returned HTTP 403. Crossref verified for title, journal, volume and article number; co-authors not retrieved.}, +} diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_W_spacing_oracy_literacy.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_W_spacing_oracy_literacy.md new file mode 100644 index 00000000..6e9cab13 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_W_spacing_oracy_literacy.md @@ -0,0 +1,142 @@ +# W — Word spacing, oracy and literacy + +*Facet W, round 5, 2026-10-15. Agent-drafted (Fable). The author has not reviewed it, and none of the text is the author's.* + +**The abstract sentence this facet serves.** "Seventh-century Irish scribes ... broke up scriptio continua by adding spaces between words (Tunbridge, 1992; Saenger, 1997). Unspaced text forced a reader to sound out the line just to hear where each word ended, scaffolding literacy with oracy; word separation freed literacy by moving boundary detection from the voice to the page, exploiting a representational effect (Zhang & Norman, 1994) that enabled readers to internalize text." + +**What this facet covers.** (1) Post-1997 scholarship on the Insular introduction of word separation and on the oralisation thesis, and what of it could be opened. (2) The evidence for and against the claim that unspaced text required reading aloud, from the two founding texts of the reading-aloud consensus (Balogh 1927, Hendrickson 1929) to the classicists' replies already carded in facet B. (3) Reading science on spacing and on the voice in reading: what spaces do for readers of spaced and unspaced scripts, whether boundary detection without spaces is voice work or eye work, and what the inner voice does in silent reading. (4) The oracy/literacy framing: where the word "oracy" comes from, what the literacy theorists claim, and the evidence that word boundaries are partly a literate concept. Facets B, C and D already hold the classicists' dispute, the historians of literacy and the spacing experiments; this facet reads around them and does not repeat their entries. + +## Search log + +All searches ran on 2026-10-15. "Hits" is the count the source reported. + +| # | Source | Exact query or request | Hits | Notes | +|---|---|---|---|---| +| 1 | Repo | `round5_items.json` (facet W), `ABSTRACT_v4.md`, `library/README.md`, facets B (structure), C and D (in full), cards `saenger1997space`, `tunbridge1992study`, `garrett1998saenger`, `odonnell2000saenger`, `knox1968silent`, `busch2002lautes`, `heilmann2020reading`, `winskel2009eye`, `bai2008reading`, `butskhrikidze2022what`, `scribner1981psychology`, `parkes1992pause`, `rayner1998unspaced` | n/a | Existing cards for Gavrilov, Burnyeat, Parkes 1992, Rayner 1998, Clanchy are partial; facet B lists Johnson 2000, Vatri 2012, McCutcheon 2015 as abstract-only. | +| 2 | Consensus | `word separation Insular manuscripts seventh century Irish scribes scriptio continua origin` | 20 | Bronner et al. 2018; Doyle & McCrae 2025 (opened); Moffat 2011; Busch & Fleischer 2015; Moran 2025; O'Sullivan 2017 surfaced in a later query. | +| 3 | Consensus | `illiterate adults awareness of words as units of speech segmentation literacy effect` | 20 | Morais et al. 1986, 1987, 1989; Kolinsky, Cary & Morais 1987; Havron et al. 2016; Veldhuis & Kurvers 2012; Barton & Hamilton 1980 (opened via ERIC). | +| 4 | Consensus | `phonological coding during silent reading inner speech skilled readers review` | 20 | Leinenger 2014 (opened via PMC); Abramson & Goldinger 1997; Costello et al. 2021; Brysbaert et al. 2023. | +| 5 | Consensus | `oracy concept origin Wilkinson 1965 definition spoken language competence` | 20 | Wilkinson 1965 (Crossref record; T&F closed); Cushing 2024; Stubbs 1981; Alexander 2020; Mercer et al. 2017. | +| 6 | Consensus | `Olson writing as a model for speech: literacy brings awareness of words and phonemes, world on paper` | 20 | Olson 1996 (Cognition); Homer & Olson 1999; Davidson 2019; Morais 2021. All closed. | +| 7 | Consensus | `continuous speech has no reliable pauses between words; listeners segment words using lexical and prosodic cues` | 20 | Matzinger et al. 2021 (opened via PMC); Johnson & Jusczyk 2001; Mattys et al. 2005; McQueen 1998; Saffran et al. 1996. | +| 8 | Consensus | `medieval reading aloud murmur monastic lectio Latin as foreign language word separation Parkes grammar of legibility silent reading Saenger critique` | 20 | Parkes 1997 (*Anglo-Saxon England*); O'Sullivan 2017; Saenger 2020; Coleman 1996; Kukkonen et al. 2024; Feldt 2023. | +| 9 | Consensus | `eye movements oral reading versus silent reading comparison fixation durations eye-voice span` | 20 | Laubrock & Kliegl 2015 and Krieber et al. 2017 (opened via PMC); Vorstius et al. 2014; Inhoff et al. 2011; Ashby et al. 2012. | +| 10 | Consensus | `Goody Watt consequences of literacy critique great divide orality literacy theory Ong Havelock Street ideological model` | 20 | Street 1984, 1988; Collin 2013; Maddox 2007; Serpell 2013. No open full text of Goody & Watt 1963 found. | +| 11 | WebSearch (12 queries) | Johnson 2000 pdf duke.edu; Vatri 2012 pdf; Perea & Acha 2009 pdf uv.es; Bronner 2018 pdf; Kolinsky 1987 DI-fusion; Gavrilov Hyperboreus; Parkes 1987 "grammar of legibility" pdf; McCutcheon 2015 pdf; Rayner 1998 pdf; Kasisopa 2013 Western Sydney; Veldhuis 2012 Tilburg; Cushing e-space | — | Found: Perea's publication list (uv.es/mperea), Kasisopa et al. 2016 at Maynooth MURAL, the Tilburg record of Veldhuis's 2015 thesis, the MMU record of Cushing. McCutcheon 2015 exists only as a copy on a third party's personal site; not used. Vatri 2012 only through a university proxy; not used. | +| 12 | Semantic Scholar API | `paper/DOI:` for 15 DOIs (Gavrilov, Burnyeat, Johnson 2000, Vatri, McCutcheon, Rayner 1998, Epelboim 1994, Rayner & Pollatsek 1996, Perea & Acha 2009, Sainio 2007, Veldre 2017, Kasisopa 2013, two mistyped) | 15 | Open PDFs reported only as ScienceDirect "bronze" links, which returned HTTP 403 (row 17). Title searches hit the rate limit (HTTP 429) after one query and were abandoned. | +| 13 | Crossref `/works/{doi}` and `query.bibliographic` | 32 records for every source cited below | 32 | Bib fields taken from these. Balogh 1927 has two Crossref records whose pagination (204–242 for part IX) differs from the print (202–240). | +| 14 | NCBI PMC esearch/efetch | by title and DOI: Leinenger 2014 (PMC4211933), Laubrock & Kliegl 2015 (PMC4585246), Matzinger et al. 2021 (PMC8007974), Li et al. 2014 (PMC3885613), Krieber et al. 2017 (PMC5289712); Li, Rayner & Cave 2009, Ashby et al. 2012, Havron 2016, Kukkonen 2024, Cushing 2024 | 5 of 10 | Five full texts read. | +| 15 | archive.org `advancedsearch`, `metadata`, `_djvu.txt` | *Classical Journal* 1929 (item `sim_classical-journal_1929-12_25_3`); *Philologus* 82, 1927 (item `philologuszeitschrift82`) | 2 | Both unrestricted OCR texts; Hendrickson read in full, Balogh in part. | +| 16 | ERIC API | `title:"Awareness of the Segmental Structure of English"` | 1 | ED222858, full text; read. | +| 17 | Direct fetch | people.duke.edu/~wj25/pubs/ (four reviews, no AJP 2000); De Gruyter PDF for Bronner 2018 (HTTP 202, empty); ScienceDirect PDF for Rayner 1998 (403); tandfonline PDF for Cushing (403); e-space.mmu.ac.uk (202, empty); figshare API (file found; the pre-signed S3 link expired in 10 s before fetch); uv.es/mperea/coloursPTW.pdf (opened); aclanthology.org/2025.cltw-1.1.pdf (opened); mural.maynoothuniversity.ie RR-Child-2016.pdf (opened); research.tilburguniversity.edu (thesis PDF link found, not read) | — | Nothing was fetched through a proxy, a lending copy or a shadow library. | +| 18 | CORE API, Scholar Gateway | CORE returned errors without a key; Scholar Gateway was not called this round (facets C and D report `ACCESS_DENIED`) | 0 | — | + +## Read-status legend + +- *Read in full*: whole text read this session from an open copy; a card exists. +- *Read in part*: open full text; named pages read; card marked "⚠ Not read in full" with the pages itemised (Balogh 1927 only). +- *Abstract only*: database abstract or publisher extract read; no card; listed in the last section. +- *Not accessible*: record verified; no text or abstract reached. +- *Carded earlier*: a card from an earlier round, read again this session, with an "Also used by facet W" section appended where this facet takes something new. + +Full texts for this facet are in `~/iit-playground/dxd_foundation_local/pdfs/W/`. + +## New cards (11) + +| Bib key | Source | Status | +|---|---|---| +| `leinenger2014phonological` | Leinenger (2014), *Psychological Bulletin* 140, 1534–1555 | read in full | +| `hendrickson1929ancient` | Hendrickson (1929), *Classical Journal* 25, 182–196 | read in full | +| `balogh1927voces` | Balogh (1927), *Philologus* 82, 84–109, 202–240 | read in part (pp. 84–92, 202–208, 226–228, 236–240) | +| `perea2015colours` | Perea, Tejero & Winskel (2015), *Acta Psychologica* 159, 8–13 | read in full | +| `kasisopa2016child` | Kasisopa et al. (2016), *Vision Research* 123, 8–19 | read in full | +| `laubrock2015eyevoice` | Laubrock & Kliegl (2015), *Frontiers in Psychology* 6, 1432 | read in full | +| `li2014reading` | Li, Bicknell, Liu, Wei & Rayner (2014), *JEP: General* 143, 895–913 | read in full | +| `matzinger2021influence` | Matzinger, Ritt & Fitch (2021), *Frontiers in Psychology* 12, 622042 | read in full | +| `barton1980awareness` | Barton & Hamilton (1980), ERIC ED222858 | read in full | +| `krieber2017eye` | Krieber et al. (2017), *PLOS ONE* 12, e0170986 | read in full | +| `doyle2025assessment` | Doyle & McCrae (2025), CLTW 5, 1–11 (International Committee on Computational Linguistics) | read in full | + +Appended "Also used by facet W" sections: `knox1968silent`, `busch2002lautes`, `heilmann2020reading`, `winskel2009eye`, `bai2008reading`, `butskhrikidze2022what`, `scribner1981psychology`, `rayner1998unspaced`, `tunbridge1992study`, `saenger1997space`. + +## Findings by question + +### Q1. The best current scholarship on the Insular introduction of word separation and on the oralisation thesis + +**What could be opened.** Almost none of the named post-1997 scholarship is open. Parkes's "grammar of legibility" essay (1987, in Maierù, *Grafia e interpunzione*; reprinted in *Scribes, Scripts and Readers*, 1991) and his "Rædan, areccan, smeagan: how the Anglo-Saxons read" (*Anglo-Saxon England* 26, 1997, 1–22) are behind publishers; the Cambridge extract of the latter gives its frame: the inherited *grammaticae officia*, in which *lectio* "was the process whereby a reader had to identify elements of the text — letters, syllables, words and sentences (discretio) — in order to read it aloud according to the accentuation required by the sense (pronuntiatio)" (abstract only). That sentence is the closest thing in the record to the abstract's "sound out the line", and it describes a schooled discipline, not a physiological need. Johnson 2000 (*AJP*), Johnson 2010, Vatri 2012, Gavrilov 1997 and Burnyeat 1997 remain as facet B left them: openings and footnotes only. Bronner, Busch, Fleischer & Poppe (2018), the one recent study of word division in early Irish and German manuscripts, is a De Gruyter chapter whose PDF request returned an empty page; its abstract restates the standard account ("conventionally been associated with the writing habits of Irish scribes and scholars, who learnt Latin as a foreign language and who applied to their writing practice their grammatical knowledge about word classes") and promises counts of division patterns in vernacular manuscripts. Moffat (2011) applies "the principles of Saenger and Parkes" to ogam; Busch & Fleischer (2015) find that in Old High German and Old Saxon before c. 975 "clitics were often written together with their hosts" and that scribal habit mattered more than date or dialect; Moran (2025) treats word separation as a practical problem in ninth-century Irish Greek-Latin bilinguals; O'Sullivan (2017) sets heavily glossed texts against Parkes's grammar of legibility as a different kind of reading (all abstract only). + +**What was opened.** Two things. `doyle2025assessment` (read in full) quotes Thurneysen (1946, p. 24): in Old Irish manuscripts "words which are grouped round a single chief stress and have a close syntactic connexion with each other are written as one", and shows editors still disagreeing about where the words are. With `tunbridge1992study` (pp. 89–98) this fixes the Insular move as a learner's analysis of Latin, not a native habit. And the founding texts of the oralisation thesis, `balogh1927voces` and `hendrickson1929ancient`, which were second-hand in facet B, are now read. Both place the end of voiced reading at the Renaissance and the press (Hendrickson pp. 193–194; Balogh pp. 236–237), and neither credits any change in the page before print. Balogh's one passage on *scriptio continua* (pp. 227–228) explains its difficulty by whole-word recognition and "lange Übung", and treats silent reading as the suppression of an instinct, not as the fruit of a layout. + +**Verdict for Q1.** The seventh-century Insular dating and the learner's motive stand (Tunbridge; Saenger's ch. 4 per facet A; Bronner's abstract). The oralisation thesis, in its founders' hands, dates the change to c. 1500 and gives it a different cause; its modern form (Saenger) is the one the abstract uses. No post-1997 palaeographer could be read in full this round; Parkes 1997 and Bronner 2018 are the two to get from a library. + +### Q2. Did unspaced text require reading aloud? Evidence for and against + +**For.** (a) The testimonia: Balogh's collection, Hendrickson's survey, and Busch's 2002 defence (`busch2002lautes`, pp. 8–40) establish that voiced reading was the ancient default and that silence is marked as special (*tacitus*, *sine murmure*). (b) A mechanism: Busch after Raible (p. 20) and Saenger (facet A) hold that voicing gives an unspaced text a second channel. (c) A field case: Scribner & Cole's Vai readers say syllables aloud "until they click into meaningful units" (1978 proxy, p. 456; card `scribner1981psychology`). (d) Cost data: readers of spaced scripts slow by 33–50% without spaces (`winskel2009eye` p. 344; `rayner1998unspaced` abstract) and by 69% in Perea et al. (2015, p. 10, 3,701 against 2,188 ms). + +**Against.** (a) The classicists: Knox's unspaced letters read silently in classical Athens (pp. 432–434); Heilmann's "quasi-dogma" and his reading of the laboured-reading passages as being about learners (p. 181, n. 28); Hendrickson's own "only strange, not necessarily difficult" (p. 186) and his "many gradations" (p. 193). (b) The founders' mechanism: Balogh grounds the difficulty of *scriptio continua* in practice and whole-word recognition (pp. 227–228), not in a need to hear. (c) Living unspaced scripts: 46 Chinese readers segment silently by lexical knowledge with the same word-based eye-movement signature as German readers (`li2014reading`); Chinese word-spaced and unspaced text tie (`bai2008reading`, 5,239 against 5,221 ms); Thai adults target word centres by boundary-character frequency (`kasisopa2013eye`, abstract; `kasisopa2016child`, Table 7) and Thai children do the same from Grade 6 at 88–93% comprehension, silently (`kasisopa2016child`, pp. 12–13). (d) The voice is not the detector even when it is used: in reading aloud the eye leads the voice by 16.2 letters at fixation onset and 9.7 at offset, about 561 ms (`laubrock2015eyevoice`), so words are found before they are said. (e) Speech does not mark word ends: "continuous speech lacks directly accessible cues to word boundaries" and pauses "are rare and occur only inconsistently" (`matzinger2021influence`, Introduction); listeners segment by statistics, prosody and the lexicon, and the cue that helps most in the experiments is a 250 ms pause, the auditory analogue of a space. (f) Word boundaries are partly literate: 60 American adults disagreed with written convention about where words end in proportion to their reading level (`barton1980awareness`, Tables 1–3; "number of words" errors 52.5% basic against 17.5% high). + +**Verdict for Q2.** The evidence supports "voiced reading was the norm and unspaced text was costly for the unpractised"; it does not support "unspaced text forced a reader to sound out the line just to hear where each word ended". Sounding out finds boundaries only for a reader who already has the words by ear, which is a claim about learners of Latin (Parkes 1997 abstract; Tunbridge; Heilmann n. 28), and for them the mechanism is matching a sounded string to a known form, not hearing a boundary in the signal. + +### Q3. Reading science on spacing, and whether spacing shifts segmentation from phonological to visual processing + +**Spacing.** Facet D holds the inventory. This round adds: `perea2015colours` (alternating colour recovers all but 4.7% of the cost of removing spaces; the word-frequency effect falls from 85 to 45 ms; landing sites still shift; a 3,000-word training leaves a 9 wpm gap; Thai readers unaffected by colour but slowed by spaces, pp. 10–13), which is the cleanest demonstration that the page needs a visible boundary, not space as such; `kasisopa2016child` (spaces scaffold Thai children's use of first-character cues, pp. 13–18); `li2014reading` (word-based control without spaces). `rayner1998unspaced`, Perea & Acha 2009, Epelboim 1994, Rayner & Pollatsek 1996, Sainio 2007, Veldre 2017 and Kasisopa 2013 stayed closed (abstract only). + +**Phonological or visual.** No study was found that measures inner speech or vocalisation as a function of spacing; the question as posed in the brief has no direct test in the open literature I reached. What exists cuts the other way round. Leinenger's review (`leinenger2014phonological`) shows that skilled silent readers of spaced text generate phonological codes within 80–125 ms of a word, before fixation through preview, and later as an articulatory inner voice that rises with difficulty; deaf readers use phonological codes too. The voice therefore did not leave reading when spaces arrived; what spaces remove is a segmentation task that fluent readers of unspaced scripts do visually and lexically (Li 2014; Kasisopa 2013, 2016) and that readers of spaced scripts cannot do without a cue (Perea 2015). Oral reading remains the costlier mode on a spaced page: 31% longer text time and 1.62 against 1.24 fixations per word in adolescents (`krieber2017eye`, Table 1); fixation 253 against 209 ms in young adults (`laubrock2015eyevoice`, Table 1). The learner side is where a phonological-to-visual shift is documented in general, not for spacing: Leinenger's dual-route summary (beginning readers assemble phonology for every word) and Kasisopa's Thai children moving from spaced to unspaced text. + +**Verdict for Q3.** "Moving boundary detection ... to the page" is supported as a representational claim with a measured cost; "from the voice" is not supported for fluent readers and untested for learners; "enabled readers to internalize text" is contradicted if it means text without inner speech, and unsupported if it means something else, since no source defines it. + +### Q4. The oracy/literacy framing + +**The word.** "Oracy" is Andrew Wilkinson's coinage of 1965 ("The concept of oracy", *Educational Review* 17(4), 11–15, abstract only; the Consensus extract shows the piece is a short educationalist's essay) for competence in speaking and listening, by analogy with literacy and numeracy; later usage (Stubbs 1981; Mercer et al. 2017; Alexander 2020; Cushing 2024, all abstract only) keeps that sense, and Cushing traces its 1960s "deficit" foundations. None of the sources read uses "oracy" for the voiced reading of a text; they say "reading aloud", "oral reading", "vocalised reading" (Heilmann), "lautes Lesen" (Balogh, Busch). The abstract's "scaffolding literacy with oracy" therefore uses the word in a new sense, and a reviewer from education will read it as speaking-and-listening skill. "Oralisation" (Saenger, per facet A) or "voiced reading" is what the sources mean. + +**The theory.** Ong 1982, Havelock 1963, Goody & Watt 1963, Olson 1994 and Street 1984 are books or closed articles; nothing new was opened. Facet C already records that none of them credits spacing, and that Ong ties legibility and word spacing to print. What this round adds is the empirical side of Olson's "model" thesis (writing provides the categories we use to hear speech): Barton & Hamilton 1980 on adults (read in full); Kolinsky, Cary & Morais 1987 (illiterate adults poor at judging word length), Morais et al. 1986 (phoneme awareness depends on alphabetic literacy), Homer & Olson 1999 (children's concept of word predicted by awareness of the written word), Havron et al. 2016 (literacy improves lexical segmentation in a second language), Veldhuis & Kurvers 2012 (literacy affects off-line word segmentation in Dutch children), all abstract only. The direction is uniform: the word as a segment of speech is in part learned from writing. Street's "ideological model" (facet C; `street2003whats`) stands as the caution against reading any of this as a cognitive great divide; Barton & Hamilton themselves conclude for specific skills, not a changed mind (§5). + +**Verdict for Q4.** Use "voiced reading" or "reading aloud" where the abstract says "oracy", or define the term; and note that the boundary the abstract has the voice supply is one the sources show writing taught. + +## What this means for the abstract's sentence + +| Claim in the sentence | Verdict | Why (sources) | +|---|---|---| +| "Seventh-century Irish scribes ... broke up scriptio continua by adding spaces between words" | **Supported, with a narrowing** | Tunbridge (pp. 193–216: separation in the earliest Irish manuscripts, Bangor c. 680–691, Adomnán before 713); Saenger ch. 4 (facet A); Bronner 2018 abstract. Narrowing: the same scribes wrote Old Irish in stress-groups (`doyle2025assessment`, quoting Thurneysen p. 24; Tunbridge pp. 89–92), separation was a learner's analysis of Latin, and it came as a bundle with points, hyphens and letter forms (Tunbridge pp. 165–175). Latin had used interpuncts before the second century AD (`busch2002lautes` p. 19; `heilmann2020reading` pp. 182–183). | +| "Unspaced text forced a reader to sound out the line" | **Cuts against** | Knox (pp. 432–434); Hendrickson "only strange, not necessarily difficult" (p. 186) and "gradations" (p. 193); Heilmann (pp. 178–183); Balogh locates the difficulty in practice and whole-word recognition (pp. 227–228); silent skilled reading of unspaced Chinese and Thai (`li2014reading`; `bai2008reading`; `kasisopa2013eye`; `kasisopa2016child`). Supported only as "the unpractised reader of a foreign language was helped by voicing" (Parkes 1997 abstract; Heilmann n. 28; Busch p. 25). | +| "just to hear where each word ended" | **Cuts against as mechanism; narrower form possible** | Speech does not mark word ends (`matzinger2021influence`); the eye finds words before the voice says them (`laubrock2015eyevoice`, 9.7–16.2 letters, 561 ms); word boundaries are partly a literate convention (`barton1980awareness`). Narrower form: a reader who has the words by ear can match a sounded string to a known form (Busch p. 20; Vai readers, `scribner1981psychology`), untested. | +| "scaffolding literacy with oracy" | **Narrower, and a terminology problem** | "Oracy" (Wilkinson 1965) means speaking-and-listening competence; the sources say reading aloud or vocalised reading. The scaffold the sources document is spaces for learners (Thai Grade 1, `kasisopa2016child` p. 9; Insular learners, Tunbridge; Heilmann n. 28), and inner speech for all readers (`leinenger2014phonological`). | +| "word separation freed literacy" | **Cuts against (facet C); narrower form supported** | No historian of literacy credits spacing (facet C §(c)); Balogh credits print (p. 236); spacing is standard c. 1050 and English male literacy is about 20% in 1558 (facet C). Narrower form: separation removed a segmentation cost for readers of Latin as a foreign language, and for any reader of a spaced script deprived of the cue (33–69%: `winskel2009eye`; `perea2015colours`). | +| "by moving boundary detection from the voice to the page" | **"to the page" supported; "from the voice" cuts against for fluent readers, untested for learners** | Page: `perea2015colours` (any salient mark does the job; 4.7% residual cost for colour); `bai2008reading` Exp. 2 highlighting; `kasisopa2016child` (spaces potentiate visual cue learning). Voice: `li2014reading`, `kasisopa2013eye`, `laubrock2015eyevoice`, `leinenger2014phonological`. | +| "exploiting a representational effect (Zhang & Norman, 1994)" | **Supported as an analogy** | The space is one of several visible marks that make the word a perceptual object and relieve an inferential task (Perea 2015; Bai 2008; Li 2014's "model of word segmentation"); the word itself becomes an object of awareness through writing (Barton & Hamilton; Olson 1996 and Homer & Olson 1999, abstract only). The card `zhang1994representations` carries the construct; nothing here tests it on reading. | +| "that enabled readers to internalize text" | **Cuts against if it means reading without inner speech; undefined otherwise** | Skilled silent readers of spaced text run pre- and post-lexical phonological codes and move the articulators on hard text (`leinenger2014phonological`); silent reading predates spacing (Knox) and, for the founders of the oralisation thesis, follows print (Balogh p. 236; Hendrickson pp. 193–194). Two reading-science papers restate the dating differently: "around 1000 A.D." on Saenger's authority (`perea2015colours` p. 8), "around the 8th century" on Manguel's (1996) (`laubrock2015eyevoice`, Introduction). | + +**The most defensible rewrite the sources allow** (agent-drafted for the author to accept, revise or discard; not the author's words): + +> Seventh-century Irish scribes, reading Latin as a foreign language, began to separate its words with space (Tunbridge, 1992; Saenger, 1997). For an unpractised reader, an unspaced line was parsed by voicing it and matching the sound to words already known by ear (Parkes, 1997; Busch, 2002); separation put the boundary on the page, where any salient mark will serve (Perea, Tejero & Winskel, 2015), and so removed a segmentation task from the reader without removing the inner voice from reading (Leinenger, 2014). + +## Abstract-only and inaccessible leads + +Opened at abstract, extract or record level only; no card. + +- **Parkes, M. B. (1997).** *Rædan, areccan, smeagan*: how the Anglo-Saxons read. *Anglo-Saxon England, 26*, 1–22. doi:10.1017/s0263675100002088. Cambridge extract only (the *grammaticae officia* paragraph). The single most relevant closed source for Q1–Q2; get from a library. +- **Parkes, M. B. (1987/1991).** The contribution of Insular scribes of the seventh and eighth centuries to the "grammar of legibility". In A. Maierù (Ed.), *Grafia e interpunzione del latino nel medioevo* (pp. 15–30); reprinted in *Scribes, scripts and readers* (1991). No open copy; not in Crossref. +- **Bronner, D., Busch, N., Fleischer, J., & Poppe, E. (2018).** (Non-)separation of words in early medieval Irish and German manuscripts and the concept "word". In C. Ulbrich, A. Werth & R. Wiese (Eds.), *Empirical approaches to the phonological structure of words* (pp. 45–70). De Gruyter. doi:10.1515/9783110542899-003. Abstract only; PDF request returned empty (HTTP 202). Already in the bib as `bronner2018nonseparation`. +- **Moffat, K. (2011).** The "grammar of legibility": word separation in ogam inscriptions. *Peritia, 22–23*, 281–294. doi:10.1484/j.perit.1.103290. Abstract only. +- **Busch, N., & Fleischer, J. (2015).** Zusammenschreibung im Althochdeutschen und Altniederdeutschen bis ca. 975. *Beiträge zur Geschichte der deutschen Sprache und Literatur, 137*(4). doi:10.1515/bgsl-2015-0063. Abstract only. +- **Moran, P. (2025).** Learning Greek in the ninth century: the origins of three Irish bilingual manuscripts. *Sacris Erudiri, 64*, 103–140. doi:10.1484/j.se.5.153405. Abstract only. +- **O'Sullivan, S. (2017).** Reading and the lemma in early medieval textual culture. *Utrecht Studies in Medieval Literacy*, 371–396. Brepols. doi:10.1484/m.usml-eb.5.115028. Abstract only. +- **Johnson, W. A. (2000)**, **Vatri, A. (2012)**, **McCutcheon, R. W. (2015)**, **Gavrilov, A. K. (1997)**, **Burnyeat, M. F. (1997)**: as facet B left them (openings, footnotes, reviews). McCutcheon circulates as a PDF on a third party's personal site and Vatri through a university proxy; neither was used. Johnson's Duke page holds only reviews. +- **Wilkinson, A. (1965).** The concept of oracy. *Educational Review, 17*(4), 11–15. doi:10.1080/0013191770170401a (Crossref). T&F closed; Consensus extract read. +- **Cushing, I. (2024).** Social in/justice and the deficit foundations of oracy. *Oxford Review of Education, 51*(3), 396–413. doi:10.1080/03054985.2024.2311134. CC BY-NC-ND; the MMU repository file exists (figshare id 65058063) but its pre-signed download link expired before it could be fetched. Retry from a browser. +- **Olson, D. R. (1996).** Towards a psychology of literacy: on the relations between speech and writing. *Cognition, 60*(1), 83–104. doi:10.1016/0010-0277(96)00705-6. Abstract only. +- **Homer, B., & Olson, D. R. (1999).** Literacy and children's conception of words. *Written Language & Literacy, 2*(1), 113–140. doi:10.1075/wll.2.1.07hom. Abstract only. +- **Kolinsky, R., Cary, L., & Morais, J. (1987).** Awareness of words as phonological entities: the role of literacy. *Applied Psycholinguistics, 8*(3), 223–232. doi:10.1017/s0142716400000278. Abstract only. +- **Morais, J., Bertelson, P., Cary, L., & Alegria, J. (1986).** Literacy training and speech segmentation. *Cognition, 24*(1–2), 45–64. doi:10.1016/0010-0277(86)90004-1. Abstract only. (Morais, Cary, Alegria & Bertelson 1979, *Cognition* 7, 323–331, record only.) +- **Havron, N., & Arnon, I. (2016).** Reading between the words: the effect of literacy on second language lexical segmentation. *Applied Psycholinguistics, 38*(1), 127–153. doi:10.1017/s0142716416000138. Abstract only. +- **Veldhuis, D., & Kurvers, J. (2012).** Offline segmentation and online language processing units: the influence of literacy. *Written Language & Literacy, 15*(2), 165–184. doi:10.1075/wll.15.2.03vel. Abstract only; Veldhuis's 2015 Tilburg thesis has an open PDF (research.tilburguniversity.edu, "Effects of literacy, typology and frequency on children's language segmentation and processing units"), not read. +- **Mattys, S. L., White, L., & Melhorn, J. F. (2005).** Integration of multiple speech segmentation cues: a hierarchical framework. *Journal of Experimental Psychology: General, 134*(4), 477–500. doi:10.1037/0096-3445.134.4.477. Abstract only. +- **Johnson, E. K., & Jusczyk, P. W. (2001).** Word segmentation by 8-month-olds: when speech cues count more than statistics. *Journal of Memory and Language, 44*(4), 548–567. doi:10.1006/jmla.2000.2755. Abstract only ("Fluent speech contains few pauses between adjacent words"). +- **Vorstius, C., Radach, R., & Lonigan, C. J. (2014).** Eye movements in developing readers: a comparison of silent and oral sentence reading. *Visual Cognition, 22*(3–4), 458–485. doi:10.1080/13506285.2014.881445. Abstract only (N = 632 children). +- **Davidson, A. (2019).** Writing: the re-construction of language. *Language Sciences, 72*, 134–149; **Kukkonen, K., et al. (2024).** Voices in reading literature. *Language Sciences, 106*, 101664; **Saenger, P. (2020).** Orality and visible language. In *The Oxford Handbook of Roman Studies* (doi:10.1093/oxfordhb/9780195336948.013.112). All abstract only. +- **Rayner, Fischer & Pollatsek (1998); Perea & Acha (2009); Epelboim et al. (1994); Rayner & Pollatsek (1996); Sainio et al. (2007); Veldre et al. (2017); Kasisopa et al. (2013):** still abstract only; ScienceDirect and APA refused plain requests; no repository copies found. +- **McKitterick (1989), Ong (1982), Havelock (1963), Goody & Watt (1963), Olson (1994), Street (1984):** books or closed articles; not retried beyond facet C's search-inside record. + +## S2 adversarial verification of the new cards (2026-10-15) + +All eleven cards were re-checked against the local full texts and Crossref (`literature/audit/AUDIT_16.md`). Three findings touch this file and were fixed above: Laubrock & Kliegl's "8th century" sentence cites Manguel (1996), not Saenger; the Perea et al. historical sentence and their restatement of Rayner et al. are on p. 8, not p. 9; Doyle & McCrae's publisher is the International Committee on Computational Linguistics (per the Anthology record), not the ACL. The verdict table and the rewrite above are otherwise unchanged by the verification. One "Also used by facet W" quotation (Heilmann p. 179) was a paraphrase and has been replaced with the printed words in the host card. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_X_counter_case.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_X_counter_case.md new file mode 100644 index 00000000..68909094 --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_X_counter_case.md @@ -0,0 +1,409 @@ +# X: The counter-case against the foundation thesis + +*Corrected on 2026-10-12 against `../audit/AUDIT_03.md`: 26 factual fixes; verbatim quotation of the Nadella post cut to 26 words.* + +*Facet X of the DXD foundation paper (PHD1750-3 / HF700-1). Agent-drafted on 2026-10-12 (Sonnet 5.5, local session). The author has not reviewed it, and none of the text is the author's. Wording marked **[Agent wording]** is a proposal for the author to accept, change or reject; it is not his text.* + +## What this facet covers + +I am opposing counsel. `CLAIM.md` states four parts: a precedent (scribes moved work from reader to page), a consequence (H-cost or H-cause), a transfer (algorithmic coordination is "unspaced script"; DXD looks for the move that takes the work off the person) and a foil (Nadella's post as "the exact opposite"). I build the strongest case against each part, in ten objections, using sources I opened where I could and saying so where I could not. I do not survey the history of spacing (facets A, B), literacy history (C), reading science (D), the human-factors frame (E), parallel cases (F), Nadella's lineage (G) or candidate moves (H). Where an objection needs one of their findings I cite the finding and its holder. + +**The question the synthesis answers.** After all ten objections are heard, which version of the thesis is still standing, and what does the author have to give up or reword to keep it? + +## Legend for read status + +- **Read in full**: I opened the whole text. +- **Read in part**: I opened the named pages or sections only. +- **Abstract only**: I read a database abstract and claim nothing beyond it. +- **Relayed**: I did not open the source; I am quoting the lab's own card or an earlier facet, and I say which. A relayed quotation carries the earlier reader's check, not mine. +- **Not accessible**: the record exists; I could not open the content. The entry says where a human with library access would look. +- **Argument only**: no source supports the step; I am arguing, and I say so. + +--- + +## (a) Search log + +All rows ran on 2026-10-12. + +| # | Source | Exact query or request | Hits | Notes | +|---|---|---|---|---| +| 1 | Crossref `query.bibliographic` | 23 title strings (Jensen and Meckling; Eisenhardt; von Hippel; Street; Star and Ruhleder; David; Liebowitz and Margolis; Reason; Seaver; Dekker and Woods; Bjork; Rosenblat and Stark; Mathur; Gray; Buçinca; Saltzer and Schroeder; Parasuraman et al.; Sarter et al.; Cox et al.; Scribner and Cole; Graff; Hirschman; LaPorte and Consolini) | 3 of 23 | The first batch hit HTTP 429 after three requests. Another agent's script overwrote my helper in the shared scratchpad; I moved to a private subfolder and reran in smaller batches (rows 2 to 4). | +| 2 | Crossref, polite pool | 12 titles (agency theory; lead users; Star and Ruhleder; David; Liebowitz and Margolis; Seaver; Bjork; Rosenblat and Stark; Parasuraman et al.; Sarter et al.; Reason; Jensen and Meckling) | 9 resolved | Not found as titled: David (1985) has no Crossref record; Sarter, Woods and Billings is a handbook chapter, not resolved; Bjork and Bjork's chapter surfaced only as a PsycEXTRA record, dropped. Rosenblat and Stark resolves only to the SSRN working paper; the journal article is verified from its PDF. | +| 3 | Crossref | 11 titles (Scribner and Cole 1978; Graff; Dekker and Woods; Star and Ruhleder 1996; David; Lee and See; Hurwicz and Myerson; Bainbridge; Zuboff; Edgerton; Holmström) | 8 resolved | Edgerton, *The Shock of the Old*: Crossref returns only two book reviews; no book record. Zuboff: only a 2023 chapter and a 2019 interview. Dropped both. | +| 4 | Crossref direct DOI | 10.2307/1805621 (David); 10.1287/isre.7.1.111; 10.1007/s101110200022; Perrow; Roberts; Saltzer and Schroeder; von Hippel; Gray; Vasconcelos; Noessel | 9 of 10 | David's JSTOR DOI returned nothing. Star and Ruhleder carries a Crossref abstract; Dekker and Woods carries none. | +| 5 | Unpaywall | 16 DOIs | 8 OA | Open: Street 2003, Reason 2000, Cox 2016, Buçinca 2021, Gray 2018, Mathur 2019, Liebowitz and Margolis (SSRN), de Winter and Dodou. Closed: Seaver, Parasuraman et al., von Hippel 1986, Eisenhardt, Jensen and Meckling (journal), Scribner and Cole 1978, Star and Ruhleder, Dekker and Woods. | +| 6 | `curl` + `pdftotext` | Street 2003 (ufpr.br copy); Cox 2016 (White Rose); Buçinca (arXiv); Mathur (arXiv); Rosenblat and Stark (ijoc.org); von Hippel ch. 1, 2 (web.mit.edu); Jensen and Meckling (Illinois course copy); Saltzer and Schroeder (web.mit.edu HTML); Perseus Phaedrus 275a | 9 opened | PDFs under `dxd_foundation_local/pdfs/X/`. | +| 7 | `curl` | Reason 2000 (PMC, Europe PMC render, Europe PMC REST); Street 2003 (journal URL); Gray 2018 (ACM); Liebowitz and Margolis and Jensen and Meckling (SSRN); de Winter and Dodou (Springer, TU Delft resolver) | 0 | All returned HTML walls or reCAPTCHA pages. Not retried. | +| 8 | `curl` / `WebFetch` | Leveson, *Engineering a Safer World*: OAPEN bitstream (three paths), DOAB, MIT Press Direct chapter page, `sunnyday.mit.edu/safer-world.pdf` | 0 | OAPEN and MIT Press returned a bot page or HTTP 403; `sunnyday.mit.edu` did not respond. I use the lab's card, which holds page-verified quotations from the OAPEN text (entry `leveson2011engineering`). | +| 9 | Scholar Gateway `semanticSearch` | Did word separation in medieval manuscripts cause the spread of silent reading and literacy, or is that technological determinism? | 0 | `ACCESS_DENIED: Access denied`. Not retried. | +| 10 | Consensus `search` | `frictionless design hides the mediator; algorithmic recommender systems captivation engagement` | 20 | Seaver 2018 (abstract), Vasconcelos et al. and Liu et al. 2025 on hypertextual friction (not used). | +| 11 | Consensus `search` | `human-AI interaction friction cognitive forcing reduces overreliance on AI recommendations` | 20 | Buçinca et al. 2021 (opened by arXiv), Vasconcelos et al. 2023 (abstract), Noessel 2026 (abstract). Several 2025 and 2026 items have zero or few citations; I did not use them. | +| 12 | WebSearch | `Leveson "Engineering a Safer World" MIT Press open access pdf`; `Dekker Woods "MABA-MABA or abracadabra"…`; `Paul David "Clio and the Economics of QWERTY"…`; `Jensen Meckling 1976 … "residual loss" "bonding costs"`; two on Saenger and Wingo and the interpunct | — | Located the Illinois copy of Jensen and Meckling; confirmed Dekker and Woods's and David's bibliographic details through secondary pages only. No open copy of the Roman-abandonment passage. | +| 13 | WebFetch | `en.wikipedia.org/wiki/Scriptio_continua`; CAMWS 2015 abstract on interpuncts; PMC page for Reason; MIT Direct | partial | Wikipedia confirms only that "later on the Romans came to follow the Greek practice of scriptio continua", citing Moore 2001 and Wingo 1972 without pages. I do not cite Wikipedia. The CAMWS PDF returned compressed binary; Reason's PMC page, a reCAPTCHA. | +| 14 | OpenAlex | `search=` on interpunct decline, Saenger reviews, Irish scribes | 0 | "Rate limit exceeded. Insufficient budget" (the shared daily budget was spent by other agents). Google Books API: "Quota exceeded". Neither retried. | +| 15 | Repo reads | `RESEARCH_PLAN.md`, `CLAIM.md`, `SOURCE.md`, `literature/REUSE.md`; Part 1 (all 482 lines); Nadella capture; `part4_phygital/…/X_counter_case.md` (lines 1 to 130); Part 2 facet E (lines 54 to 63, 309 to 327); lab cards `leveson2011`, `bainbridge1983`, `parasuraman2000`, `scribner1981`, `graff1991`, `finnegan1988`, `zuboff2019`; `org_frontier/essays/scifi_interfaces_and_algorithmacy.md`; `dial_response_algorithmacy/draft.md:263,337`; `algorithmacy_ot/manuscript/PAPER.md:19-21` | — | Locators for objection 9 come from these. | + +Housekeeping: a failed `cd` made two scratch downloads land in the worktree root early in the run; I deleted both and confirmed the root listing is unchanged. I ran no git commands. + +--- + +## (b) Entries + +Bib keys are in `X_counter_case.bib`. Entries are grouped by the objection they serve (numbers match section (c)). + +### Internal documents (the target of the objections) + +`hunt2026decision` | Hunt, R. (2026-09-30). *Decision Experience Design* (Part 1). Substack. Local file `submissions/decision_experience_design/published/2026-09-30_decision-experience-design.md`. +- *Verified:* read the whole file (482 lines). Line numbers below are that file's. +- *Read status:* read in full. +- *What matters for the objections (exact words):* + - Line 118: "Written text is dyadic and asynchronous, but fundamentally passive." and "Traditional media hold still; an algorithmic intermediary does not." + - Line 22: "The cognitive demands of this environment cannot be solved through the traditional UX toolkit of legibility, visual layout, and cognitive load reduction." Line 92: "Deciding through an algorithmic system imposes cognitive demands that research on reading, visual layout, and cognitive load cannot resolve." + - Line 60: algorithmic environments "govern untrained, everyday users whose objectives conflict with the platform's bottom line." Line 64: "An algorithmic intermediary deployed by an on-demand labor platform or ad-driven network does not share its users' goals." + - Line 66: Kahneman and Klein's two conditions for expertise (a valid, predictable environment and fast feedback), which "An adaptive algorithmic intermediary systematically disrupts." Line 156: "psychometric repositories contain no validated scales indexing algorithmacy". + - Line 106: Buçinca et al.'s forcing functions "curbed uncritical acceptance far more effectively than explanatory text, even though users rated these demanding interfaces less enjoyable." (Part 1's wording; the study's result is narrower, see the Buçinca entry.) + - Line 120: Ong and Goody and Watt "posited a 'Great Divide,' arguing that the invention of writing fundamentally rewired the neurological architecture of the human mind. Brian Street (1984) dismantled this autonomous model." + - Lines 142, 144, 146, 148: DXD expertise is "counter-delegation, structural refusal, bounded operational envelopes, and backward-compatible behavioral updates"; envelopes "establish explicit bounds on algorithmic behavior rather than explaining black-box logic". + - Line 22 also calls algorithmacy "a measurable, domain-specific competence that predicts decision quality". +- *Bearing:* the source of the sharpest objections (1, 3, 7, 8, 9), because the thesis contradicts it in several places. Locators for the contradictions are in 9a to 9f. + +`nadella2026models` | Nadella, S. (2026-10-10). *Models as Insider Risks in the Super Intelligence Era*. X. +- *Verified:* read the local capture (paragraphs P1 to P21). +- *Read status:* read in full. +- *Fragments I rely on (paraphrase, with paragraph numbers):* P3, a provider's assurances do not discharge the organisation's responsibility. P5, the supply of intelligence is to be split from who holds "the authority over it". P6, non-deterministic models are to be wrapped in deterministic system design, plus human controls and procedures. P8, the enterprise already knows how to manage powerful insiders. P9, chain-of-thought transparency on its own is "not sufficient or dependable". P11, no program should be able to evade or alter the machinery that checks its permissions. P17, "Orgs" decide for themselves what a model may access or do. P19, "an authorized person" can halt a model mid-task. P20, disclosure "to those affected". P21, the best system is the one that lets the deployer trust the model least. +- *Bearing:* the foil. Its centre is structural (P6, P11, P17, P19), not interpretive. See objections 6 and 7. + +`DIAL/author texts` (no bib keys; local files, quoted with locators in 9a to 9f): `SOURCE.md` (the author's instruction and the engagement-layer sentence), `RESEARCH_PLAN.md` §3 (Nadella count 3, rulings), `dial_response_algorithmacy/sources/phd1750_topic_and_rationale.md:52-54`, `dial_response_algorithmacy/draft.md:263, 337`, `algorithmacy_ot/manuscript/PAPER.md:19-21`, `org_frontier/essays/scifi_interfaces_and_algorithmacy.md:11`, `decision_experience_design/README.md:38-52`, `decision_experience_design/part2_ethics/README.md:8`. + +### Objection 1. The analogy fails at the object + +`jensen1976theory` | Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs and ownership structure. *Journal of Financial Economics*, 3(4), 305-360. doi:10.1016/0304-405X(76)90026-X +- *Verified:* Crossref record; the Illinois course copy carries the journal running head. +- *Read status:* read in part (section 1.4, p. 308); not the rest. +- *What it argues:* a principal who delegates decisions to an agent cannot make the agent's choices match his own interest for free. "However, it is generally impossible for the principal or the agent at zero cost to ensure that the agent will make optimal decisions from the principal's viewpoint." (p. 308). The cost has three parts: monitoring by the principal, bonding by the agent, and a "residual loss" (p. 308). Footnote 9 on the same page counts "operating rules" and "compensation policies" as monitoring. +- *Bearing on objection 1:* the classic reason an interface change does not dissolve a conflict of interest. Redesign lowers the cost of monitoring or bonding; the residual loss remains whenever interests diverge. Part 1 itself says the platform's intermediary "does not share its users' goals" (line 64). **Supports** the objection; it does not refute the claim that design can shrink the loss. + +`rosenblat2016algorithmic` | Rosenblat, A., & Stark, L. (2016). Algorithmic labor and information asymmetries: A case study of Uber's drivers. *International Journal of Communication*, 10, 3758-3784. +- *Verified:* open-access PDF, running head "International Journal of Communication 10(2016), 3758–3784". No journal DOI resolved; Crossref has the SSRN working paper only. +- *Read status:* read in part (abstract; introduction to p. 3763; surge-pricing section). +- *Quotations:* "the information and power asymmetries produced by the Uber application are fundamental to its ability to structure control over its workers" (abstract, p. 3758). "the labor that Uber drivers do is shaped by the company's deployment of a variety of design decisions and information asymmetries via the application to effect a 'soft control' over workers' routines" (p. 3761). +- *Bearing:* an empirical case (nine-month study) in which the interface is the instrument of the interested party's control. This is the strongest single source for "redesign the interface" handing the remedy to the party with the conflicting interest. **Supports** objection 1 and 2. It does not say interfaces cannot be redesigned by someone else. + +`mathur2019dark` | Mathur, A., et al. (2019). Dark patterns at scale: Findings from a crawl of 11K shopping websites. *Proc. ACM HCI* 3(CSCW), Article 81. doi:10.1145/3359183 +- *Read status:* read in full (card: library/mathur2019dark.md; arXiv v2). +- *Quotation:* "Dark patterns are user interface design choices that benefit an online service by coercing, steering, or deceiving users into making unintended and potentially harmful decisions." (abstract). The study finds "1,818 dark pattern instances" on about 11,000 sites and "22 third-party entities that offer dark patterns as a turnkey solution." +- *Bearing:* in the measured population the site's own page, or a vendor hired by the operator, is the instrument of the operator's interest, and the shopper is the aggrieved party. That is a two-party setting (operator and shopper, with a third-party vendor on the supply side), not a triad with a mediator serving both; the paper measures the operator's side only. The 11.1% prevalence (1,254 sites) is a lower bound and a minority of sites, and the cognitive-bias links are unvalidated ("may be more nuanced or complex", section 6.4). **Supports** objection 1. Gray et al. (2018; `gray2018dark`, read in full: card library/gray2018dark.md) is Part 1's own source for the same point (line 40); Gray does not measure how often designers are the instrument and says designers "could easily become complicit" (abstract; pp. 9-10), which is softer than "instrument". + +The scribes' shared interest (no entry): the claim that scribes "shared the reader's interest" is **argument only**. Saenger's snippets below say the ancient elite did not want broad reading; nobody in my sources says the Insular scribes' interest and their readers' coincided, only that the readers were learners of Latin (Parkes p. 1, relayed). + +### Objection 2. Who makes the move + +`vonhippel2005democratizing` | von Hippel, E. (2005). *Democratizing Innovation*. MIT Press. doi:10.7551/mitpress/2333.001.0001 +- *Verified:* Crossref; author's open copy at web.mit.edu/evhippel. +- *Read status:* read in part (chapter 1 opening and lead-user summary; chapter 2 by search). +- *Quotation:* "Users that innovate can develop exactly what they want, rather than relying on manufacturers to act as their (often very imperfect) agents." (ch. 1, p. 1 of the book). +- *Bearing:* **cuts both ways.** It is the best published support for the Insular case as user innovation (learners of Latin altering their own copies). It is also the best published statement of what the platform case lacks: the ability to modify the artefact that mediates you. von Hippel's thesis is about users who hold the means to innovate; the author's own draft says write access to the shared mediator is "the scarce capital" and is "licensed in narrow form to the many" (`dial_response_algorithmacy/draft.md`, section 5.1). The precedent therefore predicts that the move will come from users only where users hold write access, which platform users do not. + +`saenger1997space`, `parkes1992pause` (relayed from Part 2 facet E; snippet depth, context unchecked) +- Saenger p. 11: "The notion that the greater portion of the population should be autonomous and self-motivated readers was entirely foreign to the elitist literate mentality of the ancient world." Abstract (per facet E): "the ancient world had no desire to make reading easier and swifter." +- Parkes p. 1: where Latin was "a foreign language, there was a need for conventions which made it easier to read." +- *Bearing:* the reader and the scribe were often the same learner or the same community. I found no source on who controlled scribal practice in the Insular scriptoria; that gap is real (see section d). + +### Objection 3. Technological determinism + +`street2003whats` | Street, B. V. (2003). What's "new" in New Literacy Studies? Critical approaches to literacy in theory and practice. *Current Issues in Comparative Education*, 5(2), 77-91. doi:10.52214/cice.v5i2.11369 +- *Verified:* Crossref; open-access PDF with running heads. +- *Read status:* read in full (card: library/street2003whats.md). +- *Quotations (p. 77):* "The standard view in many fields, from schooling to development programs, works from the assumption that literacy in itself--autonomously--will have effects on other social and cognitive practices." "I refer to this as an 'autonomous' model of literacy." The alternative "posits instead that literacy is a social practice, not simply a technical and neutral skill; that it is always embedded in socially constructed epistemological principles." +- *Bearing:* a literacy-studies researcher's answer (the article gives his post and names his method as ethnographic, not his discipline) to a paper that credits a page convention with ignition. The article never mentions scriptio continua, word separation or any page convention, and never mentions Ong or Goody; the step from "literacy as such" to "a layout convention" is the project's. Street does not say technology is irrelevant: he grants that outside "consolidating technologies" can unsettle local meanings, and holds only that they are ideological, not autonomous (p. 79). "Spacing ignited literacy" has the form Street calls autonomous: a feature of the technology produces social and cognitive effects "regardless of the social and economic conditions". **Supports** objection 3. It also **sits in tension with** Part 1's line 118 ("Written text ... fundamentally passive"), since Street treats literacy as practice, never neutral; Part 1's "passive" means the medium does not respond, while Street's point concerns the effects of literacy. + +`scribner1978literacy`, `graff1979literacy`, `finnegan1988literacy`: records verified (Scribner and Cole only in Crossref); **not accessible**. The lab's cards give their reported findings (effects track schooling and practice, not literacy as such; mobility did not follow literacy; oral and literate are a continuum). I treat them as leads and do not quote them. + +`saenger1997space` (relayed): the book's abstract says word separation "originated in manuscripts copied by Irish scribes in the seventh and eighth centuries but spread to the European continent only in the late tenth century." A 300-year lag between invention and continental uptake (on Saenger's date; Parkes dates Carolingian adoption to the 8th-9th century, per facet E:316) is evidence against ignition on Saenger's own account. Part 2 facet E's safe wording: "visual aids lowered the cost and widened the group for whom fluent private reading was practical" (E1, line 327 of that file). **Supports** the H-cost reading and **refutes** H-cause as the author stated it. + +### Objection 4. Survivorship and selection + +`saenger1997space` (relayed): the Romans "gave up separation by interpuncts and adopted the Greek convention of scriptura continua" (p. 12); Saenger dates the end of interpunct separation "prior to the end of the second century A.D." (p. 29). Interpuncts "were ineffective aids to the physiological process of reading when compared with the interword spaces" (p. 28). +- *Bearing:* **Supports** the objection that separation was not an obvious improvement to the people who had it. It **complicates** it too: Saenger's own reason is social (the elite did not want fast reading), and an interpunct is not a space. The step "many layout conventions came and went" has no source in my set; it is **argument only** until someone supplies a catalogue of abandoned conventions (Parkes is the likely holder; I did not open him). + +### Objection 5. The single-move framing + +`star1996steps` | Star, S. L., & Ruhleder, K. (1996). Steps toward an ecology of infrastructure. *Information Systems Research*, 7(1), 111-134. doi:10.1287/isre.7.1.111 +- *Read status:* abstract only. +- *Quotation:* "Despite high user satisfaction with the system and interface, and extensive user needs assessment, feedback, and analysis, many users experienced difficulties in signing on and use." +- *Bearing:* a case where interface quality did not decide uptake; the authors analyse "infrastructural change" as a multi-level learning problem. At abstract strength this **supports** objection 5; I cannot say what the body argues about slowness. + +`david1985clio` (not accessible; secondary citations give 75(2), 332-337) and `liebowitz1990fable` (record verified; not obtained). David is the standard source for path dependence in conventions; Liebowitz and Margolis is the standard challenge to his QWERTY example. I did not read either, so the objection on path dependence is **argument only** here, with one caution: the case it leans on (QWERTY) is itself contested. Edgerton, *The Shock of the Old* (2006): I could not verify a book record (row 3), so I drop it; Bowker and Star (1999) I did not search. + +### Objection 6 and 7. Nadella is right; the two are complementary + +`saltzer1975protection` | Saltzer, J. H., & Schroeder, M. D. (1975). The protection of information in computer systems. *Proc. IEEE*, 63(9), 1278-1308. doi:10.1109/PROC.1975.9939 +- *Read status:* read in part ("Design Principles"); the copy has no journal page numbers. Facet G holds the full treatment. +- *Quotations:* "Least privilege: Every program and every user of the system should operate using the least set of privileges necessary to complete the job." "Complete mediation: Every access to every object must be checked for authority." **"Psychological acceptability: It is essential that the human interface be designed for ease of use, so that users routinely and automatically apply the protection mechanisms correctly."** +- *Bearing:* P11 states a 1970s information-security principle (tamper-resistant enforcement) to which S&S's complete-mediation principle belongs. The same list that grounds containment contains an interface-design principle, so security engineering itself holds that containment depends on a usable human interface. **Supports objection 7** (complementarity) and **weakens objection 6's category error**. + +`leveson2011engineering` | Leveson, N. G. (2011). *Engineering a Safer World*. MIT Press. doi:10.7551/mitpress/8179.001.0001 +- *Read status:* **relayed** from the lab card (`leveson2011.md`); its adversarial pass checked each quotation against the OAPEN text. I could not reopen it. +- *Quotations (via card):* "Safety is a system property, not a component property, and must be controlled at the system level, not the component level" (p. 14). "The most basic concept in STAMP is not an event, but a constraint. Events leading to losses occur only because safety constraints were not successfully enforced" (p. 76). "Passive controls are those that maintain safety by their presence" and active controls "greatly increase" failure modes (pp. 76-77). +- *Bearing:* Leveson's passive controls (interlocks, fail-safe states) correspond to Nadella's deterministic external limits and to Part 1's "bounded operational envelopes". **Supports 6** (the tested practice is structural, not interpretive). The card records the limit: Leveson's controllers pursue the organisation's mission and she models no divergent-objective platform ("no notion of an objective that diverges from the operators'"). + +`bainbridge1983ironies` (relayed from the lab card, marked full text with page-verified quotations): "the more advanced a control system is, so the more crucial may be the contribution of the human operator" (p. 775); the remedy is for the computer to use "methods and criteria, and at a rate, which the operator can follow, even when this may not be the most efficient method technically" (p. 777). **Supports 7**: Bainbridge's answer constrains the automated system, which is Nadella's move and Part 1's line 68, and it asks for deliberate inefficiency, which cuts against "lower the load" (objection 8). + +`parasuraman2000model` (relayed): automation applies at a level to each of four functions; for decision and action automation "high levels should be implemented only for low-risk situations" (p. 293). **Supports 7**: the human-factors literature already separates what the system may do (a level of automation) from what the person must do (the stage and function allocated to them). + +`reason2000human`, `perrow1999normal`, `roberts1990some`: records verified. Reason **not accessible** (reCAPTCHA); Perrow not obtained; Roberts abstract only ("organizational processes necessary to operate safely technologically complex organizations", nuclear carriers as case). They stay leads. The claims that defence in depth and high-reliability practice endorse external controls are **argument only** in this facet beyond Leveson and Saltzer and Schroeder. + +### Objection 8. Lowering the skill needed has costs + +`bucinca2021trust` | Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think. *Proc. ACM HCI* 5(CSCW1), 188. doi:10.1145/3449287 +- *Read status:* read in full (card: library/bucinca2021trust.md; arXiv v1). +- *Quotation (abstract):* "cognitive forcing significantly reduced overreliance compared to the simple explainable AI approaches. However, there was a trade-off: people assigned the least favorable subjective ratings to the designs that reduced the overreliance the most." Benefit rose with Need for Cognition. +- *Bearing:* the study the author's own Part 1 cites (line 106). The full text gives a narrower result than the abstract suggests: forcing functions beat simple explanations only on the questions where the AI was wrong; overall performance did not rise, the two kinds of design did not differ from each other, and on the wrong-AI questions no AI condition reached the no-AI level (H1b, that forcing functions improve human+AI team performance, was not supported, p. 188:16). "Worked best" holds among the AI conditions on wrong-AI questions only. **Complicates** "change the interface until the load fell" as a general rule; **supports** objection 8. One experiment (N=199) on AI-assisted decisions, with a user as sole beneficiary; it does not test a triad. + +`cox2016design` | Cox, A. L., et al. (2016). Design frictions for mindful interactions: The case for microboundaries. CHI EA '16, 1389-1397. doi:10.1145/2851581.2892410 +- *Read status:* read in full (extended abstract, about 4,100 words; card: library/cox2016design.md). +- *Quotations (p. 1390, the Introduction, in the proceedings pagination, which the accepted text does not show; the abstract is on p. 1389):* "The standard argument is that they [frictions] should be removed from technology in order to reduce the risk of user disengagement." "The problem with this design approach is that it can also result in mindless forms of interaction that can have negative consequences." +- *Bearing:* names the design orthodoxy the thesis adopts and the position against it. A position paper, not a test. **Supports** objection 8 at argument strength. + +`seaver2018captivating` | Seaver, N. (2018). Captivating algorithms. *Journal of Material Culture*, 24(4), 421-436. doi:10.1177/1359183518820366 +- *Read status:* abstract only. +- *Abstract:* recommender makers describe their purpose as "'hooking' people – enticing them into frequent or enduring usage", with the rise of "captivation metrics – measures of user retention". +- *Bearing:* an ethnography of platform builders. **Supports** the version of objection 8 that says effortless engagement serves the platform's retention goal; the abstract does not itself test whether an easier mediator benefits users less. Zuboff (2019) I leave to the lab card, which holds three passages only; I do not cite her. + +`vasconcelos2023explanations` (read in full: card library/vasconcelos2023explanations.md; Part 1 cites it at line 102): explanations reduce overreliance when they lower "the costs of verifying the AI's prediction." The effect appeared only in the hard task, and a harder-to-check explanation did no better than no explanation (Studies 1 and 2). **Complicates** objection 8 in the author's favour: cost-lowering helps when it makes verification cheaper than doing the task. The abstract's own wording: "some of the null effects found in literature could be due in part to the explanation not sufficiently reducing the costs of verifying". + +`noessel2026deskilling` (abstract only; a 2026 paper with zero citations at retrieval): reports a "controlled design intervention that reversed [deskilling] in a navigation task" through a "Human Goes First" forcing function. An indicator that designers are already arguing for friction; not a pillar. + +`plato1925phaedrus` | Plato, *Phaedrus* 275a (Fowler tr.), read at Perseus: "this invention will produce forgetfulness in the minds of those who learn to use it ... You have invented an elixir not of memory, but of reminding; and you offer your pupils the appearance of wisdom, not true wisdom". The clause on the "appearance of wisdom" runs into 275b (card: library/plato1925phaedrus.md). The oldest form of the worry. Its subject is writing, not word spacing; the step to spacing is **argument only**. + +### Objection 9. Internal consistency + +No external sources. All locators are in the repo; they are listed with the reword in section (c), objection 9. + +### Objection 10. What an examiner would attack + +`dewinter2014fitts` (read in full: card library/dewinter2014fitts.md; the authors reject the critique's conclusion for the Fitts list, sections 3.2.1 and 4) and `dekker2002maba` (record verified; not obtained; web-search summary only, treated as unverified): the human-factors critique of static function allocation, which holds that moving work from person to machine changes the work and the person's role rather than merely subtracting it. If the web summary is right, an examiner can read "work moved from reader to artefact" as the substitution assumption the field has criticised for two decades. I cannot quote either. Part 1 cites de Winter and Dodou at line 82 and lists "function allocation" as the 1951 Fitts forecast. + +--- + +## (c) Synthesis by objection + +Each objection has four parts: the objection at full strength, the best sources, how much of the thesis it defeats, and what the author would have to concede. I say where the support is argument and not sources. Defeat is rated on `CLAIM.md`'s four parts (precedent, consequence, transfer, foil). + +### 1. The analogy fails at the object + +**The objection.** Reading is one person facing a text that does not respond. Algorithmic coordination is two people facing a mediator that adapts, reads both of them and commits outcomes neither controls. A page convention solved a perceptual segmentation problem: where does one word end. No perceptual fix addresses a counterpart with an objective of its own. The author already says so. Part 1 line 118: "Written text is dyadic and asynchronous, but fundamentally passive ... Traditional media hold still; an algorithmic intermediary does not." Line 64: the intermediary "does not share its users' goals". Add the interest asymmetry. The scribes were readers' kin: learners and teachers of Latin, with no reason to hide a word boundary from the person who needed it (argument only). The mediator's owner has a reason to hide, and in practice also owns the interface. Jensen and Meckling (p. 308) say the cost of aligning an agent to a principal is generally not zero, and a residual loss remains after monitoring and bonding (which include operating rules). Rosenblat and Stark (pp. 3758, 3761) document an interface that is the control instrument. Mathur et al. (abstract) find interface choices built to benefit the service, sold as turnkey. "Redesign the interface" therefore assigns the remedy to the party with the conflicting interest. + +A second form concerns the "trained elite". The analogy needs a class of people who can already do the thing, to whom the move extends access. Part 1 says the opposite: Kahneman and Klein's conditions for expertise are disrupted by an adaptive intermediary (line 66), the field has "no validated scales indexing algorithmacy" (line 156), and higher self-reported AI literacy goes with worse calibration (line 138, Fernandes et al.). A reader of *scriptio continua* could master a stable code with practice, and facet E records an anecdotal report that skilled modern readers read unspaced French with ease, at a 40-70% speed cost (Mirault et al., p. 22). If nobody masters the mediator, there is no unspaced-script elite for a spacing move to un-train. + +**Best sources.** `jensen1976theory`, `rosenblat2016algorithmic`, `mathur2019dark`, Part 1 lines 64, 66, 118, 138, 156. The point about the scribes' shared interest is argument only. + +**Damage.** To the transfer, serious. It does not touch the precedent (spacing did lower a cost) and not much of the foil. It leaves "an artefact-side fix" possible, but not an artefact the interested party designs. It also forces a choice about which artefact: the platform's interface (compromised by interest) or one the user side or a third party controls. + +**What the author must concede or change.** (a) Drop any suggestion that the platform's UX layer is where the move happens, or say who else builds it. (b) Say what in the spacing story the algorithmacy case reproduces: a public convention that no single party owns. Spacing belonged to no one. **[Agent wording]**: "Word separation was a convention that belonged to no one scribe. The algorithmic case needs a convention the intermediary's owner cannot unilaterally withdraw." This keeps the analogy at the level of where the work sits and who governs the convention, and it points to what Part 1 already names: counter-delegation (line 144, the user's own agent) and bounded envelopes (line 148). (c) Stop saying the training requirement disappears; say instead that it has not been shown that any training works, which is Part 1's own gap (line 156). + +### 2. Who makes the move + +**The objection.** The conventions were designed for readers who learned Latin as a foreign language (Parkes p. 1, relayed); whether the scribes were those readers is not shown. A design discipline works inside firms that own the interface. The history of a user-driven convention is weak precedent for what a firm's designers will do. von Hippel argues that users who innovate can build what they want instead of relying on manufacturers as their agents (ch. 1, p. 1). + +**Best sources.** `vonhippel2005democratizing` (ch. 1, p. 1), `rosenblat2016algorithmic` (p. 3761), `saenger1997space` (p. 11, relayed), the author's own draft (section 5.1, write access as "scarce capital"). I found no source on who governed scribal practice in the Insular scriptoria; that part is argument only. + +**Damage.** To the transfer, moderate; to the precedent, none. It reinforces objection 1 and changes the question from "what is the move" to "who may make it". + +**What the author must concede.** The precedent supports user-side and commons-side moves (counter-delegation, standards), not firm-side UX. Part 1's remedy list (lines 142 to 148) already points the same way. The foundation paper should say that the discipline's client may not be the platform. Whether that is still "design" in the HF sense is for the author. + +### 3. Technological determinism + +**The objection.** "The critical UX move that ignited literacy" is the autonomous model: a feature of a technology produces a social and cognitive outcome regardless of the conditions that surround it. Street (2003, p. 77) names it and rejects it; Part 1 invokes Street to dismantle Ong and Goody and Watt (line 120; Street 2003 itself never mentions Ong or Goody, so that target is Part 1's attribution, to Street 1984) and warns that claims about algorithmic interfaces rewiring consciousness "fall into the same trap". A sociologist of literacy reading an HCI paper that credits spacing would ask who taught whom, who was permitted to read, which institutions required reading, and why uptake took three centuries on the Continent. Saenger's abstract (relayed) says separation reached the Continent "only in the late tenth century" (Saenger's date; Parkes dates Carolingian adoption to the 8th-9th century, per facet E:316). His explanation for why the ancient world did not separate words is social (the elite wanted no broad reading, p. 11); I have no source on his explanation of the continental lag. + +**Best sources.** `street2003whats` (p. 77), Part 1 line 120, `saenger1997space` (abstract and p. 11, relayed). Scribner and Cole, Graff and Finnegan are the standard authorities but I could not open them. + +**Damage.** To "ignited" and to H-cause, near-total. To H-cost, none. This is the cheapest objection to answer and the most damaging to the wording as it stands. `CLAIM.md`'s "Do not claim" list already includes: "That word separation created silent reading, unless the research supports it." And facets A to C hold the evidence. + +**What the author must concede.** Retire "ignited". **[Agent wording]**: "Word separation lowered the cost of fluent reading of Latin and widened the group who could do it without long training; other conditions determined whether anyone did." That is facet E's safe wording (E1) restated for this paper. The author's phrase "critical UX move" needs the same treatment. "Critical" is a judgement the sources do not make. + +### 4. Survivorship and selection + +**The objection.** The paper picks the one layout convention that persisted and calls it critical. Hindsight chooses it. And the Romans had word dividers and gave them up (Saenger p. 12, p. 29, relayed), so the move was not an obvious improvement to the people who had it. + +**Best sources.** `saenger1997space` (relayed). The wider claim that many conventions came and went is **argument only**; Parkes describes a complex of conventions (RESEARCH_PLAN §4.4, paraphrase), which is the likely place to document it. + +**Damage.** To "obvious" and "critical", moderate. To the thesis as an analogy, small: the analogy needs one documented case of work moving to the page, not a prediction that such moves always win. + +**What the author must concede.** Drop any wording that implies inevitability. Use the Roman reversal as an asset: it shows the move is a choice whose adoption depended on who wanted reading widened, which feeds objections 1 and 2. The author should not say "the Romans did not need it". + +### 5. The single-move framing is a designer's heroic story + +**The objection.** Infrastructures and conventions are slow accretions. Star and Ruhleder (abstract) report a custom system for geneticists with high user satisfaction in which many users still "experienced difficulties in signing on and use", and analyse the change as multi-level infrastructural learning. David (1985) is, on my recollection, the standard account of how conventions lock in by accumulated, interlocking choices (argument only, not read). The "killer feature" is a retrospective story. Part 1 line 118 says "Literacy, numeracy, and oracy are competencies for engaging a medium; algorithmacy is a competence for engaging a coordination", which already breaks the one-move shape. + +**Best sources.** `star1996steps` (abstract only), `david1985clio` (not accessible), Part 1 line 118. Liebowitz and Margolis (1990) warn that the QWERTY lock-in story is contested; I could not read either side. Edgerton and Bowker and Star: not verified or not searched. + +**Damage.** To "a similar move", moderate; to the analogy, small. RESEARCH_PLAN §4.4 already flags the shape. + +**What the author must concede.** Say "a complex of conventions" or "a set of moves", not "a move". **[Agent wording]**: "We need conventions, not a feature." + +### 6. Nadella is right, and the paper misreads him + +**The objection.** For agentic systems with real authority, deterministic external controls, least privilege, audit and a kill switch are the tested practice of safety-critical engineering. Nadella's P6, P11, P17 and P19 are that practice: deterministic system design (P6), controls that a program cannot bypass or tamper with (P11), organisations determining independently what a model can access (P17), and a power to pause or shut a model down mid-task (P19). Leveson (via the lab card, p. 76) puts safety in enforced constraints; Saltzer and Schroeder name least privilege and complete mediation. An "engagement layer where the user reads and rewrites the shared mediator" (RESEARCH_PLAN §3) is, from that tradition, a write path into a controlled component: what security engineering forbids unless the writer is authorised, scoped, logged and reversible. And the post is about enterprise insider risk (P8, the enterprise; P17, organisations). Answering it with an interface-design argument looks like a category error, the more so because the author did not select the "enterprise, not user" reading when he listed what he opposes (`SOURCE.md`, answers table). + +**Best sources.** `nadella2026models`, `saltzer1975protection`, `leveson2011engineering` (relayed). Reason, Perrow and Roberts are leads only. + +**Damage.** To the foil, serious if it stands. To the thesis's core (precedent and transfer), none: the objection concerns the target, not the argument. + +**Is there a principled reply?** Yes, in three parts. (1) The author's two claims sit at different layers. Containment bounds what a model may do to the organisation. The engagement layer is about a person's standing against a platform's mediator that the person did not choose, which is Part 1's setting (line 60, 64). Nothing in P1 to P21 gives the person on the other side of a platform's mediator any authority; P20 names "those affected" only as recipients of disclosure. (2) The reply must also grant the security point: write access to a shared mediator is a privilege and needs least privilege, audit and rollback. The author's own draft already treats write access as licensed (section 5.1). Counter-delegation (Part 1 line 144) is the safer form, because the user's agent acts on the user's side instead of rewriting the shared component. (3) Saltzer and Schroeder's "psychological acceptability" principle answers the charge that interface argument is out of place: security engineering writes the human interface into its own principles. + +**What the author must concede.** The paper cannot say Nadella is wrong about containment; he is right inside his scope. It must say whose authority his scope covers. It also has to choose between "the user rewrites the shared mediator" and "the user's agent bargains against it", because only the second survives the objection and Part 1 already holds it. + +### 7. The two are complementary, not opposite + +**The objection.** Containment governs what the system may do; interface design governs what the person must do. Calling Nadella "the exact opposite" (`SOURCE.md`) is rhetorical excess, and the author's own Part 1 shows why. Part 1 says the remedy is not "to demand that the user study harder or read more complex displays" but to constrain the system so that its "parameters, criteria, and cadences" are ones "human perception can track" (line 68), and lists "bounded operational envelopes" and "structural refusal" (lines 146, 148). Those are containment. Nadella's P9 says chain-of-thought transparency on its own is "not sufficient or dependable", which is Part 1's line 136 to 140 in other words. His containment principle (P19, the emergency brake) is a containment control, close to Part 1's bounded envelopes; it differs from structural refusal in who holds it. P3 and P20 name a provider and affected people, and P14 adds several models, so "keeps the dyad" is also overstated. + +The counts as set out in `RESEARCH_PLAN.md` §3 are weakened one by one: + +| Author's count | What the post says | Verdict | +|---|---|---| +| Legibility paradigm: "observe, log, audit" | P6 puts the weight on deterministic system design, P11 and P17 on external control, P19 on a brake; P15 does ask for "human readable evidence" | Partly right. The post has a legibility strand (P15) inside a structural frame (P6, P11, P17, P19). It is not at its core a text that "answers opacity with more reading" (RESEARCH_PLAN §3). | +| Keeps the dyad | P3 (provider), P20 (people affected), P14 (several models) | Overstated. The post's agent is a model; the parties named include a provider and affected people. | +| Containment versus engagement | P5 splits who supplies intelligence from who holds "the authority over it" | The real difference, and it is about who holds authority, not about opposite remedies. | + +**Best sources.** `nadella2026models`; Part 1 lines 68, 136-148; `parasuraman2000model` and `bainbridge1983ironies` (relayed); `saltzer1975protection`. + +**Reading on which both hold.** Containment is the floor: nothing the engagement layer does may widen the model's authority, and the organisation keeps the brake. The engagement layer is what the person facing the mediator can do inside that floor. Parasuraman et al. (pp. 287-288, relayed) separate the level of automation from the function allocated to the person. On this reading the author's remedy list and Nadella's principles compose. + +**Damage.** To the foil, severe as phrased; the objection leaves the thesis intact in a weaker form. + +**Wording the author could defend.** **[Agent wording]**: "Nadella answers who controls a model inside an organisation. Algorithmacy asks what the person on the other side of a platform's mediator can do. The first needs deterministic containment. The second needs more than containment, because the person is not the organisation." If the author wants to keep an opposition, the defensible one is on authority (P5 against Part 1's user standing), not on remedy. + +### 8. Lowering the skill needed has costs + +**The objection.** Easier is not better. Bainbridge (via the lab card, pp. 775, 777) argues that automation raises the importance of the operator and that the remedy is deliberate inefficiency, a process the operator can follow "even when this may not be the most efficient". Buçinca et al. find that forcing functions beat simple explanations only on the questions where the AI was wrong, with no overall gain (H1b unsupported, p. 188:16), and users liked them least. Cox et al. (p. 1390) name the "standard argument" that frictions should go and argue it breeds "mindless forms of interaction". Plato's oldest version (*Phaedrus* 275a) says a convenience of reminding gives "the appearance of wisdom, not true wisdom". And the lab's own note says it: abstraction "lowers load while also hiding the mediator's agenda" (`org_frontier/essays/scifi_interfaces_and_algorithmacy.md:11`). The author's draft goes further and argues against abstracting the steering surface from a party "already steering" (`dial_response_algorithmacy/draft.md:337`), because it "hides her own contribution from her while keeping that contribution available to the operator". Seaver's abstract describes builders who want to hook people. A seamless mediator may serve retention more than the person. + +**Best sources.** `bucinca2021trust`, `cox2016design`, `bainbridge1983ironies` (relayed), `seaver2018captivating` (abstract), `plato1925phaedrus`, the essay line and the draft. In the favour of the author: `vasconcelos2023explanations` (abstract) shows that when an explanation really cuts the cost of checking, overreliance falls. + +**Damage.** To the transfer, serious in its strongest form: "take the work off the person" cannot be a general principle. It does not touch the precedent. + +**The dilemma it creates.** Spacing makes the text's own structure explicit and removes nothing. Read that way, it is a legibility device (Parkes's "grammar of legibility"; RESEARCH_PLAN §4.2), which the first ruling rejects. Read the other way, as lowering load without exposing the mechanism, it is the abstraction that the essay (line 11) and the draft (line 337) say hides the mediator. The analogy has one reading that breaks the first ruling and one that breaks the second. + +**What the author must concede.** The work to move is the work that does not teach the person the mediator's stakes. **[Agent wording]**: "Move the work of finding the joints; keep the work of deciding." The moved work must be work whose removal leaves the person's judgment and their view of the counterpart's interests intact. This is also what Bainbridge's remedy does (hold the system to a pace the person can follow). + +### 9. Internal consistency + +I list each inconsistency with its locators and the smallest reword that removes it. All rewordings are **[Agent wording]**; the author decides, and where his text is involved I flag it and leave his words in place. + +**9a. The model is a legibility innovation; the rulings reject legibility.** +- Rulings: `decision_experience_design/README.md:38-43` ("Algorithmacy is not legibility"); Part 1 line 22 (the problem "cannot be solved through the traditional UX toolkit of legibility, visual layout, and cognitive load reduction") and line 92 ("research on reading, visual layout, and cognitive load cannot resolve" it). +- The model: word separation, which Parkes calls the "grammar of legibility" (RESEARCH_PLAN §4.2); `draft.md:263` and `:337`; `CLAIM.md` precedent. +- Why it matters: a visual-layout, load-reducing innovation from the history of reading is exactly what Part 1 says cannot solve the case. Lines 56 to 64 say the same of Ecological Interface Design, which also puts a constraint into a perceptual form. +- Smallest reword: say the precedent is about *where work is placed*, and drop "the interface becomes legible". **[Agent wording]**: "The precedent shows that work can be moved from the person to the artefact. It does not show that a more readable display solves the algorithmic case." + +**9b. "Ignited" and the ruling that the history is an analogy.** +- `part2_ethics/README.md:8` (history is "an analogy, not a mechanism"); `decision_experience_design/README.md:45-52` (no Ong-style restructuring); Part 1 line 120 (Street "dismantled" the autonomous model); `SOURCE.md` ("ignited literacy"); `CLAIM.md` (H-cause). `RESEARCH_PLAN.md` §3 records the author's answer "Let the research decide", so the contradiction is open by design. +- Smallest reword: H-cost wording, as in objection 3. If the author keeps "ignited", it needs a source that supports a causal claim; none of my sources do. + +**9c. Part 1 says the remedy is a structure; `CLAIM.md` says it is a move on the reading side.** +- `CLAIM.md` transfer: "coordination ... is today in the condition of unspaced script: only the trained can do it." Part 1 lines 66, 138, 156 give no evidence that any trained group can do it. +- Smallest reword: **[Agent wording]**: "Coordination through an algorithmic third party is a task nobody has shown how to do reliably." + +**9d. "A sensibility" is a property of persons; the remedy is placed in artefacts.** +- `algorithmacy_ot/manuscript/PAPER.md:19-21`: "the sensibility to navigate opaque triadic coordinative cooptation", with competencies to interpret, specify and keep track; Part 1 line 22 calls it "a measurable, domain-specific competence". The remedy (spacing, bounded envelopes, structural refusal) is built into artefacts and structures. +- Smallest reword: **[Agent wording]**: "Algorithmacy is the person's sensibility. DXD designs the conditions under which a person without it can still act well." Part 1 line 150 already makes the matching move for designers. + +**9e. The engagement layer asks the person to read; the first ruling says not to ask the person to read.** +- `SOURCE.md` (quoting `phd1750_topic_and_rationale.md:52-54`): "where reading and rewriting the shared mediator is the fundamental mode"; `RESEARCH_PLAN.md` §3, Nadella count 3: "engagement layer where the user reads and rewrites the shared mediator". The rulings: `README.md:38-43`. +- This turns the author's first count against Nadella ("he answers opacity with more reading") on his own proposal. +- Smallest reword: this is the author's text, and I leave it. If he wants to keep both, the count against Nadella has to move to authority (objection 7), because "reads" appears on both sides. + +**9f. Abstraction lowers load but hides the mediator; the draft uses word separation both ways.** +- `org_frontier/essays/scifi_interfaces_and_algorithmacy.md:11` (abstraction "lowers load while also hiding the mediator's agenda"); `draft.md:263` (word separation changed "the interface so that reading no longer required the training") and `draft.md:337` (abstracting the steering surface "hides her own contribution from her"; word separation is "again the precedent"). The two sentences cannot both treat word separation as the model for load-lowering and the model against abstraction without the dilemma stated in objection 8. +- Related flag: `draft.md:337` credits Inman and Ribes (2019) with the point about what the precedent redistributes; `literature/REUSE.md` warns their paper is about seamful design and the attribution needs checking. I did not check it. + +**9g. The "enterprise, not user" reading.** `SOURCE.md` answers table: the author did not select it. The post's text points to the enterprise throughout (P8, P9, P17). Not an inconsistency in his text, but the research should tell him the text favours the reading he declined. + +**9h. Smaller items.** Part 1 line 120 (Ong and Goody and Watt "rewired the neurological architecture"; Street 2003 does not mention Ong or Goody) has no audit (`REUSE.md`) and sits beside the ruling against restructuring claims. The author writes "scriptura continua"; the scholarly term is *scriptio continua* (RESEARCH_PLAN §2). + +### 10. What a human-factors examiner would attack first + +An argument, not a source set. In order: + +1. **"Ignited literacy."** A causal claim in the title position with no evidence. Street (2003, p. 77) and Part 1's own line 120 supply the rebuttal. +2. **What is "the move", and what would show it wrong?** No operational definition and no disconfirming case. Part 1 does state falsifiable propositions (lines 158 to 166); this paper does not yet. +3. **Function allocation.** "Work moved from reader to artefact" is a function-allocation claim. If the Dekker and Woods critique is as the search summary says (unverified), the examiner will ask what the paper does about work that changes character rather than disappearing. `dewinter2014fitts` is in Part 1 already; the author should read both. +4. **The analogy's dyad.** The HF literature Part 1 reviews is dyadic (line 112). The history of reading is dyadic. The examiner will ask what is human-factors about the triadic step; Part 1's answer is the unit of analysis, but the move proposed here (objections 1 and 2) leaves the human-factors discipline. +5. **Novelty.** Facet Z has not reported. I claim nothing. +6. **Provenance and citation checks.** Part 1 opens with an AI-authorship disclosure (line 15). The Gemini essays are leads, not sources (RESEARCH_PLAN §6). An examiner who checks one citation will check more. + +--- + +### Ranking by damage, with the answer to each of the top five + +Damage is the harm that remains after the cheapest repair. + +| Rank | Objection | Part of the thesis it hits | Damage after cheapest repair | +|---|---|---|---| +| 1 | 1. Analogy fails at the object (interested third party; no elite to un-train) | Transfer | High. The remaining problem is who builds the artefact. | +| 2 | 9. Internal consistency (esp. 9a, 9e, 9f) | Precedent, transfer, foil | High until reworded; the repair is wording plus one decision by the author. | +| 3 | 7. Complementary, not opposite | Foil | High on "exact opposite"; low once the contrast moves to authority. | +| 4 | 3. Technological determinism ("ignited") | Consequence | Total on H-cause; none on H-cost. The cheapest repair. | +| 5 | 8. Lowering the skill has costs | Transfer | Medium. It removes "take the work off the person" as a principle. | +| 6 | 6. Nadella is right | Foil | Medium. A reply exists (scope plus least privilege). | +| 7 | 2. Who makes the move | Transfer | Medium; folds into 1. | +| 8 | 10. Examiner attacks | All | Medium; most are answered by the other repairs. | +| 9 | 5. Single-move framing | Wording | Low. | +| 10 | 4. Survivorship | Wording | Low. | + +**Concessions and rewordings for the top five. All [Agent wording].** + +1. *Objection 1.* Concede that the platform's UX layer cannot carry the move. "Word separation was a convention that belonged to no one scribe. The algorithmic case needs a convention the intermediary's owner cannot unilaterally withdraw." Say who builds it: the user's own agent (Part 1 line 144), a public standard, or a regulator. Also concede that no one has shown a trained elite exists (line 156). +2. *Objection 9.* Choose one of two roads and say so. Road A: the precedent is about placing work, and the paper makes no claim that a readable display solves the case (9a). Road B: the paper accepts a legibility strand and revises the ruling. Either way, take "reads" out of the engagement-layer description or move the count against Nadella to authority (9e); and settle the abstraction dilemma (9f). +3. *Objection 7.* Replace "exact opposite" with the contrast on authority (wording under objection 7). Acknowledge P6, P11, P17 and P19 as the containment floor and Part 1's bounded envelopes as the same kind of control. +4. *Objection 3.* Replace "ignited" and "critical" with the H-cost wording (objection 3). Cite Street (2003, p. 77) for why. +5. *Objection 8.* Replace "take the work off the person" with "move the work of finding the joints; keep the work of deciding" and state the condition (the moved work must not hide the counterpart's stakes from the person). + +### What survives + +Stated plainly, in terms the evidence in hand supports. + +A page convention once moved a burden from the reader to the page. That lowered the cost of reading Latin fluently for people who were not long-trained, and it spread over several centuries, on dates that Saenger and Parkes assign differently. It is a fair example of a design principle in human factors: work can be placed on the artefact so the person need not do it. It does not show that design caused mass literacy, and it does not show that any one move is "the" move. + +Coordination through an adaptive, interested intermediary lacks such a convention, and nothing in the sources says one lives on the platform's interface. The candidates are conventions the platform's owner cannot withdraw alone: the user's own agent, bounded and predictable behaviour from the mediator, a public standard, or exit. The work to move is the work of finding the joints, not the work of deciding, and the move has to leave the person able to see what the counterpart gains. + +Containment, as Nadella describes it, is not the opposite of this. It is the floor under it. The two differ in who holds authority: the organisation over a model, or the person over a platform's mediator. The paper can defend that difference. + +### Claims the sources contradict + +1. "Written text is ... fundamentally passive" (Part 1 line 118) sits in tension with Street 2003 p. 77: literacy is "a social practice, not simply a technical and neutral skill". The two are different claims: Part 1 means the medium does not respond; Street speaks of literacy's effects. +2. "Ignited literacy" against Saenger's abstract (relayed): separation reached the Continent "only in the late tenth century", and against Street p. 77. +3. "Only the trained can do it" (`CLAIM.md` transfer) against Part 1 lines 66, 138 and 156. +4. "The exact opposite" (`SOURCE.md`) against Part 1 lines 68, 146, 148 and Nadella P6, P11, P17, P19. +5. "Keeps the dyad" against Nadella P3, P14 and P20. +6. "Lower the load" as a general design rule against Buçinca et al. (narrower than the abstract: wrong-AI questions only; card) and Bainbridge p. 777 (relayed). +7. "Answers opacity with more reading" against Nadella P9, which says chain-of-thought transparency on its own is "not sufficient or dependable". P15 does ask for human-readable evidence. + +--- + +## (d) Gaps and unverified leads + +- **Who governed Insular scribal practice** (objection 2). No source obtained. A human with library access should look in Parkes, *Pause and Effect*, chapter 3, and in Saenger, chapters on Insular scribes (pp. 34, 90 to 97 per facet E), and in Tunbridge's 1992 Oxford DPhil. Facet A holds the leads. +- **A catalogue of layout conventions that failed** (objection 4). Not found. Parkes is the likely holder; I did not open him. +- **Scribner and Cole (1981; and the 1978 article), Graff (1979), Finnegan (1988)** (objection 3). Records verified or carded; not opened. Scribner and Cole 1978 is in *Harvard Educational Review* 48(4); no open copy. Graff's open edition is on the WAC Clearinghouse site behind a Cloudflare challenge. Finnegan: library copy. +- **Leveson (2011)** is relayed from the lab card because the OAPEN file was walled. Anyone reopening it should check pp. 11-14, 63-67, 76-77. +- **Reason (2000), Perrow (1984/1999)**, and the high-reliability literature (Roberts 1990 abstract only; Weick and Sutcliffe not searched). The claim that defence in depth and HRO practice endorse external controls rests on Leveson and Saltzer and Schroeder alone. +- **David (1985), Liebowitz and Margolis (1990), Edgerton (2006), Bowker and Star (1999).** Not obtained, no book record found, or not searched. Objection 5 is thin on sources. +- **Dekker and Woods (2002), de Winter and Dodou (2014).** De Winter and Dodou has since been read in full (card); I could not open Dekker and Woods. Objection 10's point on function allocation depends on both; the Dekker and Woods description is from a web-search summary and unverified. +- **Seaver (2018)** abstract only; **Zuboff (2019)** not used. For objection 8 the body of Seaver is the next read. +- **The lab's own essay `org_frontier/essays/the_interested_third_party.md`** is named in `scifi_interfaces_and_algorithmacy.md` and bears on objection 1; I did not read it. +- **Gray et al. (2018)** and **Vasconcelos et al. (2023)**: both since read in full (cards). +- **Hirschman (1970), exit and voice**: the lab holds a local PDF (earlier Part 4 facet X, row 27). I did not use it; it bears on exit as the structural-refusal analogue in objections 1 and 7. +- **Facet Z (novelty)** has not reported. Nothing here is a claim of novelty or its absence. +- **Facets A, B, C, D, E, F, G, H**: where I cite their territory I rely on the existing Part 2 facet E and the lab cards, not on this run's other agents, whose files I did not read. + +## Changes after full-text cards (2026-10-13) + +- `bucinca2021trust`: result narrowed (forcing functions beat simple explanations only on wrong-AI questions; no overall gain; the two designs did not differ; no AI condition reached the no-AI level; H1b unsupported, p. 188:16); fixed in the entry, in objection 8, in the list of sources the claims contradict (item 6), and flagged beside Part 1's line 106 quotation; read status "read in full (card)". +- `street2003whats`: added that the article never mentions scriptio continua, word separation or any page convention, and never mentions Ong or Goody (the Ong and Goody target is Part 1's attribution to Street 1984), in the entry, objection 3 and item 9h; "a sociologist's answer" replaced; Street does not call technology irrelevant (p. 79); read status "read in full (card)". +- `mathur2019dark`: marked as a two-party setting (operator and shopper, vendor on the supply side), not a triad; 11.1% is a lower bound; bias links unvalidated (section 6.4); read status "read in full (card)". +- `gray2018dark`: "designer as the interested party's instrument" qualified: the paper says designers "could easily become complicit" (abstract; pp. 9-10) and does not measure how often; "not opened" updated to "read in full (card)" (here and in the open-items list). +- `vasconcelos2023explanations`: effect only in the hard task; a harder-to-check explanation did no better than none; read status "read in full (card)". +- `cox2016design`: p. 1390 marked as proceedings pagination not shown in the accepted text; `plato1925phaedrus`: the "appearance of wisdom" clause runs into 275b; `dewinter2014fitts`: read in full, authors reject the critique's conclusion for the Fitts list (3.2.1, 4); `hunt2026decision`: line 40 claim confirmed, no change. + +Synthesis verdicts changed: none to the objection verdicts. The wording of two bearing verdicts changed: Buçinca "shows the design that demanded more of the user did better" became "better only on wrong-AI questions, no overall gain"; Mathur "the designer ... often the interested party's instrument" became a two-party, operator-side finding. diff --git a/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_Z_gap_search.md b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_Z_gap_search.md new file mode 100644 index 00000000..0ee40cba --- /dev/null +++ b/submissions/berthon_ethics_paper/nadella/sources/dxf_facet_Z_gap_search.md @@ -0,0 +1,453 @@ +# Z: Gap search + +*Corrected on 2026-10-12 against `../audit/AUDIT_03.md`: 30 factual fixes.* + +*Facet Z, 2026-10-12. Agent-drafted. The author has not reviewed it, and none of the text is the author's.* + +## What this facet covers + +A documented novelty check on one claim: the scribes' move from *scriptio continua* to word separation is the design precedent for what coordination through algorithmic intermediaries now needs, and a discipline called Decision Experience Design (DXD) should look for that move. The facet asks who has already said something like this, in seven parts: (1) the analogy applied to AI, algorithms, HCI, code or explainability; (2) the history of the page treated as the origin of UX; (3) "incunabula" or pre-convention arguments about AI interfaces; (4) the term "decision experience design" and its variants; (5) the term "algorithmacy" and the critiques of "algorithmic literacy"; (6) responses to the Nadella post of 10 October 2026; (7) scholarship arguing that oversight, observability or audit-log remedies reproduce a reading model. + +## The question the synthesis answers + +What exact sentence of the form "I found no work that does X in [sources]" can the paper print, and which predecessors must it cite to avoid looking unaware? + +## Legend for read status + +- **read in full**: whole text read this session. +- **read in part**: sections named. +- **abstract only**: database abstract or publisher blurb; paraphrase only. +- **summary only**: I received a model-written summary of the page (WebFetch or WebSearch) and not the raw text. Quotations marked "via fetch" came out of that summary and must be checked against the page before use. +- **not accessible**: blocked or not opened. + +All searches ran on **2026-10-12**. "Hits" is the count the source reported, or the number of results returned for web queries. OpenAlex compound Boolean queries did not behave reliably (the same count, 45, came back for two different second clauses), so I inspected the returned records by hand and treat the counts as indicative only. + +--- + +## (a) Search log + +### OpenAlex, `filter=title_and_abstract.search` + +| # | Exact query | Hits | Notes | +| --- | --- | --- | --- | +| 1 | `"scriptio continua"` | 87 | Mostly NLP word segmentation (Javanese, Balinese, Japanese, Thai), epigraphy, papyrology. | +| 2 | `"scriptura continua"` | 16 | Pflug lexicon entry; Cordingley 2006 (Beckett); Slote 2014; Okasha 2002; no technology analogy. | +| 3 | `("scriptio continua" OR "scriptura continua") AND (algorithm OR algorithmic OR "language model" OR chatbot OR prompt OR "artificial intelligence" OR "machine learning")` | 14 | All technical: tokenizers, OCR, restoring papyri (Fujii 2023; Cullhed 2024, 2025; Tolmachev 2018). Also Rosehill Zenodo episode 2736. No analogy. | +| 4 | same stem `AND (interface OR "user experience" OR usability OR "human-computer" OR typography OR "user interface")` | error | No count returned (query too long for the parser). | +| 5 | same stem `AND (interface OR usability)` | 45 | Top 45 inspected. Nothing draws the analogy. HCI-venue hit: Cho 2021 (DIS), see entry. Also Rosehill 2736, 2737; Stenberg 2017 (music scores). | +| 6 | same stem `AND ("user experience" OR usability OR "user interface")` | 45 | Same 45 records as #5; the count is not trustworthy. | +| 7 | same stem `AND (typography OR "human-computer" OR HCI)` | 0 | | +| 8 | same stem `AND (code OR programming OR "programming language" OR whitespace OR syntax OR "data format")` | error | Too many terms. | +| 9 | same stem `AND (code OR programming OR whitespace OR syntax)` | 7 | Robert-Hayek 2025; Slote 2014; Stenberg 2017; none about code as writing. | +| 10 | same stem `AND ("language model" OR chatbot OR "artificial intelligence")` | 7 | Tokenizer, Thai sentiment, papyri; Rosehill 2736. | +| 11 | same stem `AND (algorithm OR algorithmic OR platform)` | 6 | Javanese and Balinese transliteration tools; Thai sentiment. | +| 12 | same stem `AND (platform OR "recommender" OR "social media" OR digital)` | 6 | Cho 2021; Willis 2017 *Reading*; Helmke 2026. | +| 13 | same stem `AND (design OR designer)` | 4 | Murawaki 2016; Cho 2021; Syauqi 2026; Stenberg 2017. | +| 14 | same stem `AND (transparency OR explainability OR opacity OR legibility)` | error | | +| 15 | same stem `AND (transparency OR explainability OR opacity)` | 7 | Runes, Glagolitic, Metaphysics manuscripts; no XAI hit. | +| 16 | same stem `AND ("silent reading" OR Saenger)` | 10 | Rosehill 2737 and 2729; Okasha; Slote; Ramírez 2007; Kasisopa 2016 (Thai eye movements); Stenberg. | +| 17 | same stem `AND ("human factors" OR ergonomics OR cognitive)` | 4 | Slote; Willis; Stenberg; Rosehill 2729. | + +### OpenAlex, `filter=fulltext.search` (top 10 inspected per query; counts are for the whole index) + +| # | Exact query | Hits | Notes | +| --- | --- | --- | --- | +| 18 | `"scriptio continua" "large language models"` | 50 | All tokenization/multilingual-LLM papers (Feher 2025; Gu 2024; Inoue 2026) and Rosehill 2736. Technical sense only. | +| 19 | `"scriptio continua" chatbot` | 18 | Japanese/Thai NLP. | +| 20 | `"scriptio continua" "user interface"` | 24 | Transliteration tools, Liesaputra and Witten 2012 (*Realistic electronic books*), Lecolinet 1998. None uses the analogy in the abstract; full text not opened. | +| 21 | `"scriptio continua" "human-computer interaction"` | 11 | Same group. | +| 22 | `"scriptio continua" "explainable AI"` | 3 | Syauqi 2025 survey; two Niwa records. | +| 23 | `"scriptio continua" algorithmic` | 156 | Top 10 all NLP. | +| 24 | `"scriptio continua" "interface design"`, `"scriptio continua" "user experience"`, `"scriptio continua" "algorithmic" "literacy"`, `("Space Between Words" "Saenger" "artificial intelligence")` | not run | OpenAlex free budget exhausted ("Rate limit exceeded", no API key). The full-text leg is therefore incomplete for UX and literacy. | + +### OpenAlex, other parts + +| # | Exact query | Hits | Notes | +| --- | --- | --- | --- | +| 25 | `"In the Vineyard of the Text"` (title_and_abstract) | 11 | Illich 1993 record; reviews by Saenger 1995, von Hagen 1994, Margadant 1994; Babich 2014. | +| 26 | `"Illich" AND ("vineyard of the text") AND (interface OR "human-computer" OR design)` | 0 | | +| 27 | `"history of the page" AND ("user experience" OR "interface design" OR "human-computer interaction")` | 2 | Unrelated (Persian museum web pages; Irish digital history). | +| 28 | six further Q2 queries on Hugh of St Victor, Mak, Drucker, punctuation and interface, medieval manuscripts and UX | error | Unparenthesised OR/AND; not rerun before the budget ended. Covered instead by Crossref (#47-50) and WebSearch. | +| 29 | `"decision experience design"` (title_and_abstract) | 0 | | +| 30 | `search="decision experience design"` | 0 | | +| 31 | `"decision experience" AND (design OR designer)` | 61 | Top 15 inspected: nursing, MBA and educational-game papers. No design-discipline use. | +| 32 | `"DXD" AND (decision OR design)` | 560 | Pharmaceutical DXd conjugates; "DxD Design for District" (Collina and Simonelli 2000-2001, Italian industrial design). | +| 33 | `"decision design" AND (algorithm OR AI OR "artificial intelligence")` | 134 | New: Morii 2026 (SSRN), Haid 2023 (AHFE). See entries. | +| 34 | `"decision-centred design"` | 13 | O'Hare 1998; Schnittker 2019; **Sun 2026 (arXiv), "Decision-Centric Design for LLM Systems"** (its spelling is "decision-centric", so not an exact match for the query). | +| 35 | `"choice experience design"` | 0 | | +| 36 | `"decision architecture" AND ("artificial intelligence" OR algorithmic)` | 440 | Top 6: 2025-26 Zenodo/SSRN governance papers. No design-discipline claim. | +| 37 | `"decision intelligence" AND (design OR "user experience")` | 385 | O'Callaghan 2023 and Zenodo preprints. | +| 38 | `algorithmacy` (title_and_abstract) | 0 | | +| 39 | `search=algorithmacy` and `fulltext.search:algorithmacy` | 1 and 1 | Oberschelp 1995, IFIP chapter "Developing analytical competency through informatics teaching". Crossref has no abstract. **Resolved by the card (library/oberschelp1995developing.md): the chapter prints the word once, p. 121; see section 5.** | +| 40 | `"algorithmic numeracy" OR "algorithmic fluency"` | 21 | Applied-field papers (ICU, education). No construct competing with algorithmacy. | +| 41 | `"human oversight" AND (critique OR illusory OR fallacy OR "meaningful")` | 1,182 | Top 8 inspected: Zhu 2026; Lermann Henestrosa 2026; Rubegni 2026. | +| 42 | `"human oversight" AND "interface" AND (AI OR algorithmic) AND "audit log"` | 8 | Jerry 2026 (oversight-by-design for generative UIs). | +| 43 | `"legibility" AND "James C. Scott" AND (algorithm OR AI OR "machine learning")` | 2 | Cordeiro 2026, "What the Control Room Knows" (democratic legibility, journalism). | +| 44 | `"seeing like a state" AND ("artificial intelligence" OR algorithmic OR "language model")` | 13 | McQuillan 2016; Liu 2021; Willis 2025; Lee 2025; Golumbia 2015. | +| 45 | `"observability" AND (critique OR limits) AND ("AI governance" OR "algorithmic accountability")` | 39 | Meyman 2026 (Zenodo). | +| 46 | `"transparency ideal" OR "limits of transparency" AND algorithmic` | 26 | Descampe 2021; Brauneis and Goodman 2017; dos Santos 2026. | +| 46a | `"algorithmic literacy" AND (critique OR "wrong frame" OR "beyond literacy" OR "individualizes")` | 23 | Dasgupta and Hill 2023; Tanksley 2024; Vindigni 2026. | +| 46b | `"algorithm awareness" AND "rhetoric"` | 1 | Unrelated (Van Zanen 2024). | +| 46c | `"human oversight" AND "interface conventions"` | 0 | | +| 46d | `"alert fatigue" OR "automation complacency" AND "audit logs" AND AI` | 0 | | + +### Crossref (`query.bibliographic`) + +| # | Exact query | Hits | Notes | +| --- | --- | --- | --- | +| 47 | `scriptio continua artificial intelligence interface` | relevance count only | Top 6: reference entries (Brill's New Pauly; OED), Robert-Hayek 2025, Widiarti 2020, Butskhrikidze 2022. No analogy. | +| 48 | `scriptio continua large language model prompt` | same | Reference entries; Juman++. | +| 49 | `scriptio continua user interface design human-computer interaction` | same | HCI-series books unrelated to the term. | +| 50 | `word spacing history interface design silent reading Saenger technology` | same | Saenger 1997, 1982, 2020; reviews (Mulchahey 2001, Hanna 1999, Richter 2001); Reynolds and Walker 2004 (children's reading books). | +| 51 | `scriptio continua programming languages whitespace` | same | Nothing on code. | +| 52 | `decision experience design` | same | Skavronskaya 2017 (*Visitor experience design*), as on 30 Sept. | +| 53 | `algorithmacy` | 0 | | +| 54 | `decision experience (DX) design algorithmic intermediary` | same | Nothing relevant. | +| 55 | record verification, 17 lookups (Illich review, Mak, Bolter and Grusin, Drucker, Seaver, Bucher, Dogruel, Oeldorf-Hirsch and Neubaum, Gran et al., Koenig, Green, Ananny and Crawford, Norman, Dembeck) | n/a | Each target matched at rank 1 or 2, except Cotter and Reisdorf 2020 (no Crossref record found) and Houston *Shady Characters* (Crossref returned an error). | + +### Consensus (three calls, no filters) + +| # | Exact query | Hits | Notes | +| --- | --- | --- | --- | +| 56 | `scriptio continua word separation as analogy for artificial intelligence or user interface design` | 20 | NLP segmentation; Cho 2021; Buschek 2024 (*Collage is the New Writing*). No analogy. | +| 57 | `history of word spacing and silent reading as precedent for human-computer interaction or interface conventions` | 20 | Saenger 1997 (blurb), reviews (Aronoff 2002, Baeuml 2002), Mirault 2019, Petrick 2020 (historiography of HCI), Grudin 1989, Dyson 2004, Hsu 2000 (word spacing in Chinese on a VDT). No paper uses the history as an HCI precedent. | +| 58 | `human oversight of AI critique: interface conventions or structural design instead of monitoring logs and audit trails` | 20 | Green 2022; Koulu 2020; Zhu 2026; Holzinger 2024; Abulibdeh 2026; Mitchell 2026; Grunde-McLaughlin 2026; Laux 2023, 2025; Sterz 2024. | + +### Failed or blocked sources + +| # | Source | Result | +| --- | --- | --- | +| 59 | Scholar Gateway `semanticSearch` | `ACCESS_DENIED: Access denied`. | +| 60 | Semantic Scholar `/paper/search` (four queries: scriptio continua with LLM, with HCI, "decision experience design", "algorithmacy") | HTTP 429 on every call. | +| 61 | arXiv API (`all:"scriptio continua"`, `all:"decision experience design"`, `all:algorithmacy`, and three more) | HTTP 429 / timeout. Not run. arXiv content reached only through OpenAlex and Consensus. | +| 62 | OpenAlex | Free daily budget exhausted partway; later queries returned "Rate limit exceeded" (see #24, #28). | +| 63 | Forbes (Zainescu, 11 Oct 2026) via WebFetch | HTTP 403. Summarised from WebSearch snippets only. | +| 64 | uxdesign.cc "The chat box isn't a UI paradigm" via WebFetch | HTTP 403. Snippet only. | + +### WebSearch (standard unless marked extended; each returned 9-10 results) + +| # | Exact query | Result | +| --- | --- | --- | +| 65 | `"scriptio continua" AI prompts OR "language models" OR chatbot space between words analogy` | Only Fujii 2023 and tokenizer pages. No analogy. | +| 66 | `"scriptio continua" user interface design history word spacing UX analogy` | Wikipedia; pen-computing patents; an art-school blog. No UX analogy. | +| 67 | `"scriptio continua" programming whitespace OR "code" OR "JSON" analogy` | Tildes thread on spaces and ambiguity; Unicode reference. Nothing on code. | +| 68 | `"scriptio continua" LLM tokens prompt "like" medieval scribes spaces blog` | Clérice 2020; popular history pages (The Collector; Freewrite marketing blog). | +| 69 | `"we are in the scriptio continua era" OR "scriptio continua of AI" OR "scriptio continua" "AI interfaces"` | No exact phrase. Wintersberger 2022 (CHI SIG, "continuous interaction", unrelated sense). | +| 70 | `Saenger "space between words" lessons for AI interface design OR user experience OR "design history"` | Wikipedia; arXiv 2508.16914 (visual segmentation and word recognition). Nothing. | +| 71 | `"scriptio continua" Hacker News` | No HN thread surfaced. McGinley Substack (2025-08-06) and reference pages. | +| 72 | `"scriptura continua" AI OR algorithm OR interface OR UX essay` | Led to Boris Mueller's Medium essay (see entry). | +| 73 | `"Rosehill" "How Word Spacing Changed Human Thinking" OR "Why AI Flagged Your Em Dash" episode` | No web record of the podcast; found through OpenAlex/Zenodo instead. | +| 74 | `"spaces between words" invention analogy "AI" "prompt" OR "agents" OR "LLM" interface needs its own "punctuation" essay Substack` | UBC course blog on Saenger; Heather Radke, "On Spacebars" (Substack); AI em-dash essays. No interface argument. | +| 75 | `AI chat interfaces skeuomorphic early stage conventions "we are still" "early days" interface design history typography analogy Gutenberg "AI UX"` | Chat-as-default critiques (Designative 2026-03-19; setproduct). No Gutenberg or typography analogy. | +| 76 | `Medium OR Substack "scriptio continua" "agents" OR "algorithms" OR "platforms" OR "LLMs" reading without spaces metaphor opaque systems literacy` | **Dembeck 2023** (see entry); arXiv 2607.27547 "AI Literacy: An Exercise in Power-Knowledge"; Roe et al.; Gupta et al. | +| 77 | `Amelia Wattenberger "Why Chatbots Are Not the Future"` | Willison link post, 2023-05-15; fetched original. | +| 78 | `Maggie Appleton "Language Model Sketchbook, or Why I Hate Chatbots"` | Fetched. | +| 79 | `Pete Koomen "AI Horseless Carriages" essay` | Coverage; fetched koomen.dev. | +| 80 | `Geoffrey Litt "Malleable software" Ink & Switch essay LLM` | Litt 2023 post; Litt et al. 2025 essay (summary only). | +| 81 | `Jakob Nielsen "AI: First New UI Paradigm in 60 Years" intent-based outcome specification` | Nielsen's Substack and UX Tigers; fetched. | +| 82 | `chat is the command line of AI "command line" era interfaces "GUI moment" for AI essay` | Tunguz; Fazi "Prompt in the Shell"; AltaML; The Register 2026-03-11; Speedinvest; uxdesign.cc. Snippets only. | +| 83 | `AI interfaces "incunabula" OR "incunable" chat interface early conventions not yet settled printing press analogy design` | No incunabula framing. Huttenlocher quote (movable type as nearest analogue to AI); Roosen, "AI is not a Printing Press". | +| 84 | `"pre-Gutenberg" OR "early printing" analogy for large language model interfaces design conventions page layout` | No such analogy. Generative-interfaces preprint (arXiv 2508.19227); deRSE25 workshop paper. | +| 85 | `Linus Lee "Prism" OR "text interface" language models "interfaces" essay thesis "Is chat the right interface"` | The search did not locate an essay by that name; his site and talks surfaced. Not read. | +| 86 | `Bret Victor "Humane Representation of Thought" OR "Magic Ink" large language models critique AI interface` | Victor's own texts only; no LLM commentary by him found. | +| 87 | `Illich "In the Vineyard of the Text" Hugh of St Victor page layout index "user interface" OR "user experience" OR "technology of reading" interface design` | Publisher pages, Goodreads, secondary essays. No UI reading of Illich found in these results. | +| 88 | `history of user experience design origins medieval manuscript page layout index punctuation "first interface" OR "first UX" scribes` | Getty, *Graphic Design in the Middle Ages* (2023): scribes as early graphic designers. Not called UX. | +| 89 | `Johanna Drucker interface "Graphesis" book as interface page design history humanities interface theory` | Publisher blurbs; reviews (Enslen 2016; Sattler 2014). | +| 90 | `Bonnie Mak "How the Page Matters" argument page as technology of reading` | Illinois project page; Monoskop; reading-course notes. | +| 91 | `"algorithmacy"` (extended) | Hunt's Substack posts; algorithmacy.org (the author's conference, 28-31 Oct 2026); two GitHub copies of the lab repo. No other author. | +| 92 | `"decision experience design" OR "Decision Experience Design (DXD)"` (extended) | Hunt's Substack post (the published Part 1); "Digital Experience Design (DXD)" job listings (Gensler via builtin). No other use. | +| 93 | `"decision experience" "DX" designer algorithm "decision design" 2026 Substack OR Medium OR LinkedIn` | Kobetz, "DX Perspective" (DX there is the newsletter's own name; no "decision experience"); Serrault, "AI & Decision Design" seminar (30 Jan 2026); Medium essay "UX Design 2026+: From producing to deciding". None uses the term. | +| 94 | `Nadella "Models as Insider Risks" Super Intelligence Era critique OR response OR analysis` (extended) | Nine news and explainer pages (Business Standard; Forbes; officechai; others). Forbes reports the post's argument; it is not a critique. | +| 95 | `Nadella insider risk post models logging audit "human oversight" OR "alert fatigue" OR "usability" OR "human factors" criticism` (extended) | Eight explainers. One source calls the post a statement of intent rather than a technical specification [unverified: see AUDIT_03]. Nothing on logs, usability or human factors. | +| 96 | `Nadella "insider risk" models Hacker News OR LessWrong OR "AI Alignment Forum" OR Schneier OR Lawfare reaction` | None of the named venues returned. Digg aggregator page shows engagement; All-In Podcast background. | +| 97 | `critique of "algorithmic literacy" OR "algorithm awareness" individualizes responsibility onto users structural alternatives design platform` | Korean-language study of recommender literacy limits; arXiv 2507.23585 ("Agency Among Agents", hypertextual friction); MSU dissertation on critical algorithmic literacy. Snippets only. | +| 98 | `"grammar of legibility" Parkes OR "legibility" interface conventions AI governance` | No source uses "grammar of legibility" for AI. Singapore Consensus 2026 (arXiv 2608.14611) and Lancaster "Legible AI by design" treat legibility as a goal [Singapore gloss unverified: see AUDIT_03]. Snippets only. | +| 99 | `Lorien Pratt "decision intelligence" definition Link book; Cassie Kozyrkov "decision intelligence" discipline definition` | Kozyrkov: "turning information into better actions" (secondary). Pratt: talk summary. | +| 100 | `"human oversight" of AI illusion OR fallacy "rubber stamp" interface design "structural" rather than oversight Green Enqvist Koulu critique 2024 2025` | Green 2021/2022 and coverage. | +| 101 | `Oberschelp "Developing analytical competency through informatics teaching" algorithmacy` and `"algorithmacy" Oberschelp OR "algorithmics" numeracy literacy "algorithmacy" 1990s informatics education` | The chapter was not found online; no source links it to "algorithmacy". The second search returned Hunt's Substack, which says the word appeared once before, in an unpublished talk (Leese and Matzner 2016). I could not verify that. | + +### Fetches and downloads + +| # | Item | Result | +| --- | --- | --- | +| 102 | Mueller, Medium feed `@borism` | Read first 100,000 of 122,799 characters, through the fetch tool. | +| 103 | McGinley Substack, "Scriptio Continua" | Summary only. | +| 104 | Wattenberger, wattenberger.com/thoughts/boo-chatbots | Summary only. | +| 105 | Koomen, koomen.dev/essays/horseless-carriages | Summary only. | +| 106 | Nielsen, jakobnielsenphd.substack.com | Summary only. | +| 107 | Appleton, maggieappleton.com/lm-sketchbook | Summary only. | +| 108 | Dembeck 2023 PDF, orbilu.uni.lu/handle/10993/59372 | Downloaded to `~/iit-playground/dxd_foundation_local/pdfs/Z/dembeck2023.pdf`; `pdftotext`; **read in full** (3,387 words, German). | +| 109 | Zenodo records of Rosehill episodes 2729, 2736, 2737 | Abstracts via OpenAlex (Zenodo's own API failed). | +| 110 | Abstracts via OpenAlex for Lermann Henestrosa, Zhu, Rubegni, Meyman, Cordeiro, Willis, Morii (none), Sun | Database abstracts. | +| 111 | algorithmacy.org | Page metadata only (event, dates). | + +--- + +## (b) Entries + +Ordered by closeness to the author's use. "Same claim" would mean: word separation moved work from reader to page, and that is the precedent for a design move that makes algorithmic mediation workable. **No entry is "same claim."** + +### 1. Dembeck (2023), "Im Flow: Skizze über Ausdruck und Beitrag im Digitalen" +- **Key:** `dembeck2023flow`. Till Dembeck, University of Luxembourg, *andererseits* 11/12 (2023); reprinted in the 2024 yearbook, DOI 10.14361/9783839469811-020 (Crossref record found). PDF at orbilu.uni.lu/handle/10993/59372. +- **Verified by:** Crossref record for the 2024 reprint; orbilu metadata; PDF downloaded. +- **Read status:** read in full (German; translations below are mine). +- **What it argues:** Following Illich, the shift from *scriptio continua* to the two-dimensional page around 1200 founded what Dembeck calls "biblionome Literalität" (book-bound literacy: abstraction, independent expression, the standalone "Beitrag"). He then asks what "digital literacy" is, and observes that current media-literacy and fake-news education take the book paradigm as their starting point, and suggests, in a hedged conjecture ("Es könnte aber sein", §5), that this "could miss the point entirely" ("an der Sache völlig vorbeigeht"). +- **Quotations (original German, section locators):** + - §1: "Der Übergang von der Schreibweise der scriptio continua, die Buchstaben aneinanderreiht, ohne sie durch Wortzwischenräume, Interpunktion, Absatzstrukturen usw. zu organisieren, hin zur zweidimensional disponierten Seite ging dabei Hand in Hand mit einer ganzen Reihe von Entwicklungen, die sich Jahrhunderte später als zentral für das moderne Bildungsverständnis erwiesen." + - §1: "man gerät in eine Art Tunnel" (on reading unspaced text). The modern text "erzeugt Übersicht und keinen Tunnelblick." + - §4: the digital flow is "durch algorithmische Rechenarbeit im Hintergrund zumindest mitgesteuert" (co-steered by algorithmic computation in the background). This is the only algorithm mention; it is a relative clause in the main argument, not a parenthesis (the parenthesis that follows concerns Raymond Williams's "flow"). +- **Closeness:** adjacent, and the nearest in print. It uses the same historical hinge (Illich, Parkes, Saenger in n. 1) to argue about a digital-age literacy, and it doubts, in a hedged conjecture (§5), that the goals book-bound literacy passes on to media-literacy education fit the digital. It does not treat word separation as a design move for a mediator, does not discuss algorithmic third parties, and does not propose a design remedy. +- **Bearing:** complicates a bare novelty claim for "history of the page as a lens on digital literacy". Supports the author's suspicion that book-bound literacy goals may not fit the digital (hedged in Dembeck; rulings 1 and 2). Cite as predecessor. + +### 2. Mak (2011), *How the Page Matters* +- **Key:** `mak2011page`. University of Toronto Press. Existence verified by Crossref review records (Driver 2013, *Speculum*, DOI 10.1017/s0038713413001322). +- **Read status:** summary only (publisher and course-note summaries; I did not open the book). +- **What it argues (secondary):** The page is an interface that has developed since the scroll. Mak follows one fifteenth-century text (*Controversia de nobilitate*) through manuscript, print and digital forms. Chapters include "Architecture of the Page" and "The Digital Page" (Illinois news notice, via WebSearch). +- **Closeness:** adjacent. It treats the page as an interface across media, which is the claim "history of the page as human-factors history" in general form. It is not about word separation specifically or about algorithmic mediation. +- **Bearing:** the author cannot claim to be first to call the page an interface. Needs a full read before anything is said about its chapters. + +### 3. Illich (1993), *In the Vineyard of the Text: A Commentary to Hugh's Didascalicon* +- **Key:** `illich1993vineyard`. Existence verified through OpenAlex (W-record, 1993) and Saenger's review (*Library Quarterly* 65, 1995, DOI 10.1086/602782). Publisher not shown in the records I opened. +- **Read status:** summary only (Equinox table of contents and secondary essays); also seen through Dembeck's page references (Illich 1991 German edition, pp. 55-60, 101-104, 121-129, as Dembeck cites them). +- **What it argues (via Dembeck, read in full):** The twelfth-century page, with its two-dimensional layout, made the text an object one could navigate and compare; Illich ranks it above the printing press in media-historical force. Dembeck reports Illich as saying the layout operation is "exakt diejenige Operation ... die man durchführen muss, um einen Text in den Druck umzusetzen" (§1). +- **Closeness:** adjacent. Illich is the strongest predecessor for "a change of page layout as a technology of reading". Nothing in searches #87 and #26 links him to UX or interface design; a WebSearch summary of a notice of the book (reported in AUDIT_03, not checked by me) says he extended the argument to the screen replacing the page. +- **Bearing:** supports the idea that layout, not only word spacing, was the move (problem 4 in the plan). Must be cited. + +### 4. Buschek (2024), "Collage is the New Writing" +- **Key:** `buschek2024collage`. DIS 2024, DOI 10.1145/3643834.3660681 (Consensus record). +- **Read status:** read in full (card: library/buschek2024collage.md). +- **What it argues:** Literary concepts (collage) offer lenses for designing AI writing tools; "creators of future writing tools" should "engage not only with new technological possibilities, but also with past writing innovations" (abstract). +- **Closeness:** adjacent in method (writing history as a source of AI-interface design), distant in content (collage, one writer, one tool). The full text confirms this: his history is literary Expressionism and Manovich's 1999 "avant-garde as software" (hyperlinking as a parallel), and he explicitly limits the claim to interface form. +- **Bearing:** shows the method "learn from writing history for AI interfaces" is already in HCI. Cite as a methodological neighbour. + +### 5. Mueller (2026), "The Interface is Content" +- **Key:** `muller2026interface`. Medium, 2026-02-21, borism.medium.com/the-interface-is-content-b13f4d693d66. +- **Read status:** read in part (first 100,000 of 122,799 characters, through the fetch tool). +- **What it argues:** Typography is the interface of text and cannot be separated from content. *Scriptio continua* appears once, as a counterexample to a stripped-down text; spacing "is crucial to fast reading" (via fetch). +- **Closeness:** decorative mention. It uses the history to say removing interface cues makes content less meaningful. No AI claim in the portion I saw (the unread 22,799 characters could change that). +- **Bearing:** shows a UX-side writer reaching for the same example. Not a precedent. + +### 6. Rosehill, Gemini 3.1 (Flash) and Chatterbox TTS (2026), podcast episodes 2737, 2736, 2729 +- **Keys:** `rosehill2026wordspacing` (DOI 10.5281/zenodo.20182987), `rosehill2026emdash` (10.5281/zenodo.20182971), `rosehill2026beehives` (10.5281/zenodo.20182929). Zenodo, May 2026. +- **Read status:** abstract only (OpenAlex records of the Zenodo deposits). The creator field names Gemini and a text-to-speech tool; DataCite also lists Daniel Rosehill as a creator (AUDIT_03), so "AI-generated" is not established. +- **What they say:** 2737 summarises Saenger's thesis that word separation "enabled silent reading, private reflection, and ultimately the modern sense of self". 2729 says Irish monks "accidentally rewired human cognition". 2736 links punctuation history to AI em-dash detectors. +- **Closeness:** decorative. None applies the analogy to interface design or algorithmic mediation. +- **Bearing:** do not cite. They repeat the "rewired cognition" claim that the standing ruling excludes. + +### 7. Cho, Cheon and Kang (2021), "Giving Space to Your Message" +- **Key:** `cho2021giving`. DIS 2021, DOI 10.1145/3461778.3462078. +- **Read status:** read in full (card: library/cho2021giving.md). +- **What it says:** A real-time word-segmentation (spacing) aid for Korean typists. The abstract mentions *scriptio continua* only as the contrast case ("far from scriptio continua"); the body does the same and no more (Section 1, p. 2; Section 2, p. 3, on "indecisive spacings" in Chinese, Japanese and Korean). +- **Closeness:** an HCI paper about inserting spaces, so literal, not analogical. + +### 8. NLP and digital-humanities uses of "scriptio continua" (technical sense) +- Clérice 2020 (`clerice2020evaluating`, DOI 10.46298/jdmdh.5581); Fujii, Shibata and Yamaguchi 2023 (`fujii2023how`, DOI 10.48550/arxiv.2306.09572); Butskhrikidze 2022 (`butskhrikidze2022what`, DOI 10.2478/jazcas-2022-0006); Widiarti and Pulungan 2020; Cullhed 2024/2025. +- **Read status:** Clérice, Fujii and Butskhrikidze read in full (cards); the others abstract only. +- **Bearing:** these are the "AI and *scriptio continua*" results a reviewer will find. They concern tokenization or typology of unspaced scripts, not an analogy. Fujii tested Japanese only (Thai and Chinese appear only in its introduction's list; other languages are named as future work, p. 43); Butskhrikidze is a linguistic-typology paper on Burmese, Thai and Lhasa Tibetan, not NLP; Clérice's one-line aside says *scripta continua* "disappeared around the 8th century" while "spacing remained erratic" (Section I, p. 1), a dating sentence the abstract does not show. Cite one (Fujii 2023) to show the reviewer the collision was seen. + +### 9. Wattenberger (2023), "Why Chatbots Are Not the Future" +- **Key:** `wattenberger2023chatbots`. wattenberger.com/thoughts/boo-chatbots; dated 15 May 2023 by Willison's link post. +- **Read status:** read in full (card: library/wattenberger2023chatbots.md). +- **Argues (via fetch):** "Good tools make it clear how they should be used"; "the burden to learn what works still lies with every single user." No writing-history analogy. Partly about one user and a tool; also criticises automation that turns people into "machine operators". +- **Closeness:** shares the author's premise that the burden has been put on the user; different remedy (controls and affordances), different structure (one user). The full text adds that she wants the interface to carry the user's burden, the same direction as the spacing precedent, differing in scale and parties; the essay models one user and one chatbot and never mentions an interested third party. + +### 10. Appleton (2023), "Language Model Sketchbook, or Why I Hate Chatbots" +- **Key:** `appleton2023sketchbook`. maggieappleton.com/lm-sketchbook; undated on the page, mid-2023 by backlink. +- **Read status:** read in full (card: library/appleton2023sketchbook.md). +- **Argues (via fetch):** "We are irreversibly anchored to this text-heavy, turn-based interface paradigm"; chat is "the lazy solution". Prototypes for one writer and one model. No history analogy. The design aim is to keep the writer's own reasoning active (card). + +### 11. Koomen (2025), "AI Horseless Carriages" +- **Key:** `koomen2025horseless`. koomen.dev/essays/horseless-carriages; the essay cites April 2025. +- **Read status:** read in full (card: library/koomen2025horseless.md). +- **Argues (via fetch):** AI apps "mimic old ways of building software"; users should edit the system prompt of an agent "acting on my behalf". This is the only neighbour with an agent acting for a person, but the setting is one user and an agent (an email scenario involves a boss as recipient, not as a party the mediator reads). +- **Closeness:** the "horseless carriage" structure of the author's pre-convention claim, without the page history. + +### 12. Nielsen (2023), "AI Is First New UI Paradigm in 60 Years" +- **Key:** `nielsen2023paradigm`. Substack mirror of UX Tigers. The fetch gave 22 June 2023; a search snippet gave 20 May 2023 for the original (AUDIT_03 confirms: Substack mirror 22 Jun 2023, UX Tigers original 20 May 2023). The words "articulation barrier" are the title of a separate Nielsen article linked from this one, not words in this article; this article says half the population in rich countries is "not articulate enough" to get good results from current AI bots. +- **Read status:** read in full (card: library/nielsen2023paradigm.md). +- **Argues (via fetch):** a third UI paradigm, "intent-based outcome specification", with an "articulation barrier": "half the population in rich countries is not articulate enough" to get good results. No conventions, no writing history. +- **Closeness:** the nearest claim about *who can use* AI interfaces (articulation as the new literacy), pointing the opposite way: Nielsen locates the barrier in users' writing skill and does not propose an interface convention to remove it. + +### 13. Litt et al. (2025), "Malleable Software"; Litt (2023) +- **Key:** `litt2025malleable`. Ink & Switch. Read in full (card: library/litt2025malleable.md). Argues users need agency to modify software. It is not only "one person and their tools": a whole section ("Communal creation") argues that malleability is a group matter, with shared tools for teams, hospital departments and Mastodon instances. No history analogy (confirmed). + +### 14. Bret Victor; Linus Lee +- Victor's "Humane Representation of Thought" (2014) and "Magic Ink" (2006): found his own texts; no commentary by him on LLMs. Lee: the search found his research on manipulable model internals; I did not locate or read an essay on chat interfaces. Neither is a precedent for the page-history analogy. + +### 15. Other chat-interface commentary (snippets only) +- Tunguz, "The Text Box Isn't Enough" (DOS analogy); Fazi, "Prompt in the Shell"; AltaML, "The Chat Box is the New Command Line"; UX Collective, "The chat box isn't a UI paradigm; it's what shipped" (403 on fetch); Designative, "The Conversation Trap" (2026-03-19). The command-line and DOS analogies are common and are about one user and a tool. None used a writing-history analogy. + +### 16. The oversight-fallacy literature (Q7) +All at abstract level unless noted. +- Green (2022), `green2022flaws`, *Computer Law & Security Review*, DOI 10.1016/j.clsr.2022.105681. Surveys 41 policies; finds people cannot perform the oversight functions and proposes "a shift from human oversight to institutional oversight". +- Koulu (2020), `koulu2020proceduralizing`, DOI 10.1177/1023263x20978649: oversight risks becoming "an empty procedural shell". +- Lermann Henestrosa, Grinbaum and Yamshchikov (2026), `lermannhenestrosa2026human`, PsyArXiv, DOI 10.31234/osf.io/u68y2_v1: proposes "a conceptual reorientation of human oversight away from reactive approval and toward structurally embedded control". The abstract's remedies are incentive-based ethical stress testing, normative clarity and accounting for the limits of human oversight, an institutional remedy rather than taking the work off the person. The full text also says general-purpose AI users are not served by "AI literacy" alone (section 3.3), which is close to ruling D3, but it makes the point for regulatory reasons (card). +- Zhu, Lu, Ding et al. (2026), `zhu2026designing`, *AI and Ethics*, DOI 10.1007/s43681-026-01147-7: design outputs "so that humans can efficiently check and contest them without having to resolve the task". The full text keeps the human as evaluator at run time but pushes much of the checking to design time and to sampling (pp. 5-6), and asks how to make the person's check cheaper, not what task the person keeps. +- Mitchell et al. (2026), `mitchell2026agents`, arXiv 2608.23642: agent designs "impede effective human oversight"; calls for "design-level affordances". Its solution inventory (Table 1, p. 7) includes structural items (bounded autonomy, automated pre-checks, separated roles), and it argues that design both obstructs oversight (Section 3) and erodes it (Section 4). +- Grunde-McLaughlin et al. (2026), arXiv 2602.16844: action traces "cumbersome"; a redesigned interface cut error-finding time but not final accuracy. +- Holzinger et al. (2024), `holzinger2024human`; Abulibdeh et al. (2026), `abulibdeh2026who` (*Lancet*): human-in-the-loop as "symbolic reassurance"; Meyman (2026), `meyman2026impossibility` (Zenodo, version 2.0.0): the abstract-level "formal claim that observability cannot produce ex-ante authorization" is overstated for the full text. Version 2.0.0 withdraws the universal argument of earlier editions and keeps a conditional information-sufficiency result on a stipulated model (Proposition 1, p. 11), and says observability and enforcement "remain complementary layers" (p. 18). It is a self-published, non-peer-reviewed note by an author affiliated with FERZ, Inc., the company whose doctrine it supports; its remedy is a deterministic external gate with a person only at ABSTAIN. +- **Closeness to the author's critique of observe-log-audit:** same direction. None targets the Nadella post (all predate it except Rubegni 2026-10-11, which is clinical). None draws on the history of reading. None uses "legibility" in Scott's or Parkes's sense for oversight; the only Scott hits are Cordeiro 2026 (journalism; "democratic legibility" as the capacity to read information systems; a programme statement for a newsletter, not a study) and Willis 2025 (working paper on robotics and high modernism), both abstract only. +- **Bearing:** the claim "oversight remedies reproduce a reading model" is not in these abstracts in those words. The narrower claim that human oversight as currently designed fails and needs redesign or institutional backing is established and must be cited; none of these sources replaces the human check with structure, with two qualifiers from the full texts: Zhu et al. (2026) push much of the checking to design time and sampling, so structure does replace some run-time checks, and Mitchell et al. (2026) include structural items in their inventory (bounded autonomy, automated pre-checks, separated roles), so they sit closer to the structural camp than this summary suggests. + +### 17. Literacy critiques (Q5) +- Seaver (2019), `seaver2019knowing`, DOI 10.1515/9780691190600-028; Bucher (2017), `bucher2016algorithmic`, DOI 10.1080/1369118x.2016.1154086; Gran, Booth and Bucher (2021), `gran2020aware`, DOI 10.1080/1369118x.2020.1736124; Dogruel, Masur and Joeckel (2022), `dogruel2021development`, DOI 10.1080/19312458.2021.1968361; Oeldorf-Hirsch and Neubaum (2023), `oeldorfhirsch2023what`, DOI 10.1177/14614448231182662; Koenig (2020), `koenig2020algorithms`, DOI 10.1016/j.compcom.2020.102611; Ananny and Crawford (2018), `ananny2018seeing`, DOI 10.1177/1461444816676645. All verified as records in Crossref (compare titles). Read status: **not read this run**; the repo already holds cards for several (see REUSE.md). +- Cotter and Reisdorf (2020), "Algorithmic knowledge gaps": no Crossref record found; see (d). +- Dasgupta and Hill (2023), `dasgupta2023designing`: "designing for critical algorithmic literacies" (from F1; designing for learners' literacies; the full text shows the learners are children and the evidence is the authors' own case work). +- **Whether any says literacy is the wrong frame:** I did not find a paper that says so for the triad. The structural critique of individual literacy exists in scattered form (Korean recommender-literacy study; "Agency Among Agents" arXiv 2507.23585, which proposes interaction-level friction; both snippet level). Ananny and Crawford is the canonical transparency critique. Dembeck 2023 is the only one that says the *literacy frame* may be inherited from the book. Oeldorf-Hirsch and Neubaum's review (read in full, card) extends the literacy frame and does not question it. + +### 18. "Algorithmacy" (Q5) +- **Uses by others found: one, Oberschelp (1995), in a different sense (see below); none in the sense the lab uses.** OpenAlex (title/abstract) 0; Crossref 0; WebSearch (extended) returned only Hunt's Substack posts, algorithmacy.org (the author's conference, 28-31 Oct 2026, Port of Spain) and two GitHub copies of the lab repository. +- **A prior printed use of the word: Oberschelp (1995), `oberschelp1995developing`**, IFIP chapter "Developing analytical competency through informatics teaching", DOI 10.1007/978-0-387-34842-1_12, p. 121 (read in full, card): "we have to preserve logicacy and algorithmacy by continuously training simple mental activities without artificial tools." The word occurs once and is not defined. It names a mental skill of carrying out algorithmic steps by hand, to be protected against tools that do the work; that is a different sense from the lab's, which is neither a capacity to navigate an algorithmic third party nor a coordination competence. +- **Author claim, unverified:** Hunt's Substack says the word appeared once before in an unpublished conference talk on algorithmic threat reasoning, citing Leese and Matzner (2016). I could not check it (WebSearch #101). +- **Competing definitions:** none found. "Algorithmic numeracy" and "algorithmic fluency" occur only in applied papers (OpenAlex #40). + +### 19. "Decision experience design" and variants (Q4) +- **Scholarly use of the exact phrase:** 0 in OpenAlex (two forms), Crossref, Consensus (this run covers Consensus through three other queries only; I did not rerun the 30 Sept Consensus queries). Unchanged from 30 Sept. +- **New since 30 Sept, near collisions (abstract or title only):** + - Morii (2026), `morii2026decision`, SSRN, DOI 10.2139/ssrn.6341998, "Decision Design as Judgment Architecture: Structuring Authority and Accountability in Human-AI Systems". No abstract in OpenAlex; not opened. **A prior or parallel use of "decision design" for human-AI systems; read before the paper uses that phrase.** + - Sun (2026), `sun2026decision`, arXiv, DOI 10.48550/arxiv.2604.00414, "Decision-Centric Design for LLM Systems". Abstract read: an architecture that separates decision signals from the action policy so control is "explicit and inspectable". Software architecture, not user experience. It does use "decision-centric design" for AI systems and shares the word "inspectable" with the legibility paradigm the paper opposes, but in Sun "inspectable" means that developers can read the logged signal and the policy for debugging (pp. 1-3, 8); no end user or person acted on appears in the paper (card; read in part, appendices B-C skimmed). + - Haid, Lang and Fottner (2023), AHFE, "Explaining algorithmic decisions: design guidelines for explanations in User Interfaces" (title only). +- **Practitioner use since 30 Sept:** I found none new. The author's own Substack post "Decision Experience Design" (published Part 1) is the only exact-phrase hit besides the seven practitioner uses F1 listed (G1-G7, not refetched). Kobetz's "DX Perspective" and Serrault's "AI & Decision Design" seminar (30 Jan 2026) were opened and do not use the term. +- **Acronym collision:** "DXD" is also "Digital Experience Design" (Gensler studio; job listings) and "Digital eXperience Design (DXD)" (Emakina, per AUDIT_03), "DxD Design for District" (2000-2001) and the pharmaceutical payload DXd. +- **Prior claim in decision-centred design:** Wolf, Klein and Thordsen 1991, O'Hare 1998, Thordsen 2003, Militello and Klein 2013 (from F1; keys `wolf1991decision`, `ohare1998cognitive`, `thordsen2003decision`, `militello2013decision`). I did not reopen them. The term is "decision-centered design", and F1 already concluded DXD must distinguish itself from it. +- **Prior claim in decision intelligence:** Kozyrkov: "turning information into better actions at any scale" (search snippet, secondary). Pratt: talk summary only. Neither uses "decision experience". + +### 20. Nadella responses (Q6) +- **Post:** `nadella2026insider`, x.com/satyanadella/status/2108931348857827686 (URL as surfaced by search; another agent covers the text). +- **Critiques found in print:** none on human-factors grounds. Forbes, `forbes2026emergency` (Zainescu, 2026-10-11), reports the post's argument and is not itself a critique: the essay "moves responsibility for AI safety from the company that builds a model to the company that deploys it" and policing models with models "can stack one opaque system on top of another" (snippet-level paraphrase from WebSearch; page returned 403). One source (not identified in the summary) calls the post a statement of intent without a technical specification [unverified: see AUDIT_03]. +- **Human-factors or design critique:** none found in tech press, security blogs or AI-safety forums. No hit from Hacker News, LessWrong, the Alignment Forum, Schneier or Lawfare. The search tools do not index X replies, so absence there is untested. +- **Related items that already carry the author's human-factors point without naming Nadella:** Green 2022; Mitchell 2026; Grunde-McLaughlin 2026; Zhu 2026 (entry 16). + +--- + +## (c) Synthesis + +### 1. Has anyone drawn the analogy? + +**(a) AI, LLMs, prompts, chat, agents.** I found no work that presents the move from *scriptio continua* to word separation as a design precedent for AI. The only AI-adjacent hits are tokenization papers (technical sense) and podcast episodes whose Zenodo records name Daniel Rosehill with Gemini and a TTS tool as co-creators (decorative). Printing-press analogies for AI exist (#83) but are not about word separation. The "chat is the command line" and "horseless carriage" analogies are common (entry 15, entries 9 to 12), but none reaches for writing history in the summaries and snippets I saw. + +**(b) Algorithms, platforms, recommender systems.** None found. Dembeck 2023 mentions algorithmic co-steering of "flow" in one relative clause (§4). + +**(c) HCI, UX, usability in general.** Two instances: Mueller 2026 (decorative; partly read, the final 22,799 characters unread) and Cho 2021 (literal spacing aid). The page-as-interface idea is in print without the analogy (Mak, project page; Illich, a critical notice; Drucker not checked). + +**(d) Code, data formats, programming.** None found. Search #67 returned only a forum thread on spaces and ambiguity. I did not search the camelCase versus underscore readability literature, which is the obvious place a programmer would look for word separation in identifiers (see (d)). + +**(e) Explainability, transparency.** None found. The OpenAlex hit #15 returned no XAI work; #22's three full-text hits (Syauqi 2025, two Niwa records) were not opened. + +Summary: **zero uses of the analogy as a design precedent, one adjacent use for digital literacy (Dembeck), decorative mentions in Mueller 2026 and three Rosehill episodes, and one literal spacing aid (Cho 2021).** + +### 2. The page as the origin of UX + +No source I found calls the medieval page the origin of UX. Mak (2011) treats the page as an interface that has developed since the scroll (project description; book not read), Drucker's *Graphesis* (2014) treats the interface as a visual form of knowledge production (blurb) [unverified: see AUDIT_03], and Getty's 2023 exhibition calls scribes early graphic designers (press page). Illich ties layout to a new technology of reading, and a WebSearch summary of a notice of his book (reported in AUDIT_03) says he extended the argument to the screen replacing the page. I verified existence but did not read these books; their claims here are from summaries. Candidates in the brief that I did not search in a way that returned evidence: Piper, Manovich, Bolter and Grusin (record verified only), Kirschenbaum, Grafton, Blair, Houston, Bringhurst, Norman. Norman's *Things That Make Us Smart* (not the 2013 *Design of Everyday Things* listed in (d) 9) exists (reviews in Crossref) but I did not test it for writing history. The claim "the author has the first human-factors reading of the page" cannot be made on this evidence; "the first reading that I found" can. + +### 3. Pre-conventional AI interfaces + +The nearest are Nielsen 2023 (the third-paradigm article, which says half the population in rich countries is "not articulate enough" for current AI bots; "articulation barrier" is the title of a separate Nielsen article), Wattenberger 2023, Appleton 2023, Koomen 2025 and Litt 2025. Strongest for the author's purposes: Koomen, because the agent acts for the user, and Nielsen, because his barrier is a literacy barrier. They differ in four ways: (1) one user and a tool, never two people and a mediator (Litt 2025 has a communal-creation section on shared tools for groups); (2) the remedy is more controls, a better paradigm or a malleable tool, not a convention that takes work off the person; (3) no history of reading or writing; (4) none uses "incunabula", "pre-Gutenberg" or "scriptio continua" in the summaries and search snippets I saw (searches #69, #83, #84). The deRSE25 workshop paper (eceasst 2716) and a generative-interfaces preprint (arXiv 2508.19227) also treat chat as an unsettled form (snippets only). + +### 4. The term DXD + +No scholarly use of the exact phrase. The new collisions are Morii 2026 ("decision design", unopened) and Sun 2026 ("decision-centric design", software architecture). The practitioner label from F1 stands. The decision-centred design lineage (Klein, Militello) holds the older claim to design organised around decisions (my gloss). + +### 5. "Algorithmacy" + +Not only the author: Oberschelp (1995, p. 121) prints "logicacy and algorithmacy" once, in a different sense (a mental skill of working through algorithmic steps unaided, kept alive by tool-free practice). One author-reported earlier mention (Leese and Matzner 2016) remains unchecked. The literacy-critique family exists (Seaver, Bucher, Cotter and Reisdorf, Gran et al., Dogruel, Oeldorf-Hirsch and Neubaum, Koenig), but I did not read them this run, so I cannot say which would call literacy the wrong frame. + +### 6. Nadella responses + +I found no published critique that targets the post on human-factors grounds. Forbes (Zainescu 2026-10-11) summarizes the post's accountability shift. No human-factors critique in the sources searched. + +### 7. Oversight remedies reproduce a reading model + +The prediction "human oversight as currently designed fails and needs redesign or institutional backing" is established (Green 2022 onward); none of these sources replaces the human check with structure, with two qualifiers from the full texts: Zhu et al. (2026) push much of the checking to design time and sampling, so structure does replace some run-time checks, and Mitchell et al. (2026) include structural items in their inventory (bounded autonomy, automated pre-checks, separated roles), so they sit closer to the structural camp than this summary suggests. The sharper claim, that audit logs ask people to read and so repeat the literacy model, is not in the abstracts I saw, and I found no "grammar of legibility" use for AI. + +### (i) Nearest prior uses, ranked + +| Rank | Source | Closeness | What it leaves undone | +| --- | --- | --- | --- | +| 1 | Dembeck 2023 | Adjacent: Illich's *scriptio continua*-to-page history applied to digital literacy; says the book-literacy goals inherited by media-literacy education may miss the digital case (§5, hedged: "könnte") | No algorithmic third party; no design move; no human-factors frame; German-language, small venue | +| 2 | Lermann Henestrosa et al. 2026; Zhu et al. 2026; Green 2022 | Adjacent in remedy: from oversight to "structurally embedded control" (the abstract's remedies are incentive-based stress testing and normative clarity) or institutional oversight | No history of reading; no triad; none addresses Nadella | +| 3 | Mak 2011; Illich 1993 | Adjacent in frame: the page as interface or technology of reading | Not about AI or algorithms; not word separation as a design move (read status: summary only) | +| 4 | Koomen 2025; Nielsen 2023 | Adjacent in claim: AI interfaces await a new design; articulation as barrier | One user and an agent or tool; no history; no convention proposal | +| 5 | Buschek 2024 | Method neighbour: literary concepts (collage) from avant-garde writing inform AI-writing-tool design | Collage, not word separation; one writer | +| 6 | Wattenberger 2023; Appleton 2023; Litt 2025 | Anti-chat | Mostly a single user (Litt 2025 has a communal-creation section on shared tools for groups); no history | +| 7 | Mueller 2026 | Decorative mention | Typography only | +| 8 | Rosehill 2026 (human host with Gemini and TTS co-creators) | Decorative | Popular summary of Saenger | +| 9 | Fujii 2023; Clérice 2020 and other NLP | Technical sense | Not an analogy | + +### (ii) Draft gap statements + +*Drafted by Claude for the author to accept, change or discard.* + +**Strong form (what the searches support):** + +> On 12 October 2026 I searched OpenAlex, Crossref, Consensus and the open web (WebSearch). In OpenAlex title and abstract (queries 1-3, 5-7, 9-13, 15-17) I searched "scriptio continua" and "scriptura continua" alone and with algorithm, language model, chatbot, prompt, artificial intelligence, machine learning, interface, usability, user experience, typography, human-computer, code, programming, whitespace, syntax, platform, recommender, social media, digital, design, transparency, explainability, opacity, silent reading, Saenger, human factors, ergonomics and cognitive. In OpenAlex full text (queries 18-23, first ten hits each) I searched "scriptio continua" with "large language models", chatbot, "user interface", "human-computer interaction", "explainable AI" and algorithmic; the full-text queries for "interface design", "user experience" and "algorithmic literacy" were not run (query 24), and prompt, platform, usability and code were not run in full text. In Crossref (queries 47-51), Consensus (56-57) and WebSearch (65-76) I searched "scriptio continua" or "scriptura continua", or "word spacing" or "space between words", with AI, language model, prompt, chatbot, user interface, interface design, human-computer interaction, silent reading, Saenger, programming, whitespace, code, agents, algorithms and platforms. Four OpenAlex queries returned errors (4, 8, 14, 28). I found no work in these results that presents the move from continuous script to word separation as a design precedent for AI interfaces, algorithmic intermediaries or human-factors practice. The uses I found are technical (tokenizing unspaced scripts), decorative (a typography essay; podcast episodes whose records name a human host with Gemini and a TTS tool as co-creators), or adjacent: Dembeck (2023) draws on Illich's account of the page to argue about digital literacy, with algorithmic co-steering of the flow mentioned once and not developed. + +**Algorithmacy (add to the gap statement wherever it names the term):** + +> The word "algorithmacy" is not new. Oberschelp (1995, p. 121) prints "logicacy and algorithmacy" once, in a different sense: a mental skill of carrying out algorithmic steps without tools, to be preserved by tool-free practice. I found no other use. The sense used in this paper is a different one. + +**Fallback if the author or a reviewer accepts Dembeck as a close precedent:** + +> Dembeck (2023) uses the same Illich-based history of the page to question whether book-bound literacy is the right frame for the digital. I found no work that carries the history of word separation into the design of coordination through an algorithmic third party. + +**Limits the sentence must carry:** Semantic Scholar (#60), the arXiv API (#61) and Scholar Gateway (#59) were not available in this run, and Google Scholar was not searched (the audit's rerun of five arXiv API queries on 12-13 October 2026 found 0 hits; AUDIT_03); OpenAlex's full-text leg was cut short (#24); four OpenAlex queries returned errors; X, Hacker News and podcasts are not searchable by these tools. Do not write "no one has". + +### (iii) Sources the author must cite as predecessors + +1. Saenger 1997, Parkes 1992, Illich 1993 (the history itself; see facet A). +2. Dembeck 2023 (closest digital-literacy use of the same history). +3. Mak 2011 (page as interface) and Drucker 2014 (interface as visual knowledge), after reading. +4. Nielsen 2023, Wattenberger 2023, Koomen 2025 (current AI-interface critiques the paper should position itself against). +5. Green 2022, Lermann Henestrosa et al. 2026, Zhu et al. 2026, Mitchell et al. 2026 (oversight critique). +6. Morii 2026 and Sun 2026 (name collisions), after opening. +7. Wolf, Klein and Thordsen 1991; Militello and Klein 2013 (decision-centred design). +8. Ananny and Crawford 2018 (transparency ideal); Seaver, Bucher, Gran et al. (algorithmic literacy family). + +### (iv) What undercuts a novelty claim + +- **Dembeck 2023** already joins the history of the page to the question of digital literacy and doubts that book-bound literacy fits the digital. +- **Saenger's own blurb** states the cost-shift in nearly the author's terms: "the task of separating the words in continuous written text, which for half a millennium had been a function of the individual reader's mind and voice, became instead a labor of professional readers and scribes" (publisher abstract of *Space Between Words*, Consensus record, DOI 10.1515/9781503619081). That supports H-cost and is not the author's invention. The same blurb says "greater diffusion of literacy" was not seen as an advantage by the ancient world, which also bears on H-cause. +- **Oversight critique is already established** (Green 2022 and after). A human-factors critique of observe-log-audit regimes is not new as a direction; I found none addressed to the Nadella post. It may be new as an application to the post and as a reading-model diagnosis. +- **"Algorithmacy"** has an earlier printed use: Oberschelp 1995, p. 121, in a different sense (read in full; card). The Leese and Matzner 2016 claim is unchecked. +- **Name collisions** with "decision design" and "decision-centric design" (2026). +- **The page-as-interface idea** is in print (Mak, Illich, Drucker), so "history of the page as human factors" is not new in general form. + +### Claims the sources contradict or complicate + +- That nobody has linked *scriptio continua* to digital literacy: Dembeck 2023 has. +- That "decision experience design" is entirely unclaimed by neighbours: "decision design" (Morii 2026) and "decision-centric design" (Sun 2026) appear in AI contexts. +- That the "rewired cognition" line is safe: Rosehill's episode 2729 (human host with Gemini and TTS co-creators) repeats it [quote unverified: see AUDIT_03]; the standing ruling excludes it. + +--- + +## (d) Gaps and unverified leads + +1. **Morii 2026 (SSRN 6341998).** No abstract; open it. Library or SSRN download. +2. **Oberschelp 1995** (Springer, IFIP, DOI 10.1007/978-0-387-34842-1_12). Resolved: read in full (card); the word appears once, on p. 121, in a different sense. +3. **Leese and Matzner (2016)** unpublished talk cited in the author's Substack. Ask the author for the citation. +4. **Cotter and Reisdorf (2020), "Algorithmic knowledge gaps"** (*International Journal of Communication* 14). No Crossref record located; fetch from ijoc.org. +5. **Semantic Scholar, arXiv search, Google Scholar, Scholar Gateway.** Blocked (429 or access denied). Rerun when the limits clear: `"scriptio continua"` with LLM, interface, UX; `"decision experience design"`; `algorithmacy`. +6. **OpenAlex full-text leg** stopped at the free budget. Rerun #24 with an API key. Also rerun #28 with parentheses fixed. +7. **Social media.** X threads, Hacker News comment search (Algolia), Reddit, podcasts and conference talks are not covered. A person should search Algolia HN for "scriptio continua" and X for "spaces between words" with AI terms. +8. **Programming angle.** I did not search the camelCase versus underscore identifier readability literature (e.g., eye-tracking studies of identifier styles) or minified code as unspaced text. That is where a "word separation in code" paper would sit. +9. **Books not opened:** Mak 2011, Drucker 2014, Illich 1993, Manovich, Bolter and Grusin 1999, Kirschenbaum, Piper, Houston, Bringhurst, Norman 2013 (*The Design of Everyday Things*, a different book from *Things That Make Us Smart*). Needs library access. Mak and Illich are the priority. +10. **Mueller 2026**: unread final 22,799 characters; the essay could link spacing to AI. +11. **Pages read only through a fetch summary** (Wattenberger, Appleton, Koomen, Nielsen, McGinley): re-read the raw pages before quoting. Nielsen's date is settled by AUDIT_03: Substack mirror 22 June 2023, UX Tigers original 20 May 2023 (search snippet). +12. **Snippet-only items:** Forbes (403); uxdesign.cc; Tunguz; Fazi; AltaML; Singapore Consensus 2026 (arXiv 2608.14611); "Legible AI by Design" (Lancaster); "Agency Among Agents" (arXiv 2507.23585); Korean recommender-literacy study; Litt 2025. None may be relied on. +13. **Not run:** DXD in LinkedIn and Medium for items after 30 September beyond what WebSearch indexed; the seven F1 practitioner URLs were not refetched. +14. **Not verified:** Pratt's definition of decision intelligence; Kozyrkov's exact wording (secondary). +15. **Linus Lee** essay on interface thesis not located; **Bret Victor** on LLMs not found. +16. **Beyond-the-brief candidates not searched at all:** McLuhan; Kittler; Hutchins on cognitive artefacts; Eisenstein on print and typographic fixity (facets C and E cover these). + +## Changes after full-text cards (2026-10-13) + +- `oberschelp1995developing`: the "unresolved database hit" is resolved. The chapter prints "logicacy and algorithmacy" once, p. 121, in a different sense (a mental skill of carrying out algorithmic steps without tools, preserved by tool-free practice); "Uses by others found: none" corrected; section 5 summary, the novelty-undercut list and the open-leads item updated; a plain statement of the prior use added to the draft gap statements (marked as drafted by Claude). +- `meyman2026impossibility`: "a formal claim that observability cannot produce ex-ante authorization" replaced; version 2.0.0 withdraws the universal claim and keeps a conditional result on a stipulated model; self-published note by an author affiliated with FERZ, Inc. +- `sun2026decision`: "inspectable" means readable by developers (pp. 1-3, 8); no end user appears; read status "read in part (card)". +- `litt2025malleable`: "one person and their tools" corrected (a "Communal creation" section); read status "read in full (card)"; the summary table row and the "one user and a tool" difference list carry the qualifier. +- `mitchell2026agents`, `zhu2026designing`: structural items in Mitchell's inventory; Zhu moves checking to design time and sampling (pp. 5-6); "none of these sources replaces the human check with structure" qualified in the entry and in section 7. +- `lermannhenestrosa2026human`: added the section 3.3 point (close to ruling D3, argued on regulatory grounds). +- `buschek2024collage`, `cho2021giving`, `butskhrikidze2022what`, `fujii2023how`, `clerice2020evaluating`: read in full; Buschek's history is literary Expressionism and Manovich 1999 and he limits the claim to interface form; Cho's body matches the abstract; Butskhrikidze is a typology paper not NLP; Fujii tested Japanese only; Clérice's dating aside noted. +- `wattenberger2023chatbots`, `appleton2023sketchbook`, `koomen2025horseless`, `nielsen2023paradigm`, `cordeiro2026control`, `dasgupta2023designing`, `oeldorfhirsch2023what`, `dembeck2023flow`: read status updated to "read in full (card)"; Wattenberger wants the interface to carry the user's burden and never mentions an interested third party; Appleton's aim is to keep the writer's reasoning active; Cordeiro is a newsletter programme statement; Dasgupta's learners are children; Oeldorf-Hirsch and Neubaum extend the literacy frame. + +Synthesis verdicts changed: (1) section 5 "Algorithmacy": "Only the author uses it. One unresolved database hit ... remain unchecked" became "Not only the author: Oberschelp (1995, p. 121) prints 'logicacy and algorithmacy' once, in a different sense". (2) Section 7 and entry 16 bearing: "none of these sources replaces the human check with structure" now carries two qualifiers (Zhu, Mitchell). (3) Section 4 differences list: "(1) one user and a tool" qualified for Litt 2025. (4) Novelty-undercut bullet: "Algorithmacy may have an earlier use ... Unresolved" became "has an earlier printed use ... in a different sense".