Study 019: freeze — authorship across representations (12 review rounds; freezable as written)

.github/workflows/ci.yml

@@ -350,6 +350,58 @@ jobs:
         working-directory: studies/018-transition-rules
         run: python -m pytest harness/tests -q
 
+  study-019-harness:
+    name: Study 019 · deterministic harness
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out source
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - name: Set up Python
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
+        with:
+          # The fixed CI runtime, and the honest statement of what it is
+          # (round-5 finding R5-5). harness/PINS.json registers the CPython
+          # 3.12 SERIES and integrity.verify_interpreter() enforces
+          # implementation and series; the patch level is deliberately
+          # reported at run time and not required (Study 012's round 3,
+          # finding 20). Naming 3.12.11 here fixes the runner so this job is
+          # reproducible, and it is not a claim that another patch would be
+          # refused.
+          python-version: "3.12.11"
+      - name: Install pytest
+        # Fully offline apparatus: the harness, the design generators and the
+        # controls are stdlib-only, so pytest is the whole install. No evaluator
+        # binary, no external clone.
+        run: python -m pip install --disable-pip-version-check pytest
+      - name: Verify the port chain, the pins and the study manifest
+        # `integrity.py` refuses to run without PYTHONSAFEPATH=1: invoking a
+        # script by path puts that script's own directory first on `sys.path`
+        # before any byte of the file runs, so the operator applies -P and this
+        # step establishes that they did.
+        working-directory: studies/019-authorship-across-representations
+        env:
+          PYTHONSAFEPATH: "1"
+          PYTHONDONTWRITEBYTECODE: "1"
+        run: python harness/integrity.py
+      - name: Run the harness, design and control suite
+        # Deterministic and offline: the registered partition and its per-code
+        # reachability, the transcript binding, the leak-token screen, the
+        # scorer's population and publication rules, the ports chain, the
+        # manifest's currency with the tree, and the currency assertions that
+        # recompute every count the preregistration states from the artifact it
+        # states it about. THE MATRIX ADJUDICATION NEVER RUNS HERE: it is an
+        # ATTEMPT, not a test (§7), and neither it nor anything that invokes
+        # `codex`, `jpack` or `opa` is part of this job. The engine-backed
+        # pipeline module skips itself unless the pinned binaries are present
+        # and hash to their pins, which in CI they are not.
+        # `PYTHONDONTWRITEBYTECODE` is not decoration: `integrity.verify()`
+        # refuses any `.pyc` the running interpreter did not produce, so a test
+        # run that writes bytecode would break the step above on the next run.
+        working-directory: studies/019-authorship-across-representations
+        env:
+          PYTHONDONTWRITEBYTECODE: "1"
+        run: python -m pytest harness/tests -q
+
   python:
     name: ${{ matrix.os }} · Python ${{ matrix.python }}
     strategy:

README.md

@@ -61,6 +61,7 @@ records; it remains the canonical matrix, with the external source each study bu
 | [016](studies/016-policy-currency-anchor/) | Interoperability | Frozen + run — R1 holds (both strata) | [`ANALYSIS.md`](studies/016-policy-currency-anchor/ANALYSIS.md), [`results/`](studies/016-policy-currency-anchor/results/) |
 | [017](studies/017-witnessed-currency/) | Currency governance | Frozen + run — R1 holds (both strata) | [`ANALYSIS.md`](studies/017-witnessed-currency/ANALYSIS.md), [`results/`](studies/017-witnessed-currency/results/) |
 | [018](studies/018-transition-rules/) | Currency governance | Frozen + run — R1 holds; reviewer holdout diverged on three preregistered cells | [`ANALYSIS.md`](studies/018-transition-rules/ANALYSIS.md), [`results/`](studies/018-transition-rules/results/) |
+| [019](studies/019-authorship-across-representations/) | Blinded authorship | **Design draft — not preregistered** | [`PREREGISTRATION.md`](studies/019-authorship-across-representations/PREREGISTRATION.md) (draft), [`design/`](studies/019-authorship-across-representations/design/) |
 
 This repository claims **no JPS conformance** for anything in it, and the table above adds no
 aggregate headline: each study answers a different preregistered question and must be read with

studies/019-authorship-across-representations/.gitignore

@@ -0,0 +1,2 @@
+__pycache__/
+.pytest_cache/

studies/019-authorship-across-representations/CORRECTION-TARGETS.md

@@ -0,0 +1,76 @@
+# Correction targets — Study 019 (registered before the freeze)
+
+Preregistration §10 pins, before the freeze, where a correction must land and what it must
+correct — verbatim wording, venue, URL, retrieval date — so that a failed, corrected, or
+retracted claim has a pre-committed place to be corrected IN, with the same prominence as
+the claim (the Study 012 `CLAIM.md` discipline; round-7 finding R7-9 is why this document
+is a registered, freeze-gated artifact rather than a declared intention).
+
+The study publishes in-repository. Every venue below is a file in
+`Judgment-Pack/judgment-pack-evaluator-experiments`, referenced at `main`; URLs retrieved
+2026-08-19. A correction to any target lands in the SAME file, at the SAME prominence
+(head-of-file banner for documents; the row itself for index rows), and — for a corrected
+or retracted R1 — additionally as a banner at the head of `ANALYSIS.md` and an entry in
+`DEVIATIONS.md`, which is freeze-excluded precisely so it can receive one.
+
+| # | Claim that may need correction | Venue | URL (retrieved 2026-08-19) |
+|---|---|---|---|
+| T1 | The R1 verdict sentence `ANALYSIS.md` will publish — one of the §5 decision table's registered rows, quoted verbatim from `RESULTS.json`'s `verdict` member at the primary attempt | `studies/019-authorship-across-representations/ANALYSIS.md` (does not exist until the attempt; the target binds the file by registered path) | https://github.com/Judgment-Pack/judgment-pack-evaluator-experiments/blob/main/studies/019-authorship-across-representations/ANALYSIS.md |
+| T2 | The study-index row — Question, Theme, Status columns — whose Status cell will carry the R1 verdict | `studies/README.md`, the Study 019 row | https://github.com/Judgment-Pack/judgment-pack-evaluator-experiments/blob/main/studies/README.md |
+| T3 | The repo-root index row for Study 019 (same Status discipline) | `README.md`, the Study 019 row | https://github.com/Judgment-Pack/judgment-pack-evaluator-experiments/blob/main/README.md |
+| T4 | The study README's rendered round-status sentence and its verdict history, should any round's record require correction | `studies/019-authorship-across-representations/README.md` | https://github.com/Judgment-Pack/judgment-pack-evaluator-experiments/blob/main/studies/019-authorship-across-representations/README.md |
+| T5 | The pilot readings quoted in the preregistration's Design provenance (A 0.878, B 0.897, C 0.806; high-kill 1/5, 0/5, 0/5) — non-citable, but published in a frozen document and correctable only by `DEVIATIONS.md` entry once frozen | `studies/019-authorship-across-representations/PREREGISTRATION.md`, Design provenance | https://github.com/Judgment-Pack/judgment-pack-evaluator-experiments/blob/main/studies/019-authorship-across-representations/PREREGISTRATION.md |
+
+The verbatim wording of T1 cannot exist before the attempt; what is pinned is its closed
+vocabulary — the §5 decision table's row strings as implemented by
+`harness/e4lib/decision.py` and published by `harness/score.py`, which are freeze-covered
+bytes — and the rule that ANALYSIS.md quotes the published string unedited. For T2–T4 the
+current wording is whatever those files carry at the freeze commit; the correction
+obligation attaches to the row/sentence, not to a snapshot of it. A correction is written
+in every branch of §5's outcome space, including "no correction needed" being visibly
+distinguishable from "correction owed and absent": if the primary attempt publishes
+without incident, `CORRECTION.md` is still created, stating that no target required
+correction — an absent file is a failure to publish, never an outcome.
+
+## Target bound by the parallel ceremony line (merged 2026-08-19, both lines' work kept)
+
+The parallel fill line searched for claims beyond the T1–T5 in-repo surfaces and bound one
+this table lacked, kept here verbatim:
+
+## Target 1 — the design-phase pilot's directional read
+
+**Venue:** `design/mutants/E4-NOTES.md` in this repository — a committed design record on
+`main` of `Judgment-Pack/judgment-pack-evaluator-experiments`, which is a published venue
+in exactly the sense Study 012's first target was: public, attributable, and quotable by a
+reader who never opens the harness.
+
+**URL (commit-pinned):**
+<https://github.com/Judgment-Pack/judgment-pack-evaluator-experiments/blob/c7ce3269c3d57d29da957b2c2e3556dcbf7e80df/studies/019-authorship-across-representations/design/mutants/E4-NOTES.md>
+
+**Retrieved:** 2026-08-19, from the working tree at that commit.
+
+**Verbatim wording:**
+
+> **E4 discriminates, and in this pilot the direction is B/C above A.** The earlier surface
+> read (35–49 authored rows vs 1–4 test rules) was misleading: the Rego test rules are
+> table-driven and carry many assertions. Small N; non-citable; but the endpoint has
+> headroom and variance, which is what the pivot needed.
+
+**What would require correcting it:** the sentence already labels itself small-N and
+non-citable, so a divergent registered result does not falsify it — but a registered result
+in the SAME direction must not be reported as confirming a prediction, because the
+preregistration (§11) states the anchor "no longer supports either way" after the reference
+repair. The correction this file binds is therefore about *use*, in either direction: if
+any post-freeze surface cites the pilot direction as evidence — for A or against A — the
+correction lands at the head of `CORRECTION.md`, in this venue, quoting this wording.
+
+## The recorded decision: no external venue carries a bound claim
+
+Searched before the freeze: the maintainer's external posts on the program (dev.to,
+retrieved 2026-08-19) concern the evaluator's authorship-coverage line (Studies 009–012)
+and the trust-input line; none asserts a claim about what a representation's authored test
+suite pins — this study's subject. Study 012's own published claim has its own correction
+machinery (`studies/012-policy-perturbation/CORRECTION.md`) and is not re-bound here. If a
+reviewer of the freeze PR knows an external claim this list misses, adding it is a one-line
+edit to this file **before** the squash-merge; after the freeze, a missed target is a
+DEVIATIONS.md entry, not a quiet edit.

studies/019-authorship-across-representations/DEVIATIONS.md

@@ -0,0 +1,6 @@
+# Deviations — Study 019
+
+Deviations from the frozen preregistration land here with a reason and a date — never by
+editing the preregistration or any frozen artifact. Nothing is frozen yet.
+
+(none)