@@ -1,3 +1,19 @@
+## 2026-07-28 — PROBE-BABEL-STANCES slice 2: two Rosetta stones (Opus orchestrator + 3 Sonnet workers)
+
+- **Model split per operator directive** ("cheaper Sonnet 5 agents for grindwork, opus for filigrane planning"): three Sonnet workers (fixture-field augmentation, CI wiring, board hygiene) under the §1 guardrails preamble; Opus designed the grid/stone/channel machinery, ran all gates, adjudicated, and is sole writer of this log.
+- **Sonnet W1 (workflow `babel-slice2-grindwork`):** added `pos`/`morph`/`prag` level-split fields to all 30 `lanes.rs` rows; flagged one deviation honestly (`morph: "er|kennen"` concatenates to `erkennen`, which does not equal the row's stored `root: "kennen"` — the bare semantic stem vs the inseparable-prefix lemma). Left the pre-existing cs-bkr DIE root deviation untouched as instructed.
+- **Sonnet W2 (same workflow):** wired both probe examples into `.github/workflows/rust-test.yml` — closing codex P2 (`cargo test` never executes an example's `main()`, so every `assert!` in these probes had only ever run locally). Read the whole workflow first and chose the general `test` job over the Codecov-only job, with rationale.
+- **Sonnet W3 (board hygiene):** the new `E-TWO-ROSETTA-STONES-AND-THE-FOUR-CHANNEL-SPLIT-1` entry + a STATUS_BOARD section. It also inserted a full ⊘ RETRACTED block into the merged slice-1 entry; CodeRabbit correctly flagged that as an append-only violation (prior entries may amend only Status/Confidence), so the orchestrator collapsed it to a Status-line amendment pointing at the new entry.
+- **Opus adjudication — the retraction, not a patch.** Codex's P1 was correct and headline-invalidating: the coordinate was a LEMMA while the file's own `senses` field already recorded `know.v.01` ≠ `know.v.05`, so slice 1's "residual" measured English surface-lemma reuse. Second P1: `en-kjv` polluted its own resultant. Rebuilt as a three-axis synset×POS×frequency-band convergence with en excluded from every resultant; the finding INVERTS (the grid sees more than most lanes) rather than being repaired in place. **Second instance in one arc of measuring an artifact of my own construction** — the standing lesson is that a grid needs an axis independent of the thing it measures BEFORE a residual is claimed.
+- **Built:** per-lane `LanguageStone` (private `BeliefArena` + interner, one writer per stone, DN-shaped term paths, CHECK rows at c = 0.5); the 4-channel phase split (morph/syn/sem/prag) replacing pooled divergence; valency frames; the passion composite; a CHAODA-shaped population-relative outlier read.
+- **Measured (asserted, verified rows):** German alone lexically tracks KNOW's two grid coordinates (three channels at once: morph π, syn π, sem 0.833π); Latin silent on all three (`co|gnosco` both); pragmatic channel antiphase in **all six** lanes at coherence 1.000 (the ידע calque — translationese, not convergence) and silent at NAKED/DIE; sem-channel coherence KNOW 0.413 vs NAKED 1.000; passion peaks (German, KNOW) = 0.3130 and is structurally zero elsewhere. Czech DIE morphology-only (sem 0.000, morph π) is CHECK-row-dependent → report-only.
+- **Two defects self-caught mid-run:** (a) the outlier read divided by a MAD of exactly 0, printing robust-z 2.6e6 — replaced with a degenerate-population branch + strict-minority assert; (b) the quale term conflated **surprise with uncertainty** (a CHECK row's low confidence produced quale 0.25 with nothing seen, ranking two reconstruction lanes above four verified ones) — passion now runs over VERIFIED cells only; a confidence-independent surprise term is queued.
+- **Honest deferrals:** the `prefix|stem` carving is typologically biased (a Finnish suffix switch is invisible to it) — spine slot 6 left RESERVED rather than filled with a reconstruction; full HHTL+helix/CLAM routing needs a trained lexeme codebook (30 rows are not a codebook) → slice 3.
+- **Gates:** `cargo fmt --check` clean · `clippy -p lance-graph-planner --examples` zero warnings · `cargo test -p lance-graph-planner` green · both probe examples green (and now in CI).
+- **Branch note:** PR #861 merged mid-work, so the branch was restarted from `origin/main`; this is a NEW PR, not a continuation.
+- **Review round (same PR, second commit pair):** codex P2 — the pragmatic-calque gate iterated EVERY occupied lane, so the two CHECK reconstructions (cs-bkr, arc-onkelos) were being ASSERTED on; with the example now running in CI that promoted unverified philology to a release gate. This is the slice-1 DIE defect one channel over. Fixed: a single `pair_verified` gate, asserts over verified pairs only, CHECK lanes printed as consistent-but-not-claimed, and the "all six traditions" headline softened to the four verified lanes. CodeRabbit — (a) append-only violation: a 22-line ⊘ RETRACTED block had been inserted into the MERGED slice-1 entry, where the rule permits only a Status/Confidence amendment; collapsed to a Status-line amendment pointing at the new entry, which carries the full retraction. (b) STATUS_BOARD newest-first ordering + a worker-recorded "dashboard exception" that the rule does not support — section moved to the top, exception removed. (c) stale `#861` provenance in the agent-tag file (the PR flipped to #862 mid-work when #861 merged) and non-canonical `examples/…` paths — corrected. (d) nitpick: an `unwrap` in the prag loop that would panic the moment the RESERVED Finnish slot is minted — now follows the file's silent-lane convention. Sonnet took the two mechanical board files; the append-only judgment and the code were main-thread.
+- **Board (this commit):** EPIPHANIES `E-TWO-ROSETTA-STONES-AND-THE-FOUR-CHANNEL-SPLIT-1` + ⊘ retraction inside `E-THE-GRID-COLLAPSES-WHAT-A-LANGUAGE-SPLITS-1` · STATUS_BOARD PROBE-BABEL-STANCES section. Task #52.
+
 ## 2026-07-28 — PROBE-BABEL-STANCES slice 1 (Opus orchestrator + 1 Sonnet curator)
 
 - **Model split per operator directive** ("Sonnet 5 agents for grindwork, opus for filigrane planning"): Sonnet curated the 5-lane verbatim lexicon under the §1 guardrails preamble; Opus designed the grid/phase/amplitude machinery, gated, and adjudicated.
 
@@ -1,6 +1,34 @@
+## 2026-07-28 — E-TWO-ROSETTA-STONES-AND-THE-FOUR-CHANNEL-SPLIT-1 — PROBE-BABEL-STANCES slice 2: two Rosetta stones (a CORPUS convergence grid + per-lane LANGUAGE stones that LEARN), a four-channel phase split (Morphologie/Syntax/Semantik/Pragmatik) that replaces pooled divergence, and a tracking census that INVERTS the slice-1 headline — only **German** lexically tracks the grid's two `KNOW` coordinates, and the pragmatic channel's coherent antiphase (1.000 across all six lanes) exposes translationese: a shared calque, not independent convergence.
+
+**Status:** IN PR (`probe_babel_stances`, self-asserting, running in CI via `.github/workflows/rust-test.yml` — closing slice 1's P2 finding that the example's asserts never ran). **Confidence:** High on VERIFIED-row asserts (the falsifier gates on them); the Czech `DIE` morphology-only reading is explicitly CHECK-row-dependent and printed report-only, never asserted.
+
+**Retracts and replaces slice 1's headline mechanism** (the slice-1 entry's Status line below is amended to RETRACTED and points here; per the append-only rule only that line was touched — the retraction is stated in full HERE, not inside the historical entry). Codex flagged two P1s: the coordinate was a LEMMA where WordNet's own `senses` field distinguishes `know.v.01` from `know.v.05` as separate synsets, so the "residual" measured English surface-lemma reuse, not grid escape; and `en-kjv` sat inside its own resultant, contributing a guaranteed-constructive phasor to a measurement made against itself. Both are structurally impossible in slice 2.
+
+**Stone 1 — the CORPUS stone: a stratified convergence grid.** Per the operator's steer (*"every other codebook also has frequency, POS and meaning — when you use POS and frequency and WordNet you at least have a convergence"*), a coordinate is the CONVERGENCE of three independently-motivated axes: synset (WordNet — meaning), POS (syntactic category), frequency band (COCA-shaped `top100`/`common`/`rare`). Two verse addresses are distinct coordinates only when ≥ 2 of the 3 axes agree they differ. `KNOW`'s two addresses differ on synset AND band (the carnal sense is archaic-rare); `NAKED` and `DIE` differ on neither — the grid genuinely files them together, so lane silence there is the EXPECTED reading, not a defect. This is the mechanism that inverts the headline: **the grid now sees MORE than most lanes**, and the live question becomes "which lane TRACKS the grid's distinction, and at which linguistic level?"
+
+**Stone 2 — the LANGUAGE stone: per-lane linguistic structure that LEARNS.** Per the operator's steer (*"the second Rosetta stone is for linguistic representation and relationships WITHIN a language, and it learns while using the corpus Rosetta stone"*), each lane owns a private `BeliefArena` (its own interner, its own stamps — one writer per stone, the mailbox discipline applied to reasoning state). Traversing the corpus stone, a lane admits within-language `Sim` beliefs between its two surface occurrences at each contrast, ONE PER CHANNEL (`sem`/`morph`/`syn`/`prag`), CHECK rows entering at reduced confidence `c = 0.5` so curatorial uncertainty is carried in the belief, never laundered. `en-kjv` NEVER enters any cross-lane resultant (the slice-1 pollution fix); it appears only in the tracking census and owns a language stone like any lane.
+
+**The four-channel phase split (Morphologie/Syntax/Semantik/Pragmatik).** Divergence is now ATTRIBUTED per level, not pooled — each φ ∈ [0, π] graded by the slice-1 LCS-radix fix (`φ = π·(1 − shared_path)`, kept verbatim). Measured: German `KNOW` (`gewar` adj ↔ `erkandte` v) diverges on all THREE live channels — φ_sem = 0.833π (suppletive stems), φ_syn = π (adj vs verb), φ_morph = π (∅- vs er- prefixation) — VERIFIED, asserted. Latin `KNOW` (`co|gnosco` ↔ `co|gnosco`) is silent on all three — same prefix, same stem, nothing to attribute. Czech `DIE` (`u|mříti` ↔ `ze|mříti`) is morphology-ONLY — stems identical (φ_sem = 0), perfectivizing prefixes wholly divergent (φ_morph = π); a pooled measure would have read this as "0.286π mild semantic divergence" (slice 1's number), the split shows it is pure aspect morphology. **This Czech reading is CHECK-row-dependent — report-only, never asserted**, same status as slice 1's honest-abstention treatment of the row.
+
+**Translationese finding.** `KNOW`'s pragmatic channel is antiphase (φ_prag = π) in EVERY one of the six lanes, cross-lane coherence 1.000: every tradition, English included, carries the Hebrew ידע euphemism at Gen 4:1 as a calque. **Coherent antiphase = inheritance of one convention, not independent convergence** — the pragmatic channel stays silent at `NAKED`/`DIE` (fires-and-can-stay-silent, per the falsifiability rule's twin test).
+
+**Tracking census — the inversion, measured.** Of the five non-English lanes, exactly ONE (German, Luther 1545: `gewahr` adj vs `kennen` v) lexically tracks the grid's two `KNOW` coordinates. sem-channel coherence: `KNOW` 0.413 (contested — the incoherence IS the German signal) vs `NAKED` 1.000 (agreed). This is the direct measured inversion of slice 1's "German is the sole splitter" framing: slice 1 asked which lane splits what the grid collapses; slice 2 shows the grid mostly does NOT collapse it — German is the one lane that answers the grid's distinction in the lexicon, not the one lane that adds a distinction the grid lacks.
+
+**Valency: the distinction moves levels, it never disappears.** Per the operator's steer (*"reflexiv, Relativpronomen, transitive intransitive"*), English does not fail to mark `KNOW`'s two coordinates — it marks them by subcategorization FRAME instead of lexeme: at 3:7 `knew` takes a clausal complement whose subject the matrix subject binds (*"they knew that THEY were naked"*, reflexive-shaped), at 4:1 a direct-object NP (*"knew Eve"*, plain transitive). Luther marks the same split TWICE — by frame AND by lexeme.
+
+**Passion (quale × magnitude × phase) peaks at (German, KNOW) = 0.3130**, asserted, and is structurally ZERO wherever the grid's axes agree (`NAKED`/`DIE`) — no Erkenntnis without a distinction to see.
+
+**Two self-caught defects, fixed during the run (kept as findings, not silently patched).** (a) The CHAODA-shaped outlier read divided by a MAD of exactly 0 (7 of 8 points identical), printing a meaningless robust-z of 2.6e6 — replaced with an explicit degenerate-population branch plus a strict-minority assert. (b) The quale term conflated surprise with uncertainty: a CHECK row's low confidence (`c = 0.5` → expectation 0.75) produces quale 0.25 with nothing actually seen, and the first run of the passion composite ranked cs-bkr/arc-onkelos above four verified one-lexeme lanes purely on curatorial doubt. Fix: passion is computed over VERIFIED `(lane, contrast)` cells only; CHECK cells print report-only. Properly separating confidence from surprise (a confidence-independent surprise term) is queued, not built.
+
+**Honest limitation — the morph field is typologically biased, and the fix is deliberately deferred.** The fixture's `morph` field carves `prefix|stem`, biased toward prefixing languages. An agglutinative lane (Finnish `tietää`/`tuntea`, Biblia 1776, public domain) carries its meaning switch on a SUFFIX where this carving cannot see it. Spine slot 6 is left explicitly RESERVED rather than filled with a reconstruction — inventing an unverified lane to make a point is exactly the failure mode this probe keeps self-catching.
+
+**Metric check (operator steer: "newer metrics than Levenshtein; HHTL+CLAM, HHTL+helix").** `ndarray::hpc::clam` trees fixed-length `&[u8]` fingerprints under a pluggable Hamming-family `DistanceFn`; helix `DistanceLut` is a 256×256 L1 table over a linear codebook order — a true metric, safe for CLAM/CAKES pruning where edit distances are not. The substrate-shaped successor to Levenshtein is the q-gram character profile (a profile IS a sparse fingerprint). Measured caveat: at stem scale (≤ 8 chars) pure bigrams QUANTIZE (`gewahr` ∩ `kennen` share zero bigrams → φ snaps to π → dead phase again, the slice-1 trap reappearing through the metric door) — so the graded LCS radix path stays the phase GENERATOR, and the profile cosine runs as a fingerprint-space cross-check; both metrics must agree on the flagship ordering (asserted). Full HHTL+helix routing needs a trained lexeme codebook — 30 curated rows are not a codebook — so it stays the queued slice-3 probe.
+
+**Queued next:** slice-3 (trained lexeme codebook → `DistanceLut` L1 → CLAM cross-lane cognate/false-friend search); the Finnish suffix-carving fix (`prefix|stem|suffix`); a confidence-independent surprise term for passion; verifying the remaining CHECK rows against critical editions. Refs: `E-THE-GRID-COLLAPSES-WHAT-A-LANGUAGE-SPLITS-1` (the retracted slice-1 entry, mechanism this supersedes), `E-EYES-OPENED-PRINTS-BLIND-1` (the 3:7 convergence that survives, unrelated to the retracted residual number), `crates/lance-graph-planner/examples/probe_babel_stances.rs` module doc (the authoritative slice-2 statement), task #52.
+
 ## 2026-07-28 — E-THE-GRID-COLLAPSES-WHAT-A-LANGUAGE-SPLITS-1 — PROBE-BABEL-STANCES slice 1: with English/WordNet as the COORDINATE SYSTEM (not a sixth lane) and linguistic resonance read as PHASE, the Genesis `KNOW` coordinate carries a measured residual of **0.500** that the grid cannot address — and the sole splitting lane is **German** (Luther 1545 `gewar`/gewahr @3:7 vs `erkandte`/kennen @4:1), while `NAKED` stays fully constructive at 1.000. The split coordinate's awareness address is **3:7** — the same verse the monolingual reflexive-lift blade crowns independently. Two machineries, one address.
 
-**Status:** SHIPPED (`probe_babel_stances`, self-asserting; 6-lane spine, verified-row-gated asserts). **Confidence:** High on the asserted rows (en-kjv machine-verified from Gutenberg #10; de/la/el KNOW+NAKED are anchor facts). The `DIE` finding is **explicitly NOT claimed** — see the honesty correction below.
+**Status:** SHIPPED, then **HEADLINE RETRACTED 2026-07-28 (codex P1)** — the residual mechanism below is superseded; the retraction and what survives it are stated in full in `E-TWO-ROSETTA-STONES-AND-THE-FOUR-CHANNEL-SPLIT-1` (prepended above). Entry text otherwise unchanged (append-only: only this Status line is amendable). **Confidence:** High on the asserted rows (en-kjv machine-verified from Gutenberg #10; de/la/el KNOW+NAKED are anchor facts); the headline residual number is WITHDRAWN. The `DIE` finding was **explicitly NOT claimed** — see the honesty correction below.
 
 **The three operator steers became one mechanism.** (1) *"English meaning as a coordinate system (wordnet) vs linguistic resonance"* — English is the GRID, its synsets are addresses, the five other lanes are measured AGAINST it (asymmetric by design: chart vs territory). Machine-verified: the KJV uses ONE root at both addresses of all three pairs — `knew`/`knew`, `naked`/`naked`, `die`/`die` — so the grid genuinely cannot address what a lane may split. (2) *"treat linguistic resonance as phase"* — per (lane, coordinate): shares one root = **phase 0** (constructive, agrees with the grid), two roots = **phase π** (antiphase). Amplitude = `|Σ e^{iφ}|/N`; **residual = 1 − amplitude IS the expressiveness escaping the coordinate system** — the multilingual form of `AddressAbsent`. Deterministic from lanes + grid, never stored (D-QUANTGATE-clean). (3) *"use the richness of our 32 lane SoA"* — a 32-slot spine, 6 occupied, 26 `None` (RESERVE-DON'T-RECLAIM). **Honest scope: lane-SHAPED, not SoA-WRITTEN** — no envelope, no tenant, no `ENVELOPE_LAYOUT_VERSION` (a real value tenant is `v3-envelope-auditor`-gated, not this slice); the columnar-read payoff is stated as a CLAIM, not a benchmark.