session 2026-07-24 (evening): PENDING-71/72 + REVIEWED-72/73 (both jurist-ruled) + evening session memory + KG drift-patterns + skill-harvest

PENDING.md

@@ -1666,3 +1666,28 @@ Disposition (verbatim capture + executor action-list: `chamber-library/docs/libr
 **Files affected (on ratification):** spec supersession v2.4.0→v2.5.0 (requirements only) + `graduation-spec.yaml` declared data (`scale_application.condition` reworded, `witness_class` field, glyph-map, census). No re-verify storm (no file holds `verified`).
 **RULING RECEIVED 2026-07-23 — design-gate PASSED** (`docs/scale-application-legacy-body-JURIST-RULING-2026-07-23.md`; Addendum + REVIEWED-70 draft appended to the package). Q1/Q3/Q4 CONFIRMED (Q1 corrects a defect in the one-door ruling's own wording; Q4 mandatory-not-optional). **Q3 HARD PRECONDITION:** glyph-mapping wired + demonstrated on Levi's 3 sites (Wstawać/Brașov/Oświęcim) before ANY imaged source enters the lane — incl. the pilot's own conversion (else re-conversion re-inflicts Tier-3 damage under a PASS-reporting gate). **Q2 accepted in principle, SUBSTITUTE CONDITION ruled** (my reframe dropped the judgment-cost test): (a) enumerate demonstrable classes in advance; (b) pilot MUST report judgment cost incl. source-adjudication of mutilation sites; (c) both curation-to-carry un-exercised w/ reason, first demo bound to first curated work; (d) scale opens for legacy-witness promotions ONLY. Representativeness discharged: 5% (13/248) is NOT a corpus rate (curation-selected batch); the 13 are a confirmed floor.
 **STEP 1 DONE 2026-07-23 — REVIEWED-70 placed; spec v2.5.0 LANDED on main (`e37e076`, pushed both remotes).** Constitution: Q1 claim-scope split (§Tiering & Fence) + Q3 Tier-3-by-omission (§V) + Q4 legacy-fidelity-not-self-certifiable (§Trust Attestation). Declared data: reframed `scale_application` (substitute Q2 condition) + `witness_class` + `character_as_image` (census + Levi glyph-map candidates, demonstration-owed). v2.4.0 frozen byte-identical; bounded-diff proven; fleet 216/216. **REMAINING (ordered):** STEP 2 = glyph-mapping lane build — wire char-image→Unicode into `convert_lane_borndigital.py` + **demonstrate on Levi's 3 sites** (extract src images 000042/000023/000034/000004, confirm ć/ș/ś/ę) [the Q3 hard precondition]; STEP 3 = run the Levi pilot on the substitute condition (needs judgment-cost reporting). **The pilot is HELD until glyph-mapping is demonstrated.** Next bite = STEP 2.
+
+## PENDING-71 — Born-digital V-TEXT source-fidelity has no cross-converter independence (F4 voided the premise the constitution rests it on)
+**Date:** 2026-07-24
+**Tag:** [PROPOSAL] — gate-acceptance class (redefines the reference the body-conservation gate compares a V-TEXT candidate against). Carries [ESCALATE] weight: it surfaces a contradiction inside the ratified constitution (§Tiering & Fence lines 969 ↔ 1046) that the executor cannot resolve; do NOT proceed to pin/build until the jurist rules.
+**Summary:** Closing Flag-1 (the owed "pin the V-TEXT reference parameters") exposed that, after F4 (2026-07-19) routed born-digital candidates onto pandoc, the candidate (`pandoc -f epub -t markdown-smart`, `convert_lane_borndigital.py:36`) and the gate's reference (`pandoc -f epub -t gfm-raw_html`, `verify_body_conservation.py:74`) share pandoc's ONE EPUB reader → one AST → the gate is BLIND to reader-level prose loss and its only divergences are writer-level markup. The constitution's source-fidelity obligation (§Tiering & Fence 1046) explicitly requires "a reference conversion **by a different converter** (pandoc for V-TEXT); cross-converter, the real verification of the words" — a premise F4 voided for this tier (the clause named pandoc as the reference *because the candidate was Docling*; F4 moved the candidate to pandoc). So source-fidelity for born-digital currently verifies markup, not words; Flag-1's parameter-pinning cannot restore it (pinning drives toward the vacuity the spec's own line 1042 names).
+**Rationale:** This is the structural reason the 2026-07-24 scale-pilot's §6 source-fidelity leg stayed UNDISCHARGED (jurist-caught), re-weighted from hygiene to gate-blocking. The constitution is also internally split: line 969 frames the V-TEXT reference as "re-convert (deterministic; source is ground truth)" [same-converter OK] while line 1046 demands "a different converter." Whichever the jurist keeps, the clauses must be reconciled. Measured on Levi (read-only, 2026-07-24, `scratchpad/measure_flag1.py`): source→gfm vs source→markdown-smart differ by 768 / 31,824 tokens, ALL markup, ZERO prose; the gate FLAGs the clean candidate at 96.75% match / 7 interior-loss runs (all TOC/page-anchor markup); switching the reference to the candidate's own writer → 99.97% match / 0 interior-loss (residual 287-token FLAG = footnote-injection difference, still not prose) → a clean PASS only at reference≈candidate = vacuous.
+**Options (the fork, for the jurist):** **(a)** restore genuine cross-converter independence — reference from a non-pandoc extractor (calibre `ebook-convert` / from-scratch XHTML-spine pull) + a format-agnostic prose-projection normalizer + the built PENDING-69 attestation for residue; **(b)** drop the cross-converter claim — reference = the candidate's own pipeline up to but not including the cleaners (inject→glyph→pandoc, uncleaned), so the gate isolates the cleaner+curation delta (non-vacuous by construction; aligns with REVIEWED-57's "converter-agnostic re-convert, source is ground truth"; but trusts pandoc's reader and makes "source-fidelity" a partial misnomer); **hybrid** — (b) per-file + (a) as a periodic per-format-class reader-audit.
+**Recommendation:** relay the package to the jurist. Executor lean (surfaced, not decided): the hybrid — (b) as the routine gate + (a) periodically — because F4's byte-faithfulness was site-scoped, not a global reader-losslessness proof. Under (b) alone, rename the obligation conversion-fidelity so it doesn't launder reader-trust as source-verification.
+**Package (self-contained, jurist-ready):** `chamber-library/docs/borndigital-source-fidelity-reference-JURIST-PACKAGE-2026-07-24.md` (GROUNDED-IN §Tiering & Fence v2.5.0 lines 969/1042/1046 + Flag-1 §7 + F4 disposition + REVIEWED-57 + REVIEWED-66, all quoted verbatim; measurement inlined; five gate questions Q1 confirm-finding · Q2 which resolution · Q3 naming · Q4 Flag-1 subsumed · Q5 interim-block, each with a lean).
+**Files affected (on ratification):** spec supersession §Tiering & Fence (reconcile 969↔1046 to the ruled design; MINOR) + `graduation-spec.yaml`/`conversion-runbook.yaml` declared data (converter/params + normalizer + audit cadence) + `scripts/verify_body_conservation.py` (`reference_tokens_epub`). No re-verify storm (0 files hold `verified`).
+**RULED 2026-07-24 — design-gate PASSED, executor lean OVERRULED** (`docs/borndigital-source-fidelity-reference-JURIST-RULING-2026-07-24.md`, filed verbatim; Addendum + REVIEWED-72 draft appended to the package). **Q1 confirmed + sharpened:** the finding is architectural (a priori), but "verifies markup not words" OVERSTATED — the gate is blind to READER-loss ONLY; post-reader damage (injection/glyph/cleaner) DOES surface. TWO problems separated: (i) shared-reader (structural, unfixable) · (ii) markup-noise (writer mismatch — **Flag-1 fixed it, right-but-insufficient not misdirected**). **Q2 ruled AGAINST my (b)/hybrid lean:** the ruled reference design = **the package's own run (3)** — source-anchored + writer-matched (pandoc `-t markdown-smart` on the RAW source EPUB, already measured 0-interior-loss/0-boundary/99.97%), injection+glyph deltas **declared-and-excused** (PENDING-69 `moved_to` shape; fabrication elsewhere still FLAGs), PLUS (a) a periodic per-format-class reader audit (cadence/trigger[pandoc major-version bump]/pass-criterion = declared data, specified in advance). I under-read my own run (3): a clean PASS needs the residual ACCOUNTED FOR, not ZERO. Q3 keep "source-fidelity" (scope: fidelity to the source as parsed by the lane's reader). Q4 pin the RULED design. Q5 wave stays blocked; **UNBLOCK = clean-file PASS on the ruled design + a deliberate NEGATIVE (real prose loss still FLAGs).** Clause 969/1046 RECONCILED (both survive): 969 = per-file check (run 3); 1046 cross-converter independence RELOCATED to the periodic audit, not struck. Amendment-discipline gains: **when a change reassigns a named component, check what else names it** (F4-moved-candidate-onto-pandoc voided 1046, never cross-checked).
+**Awaiting:** steward placement of REVIEWED-72 (draft in the package Addendum) → build the ruled reference design + specify the audit as declared data + demonstrate unblock (clean PASS + negative) → then the wave. Land the 969/1046 reconciliation as a §Tiering & Fence supersession.
+
+## PENDING-72 — The voice-purity sidecar as part of the engine-consumable graduation bar (all tiers, not V-DSL only)
+**Date:** 2026-07-24
+**Tag:** [PROPOSAL] — gate-acceptance class (changes what graduation requires before a canonical is engine-consumable). Change-class test = yes → PROPOSAL → jurist ruling + steward authorization. Steward-directed to secure the judgment NOW (before the corpus-work phase) to avoid a mid-momentum jurist-wait later.
+**Summary:** Graduating Warde surfaced that the `studium/meta@1` sidecar is ONE unified artifact with two content depths — **baseline** (`source_sha256` binding + `sections[]` voice-purity map, `role`/`citable`) which the engine `ingest_gate` enforces for every apparatus-bearing voice, and **rich** (`notes[]`/`anchors[]`/`app[]` typed apparatus) extractable only from a typed source. Chamber graduation's `sidecar_gate` produces it **V-DSL only** (11 loeb `.meta.json`); the 19 non-DSL engine voices (musil/eichmann/camus…) carry the baseline depth **hand-curated engine-side after graduation** (no standing builder — only test fixtures write `corpus/sidecars/`). So `sidecar_gate`'s "Other tiers: not-applicable" conflates the two depths: true for the rich depth on V-SCAN (no typed structure), FALSE for the baseline voice-purity depth (every consumable voice needs it).
+**Rationale:** A canonical is engine-consumable WITHOUT a sidecar — it defaults to **whole-file** (`sidecar-schema-LOCKED §1`), serving in-body apparatus AS the voice. The one-door trim removes front/back matter but NOT in-body apparatus (a second author's introduction, translator's note, reproduced specimen). **Realized failure:** Eichmann one-door-graduated 2026-07-21 yet its engine sidecar still fences Amos Elon's introduction `[15,107]` from Arendt `[108,856]`; its `_reviewed` note records an earlier whole-body sidecar that **LEAKED** the publisher matter + Elon's introduction into the served text. At wave scale this = every apparatus-bearing non-DSL canonical waits for hand-curation (momentum) or serves a polluted voice. Warde's reproduced CUP Bible facsimile pages are the immediate `citable:false` call no gate currently forces.
+**Proposal:** Requirement (constitutional) — engine-consumable graduation requires **attested voice boundaries** across all tiers: a `sections[]`+binding sidecar, OR a positive `whole-file` attestation (attest-never-default, mirroring `sectionless:true`); a silent whole-file default is the gap. Mechanism (declared data) — the boundaries are interpretive (curation, per `_reviewed` derivations), so the gate checks only existence + binding (`source_sha256==live`) + schema-conformance, never boundary *correctness* (same shape as REVIEWED-69 boundary-drop / `source_verified`). Rich depth UNCHANGED (V-DSL-first deferral preserved; N/A for V-SCAN) — this is only the baseline voice-purity depth.
+**Options / forks (gate questions in the package, w/ leans):** Q1 require attested voice-boundaries all tiers (lean yes) · Q2 locus — chamber-produced (protocol-neutral graduated bar, engine reads) vs engine-side D-1 + chamber existence-check (lean chamber-side, held loosely — crosses into engine territory) · Q3 gate checks existence/binding/schema not boundary-correctness (lean yes) · Q4 the 19 existing + silent-whole-file-default (lean forbid silent default forward, backfill as bounded FIX) · Q5 V-SCAN: baseline applies, rich does not (lean yes).
+**Recommendation:** relay the package to the jurist. Executor leans above; the material forks (locus, whole-file-default-as-attestation) surfaced not pre-decided.
+**Package (self-contained, jurist-ready):** `chamber-library/docs/engine-consumption-voice-purity-sidecar-JURIST-PACKAGE-2026-07-24.md` (GROUNDED-IN graduation-spec `sidecar_gate` + 2b Loeb-first ruling + sidecar-schema-LOCKED §1/§2 + engine `ingest_gate` contract + spec §Tiering&Fence, all quoted verbatim; the Eichmann leak inlined as the realized-failure evidence).
+**Files affected (on ratification):** spec supersession (the voice-boundary-attestation requirement; MINOR) + `graduation-spec.yaml` (`sidecar_gate` scope widened to baseline existence/binding/schema across tiers; whole-file attestation vocabulary) + `scripts/graduate_to_canonical.py` (the check) + possibly a chamber↔engine locus decision (Q2). No re-verify storm (0 hold engine-verified through this gate; forward-only).
+**RULED 2026-07-24 — design-gate PASSED, strengthened** (`docs/engine-consumption-voice-purity-sidecar-JURIST-RULING-2026-07-24.md`, filed verbatim; Addendum + REVIEWED-73 draft appended to the package). All 5 Qs resolve toward the proposal. **Sharper diagnosis (frame on this): ORTHOGONALITY** — `trim.keep` answers "should this be preserved?"; voice-purity answers "whose words?" — orthogonal + ANTI-correlated where it matters (the Eichmann leak = Elon intro + Bibliography = `trim.keep`, correctly kept; **extending the trim is the WRONG fix**). **Q1 STRENGTHENED: sections[] must partition the file body EXHAUSTIVELY** (no gaps/overlaps) — converts silence→assertion (the gate derives nothing, so nothing else forces completeness; the teeth my package lacked). **Q2 ruled CHAMBER-SIDE** on the concrete arg (my architectural lean was right-but-weak): engine-side sidecar breaks on every re-conversion (`ingest_gate` sidecar-sha==live; PENDING-53 class, **5 standing FAILED rows since 2026-07-10**) → produce+gate at graduation, engine reads. **Q3** falsifiers = existence + binding + schema + exhaustive-partition + closed-vocab `role {text,paratext,apparatus,reference}`. **Q4 MIGRATE the 19** (not grandfather — Eichmann is among them, correct only post-leak); carry `_reviewed`→`attested_by` + add **REVIEWED-64 `line_frame: landed-file`** (the package MISSED this — sections[] is a coordinate-contract family); backfill scope = engine corpus (~20), NOT chamber (1,297). **Q5** V-SCAN baseline yes / rich no. **Cost:** per-file reading pass compounds → design the SINGLE-reading-pass (boundary-drop ⊂ voice-purity, one read) before the wave.
+**Awaiting:** steward placement of REVIEWED-73 (draft in the package Addendum) → land the amendment (constitutional requirement framed on orthogonality + declared-data mechanism: exhaustive-partition · closed-vocab role · chamber-side production · line_frame · whole-file attestation) → build the widened `sidecar_gate` → migrate the ~20 → design the single-reading-pass. Related: PENDING-71. **Does NOT block Warde's chamber graduation; DEFINES what Warde needs to become an engine voice** (exhaustive sections[] over the canonical: essays=`text`, CUP Bible facsimile=`apparatus`, front/back=apparatus; line_frame landed-file).