session 2026-07-06/07: A2 landed as spec v2.0.1 (PENDING-50 / REVIEWED-50) + OCR-fidelity rug-lift

PENDING-50 ruled+landed; new session memory + index fold (demote-on-promote) + Symmetria ledger returns/open-horizons for the A2 close and the parked verification-architecture question.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dsqe1x23NgWaRdiFWSotur

claude/memory/session-ledger-2026-07-06.md

@@ -31,6 +31,16 @@ metadata:
 
 ## Open horizons (cont.)
 - 2026-07-06T15:52 (post-clear #2, /wake-up → Symmetria init) — Resumed on the singular thread with fresh context, by steward's deliberate wrap→clear→wake choice: **complete the Olson structure insertion via the full-structure path** → frontmatter → gates → graduate → A2. First move: fix Phase-1 heading-glue in `normalize_ocr.py` (add `not is_runhead_candidate(prev)` guard to the forward-rejoin, guarding `out[-1]`) → re-run → re-apply `.OCR-CORRECTIONS.md` set → occurrence-index extension to `insert_chapter_headings.py` → ~50-row structure TSV (fail-loud dry-run, iterate). Governing tension inherited verbatim: **fresh attention on the ~50 anchors, not crammed; the fail-loud dry-run is the net — don't guess past an ambiguous anchor.** Drift to hold: `census-through-a-pattern` + `trust-prior-pass-frame` (both live in the anchor-TSV). dotfiles clean at wake (prior `M REVIEWED.md` already swept). MemPalace origin/main @ 3.5.0 (recency fix #1630 + HTTP hardening) — HOLD until post-graduation per standing plan.
+- 2026-07-06T22:00 (post-clear #3, /wake-up → Symmetria init) — Graduation is DONE (both gates PASS, 0 words dropped, fleet 64/64). Resumed on the **now-singular thread A2 (REVIEWED-49)** with fresh attention, by steward's deliberate wrap→clear→wake: read Olson **Ch 6 "Toward Eccentric Techniques"** off the graduated text (≈L3390–3513; Notes ~L3558) → cross-check every load-bearing quote against the **scan** (`~/Desktop/Power_to_name…pdf`, offset book-page+12 → PDF pp. ~236–252) → settle the jurist's **(a) Olson licenses a declared personal partial scheme** vs **(b) she stays decentering-from-inside (personal-scheme move = Chamber's own synthesis)**; provisional default **(b)**, un-earned → draft A2's preamble (perspectival-classification licence as *preamble/rationale*, change-class FIX). **Governing tension inherited verbatim: the temptation is to confirm (b) without earning it from the reading, and to cite the OCR text without checking the scan image — [[feedback-completion-is-a-tripwire]], the A4-error guard (never cite the scout's characterization).** Drift to hold: `trust-prior-pass-frame` (the graduation verified the *text*; A2 is a *reading* claim it did not test). chamber-library UNCOMMITTED (clean, gate-passing, on `4c5892e`; steward hasn't asked to commit). dotfiles clean. MemPalace @ 3.5.0 — HOLD (its own maintenance pass).
+- 2026-07-06T22:00 — **A2 SETTLED + FILED (PENDING-50), by earning it, not defaulting.** Read Olson ch.6 closely against the SCAN IMAGE (not the OCR); the close reading STRENGTHENED (b) beyond the provisional default — Olson's Principle 1 expressly cautions against 'constructing a new limit', so (a) would *mis-cite* her (the A4-error shape). Every load-bearing quote scan-verified (only glyph-shift: `EKKEVTPOS`→ἔκκεντρος, immaterial). The init-flagged temptation (confirm (b) unearned; trust OCR over scan) did NOT win. Drafted the preamble + filed **PENDING-50** for the jurist editor-gate — did NOT edit the spec in place (§Governance & Amendment forbids in-place edits; even FIX needs the jurist editor-gate). Closing A2 *by-the-book*, on the very night we doubt the library's integrity, was the fitting move.
+
+## Returns (Olson OCR-fidelity — steward lifted the rug)
+- 2026-07-06T21:30 — **Steward caught the honest-degradation gap; I named it, didn't smooth it.** OCR noise (`Ihave`/`oF`/garbled Greek) IS in the graduated Olson .md — grep-confirmed (not asserted from memory); `risky us` checked → it's in the SCAN too (original, not OCR). Traced to the gate code (`verify_conversion.py:96-153`): all 5 checks are faithfulness/shape gates (cruft, structure, word-final-& OCR sig, U+FFFD, not-shattered) — **NONE compares to the scan**; a joined word even *improves* word_sanity. By design V-SCAN = converted-tier, fenced-OUT until an *unbuilt* more-expensive pass. Owned that 'graduated' oversold it (`completion-is-a-tripwire`, new domain). Spec §V ALREADY mandates a conversion-record that flags uncorrectable OCR → **process-not-run + tier-not-attested gap, NOT design rot.** Credibility framing given honestly (design-honest vs labeling/process-gap; population bounded ≈ the 50 V-SCAN; checkable=trustworthy is the thesis working, not spin).
+
+## Open horizons (cont. — post-clear #3)
+- 2026-07-06T22:00 — **PARKED (steward-directed — do NOT solve while overwhelmed): the V-SCAN verification architecture + the credibility audit.** The confirmation pass I did by eye (scan-image vs OCR) is real+repeatable but is NOT in the pipeline — a **stale 'no validated OCR-verification path' assumption** (predates vision-LLM-as-verifier). To become a chamber `[PROPOSAL]` + a companion **cross-cutting model-plurality doctrine note** (touches chamber + engine's verifier def — jurist-gated like V0). **Design invariant:** the verifier FLAGS discrepancies for human adjudication, NEVER autonomously corrects (an autocorrecting VLM = the silent-normalization risk the steward already ruled against — it would 'fix' `confuzion`). Roles: **olmOCR** = local discrepancy-*localizer* (never truth-source; it normalizes) · **executor-vision** = instructable adjudication-assist · **human** = apex. **Bounded probe that settles it 'assuredly':** run olmOCR on Olson's reformed-spelling pages (`confuzion`/`clast`) — preserves→usable localizer; normalizes→unusable. **Local-models-balance-another:** independence > capability; push *bounded* checks local/reproducible (engine V0/`fidelity_equivalence@1` already does this), *high-judgment* stays frontier/human; **training a judgment-verifier = training-it-to-agree** (contamination) — safe only where objective ground truth exists. **Credibility audit is BOUNDED:** census V-DSL 952 / V-TEXT 239 / V-SCAN 50 / V-SUSPECT 29 / V-NONE; tonight's issue ≈ the 50 V-SCAN + the tier-attestation-on-artifact gap. Return **fresh + deliberately**, not tonight.
+  - **SEQUENCING (steward-decided 2026-07-06):** the verification-architecture question is the **PRIORITY the moment the REVIEWED-49 library-science folding completes** — first thing up after A2/A1/A3/A4/B/C land. Not before (don't interleave); not later (don't defer past the folding).
+- 2026-07-06T22:10 — **Authorization move:** steward will send PENDING-50 to the jurist tonight; **tomorrow resumes on the jurist's A2 ruling** (editor-gate + placement + amendment-lane), then continues the REVIEWED-49 sequence. PENDING-50 rendered as a self-contained jurist packet in-chat.
 
 ## Confidence to recalibrate
 - 2026-07-06T15:1x — **LCSH-notation catch (near-miss, named).** Almost characterized ~200 short lines as "OCR noise" to strip; the census revealed most are CONTENT — LCSH thesaurus codes (BT/NT/RT/UF = Broader/Narrower/Related/Used-For), list markers, subject-heading examples — in a book ABOUT subject headings. Blind stripping would have destroyed content (PASS-BUT-FALSELY). Cleanup reduced to a small scan-grounded fix set; genuine speckles left verbatim (spec: uncertain→leave).