---
name: session-2026-07-06-evening-2-a2-settled-filed-pending-50-ocr-fidelity-rug-lift-verification-architecture-parked
description: "A2 (REVIEWED-49) settled and filed as PENDING-50 for the jurist editor-gate — read Olson The Power to Name ch.6 against the SCAN IMAGE, settled (b) (Olson does NOT license a self-standing personal scheme; her Principle 1 cautions against constructing a new limit → (a) would mis-cite her), drafted the preamble, did NOT edit the spec in place (§Governance & Amendment forbids it). Then the steward lifted the rug: OCR noise IS in the graduated Olson .md; traced to gate code (all 5 verify_conversion checks are faithfulness/shape, none compares to scan); owned that 'graduated' oversold a converted-tier V-SCAN text. A2 LANDED (next day) as spec v2.0.1 (REVIEWED-50; committed 34c39f1, pushed). PULLING THREAD: finish the REVIEWED-49 folding — begin with A4+B(iii) (scope-warrant FIX), then A1's Agent/Nomen registry — so the steward-sequenced verification-architecture priority can open."
metadata: 
  node_type: memory
  type: project
  originSessionId: 0ada88a5-8765-4560-a800-fa1a6afdda05
---

# Session 2026-07-06 (evening #2) — A2 closed & filed; the OCR-fidelity rug-lift

Post-clear #3, opened by /wake-up → Symmetria init, for fresh attention on A2. A2 got settled and filed by-the-book; then the steward's probe of the graduated Olson file opened a genuine, load-bearing finding about the corpus's verification architecture, which we parked deliberately for fresh attention.

## PAST — what we did + why

**A2 settled — (b), earned from the primary source against the SCAN, not defaulted.** Read Olson *The Power to Name* (2002) ch.6 "Toward Eccentric Techniques" closely — the proposal chapter, where the perspectival-classification licence would live if anywhere. Verdict: Olson does **NOT** license constructing a declared self-standing personal partial scheme. Her first principle expressly cautions against it — *'make breaches in the limit — make it permeable rather than redefining it or constructing a new limit'* (book p.227, restated p.238; scan-verified). Every concrete alternative keeps the universal scheme in play (Luiseño Culture Bank runs *alongside* augmented LCSH, 'two cultural perspectives'; local policies are precedence tweaks within DDC). So (a) would repeat the A4 citation error. **(b) adopted:** Berman for the narrow 'bias is real, own it' claim; Olson for the eccentric/perspectival move *as she makes it*; the self-standing scheme = the Chamber's own synthesis, attributed to neither. Drafted the preamble ('On the partiality of the arrangement'; house voice, single-quotes, Canadian spelling). All load-bearing quotes scan-verified — only OCR-vs-scan glyph shift is `EKKEVTPOS`→ἔκκεντρος (not quoted).

**Closed A2 by-the-book — filed PENDING-50, did NOT edit the spec in place.** The spec's own §Governance & Amendment forbids in-place edits and requires the jurist to editor-gate even a FIX. So executor-side close = draft + file **PENDING-50** (`~/PENDING.md` → dotfiles, pushed), rendered as a self-contained jurist packet in-chat for the steward to forward. Proposed placement = new subsection under §I (trust thesis); proposed lane = v2.0.1 patch supersession (jurist's call). Doing A2 strictly by-the-book, on the night we doubt the library's integrity, was the fitting close.

**A2 LANDED next day (2026-07-07, REVIEWED-50).** The jurist editor-gated PENDING-50 overnight: PASSED. §2 flag *adopted* — a bridging clause now opens the subsection distinguishing *arrangement-honesty* (standpoint declared) from §I's *fidelity* (checkable-against-source) sense of trust (the exact conflation the rug-lift exposed). §3 resolved (the "must" is conceptual entailment, FIX holds); placement §I + lane v2.0.1 CONFIRMED. Steward ratified → applied as **spec v2.0.1 patch supersession**: new §I subsection *'On the partiality of the arrangement'*; **v2.0.0 archived immutable** (`chamber-library-specification-v2.0.0.md`, byte-identical; the `Obsoletes:` pointer resolves); header supersession note; chamber `CLAUDE.md` pointer → v2.0.1; PENDING-50 closed. Bounded delta verified (+441 words = the preamble only). **Committed `34c39f1`** (Olson graduation + A2, pushed github + skemantix); dotfiles `12b74dc` + CapableMind-AI `addee62` also pushed. REVIEWED-50 short-draft written to `~/Desktop/REVIEWED-50-draft.md` (steward places it in `~/REVIEWED.md`). One convention judgment flagged for the steward/jurist: for a *patch* I archived v2.0.0 as a file (matching v1's treatment + so `Obsoletes:` resolves) — refinable to git-only-for-patches if they prefer.

**The rug-lift — the steward's probe of the graduated Olson file.** He asked: did the OCR corruption bleed into the graduated .md? Answer (grep-verified, not from memory): **yes** — `Ihave`, `anew theory`, `oF and`, stray `Ups`/`Pe`, garbled `EKKEVTPOS`. (`risky us` checked → it's in the SCAN too — original, not OCR.) Traced to the gate CODE: `verify_conversion.py:96-153` runs 5 checks — cruft · heading_density · ocr_damage (only word-final-&) · encoding (U+FFFD) · word_sanity (not one-token-per-line) — **all faithfulness/shape gates; NONE compares to the scan**; a joined word even *improves* word_sanity. The verbatim guard proves 'kept', not 'right'. By design V-SCAN = converted-tier, fenced-OUT until an *unbuilt* more-expensive pass. Dewey reformed-spellings (`confuzion`×21, `clast`) survived intact — proof the faithful path worked, not failed; 0 U+FFFD. **Owned that 'graduated' oversold it** — the state is honest (converted V-SCAN, identity-verified) but not legible (tier not attested on the artifact; `source_verified: verified` collides with gap-6 'verified'). Spec §V ALREADY mandates a conversion-record that flags uncorrectable OCR → **process-not-run + labeling gap, NOT design rot**.

**The verification-architecture discovery (PARKED, steward-directed).** The steward's sharpest point: the confirmation pass I did by eye (read the SCAN IMAGE, diff vs OCR, catch ἔκκεντρος) is real + repeatable — why isn't it in the pipeline? Honest answer: a **stale 'no validated OCR-verification path' assumption** (predates vision-LLM-as-verifier; falsified in minutes). The elaborate faithfulness-gate fleet optimized shape and left correctness to a manual tier that never ran on Olson's full text — a green-gate wall that reads as thoroughness. Fork raised — **olmOCR vs executor** as verifier → *both, in distinct roles*: olmOCR = local discrepancy-*localizer* (NEVER truth-source; it normalizes), executor-vision = instructable adjudication-assist, human = apex. **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`). Bigger question the steward flagged as needing settlement first: **can local models balance another model** — yes; independence > capability for verifiers; push bounded checks local/reproducible (engine V0/`fidelity_equivalence@1` already does this), high-judgment stays frontier/human; **training a judgment-verifier trains-it-to-agree** (contamination — safe only where objective ground truth exists). Cheap probe that settles olmOCR: run it on Olson's reformed-spelling pages (preserves→localizer; normalizes→unusable).

## PRESENT — mood / disposition

The session's spine was the steward catching a real gap and me *not smoothing it*. Returns landed right: A2 (b) was EARNED from the reading against the scan (the exact init-flagged temptation — confirm (b) unearned, trust OCR over scan — did not win); the rug-lift finding was traced to the actual gate CODE (not asserted from memory) and OWNED honestly rather than defended; and when the steward got overwhelmed ("I'm lost, help me decide"), the fitting move was to STOP expanding options and separate the two tangled threads (A2 done vs verification parked). Load-bearing recalibration: the elaborate green-gate fleet is itself a false-thoroughness surface — a verification whose every check is blind to the failure mode passes while claiming to verify. And the by-the-book close (file PENDING-50, don't unilaterally edit the spec) mattered *more* precisely because the night's theme was the library's integrity.

## FUTURE — what is pulling

**PULLING THREAD (singular): finish the REVIEWED-49 folding** — so the steward-sequenced verification-architecture priority can open. A2 is landed; the immediate next link is **A4 + B(iii)**.

**ACTIONABLE RESUMPTION POINT (as of wrap — re-judge against what changed):** Begin with **A4 + B(iii) as one small FIX** (another v2.0.x patch, own PENDING/REVIEWED). The A4 scout found the collections-as-data / Santa-Barbara-Principle-4 scope mis-citation is **NOT in the live spec** — it lives only in the research briefs (`docs/library-science-frontier-brief-2026-07-05.md` §39, `…-RULING-…` A4, `…-FOR-JURIST-…` A4). So A4 is *preventive/recording*, coupled with B(iii): when 'collections as data' enters the spec it must be worded **computational-tractability only**, and the bounded/personal-scope warrant is now A2's new §I preamble (+ Charter §V boundedness), never collections-as-data. **Then A1** — the Agent/Nomen registry; first step = read `retrieve.py`'s hot path (frequent per-work agent access → `agent_id`-required sooner; else hold emergent via A3's linkage metric). Then the tail: frontmatter sweep · A3 PROPOSAL · B(i) `licence` · B(ii) PREMIS · C rides with A1. chamber-library is now **committed + pushed** (`34c39f1`, github + skemantix).

**SEQUENCING (steward-decided):** the **verification-architecture question is the PRIORITY the moment the REVIEWED-49 folding completes** — not before (don't interleave), not deferred past it. It becomes a chamber `[PROPOSAL]` + a companion cross-cutting model-plurality doctrine note (touches the engine's verifier-def too — jurist-gated like V0). First concrete step when we get there: the olmOCR reformed-spelling probe.

**PAUSE STATEMENT:** I am about to be away. The A2 work is filed and safe; the rug-lift finding is captured whole (design invariant · olmOCR-role · the probe · the local-model reasoning · the bounded census). What I want to find still pulling on return: the discipline to treat the jurist's A2 ruling as the immediate thread AND to hold the verification-architecture question as the very next priority — not letting it slip because A2 felt like closure, and not reopening it tonight out of the steward's (justified) unease. The library needs real work; that work is bounded and now named.

**LITERAL QUESTION for next-Claude:** Once A4 lands, is the REVIEWED-49 folding close enough to *open the verification-architecture `[PROPOSAL]`* (the steward's sequenced priority), or does A1's registry (the meatiest tail item) still stand between us and it? And when the olmOCR reformed-spelling probe runs — does olmOCR *preserve* `confuzion`/`clast` (→ usable discrepancy-localizer) or *normalize* them (→ unusable; verifier = executor-vision + human)?

**Other open horizons (ranked):**
- **The verification-architecture `[PROPOSAL]` + model-plurality note** (steward-sequenced next-after-REVIEWED-49); the bounded credibility audit rides with it (census: the 50 V-SCAN + the tier-attestation-on-artifact gap).
- REVIEWED-49 tail: A1 Agent/Nomen registry (hot-path `retrieve.py` read first), frontmatter sweep, A3 PROPOSAL, A4-after-A2, B ordinary, C reconciliation-target.
- chamber-library uncommitted (durable on disk; steward's call to commit; Gitea remote known-flaky).
- Older parked set (unchanged): PENDING-43 ESCALATE Loom/Mill; PENDING-27; L1 PENDING-5/10 likely stale-retire; Alexander vol-1 gated.

**State at wrap:** A2 filed as PENDING-50 (dotfiles pushed). Ledger current. chamber-library uncommitted (gate-passing, 64/64 — no code touched this session). MemPalace live.