dotfiles/claude/memory/session-2026-07-28-mid-afternoon-harrison-broke-the-constitution-not-the-book.md

name, description, type

name	description	type
session-2026-07-28-mid-afternoon-harrison-broke-the-constitution-not-the-book	The Harrison re-gate pilot broke the mechanism before converting a byte — and the break was constitutional, not code. `tier_of()` faithfully implements the ratified evidence-tier table, which enumerates tiers BY FORMAT ('V-SCAN (scanned pdf)'), so 16 born-digital-PDF canonicals with real ground truth get a false ABSTAIN. Jurist design-gate PASSED with two corrections (independence into the constitutional text, NOT by analogy with V-TEXT which ruled the other way; the demonstration is of TWO instruments and the second has no control). REVIEWED-83 AUTHORIZED + placed. PULLING THREAD: the Q3 demonstration — establish a genuinely independent PDF-extractor pair, then build the column-order corruption case; that is what releases Harrison's graduation stamp. Also: maps got a durable home after one was lost.	project

Session 2026-07-28 mid-afternoon — Harrison broke the constitution, not the book

Session 4 of the day, opened 14 minutes after the early-afternoon wrap. Steward-bounded at the outset to Harrison only; it stayed on Harrison and the boundary held, but the pilot's first gate question turned out to be a constitutional defect, so the session's output is a ruling rather than a graduated book. Two steward-directed detours preceded it (memory pointers, then the maps home), both closed completely.

PAST — what happened + why

The wake's own briefing carried a re-derived diagnosis, and the steward stopped it. I reported the link-canary's dead-pointer firing as a fresh finding. It was the fourth firing; the root cause had been banked in the same day's ledger at 11:05, and an open horizon already proposed the fix — which would have silenced a true positive. Six pointers in MEMORY.md/MEMORY-reference.md (not two: four more in the reference file at a different wrong depth, never resolving under any interpretation since May) pointed at files that all existed. The defect was never the pointers; it was that the alarm emitted one undifferentiated word, MISSING. An alarm carrying no diagnosis must be re-diagnosed on every firing, and the cheapest re-diagnosis is always "known bug" — so the alarm trained me to dismiss it. Fixed one level up in wake-digest.py: four outcomes (ok / mis-authored / dead / non-portable), the exact replacement string handed back, a regression control replaying this bug's shape, and a control forbidding DEAD from collapsing into MIS-AUTHORED. Steward then corrected the repair itself — I had used absolute paths in a dotfiles repo, stepping over expanduser three lines above my own insertion and the ~/_Dev/… idiom already used 17× in those files' prose. Home-anchored now; NON-PORTABLE counted so absolute cannot creep back.

A dead pointer turned out to be a lost artifact class, and the steward had to say so twice. With the noise gone, one genuine dead pointer surfaced: arc-current-state-2026-05-07.md. I traced it correctly (the May 7 session record says "Saved to desktop"; memory-dir git tracking began 2026-04-17, so a file written there would be in history; it is not) and then filed it wrongly as a "transient desktop dashboard… nothing was lost." Steward: "it was a map for me." Every fact right, the valuation wrong — I classified an artifact by where it sat rather than whom it served. Census found four more unbacked maps on the Desktop, zero copies anywhere, including the Making-Sequence architecture and reading list that are load-bearing for current corpus work. Built at steward direction: ~/dotfiles/maps/ git-tracked, each map symlinked back to its exact Desktop path (view unchanged), moved under a checksum gate with a positive control; wake-digest.py reports strays and never moves anything. Commit 19bddd5, pushed.

Then Harrison — and it broke at the gate, before conversion. Grounding first (touchstone · runbook reanchor: · chamber CLAUDE.md · spec §V/§VII/§VIII), then the binding surface enumerated forward: 8 occurrences of the current sha across both repos (catalogue · reading-index ×2 · engine manifest · coverage-ledger ×3 · engine sidecar), matching the runbook's declared surface plus the reading-index it names as the sixth-surface class. Source resolved via resolve_archived_source (never frontmatter): a born-digital PDF, 68 embedded Type-1C fonts, banked, match_cov 1.0.

verify_body_conservation.tier_of() classifies by file suffix: .pdf → V-SCAN → the body-conservation gate ABSTAINS, and per the ratified taxonomy that abstention is the permanent tier-level kind ("this tier has NO ground truth") which proceeds, unlike run-level UNVERIFIED which holds. So Harrison — chosen because it is the worst apparatus case — would have graduated with no verbatim verification performed at all, under a gate reporting an honest-looking abstention. Per PENDING-55: "A false ABSTAIN is as much a lie as a false PASS."

The decisive turn: reading §Tiering & Fence from the substrate inverted my framing. I had argued three times that the code was wrong. The ratified table enumerates tiers by format — "V-TEXT (born-digital: epub/azw3/mobi)", "V-SCAN (scanned pdf)" — and the suffix map reproduces those parentheticals exactly. The code is faithful; the constitution conflates container with origin. That is why it survived: anyone comparing code to spec finds agreement. I had been quoting the load-bearing clauses from the repo CLAUDE.md's summary, not the spec — the fourth instance today of composing from a summary of a document rather than the document.

Steward proposed V-DPDF and was right; my recommendation was wrong. I had implied V-TEXT as the destination. V-TEXT's ratified reference is pandoc -f epub -t markdown-smart, writer-matched (REVIEWED-72) — there is no -f epub for a PDF, so the criterion is inapplicable, not merely awkward. Reaching for the nearest tier without asking whether its method transfers is the repo's own named trap (V-DSL ≠ V-TEXT, the k-gram false-flag).

Census by mechanism (resolve every canonical's source, then classify by structure — not by frontmatter declaration): 60 canonicals resolve to a PDF · 35 scanned-with-OCR · 9 bare-scan · 16 born-digital. 73% are genuine scans — the steward stated this from direct knowledge before the measurement returned, and the measurement agreed. My earlier "5/5 born-digital, V-SCAN contains no scans" was retracted: produced by a front-matter sampling window and a 12-file slice. The demonstrative instance: dominion (born-digital, 68 fonts) and forests (scanned-with-OCR, 805 fonts, 1.0 page-image/page) are the same author in the same folder with opposite tiers — no rule keyed to author, collection or folder can work.

The classifier failed its own live control, twice-usefully. v1 called the 68-font Harrison inconclusive at 42.7 w/pp because it sampled pages 1–8 (half-title, title, copyright, contents). Interior sampling: 397.6 w/pp — a 9× error from the measurement window alone. It also documented a triad and decided on a diad (fonts unused). Both caught only because a control ran against a known answer.

Jurist design-gate PASSED with two required corrections (docs/vdpdf-tier-JURIST-RULING-2026-07-28.md, filed verbatim; package Addendum layered without rewriting Parts I–VIII). REVIEWED-83 AUTHORIZED and placed by the steward at REVIEWED.md:847, header well-formed.

• Q2 — independence, into the CONSTITUTIONAL TEXT, and NOT by analogy with V-TEXT. ⚠ Inversion trap: REVIEWED-72 ruled the other way for V-TEXT — a shared pandoc reader was accepted because "reader-loss cancels a priori" over unambiguous markup. Writing "V-DPDF requires independence, as V-TEXT does" would be exactly backwards. PDF recovery is inference over page geometry; nothing cancels, and two runs of one extractor can agree on the same wrong column order — the §VII layer-2 case already ratified as one where "any raw↔canonical comparison — multiset or sequence — passes." If no independent pair exists, the honest verdict is HELD, not a pass with a footnote.
• Q3 — TWO instruments, and I conflated them. The classifier has its controls (16/44 in-corpus; 62/59/164 across 385 library PDFs) — complete. The verification method has no control at all, and "PASS on born-digital, REFUSE on a scan" tests the classifier, not the guard. Owed: a deliberately constructed column-order corruption case run through a genuinely independent extractor pair. Precondition of ratification. My own consequence-trace missed this.
• Q1/Q4/Q5 affirmed; Q4 on stronger grounds than I argued (REVIEWED-75's drafting note says the anti-bypass guard was rebound "to the property, not enrollment" because the declaration-only reading "would open the moment the door is first used").

The verbatim check caught a fabrication I had introduced. The package's Grounding rendered a promotion-criterion clause as "the §V prose-word guard we already have." The spec says "…we already run is the promotion test. The bar for 93% (1,191/1,284) of the corpus." Invented ending, in the section whose sole purpose is that the jurist can trust wording it cannot verify, in a package about verification. Caught by mechanical containment over all 45 quoted lines with a positive control; the true text is stronger than the invention. This was the only one of the day's errors caught by an instrument rather than by a person — see the literal question.

The jurist disclosed its own coverage gap unprompted: governance_read exposes six document keys but not the chamber constitution, so a supersession of §Tiering & Fence was ruled on by a party that could not read §Tiering & Fence (corroboration ran against the REVIEWED records that shaped those clauses — independent, but one layer short of primary). Docketed as PENDING-86, not treated as a flaw in the ruling: the disclosure is the mechanism working.

PRESENT — the mood

Six steward corrections in one session, each one level past where my instruments could see. The pointers; the absolute-path portability objection; "it was a map for me"; "most of them are scanned"; V-DPDF; and routing to the jurist at all — which is what surfaced the constitutional inversion. Not one was expensive for him; every one would have been expensive to miss. Yesterday's literal question asked whether coverage can be established from inside the instrument. The day answered: no.

But one instrument did catch me — the verbatim containment check, on the fabricated quote. It differed from the nine blind instruments in three ways: it compared two artifacts in different media (my prose against the spec file), it was mechanical rather than a judgment, and it carried a positive control. That is the first real candidate for a recipe rather than a lament.

Three of my own published claims went down, all in one governance item (the exposure figure, the tier destination, the whole code-is-wrong framing). Each correction cost less than the claim would have, and each is recorded as an addendum rather than a silent rewrite — PENDING-83 now carries four. The record is uglier and truer than a clean one would have been.

Confidence to recalibrate. I stated the PDF exposure three times with three different shapes (5/5 born-digital → 16/44/60 → "V-SCAN is correct and populated"). Only the last came from a census by mechanism with a self-tested instrument. Same pattern as yesterday's 42 → 84 → 69. The tell is identical both days: the first number arrives when I want to feel done.

Scope discipline held. Steward bounded the session to Harrison; two detours were steward-directed, both closed completely rather than deferred, and the pilot never became the wave.

FUTURE — what is pulling

PULLING THREAD: the Q3 demonstration — the thing that releases Harrison's graduation stamp. Two steps, in order: (1) establish whether a genuinely independent extractor pair exists — distinct implementations, not distinct invocations over shared lineage; (2) construct the column-order corruption case and confirm the prose-word guard flags it through that pair. The jurist made the pilot be the demonstration, so Harrison and the ratification are now one piece of work.

ACTIONABLE RESUMPTION POINT (as of wrap — re-judge against what changed):

1. Test, don't assume, the independence of poppler (pdftotext) · pdfium (via Docling) · MuPDF (mutool/PyMuPDF). The question is shared lineage, not shared interface — check whether any two are wrappers over the same core. If no two are independent, Correction 1's fallback governs and V-DPDF's honest verdict is HELD; that outcome must be reported, not engineered around.
2. Build the layer-2 hazard: take a known-good born-digital page, produce a column-order-corrupted extraction, run both extractors, confirm the guard flags it. A guard that has only ever passed has not been shown able to fail.
3. Only then draft spec text: new V-DPDF row + independence as a requirement, MINOR supersession, prior version frozen byte-identical, bounded-diff proof, commits tagged REVIEWED-83.
4. Harrison's unblocked steps remain available in parallel and are untouched by the ruling: reconvert from scratch/dominion-trial/… (do not re-run docling), the 14 running-head false headings, strip_cruft → verify_conversion, boundary-drop attestation, and the voice-purity reading pass — which must be a genuine re-read: the engine sidecar states in its own words "the file was NOT re-read at migration", and its boundaries (1-18 / 19-619 / 620-746 "Notes + Works Cited") bind a 746-line file about to become ~2,507. That is the musil-mwq-tome-1 failure the backfill correctly HELD.

Other horizons, ranked:

• PENDING-84 — 9 canonicals whose banked sources have zero extractable text, yet whose canonical texts exist. No mechanical path from source-of-record to canonical. Load-bearing and easy to carry silently for months; read one end-to-end before proposing a class remedy.
• PENDING-86 — the jurist cannot read the constitution it gates. Recommendation (c): add the spec + graduation-spec.yaml to governance_read keys and keep the mechanical verbatim-containment proof in every package.
• PENDING-85 — two boundary-case classifier verdicts (arcades-project born-digital on 1,664 fonts; haydn-mozart-beethoven 4,537 w/pp) need eyeball before per-file verdicts gate anything.
• The ARC map's function has no successor — the open-work register carries the content; nothing exists the steward can open and orient by. Short to regenerate now that ARC is nearly closed.
• Chamber V1's purpose — displaced a seventh time, deliberately, and now genuinely downstream of a corpus that can be verified at all.
• 15 dormant governance items (March–May), untouched again. Governance ledger this session: 15 → 19 open. I opened four and closed none. Three were required by the ruling; that is a real cost, honestly stated.
• Skill-harvest register compaction — still owed since 2026-07-22, still over read caps.

PAUSE STATEMENT: I am about to be away and do not know what will have changed. Nothing is half-finished: the ruling is filed verbatim, REVIEWED-83 is placed, the dockets are written, the classifier is scratchpad-only and wired to nothing, no hash or binding was touched, and Harrison's canonical is exactly as it was this morning. What I want to find still pulling is the Q3 demonstration — because the whole session argued that a gate can report honestly while verifying nothing, and only a guard demonstrated to fail on a constructed hazard turns that argument into a fact. The failure mode to guard against is drafting the spec text first because it is the pleasant part; the demonstration is a precondition, and writing the row before the guard has been shown to bite would reproduce, inside the repair, the exact shape the repair exists to end.

LITERAL QUESTION for next-Claude: Today six errors were caught by the steward, one by the jurist's self-disclosure, and exactly one by an instrument — the verbatim containment check that found my fabricated quote. That one differed in three respects: it compared artifacts in different media (my prose vs the spec file), it was mechanical rather than a judgment, and it carried a positive control. Yesterday's rule (measure in a different medium than the instrument is written in) covers only the first. So: are those three properties jointly the recipe — and if they are, why did I build a nine-control selftest for the classifier and still not build one for the verification method the jurist had to demand? The uncomfortable possibility is that I reliably instrument the part I find interesting and leave uninstrumented the part I have already decided is fine — in which case the missing check is not another control but a standing question at design time: which claim here am I least curious about?

State at wrap: chamber-library 4 dirty (jurist package + ruling + last session's ledger and tool-log) — uncommitted, two remotes, steward's call. dotfiles pushed. REVIEWED-83 placed at REVIEWED.md:847. 19 open governance items. classify_pdf_origin.py + census TSV are scratchpad-only, wired to nothing.