The diff adds a new edition, its draft assessment, five blind check-seat assessments, harvested arbiter records, run records and spend lines; AGENTS.md is untouched and the citation tooling found no identifiers added under content/, so there is no fabrication surface to test. Every assessment file carries runId, model, date, promptVersion, humanReviewed: false, role (draft|check) and the ledgerHash it judged, and the edition header labels itself a candidate that replaces the incumbent only if the panel prefers it — no unratified draft is dressed as ratified (§3.15, §4). The substantive editorial act is a downgrade: DEEP-C009 moves from weakly supported to unresolved because its only record (DEEP-E023) is a context pointer to an unretrieved Vidale & Micheli study, and the article's Jiroft paragraph is rewritten to match — that strengthens §3.6/§3.7 rather than dodging any check. The article repeatedly states that all twelve records come from one unverified self-published source, grades most featured claims 'unresolved', flags prior-driven grades as such, and displays four competing accounts with named decisive tests, which satisfies §3.3, §3.5, §3.11 and §3.13. Seat disagreement is displayed rather than suppressed: four seats say unresolved, one weakly_supported, one mixed — all within one step of the draft — and each seat's own words are preserved. The one omitted file is a working model reply, deliberately excluded per the run-account convention, and nothing in the account I can see suggests it hides a violation.

The visible records carry mutually inconsistent run provenance. The actual edition execution is recorded as `2026-09-10-edition-deep-memory-120824`, while the published assessment claims runId `2026-09-10-edition-121547` and the edition uses `edition-2026-09-10-121547`; no corresponding execution record for the claimed assessment runId is shown. Likewise, the check execution is `2026-09-10-check-deep-memory-121549`, while raw seat files and installed assessments assert differing shortened or vendor-specific runIds without documenting them as sub-runs. This prevents readers from reconstructing exactly which run produced each assessment, contrary to the required model/runId/date/promptVersion stamps; the fix is to use the actual execution identifiers or explicitly separate artifact IDs from documented sub-call IDs. The omitted `reply.json` is not needed to establish the mismatch among the public records shown.

The pull request accurately updates the Deep Memory edition and adds the corresponding draft and five-seat check assessments, maintaining full alignment with §3.15 multi-model concurrence. The edition corrects DEEP-C009 from weakly supported to unresolved in strict adherence to §3.6 (recognizing that a citation pointer to an unretrieved paper constitutes context rather than direct evidence of wear traces), and updates the article text and rationale accordingly. All five check seats (Gemini, GLM, GPT, Grok, Opus) are recorded with their explicit reasoning and case verdicts, all falling within one step of the draft's unresolved standing. Provenance stamps, run records, spend logs, and harvested arbiter files are properly recorded and append-only, and no citations or constitution files are modified.

The visible change is an append-only deep-memory edition plus five stamped check assessments, run records, spend lines, and harvested arbiter logs. No AGENTS.md edit, no new citation identifiers (verification empty), and no confidence material. The draft honestly grades the thin self-report ledger unresolved, corrects DEEP-C009 to unresolved because DEEP-E023 is context-only (pointer ≠ support), keeps competing accounts and cruxes, and labels role/draft stamps with humanReviewed: false. The five-seat panel is complete; case verdicts (three unresolved, one weakly_supported, one mixed) all sit within one step of the draft on the §3.15 scale and are displayed in full. The failed report run is recorded with its credit error; the sole omitted file is the edition working reply, which the run-account rules exclude by design.

The diff adds no citations (the tooling confirms no identifiers were added under content/), edits no AGENTS.md, and every new assessment and edition file is append-only, stamped with runId/model/promptVersion/date, and labeled humanReviewed: false with role labels — nothing AI-generated is presented as a human or ratified judgment. The edition's content is exactly what the constitution asks for under a thin ledger: it grades the case unresolved, repeatedly tells the reader that all twelve evidence records are weak self-reports from one unverified source, downgrades DEEP-C009 to unresolved on the rule that a pointer to an unretrieved study is not support (a tightening, not a loosening, of §3.6/§3.7), and displays disagreement rather than hiding it — the five check seats (weakly_supported, three unresolved, mixed) all fall within one step of the draft's unresolved verdict under the §3.15 graded scale, with each seat's own words displayed beside the standing. The failed report run for pre-columbian-amazon is fail-closed (zero calls, zero cost, nothing produced) and recorded honestly, and the spend.yaml entries reconcile with the run records (the five check-seat costs sum to the reported 0.68699). Minor imperfections — abbreviated runId strings in the proposals/ seat files versus the full runId in run.yaml — are record-keeping noise, not misattribution of who did what or how it was checked.