The diff adds one edition candidate plus its run record, model reply and spend line; nothing under AGENTS.md or any check is touched, and the tooling found no citation identifiers to resolve. The edition file's own header states it replaces the incumbent only if the panel prefers it, so no unratified draft is presented as ratified, and it carries runId, producedBy, date, model and promptVersion plus ledger/inputs hashes (§3.14, §3.15 stamping). Epistemically the article is unusually scrupulous: the founding self-published page is named once as a source and thereafter attributed locally ('the page states', 'on the page's word', 'the page's captions'), self-reports are defined as weak support at most, claims with no evidence record are explicitly flagged ('captions without records', 'unresolved, for want of a record'), and unread material is marked (abstract only, transcription not checked against the printed page) — consistent with §3.7, §3.8, §3.9. Negative and undermining evidence is displayed prominently (the bundle null, the excavators' 'missing links', the later-added Rapa Nui reliefs read against a nineteenth-century story), and the verdict stays 'unresolved' with a plain statement that it does not mean balance (§3.11, §3.13). The rationale also records specific corrections — removing a detail drawn from a declined record and fixing 'two centuries after contact' to what DEEP-E071 states — which is honest housekeeping rather than fabrication. Cost, tokens and spend are recorded, and the run's notes match the run.yaml and the account. I see no rule broken; the only residual uncertainty is whether the claim IDs and grades match the ledger records, which is outside the diff but is CI's mechanical job.

The edition is explicitly marked as a candidate that replaces the incumbent only if preferred, so it does not present an unratified draft as ratified. Its run, model, date, prompt version, predecessor, assessment hash, ledger hash, and producing run are recorded, while the rationale explains exactly what changed and what remained unchanged. The article distinguishes page self-reports, primary material, untested priors, source statements, and editorial inference; it repeatedly identifies missing records, unread texts, unchecked transcription, and unresolved claims rather than laundering them into certainty. It presents four serious competing accounts, separates descriptive credibility from historical diagnosticity, and states concrete tests that could move the assessment in either direction. No new citation identifiers or source records are added, the citation tool reports no identifiers requiring resolution, no files are omitted, and neither AGENTS.md nor the constitutional checks are altered.

The pull request adds an edition candidate (`edition-2026-09-20-112834`) for the Deep Memory case that re-adopts the incumbent assessment while applying protocol v11 to treat the founding seed page strictly as a cited source rather than an editorial narrator. No new evidence or source identifiers were introduced, and provenance details (runId, model, promptVersion, hashes) are intact and consistent across the proposal files and spend ledger. Claims, priors, and anchors are clearly distinguished in the prose in compliance with §3.1, §3.8, §3.9, and §3.13. No constitutional rules are violated.

The diff adds only a new edition candidate (plus its run/spend/reply records) that re-adopts the incumbent assessment unchanged, keeps the identical featured set and crux order, and rewrites the article under edition-v11 so the seed page is introduced once as the case's founding source and thereafter attributed locally ("the page reports", "on the page's word", "the page's caption") rather than as a narrator with independent judgments. Rationale, stamps (runId/model/date/promptVersion/previous/basis hashes), and the article itself disclose priors, self-report weakness, unread primaries, declined details removed, and the unresolved verdict with explicit "what would change our mind"; no new citations, locators, or evidence records are invented, AGENTS.md is untouched, and the RUN ACCOUNT confirms honest re-adoption with no ledger movement. Omitted-files list is empty and citation tooling found nothing to resolve, so nothing hidden or fabricated is visible.

The change is an edition candidate that re-adopts the incumbent's already-assessed judgments and rewrites only the telling, which is properly presented as unratified ('Replaces ... only if the panel prefers it'), so no unratified draft is presented as ratified (§3.15). No new external citations or locators are added — the verification section found none — and all quoted figures (55%/926, 62.5% vs 9%, 37%, 263 summaries) are attributed to the founding page's self-reports and consistently labeled 'held on the page's word' with the weak-support footing disclosed, satisfying §3.8 and §3.13's honest labeling. The article displays the competing accounts, negative evidence, and the counter-argument to the featured hypothesis, and it removes one detail that had come from a declined record while correcting the birdman-myth dating to match DEEP-E071 — the opposite of silent weakening. Provenance is stamped with runId, model, promptVersion, and date, the run record and spend ledger are coherent and cross-linked via producedBy, and no AGENTS.md edit or confidence material appears anywhere in the diff.