Every identifier the diff adds under content/ resolved mechanically (Schmidt's Neo-Lithics PDF, the two Tepe Telegrams posts, the IJAS and Antiquaries Journal DOIs, the Berossos/Manetho scan, the Samizdat transcription), and each source record's verification label ('ai_verified', with retrieval date and run id) matches the run account's retrieval list; no locator in the diff contradicts its own content. Provenance is stamped throughout: the assessment header says plainly 'Append-only; NOT human reviewed', carries runId/producedBy/model/date/promptVersion and humanReviewed: false, and the edition is labelled a candidate that replaces the incumbent only if the panel prefers it — no draft is dressed as ratified (§3.15, §4). The verify pass is visibly strict rather than dodged: 55 candidates were refused with reasoned, anchored explanations, splits are recorded with readerActs showing who changed a direction or dropped a claim link and why (§3.5, §3.9, §3.14), the Wiggermann material enters only as weak support with its secondhand blog-transcription status stated and the printed page listed as blocked (§3.7, §3.8), and negative/absence records (DEEP-E073, DEEP-E074) are carried explicitly (§3.11). The article and assessment repeatedly mark anchor-only claims as such and grade them unresolved, and the synthesis names DEEP-E020 as the record whose removal would most change the verdict — calibrated rather than falsely precise (§3.13). AGENTS.md is untouched; the only docs change is none, and the harvested arbiter files are verbatim prior-verdict records. My one reservation is minor and not a violation: DEEP-C092's statement ('In short, Oannes taught men all those things conducive to a settled and civilized life') is phrased as the assertion rather than as a source-statement claim, and its sourceAnchor quote ('how to build cities, found temples, and make laws') is a neighbouring clause rather than the 'in short' phrase itself — I would like the anchor quote aligned to the sentence the statement compresses, but the quote is genuine and at the cited page, so this is ordinary imperfection.

The visible run account labels the assessment as an AI draft, preserves run/model/date provenance, records refusals, and the added citations all resolve. However, twelve changed files are omitted, including raw.json, report.md, proposal.yaml, judgments.yaml, and multiple model reply files; these public working records could contain unshown source text, inconsistent provenance, or material whose publication permission cannot be assessed. The run account summarizes the admitted content but expressly excludes those working files, so it does not decide compliance for the complete change. I would need the omitted files themselves, or a mechanically authenticated account covering their complete contents and provenance/confidence status.

The change complies with the constitution across all epistemic and operational dimensions. AGENTS.md is untouched, and mechanical citation verification confirms all added DOIs, URLs, and archive links resolve successfully. The new sources in sources.yaml are honestly marked with verification labels (`ai_verified`) accompanied by clear reliability caveats regarding abstract-only or secondhand blog access. Newly admitted evidence records and claims are atomic, accurately located, and strictly distinguished by direction (supports, qualifies, context). The new draft assessment and edition maintain calibrated uncertainty, honestly stating that the central hypothesis remains unresolved while explaining why object descriptions are confirmed without establishing diffusion or connection. Provenance stamps, spend records, append-only history entries, and comprehensive dispositions (recording both acceptances and reasoned rejections) are completely documented.

The pipeline admits 7 sources, 25 evidence, 9 claims and 2 research items after a second-reader verify that rejected 55 candidates for atomicity, relevance or unsupported anchors, with reasons and splits recorded in dispositions. Every new DOI/URL mechanically RESOLVES; sources carry honest ai_verified labels, exact locators, and secondhand status (Wiggermann via blog) stated as weak. The draft assessment is stamped role:draft, humanReviewed:false, with runId/model/date; the article ties claims to admitted records, keeps anchor-only material unresolved, and leaves the case verdict unresolved with calibrated uncertainty. History and image correction are append-only and provenance-stamped; AGENTS.md is untouched. Omitted files are working drafts whose substance appears in the run account and content diffs.

Every citation the diff adds under content/ resolves mechanically (both DOIs and all URLs, including the Wayback snapshots), and the sources are plausibly what the records claim — Schmidt's Neo-Lithics note, the Tepe Telegrams posts, the Vidale & Micheli and Pitts et al. abstracts, Verbrugghe & Wickersham's Berossus — so no fabrication surface is visible. The verify run is visibly disciplined rather than permissive: 55 of ~82 proposed candidates were refused with specific reasons (non-atomic statements, anchors that overstate the quote, records bearing on rejected claims, a source nothing cites), splits and direction changes are recorded as readerActs with model/run/date, and the Pausanias retrieval failure (HTTP 522) is blocked with a reopen route rather than quoted from memory. Provenance is stamped throughout — runId/model/promptVersion/date, reviewState: ai_extracted, humanReviewed: false, 'NOT human reviewed' headers, append-only history entries naming the runs — and secondhand or abstract-only footing is disclosed as a limitation (the unchecked Wiggermann transcription graded weak, the banduddû record marked as transcription-only). The one content correction (IMG-DEEP-P07) is applied with a reasoned history entry rather than silently. The omitted files are drafter/verify working files whose substance is already visible in the run account, and the harvested governance files are verbatim records of prior verdicts, not edits to AGENTS.md or to the checks; anchor-only claims stay unresolved at low confidence and no unratified draft is presented as human-reviewed.