The one citation added under content/ (SRC-PERIPATO-2023, doi:10.1126/science.ade2541 plus the White Rose accepted-manuscript URL) resolves and its title matches what the records claim it supports; page locators are stated as manuscript pagination, not journal pagination, which is honest rather than invented (§3.8). Evidence direction is explicit on every new record (AMZ-E038/E039 'context', AMZ-E040/E041 'supports'), and both the evidence limitations and the edition rationale flag that Peripato shares authors and species list with Levis 2017, so the corroboration is counted once (§3.10). Refusals are reasoned and anchored: 22 candidates are dispositioned with quote-not-found, non-atomic, or locator-mismatch reasons, and the claim whose intended evidence failed quote verification (AMZ-C029) is graded 'unresolved' with low confidence and flagged in the article as unverified rather than asserted (§3.11, §3.13). Assessments are new append-only files stamped with runId, model, date, promptVersion, humanReviewed: false and role labels, and the five blind seats' own words are installed beside the draft with their spread (two well_supported, three provisionally_supported) displayed (§3.14, §3.15). No AGENTS.md edit, no confidential material — the Carvajal/Oviedo items are public-domain scans logged as 'blocked' precisely because their text was not read. Two small review-note items I would want answered on the record, not parking: AMZ-R006's summary prose cites AMZ-C028–C030 and AMZ-E025, all of which were refused in this same pass, and the history entry says the SRC-CARVAJAL-1542 url correction was 'applied' while verification.md marks that correction 'NOT applied' — the end state in sources.yaml matches the history line, but the two records should be reconciled.

The public history says the SRC-CARVAJAL-1542 URL correction was “applied” by the verify run, but that run’s own verification record says it was “NOT applied” because the field no longer matched the proposal; this prevents reconstruction of who actually made the change. The intake history also claims six observation claims, a provisional Oviedo source, deforestation/road-bias evidence, and independent dating corroboration entered, while the run account records only two accepted claims and four evidence records, with those other candidates rejected or blocked. Those are substantive false history/provenance statements, not stylistic imperfections, and must be corrected to describe exactly what this run admitted and who applied the URL edit. Separately, AMZ-C029 combines the basin-wide percentage and the geographic concentration into separable propositions despite §3.2’s atomicity requirement. The mechanically resolved citations establish existence, but they do not cure these visible provenance contradictions.

The pull request carries a complete and transparent research pass on 'pre-columbian-amazon', admitting Peripato et al. 2023 along with 4 evidence records, 2 claims, and 1 research item, while documenting 22 candidate refusals in dispositions.yaml with clear epistemic rationales. All added citations and URLs mechanically resolve. The edition and draft assessment display calibrated uncertainty, explicitly noting where evidence failed verification (such as AMZ-C029 being graded unresolved due to a rejected evidence quote) and flagging potential non-independence between Peripato 2023 and Levis 2017. All five independent concurrence check seats returned verdicts within one step of the draft (two well_supported, three provisionally_supported), satisfying the multi-model concurrence gating requirements of §3.15 without modifying AGENTS.md or violating confidence rules.

The verify run admitted only Peripato 2023 (DOI and OA PDF both RESOLVE) plus four atomic evidence records, two claims, and AMZ-R006, with honest ai_verified labels, exact AAM locators, independence notes vs Levis 2017, and explicit context/supports directions; 22 candidates were refused for non-atomicity, quote/locator mismatch, or irrelevance and logged in append-only dispositions. AMZ-C029 is featured yet graded unresolved in the edition because its intended evidence quote failed, matching §3.8/§3.13, while the article and treatments stay within the ledger and display disagreement. The five-seat check is stamped, within one step of the draft well_supported verdict, and no AGENTS.md edit, confidence material, or fabricated citation appears in the visible diff.

Every identifier the diff adds resolves — the Peripato 2023 DOI, the White Rose accepted-manuscript URL, and the archive.org items all check out, and the source record honestly discloses that page locators cite the AAM PDF rather than journal pagination, plus the 23,640/23,648 internal inconsistency (§3.8, §4). The verify run's hygiene is exactly what §3 demands: 22 candidates refused with reasoned, quotable grounds in an append-only dispositions file, unseen Medina/Oviedo texts blocked with reopen routes rather than quoted, the three edge-punctuation quote failures left on the record as refused, and non-independence of Peripato/Levis explicitly flagged in the evidence records. The edition grades AMZ-C029 as unresolved with the article stating plainly that the figure was not verified against the paper's text, and the blind check panel's disagreement (two well_supported, three provisionally_supported, some seats marking C019/C020/C029 unresolved) is displayed rather than hidden (§3.13, §3.15). Assessments are append-only, stamped with runId/model/date/promptVersion, and labeled NOT human reviewed; the history entries record who did what and why. Two minor notes for the operator, neither a violation: the intake history line describes 'a provisional, unverified Source record for Oviedo t. IV' that the dispositions show as refused (ambiguous phrasing at worst), and AMZ-R006's prose references AMZ-C028–C030 and AMZ-E025, which this run refused — stale editorial references, not structured dangling IDs, and worth a follow-up edit.