Both added sources resolve mechanically (Rostain 10.1126/science.adi6317; Achmatowicz 10.12797/SI.17.2018.17.03, plus its landing page and Wayback snapshot), and their titles match what the records claim they support. The verify account is the kind of honest ledger §3 wants: 58 of 73 candidates refused or blocked with concrete, per-part reasons (non-atomic sourceStatements split into E073–E077, locators narrowed where the quote sits on p.4 or p.11 rather than the compound page range, directions reset to 'context' by the second reader with reader acts stamped), and SRC-CRUZ-1651 refused outright because nothing accepted cited it (§3.6). The Carvajal label is handled correctly rather than hidden: history carries the original raise to ai_verified, then the operator's append-only reversal to 'unverified' with a note scoped to exactly the pages read (introduction pp. i–lxxviii) and the statement that the verifier read the item page, not the OCR, answering the prior panel's park and note #341; the 44 'failed' rows are re-dispositioned 'blocked' with retrieval routes, which is the truthful label since those quotes were checked against nothing. The one FAILS in verification is the descubrimientod00carvgoog djvu.txt URL, which appears only as a future retrieval route inside blocked dispositions, not as a citation carrying weight, and its sibling archive.org text URL resolves — that is a transient reachability fact, not fabricated provenance. The editions grade six anchor-only featured claims to 'unresolved' at low confidence and say so in the prose ('the ledger holds that only on the paper's anchor'), which tightens rather than loosens the checks (§3.8, §3.13), and the founder-funding disclosure is printed in the article. One review-note-level defect I would want fixed: SRC-ACHMATOWICZ-2018's verificationNote says 'text retrieved and shown to the drafter' while its own reliabilityNotes say only the landing page/abstract was shown — the note should be scoped to the abstract, as the Carvajal note now is.

The verifier rejected AMZ-E067 because occupation dating and attribution of platform/road construction are distinct propositions that could be true or false independently, yet AMZ-C061 still combines them: it dates “the Kilamope/Upano occupation that built the earthen platforms and roads.” Its admitted evidence, AMZ-E073, states only the occupation date range, so support for that measurement is silently extended to the separate builder attribution. This compound claim is then used in the assessment and article as if the whole proposition were evidenced. Split AMZ-C061 into dating and construction-attribution claims, admit evidence for each, and revise the linked assessment/article language; the omitted working files do not prevent this determination because the conflict is explicit in the diff and run account.

The change fully complies with the constitution and demonstrates sound epistemic discipline across both cases. For pre-columbian-amazon, new sources (Rostain 2024 and Achmatowicz 2018) mechanically resolve and are accompanied by carefully split, atomic claims and evidence records, while unread or non-verifiable candidates from Carvajal and Laureano de la Cruz were appropriately rejected or marked blocked. Furthermore, the operator's corrections directly address prior panel feedback (#341) by restoring SRC-CARVAJAL-1542's verification label to 'unverified' and cleaning AMZ-E074's locator. In transients, the new edition candidate correctly applies the anchored-is-not-evidenced principle (§3.6, §3.13) by regrading anchor-only claims to unresolved at low confidence. AGENTS.md is untouched, no checks are loosened, and the run account and visible diff sufficiently decide compliance despite routine omitted proposal working files.

The admitted content is tightly gated: verification accepted only Rostain 2024 and Achmatowicz 2018 (both DOIs resolve to the claimed titles), six atomic evidence splits with second-reader direction/locator acts, five claims, and two research items, while 58 candidates were refused or blocked with concrete reasons (quotes not found, non-atomic statements, unretricable anchors, §3.6 bare sources). SRC-CARVAJAL-1542 is correctly left unverified after the operator undid the premature ai_verified stamp and rewrote the note to the pages the drafter actually read; failed Carvajal/Cruz quote-checks are corrected to blocked with re-fetch routes rather than treated as positive non-matches. Editions and draft assessments regrade anchor-only featured claims to unresolved, disclose abstract-only and unread-body limits, and are stamped role: draft / humanReviewed: false. No AGENTS.md edit, no confidence material, no fabricated locator or citation in what was admitted, and prior panel notes (#341) are answered on the append-only history. Omitted working files are non-decisive given the full content diff and run account.

The chain's content runs behave the way the constitution demands: the verify account admits only 2 sources, 6 evidence records, 5 claims and 2 research items while refusing 58 candidates with concrete reasons (verbatim-quote failures, non-atomic statements, §3.6 source-without-evidence refusals), and every admitted identifier resolves in the citation verification (Rostain 2024 and Achmatowicz 2018 DOIs both RESOLVE; the failing djvu.txt URL appears only as a re-verification route, not a citation). The most scrutiny-worthy material — the operator's by-hand corrections — is exactly the opposite of check-dodging: 44 mislabeled 'failed' dispositions are corrected to 'blocked' because the OCR text was never actually served, and the premature 'unverified'→'ai_verified' raise on SRC-CARVAJAL-1542 is reversed on the record after the panel park, with the final sources.yaml state honestly 'unverified' and a note scoping what was and was not read. Both edition candidates are stamped draft overlays (humanReviewed: false), the articles repeatedly tell anchor-only propositions as 'what the ledger holds' rather than findings, and the regrades of anchor-only claims to unresolved tighten rather than weaken the assessments. The founder funding disclosure, the recorded reader acts, and the append-only history entries answering review note #341 all preserve reconstructability. Residual notes I would log rather than park: the Achmatowicz record's year '2018' differs from Crossref's 2019 (plausibly issue-vs-online metadata, worth a confirmation), and the research item AMZ-R007 leans on a tertiary '≤5,000' snippet the ledger itself declined verbatim — but the item frames it as unconfirmed hypothesis, not evidence. AGENTS.md is untouched; no confidence material appears; no check is loosened.