The diff adds only ledger records from one verify run plus by-hand operator corrections, all of which are stamped and disclosed. The verify account is an honest ledger: 31 of the re-submitted candidates are refused or blocked with concrete, per-record reasons (non-atomic sourceStatements split, the p. ix locator on AMZ-E090 rejected as unsupported, Markham-anchored claims blocked because their anchor source was never retrievable), and every second-reader direction change (supports→context, claimId drops) is recorded as a readerAct with model, runId, promptVersion and reason — which strengthens rather than dodges §3.5/§3.8. The operator corrections answer prior seats' objections on the record and are append-only: the duplicate AMZ-E110 id is renamed AMZ-E112 with a superseding disposition row, AMZ-E107/E108 are tombstoned as a duplicate and a still-compound record, AMZ-C072 is split into AMZ-C072/AMZ-C085 with its own anchors, and the AMZ-C078 id collision with a blocked Markham candidate is corrected in a further entry rather than by rewriting the earlier ones (§3.2, §3.14). The two hand-made evidence records AMZ-E113/E114 candidly state that no second model has read them and that this panel is their check, so no unratified draft is presented as ratified (§3.15). Citation verification found no identifiers, AGENTS.md is untouched, no confidence material appears, and I see no self-contradicting locator or quote in the added lines.

The verification run retrieved and read the existing Medina source, admitted only records whose quotations, locators, relevance, and directions passed review, and recorded 31 refusals or blocks with concrete reasons. The operator then corrected the duplicate AMZ-E110 identifier append-only as AMZ-E112, tombstoned duplicate or compound evidence AMZ-E107/E108/E086, and split AMZ-C072 and its evidence into atomic C072/C085 and E113/E114 records without erasing the run’s original output. Those corrections carry model, session runId, date, prompt version, and explicit supersession history, while the original collision and rejected records remain inspectable. Evidence directions distinguish support, qualification, and context, and the records explicitly separate source statements from editor inferences and note dependence on Medina’s quotations of earlier authors. No AGENTS.md edit, confidence material, failed added citation identifier, unratified assessment, or weakening of the verification gate appears.

The change adds claims, evidence records, dispositions, and proposal/governance files for the pre-columbian-amazon case following verification and operator corrections. The added claims (AMZ-C068, AMZ-C069, AMZ-C072, AMZ-C074–C077, AMZ-C085) are strictly atomic, well-anchored to Medina's 1894 edition of Carvajal, and cleanly separated by proposition. The operator's manual interventions explicitly resolve prior panel objections from #353: fixing the AMZ-E110 ID collision by creating AMZ-E112, rejecting duplicate/compound records (AMZ-E107, AMZ-E108, AMZ-E086), splitting multi-ground propositions into atomic claims and evidence (AMZ-C072/AMZ-C085, AMZ-E113/AMZ-E114), and avoiding proposal ID clashes. All lineage, run IDs, session stamps, and second-reader determinations are preserved in history.yaml and dispositions.yaml in full compliance with §3.8, §3.14, and §3.15. No citations are fabricated, AGENTS.md is untouched, and no constitutional checks are loosened.

The verify run admits only records the second reader passed, refuses 31 candidates with concrete atomicity/relevance/locator reasons, and stamps readerActs when direction or claim links change; retrieval of the Medina OCR is on the record and citation tooling found no new identifiers to fabricate. Operator corrections answering the prior #353 park are append-only and reconstructable: the AMZ-E110 id collision becomes AMZ-E112, compound AMZ-E086/AMZ-C072 are split into atomic AMZ-E113/E114 and AMZ-C072/C085 with distinct anchors, and duplicate/compound AMZ-E107/E108 are tombstoned as rejected rather than silently rewritten. Provenance is honest (ai_extracted labels, origin lineage, session_01AiXD4y5Y91KC7iM3uBhWmu / operator-by-hand-v1 stamps, history entries naming the seat objection fixed); AGENTS.md is untouched; no confidence material appears. Harvested arbiter files are verbatim prior verdicts, not constitutional edits. Residual quality note only: E113/E114 were hand-split without a fresh second-reader pass, but the quotes sit in the same already-read passage and this panel is the stated multi-model check.

The change routes previously blocked records back through the ordinary verify verb (same second reader, same checks), which refuses 31 of 54 candidates with concrete per-part reasons — atomicity splits, locator misattributions, relevance failures, and blocks with retrieval routes — so the gate is applied, not dodged. The operator's by-hand corrections (AMZ-E110→AMZ-E112 collision fix, tombstones on duplicate/compound E107/E108, split of C072/E086 into C085/E113/E114) are each recorded as append-only history entries with reasons, and the two new records' quotes are the same Spanish spans the run's own second reader already verified verbatim at Medina p. xiii/n.6 (judgment hash bc192270…), so I can confirm them from the material rather than taking the operator's word. The history honestly discloses that no second model read the split parts and that this panel is their check, which is exactly the process §3.15 prescribes; the corrections tighten (duplicate-id check in verify) rather than weaken the checks. No AGENTS.md edit, no citation identifiers added (mechanical verification found none), no confidence material, and disagreement is displayed in dispositions and history rather than hidden.