The diff adds only new files: a candidate edition, its draft assessment, five check-role assessments, run records, raw replies, a comparison and spend ledgers. It touches no claim, evidence or source record, no tooling, and not AGENTS.md, and it adds no citation identifiers. Every assessment is a new append-only file stamped with runId, producedBy, model, date and promptVersion, with humanReviewed: false and role draft or check. The RUN ACCOUNT confirms each seat reply and the repaired drafter reply are carried whole by the installed files, and the spend ledgers reconcile exactly with both run.yaml totals (e.g. $3.004615 and 252,573 input tokens including 384 cache reads). The ratification claim checks out on §3.15's graded scale. On the case verdict, three seats say well_supported and two provisionally_supported, all within one step of the draft. All five seats are within one step on every draft load-bearing claim (C003, C002, C004, C011, C013, C014, C010). The six splits are exactly the claims held unresolved (C019–C023, C029); none is load-bearing in the draft or in any seat's own list, so they are displayed rather than parked. The fresh check (10:31) postdates the edition (09:45), and the rationale openly says the drafter answered the earlier panel, so standing does not rest on checks the drafter saw; the missing shownChecks stamp is disclosed, not misattributed. The article tells anchor-only claims as reports ('This site has not yet checked…'), consistent with §3.6/§3.9, and its founder-funding note quotes a published funding statement, not confidence material. Minor points only: the rationale should name the 2026-09-21 check run it answers, and the draft leaves C016/C017 out of loadBearing though every seat lists them — immaterial here since all seats are within one step, but worth watching because that designation controls splits.

The new assessment upgrades AMZ-C008 to well_supported while expressly acknowledging that its bundled elements have materially different support: attached records concern trees, pre-Columbian attribution is contested, sibling records carry dark earths and raised fields, and forest islands have no checked record. The article nevertheless publishes all four elements together as one claim that they record long human reshaping, allowing support for some conjuncts to grade the unsupported conjunct and violating the atomic-claim rule. A similar, smaller source-versus-inference problem appears in AMZ-C026: the article says “fishing all year,” although the check record notes that dry-season taxa establish year-round resource potential more directly than uninterrupted year-round exploitation. Split AMZ-C008 into independently assessable propositions and label AMZ-C026’s seasonality conclusion as inference or narrow it to what the remains directly establish.

The pull request presents a leaner edition of the 'pre-columbian-amazon' case along with complete governance records, spend tracking, an incumbent comparison unanimously preferring the candidate, and a fresh independent five-seat check panel. All five models on the check panel (Gemini, GLM, GPT, Grok, Opus) returned case verdicts concurring with the draft assessment ('well_supported' or 'provisionally_supported', within one step), and none of the load-bearing claims split the panel. No primary claim, evidence, or source records were modified, and the omitted raw working files are fully verified and accounted for by the tooling in the run account. The change adheres strictly to the constitutional requirements for edition updates and multi-model concurrence.

The diff adds only an append-only edition, its draft assessment, five blind multi-vendor check assessments, run/spend/comparison records, and proposal working files; no claim, evidence, source, or AGENTS.md record is touched, and citation tooling found no new identifiers. The draft is stamped role: draft with runId/model/date/promptVersion; the fresh check (after the edition) installs five independent seat assessments the drafter never saw, all carried whole from seat replies, with case verdicts within one step of the draft’s well_supported (3 well_supported, 2 provisionally_supported) and no split on a load-bearing claim. Verdict moves name the seats and the deciding records; six featured claims are held unresolved against every seat’s upgrade because none has a checked evidence record, which respects §3.6 and honest labels. The article is shortened within budget, drops internal ledger jargon, tells unchecked material as reports, and keeps competing readings and negative evidence visible; comparison is unanimous candidate-preferred. Repair of three malformed claim-span closers is recorded and the repaired reply matches the installed edition/assessment. Nothing fabricates provenance, presents an unratified draft as ratified, exposes confidence material, or weakens §3.15 checks.

No AGENTS.md edit appears anywhere in the diff, and the change touches no claim, evidence, or source records (the run account's records digest is empty, and the citation verifier found no added citation identifiers), so no fabrication or provenance surface is introduced; the article's sources ride the existing verified ledger. The ratification process is sound: the tooling confirms all five seat replies were carried whole into the installed check assessments, the case verdict is 3× well_supported / 2× provisionally_supported (all within one step of the draft), and I checked each of the draft's seven load-bearing claims seat-by-seat (C002, C003, C004, C010, C011, C013, C014) — none splits, so 'ratified, five of five' is an honest standing under §3.15's graded scale, not a presented-as-ratified draft. Disagreement is displayed rather than hidden: GPT's provisionally_supported dissent, the six claims held at unresolved against upgrades from every seat for lack of checked evidence records, and the declined upgrades are all stated in the edition's rationale and the article tells unchecked claims as reports. Append-only assessment files carry full provenance stamps (runId, producedBy, model, date, promptVersion, humanReviewed: false, basis.ledgerHash), the founder-funding of the 2026 study is disclosed in the article's own text, and the repair round (malformed bracket markers) is recorded in errors.md and verified by the tooling's field-by-field comparison. The missing shownChecks on the edition assessment is a housekeeping gap, but the fresh blind check run postdates the edition and supplies the ratification the PR claims, so nothing publishes on the unverified basis the PR body itself flags.