The change adds two edition candidates and their paired draft assessments, spend ledgers, run records, harvested prior arbiter verdicts, and one append-only history entry; AGENTS.md is untouched and the citation tooling found no new identifiers under content/, so there is no fabrication surface I can check against. Provenance is stamped throughout: each assessment carries runId, producedBy, model, date, promptVersion, humanReviewed: false, role: draft, and the header comment says plainly 'NOT human reviewed; standing derives from blind checks' — no AI record is dressed as human, no unratified draft as ratified (§3.15, §4). Epistemic discipline is visibly applied rather than loosened: six anchor-only claims (C019, C020, C021, C022, C023, C029) are graded unresolved and explicitly said to carry no weight, evidence direction and diagnosticity are separated from credibility (§3.4, §3.5), independence is treated head-on (E089 showing the two manuscript branches are not independent witnesses, Herrera as a redaction rather than a fresh witness, the tree-species study 'counts once, not twice' — §3.10), negative evidence (E010, E011) is carried as spine rather than afterthought (§3.11), and the founder's funding of the 2026 LiDAR paper is disclosed in the article itself. The known defect in the earlier assessment (E115 cited for the Upano chronology; fish weirs listed in whatIsClaimed) is not hidden: review note #358 is recorded verbatim in history.yaml with the operator's reasoned answer, and the second edition in the same PR corrects both under edition-v10 while leaving the append-only record intact (§3.14). The one omitted file is a working reply.json, whose corresponding edition and assessment are shown whole in the account, so it cannot hide a violation I would otherwise see.

The visible run account and diff honestly label both assessments as AI-generated drafts, preserve the flawed first assessment append-only, and record its correction in the later edition. The later assessment clearly distinguishes admitted evidence from anchor-only claims and corrects the misuse of AMZ-E115 without erasing history. However, `proposals/2026-09-19-edition-pre-columbian-amazon-045109/reply.json` is an added public file omitted from both the diff and the run account, so I cannot determine whether it contains inconsistent provenance, confidence material, or claims not represented in the generated outputs. I would need that complete file, or a mechanically authenticated account of all its fields, before finding the whole change compliant.

The pull request adds two successive edition candidates and draft assessments alongside transparent run and spend records, directly addressing a prior review note (#358) on the record. The succession from edition-2026-09-19-042009 to edition-2026-09-19-052132 preserves append-only integrity, correcting the attribution of AMZ-E115 strictly to the builder claim (AMZ-C086) and explicitly marking the fish-weir infrastructure (AMZ-C019) as anchor-only in the case assessment header. Both candidate articles maintain strict alignment with the ledger, distinguishing admitted evidence from anchor-only claims and qualifying population numbers as model envelopes rather than censuses. Provenance stamps, hashes, and run lineage are preserved throughout, no citations failed mechanical verification, and AGENTS.md is untouched.

Two successive edition drafts for pre-columbian-amazon: 042009 incorporates newly admitted Medina chronicle records (branch non-independence E089, invention-charge genealogy) into treatments and the article without moving the case verdict; 052132 is the append-only fix answering review note #358, correcting whatIsClaimed so fish weirs are marked anchor-only and citing E073/E074/E115 only beside the Upano claims they attach to. Assessments are stamped role:draft, humanReviewed:false, with runId/model/date/promptVersion and ledgerHash; six anchor-only limits stay unresolved; competing accounts, calibrated uncertainty, and negative evidence are explicit. History records the operator reply to the review note; AGENTS.md is untouched; citation tooling found no new identifiers; the omitted reply.json is the second run’s working file and does not hide a content violation visible in the run account or assessments.

The two edition runs are honestly stamped as AI drafts (humanReviewed: false, 'NOT human reviewed', append-only), with runId/model/date/promptVersion provenance on every record, and the mechanical verification found no citation identifiers to check, so no fabrication surface is visible. The one substantive defect — assessment 042009 spending AMZ-E115 on the Upano dating claim it is not attached to, and listing fish weirs in whatIsClaimed while grading AMZ-C019 unresolved on its anchor alone — is not hidden: it was raised as review note #358, recorded append-only in history.yaml, and cured by a second edition run (edition-v10) whose whole rationale is that correction, with the flawed assessment left unedited as the append-only rule requires. Anchor-only claims (C019, C020, C021, C022, C023, C029) are consistently named as anchor-only and graded unresolved, the founder-funding conflict is disclosed verbatim in the article's transparency block, and disagreement is displayed rather than hidden. No AGENTS.md edit appears, no check is loosened, and the harvested arbiter files are verbatim prior-verdict records, not constitutional changes. The omitted 045109/reply.json is a working file whose substance (edition, assessment) I can already see in the run account, so it does not leave the question undecidable.