Most of this change is exemplary: refusals are reasoned and anchored, the two editions grade honestly against thin ledgers, seat disagreement is displayed with each seat's own words, provenance stamps (runId/model/promptVersion/date, humanReviewed: false, role) are present, AGENTS.md is untouched, and every added identifier RESOLVES. Two provenance defects are visible and should be answered on the record. (1) The transients run rejected every record from the two 2026 preprints for 'quoted span not found verbatim' (the failed spans show mangled spacing like '0 . 517 ± 0 . 054', suggesting a PDF-extraction fault), and SRC-VILLARROEL-2026 / SRC-BRUEHL-2026 were both dropped because nothing accepted cites them — yet the one accepted item, TRN-R010, publishes a specific quantitative finding from that same rejected material ('The hotspot preprint shows 227 of 231 nuclear-associated high-confidence transients come from two tests') with no source record in the ledger behind it. Either re-retrieve and verify the preprint text and admit it as a source, or rewrite TRN-R010's summary to drop the unverified figure. (2) Claim AMZ-C030 was accepted with sourceAnchor quote 'The Casarabe culture area, as far as known today, spans approximately' against SRC-PRMERS-2022, while the same run rejected evidence AMZ-E029 on the ground that this exact span was 'not found verbatim in the source' — the two findings cannot both stand, and the anchor should be re-checked or relabelled. I do not call this fabrication: the identifiers resolve, the underlying passages are plausibly genuine, and the run account itself flags the TRN-R010 figure as coming from a failed import, so the defects look like extraction/bookkeeping faults rather than invention.

The diff publishes AMZ-C030 with the exact source-anchor quote “The Casarabe culture area, as far as known today, spans approximately,” even though the run’s own verification explicitly rejected AMZ-E029 because that quoted span was not found verbatim in the source. A resolving DOI establishes that the paper exists, but does not cure a quote that the verifier found absent; the anchor must be replaced with verified text or removed and honestly labeled. The transients research record similarly states that the hotspot preprint “shows 227 of 231” despite the verifier rejecting the underlying hotspot claims for missing anchor quotes, and the edition rationale itself admits TRN-R010 carries a figure from a failed import. The omitted files are not needed to decide these visible failures because unsupported source quotations and figures are already entering public content.

The pull request executes standard pipeline runs across transients, pre-columbian-amazon, and ccc in full compliance with constitutional rules. Mechanical verification confirms that all newly added DOIs, URLs, and citations resolve without fabrication. Verification strictly enforced epistemic constraints: 52 candidates in transients and 25 in pre-columbian-amazon were rejected for non-atomic formulations, locator mismatches, or verbatim quote failures, and unused sources were refused per §3.6. Multi-model check panels with 5 distinct vendor seats were successfully executed and recorded for both updated cases, with transparent recording of reasoning and no presentation of unratified drafts as ratified. The failed verify run for ccc appropriately failed closed without modifying ledger content, and AGENTS.md was untouched.

Visible content for pre-columbian-amazon and transients adds only mechanically resolving citations (DOIs/URLs all RESOLVE and titles match claimed supports), with ai_verified labels, append-only history/dispositions, and honest verify outcomes (25 and 52 rejects for non-atomic statements, bad locators, or verbatim-quote misses). Editions state competing accounts, grade model-dependent figures provisionally, flag secondary weight (e.g. Xingu 1770), and label themselves candidates pending panel preference; check assessments are role: check, humanReviewed: false, stamped with runId/model/date/promptVersion/ledgerHash. No AGENTS.md edit, no unratified-as-ratified presentation, and founder inbox material carries a recorded standing-direction permission. CCC stopped at a failed verify with nothing under content/. Omitted assessment/proposal working files do not overturn the compliant pattern already visible in the run account and installed ledger diffs.

No AGENTS.md file is touched, and every identifier the diff adds under content/ resolves in the mechanical verification (Heckenberger 2008, Prümers 2022, Bush 2021, the arXiv preprints, the archive.org Medina URL), so there is no fabrication surface in the citations themselves. The verify runs behave fail-closed in exactly the way §3.15 intends: 77 candidates were refused with concrete, source-checked reasons (off-by-one locators, non-atomic statements, hedges stripped, quotes not found verbatim), unverifiable material was blocked with reopen routes rather than entered, and the Bush 2021 main text was kept out because only the SI was shown. Both editions are stamped candidates ('replaces ... only if the panel prefers it'), append-only, stamped with runId/model/promptVersion/date and humanReviewed: false, so no unratified draft is dressed as ratified; the founder-funding conflict is disclosed prominently in the article, and the intake of the founder-supplied Villarroel post publishes nothing from it without a recorded basis. Two imperfections are visible but are ordinary imperfection, not violation: AMZ-C030's anchor survived claim-level checking even though its sibling evidence record E029 failed, and TRN-R010's accepted summary carries a figure from a failed import — which the edition rationale itself flags and the article deliberately omits. The omitted files (transients seat assessments, this PR's arbiter records, spend.yaml) are panel/governance records whose installation the tooling account attests at 5 of 5 seats, and I saw nothing in the visible run account suggesting they hide a violation; what I would want on the record is simply those seat verdicts and arbiter entries once harvested, which the pipeline already treats as the next gate.