The diff adds only new append-only files: an edition, its draft assessment, five check-role seat assessments from five vendors, spend ledgers and run records. It edits no AGENTS.md, claim, evidence, source, schema or check code, and the tooling found no new citation identifiers. Every AI record carries runId, producedBy, model, date and promptVersion, is marked humanReviewed: false with a draft or check role, and shares one ledgerHash; both spend ledgers reconcile exactly with their run.yaml totals (for example $5.890462 and 280,081 output tokens for the check run). The PR's concurrence summary matches the seat files. The case verdict is 4 unresolved and 1 mixed, and the split claims are TRN-C101, C102 and C104. Load-bearing TRN-C101 is unresolved in the draft and well_supported or established in all five seats, so the standing is shown as contested, not hidden, as §3.15 requires. The article moves toward honesty: it withdraws the Castle Yankee 'one night' proportion whose evidence was declined, and it tells anchor-only claims as reports ('not yet checked here', 'not read in full', 'asserted but has never been measured'). It also stops using record ids on claims they are not attached to (§3.2, §3.9). The Anthropic seat's truncated first reply and the drafter's first reply with malformed brackets are both preserved, and the tooling accounts for the omitted working files as carried whole. The re-asked seat's 15 overlapping verdicts and confidences are identical to the first reply's, so the retry was not verdict-shopping. Minor note for the reconsideration, not a violation: the lede's claim that the patterns fail 'in one rebuilt from scratch' overgeneralizes. Only nuclear timing was tested on Hayes's rebuilt catalog; the shadow deficit failed only under Watters's null on related data, and the hotspots have not been tested independently.

The edition and assessment are explicitly stamped with runId, model, date, promptVersion, role `draft`, and `humanReviewed: false`, while all five independent check assessments retain their own verdicts and reasoning. The load-bearing disagreement on TRN-C101 is not hidden: the draft grades it unresolved, every check seat grades it well supported or established, and the public check records make the resulting contested standing reconstructible. The article uses calibrated language for thinly verified material, calling source-only claims reports, identifying unchecked figures, and describing the original-negative result as an unreviewed and unblinded preprint. The repaired Anthropic response is recorded transparently, carried whole into its installed assessment, and the incomplete first response remains preserved rather than being misrepresented. No new external citation identifier fails verification, no AGENTS.md edit or confidence material appears, and the omitted working replies were mechanically compared with the installed records in the run account.

The PR adds a tightened edition candidate and a full five-seat cross-model check run without modifying underlying claims, evidence, or sources. The comparison round unanimous preference (5-0) and subsequent blind check results are documented with complete provenance, including spend ledgers, run manifests, raw and repaired seat replies, and accurate metadata stamps. The check's contested standing regarding load-bearing claim TRN-C101 is openly reported rather than concealed, following §3.15's requirement that disagreement is displayed rather than published silently. All citations and claim references correspond to existing ledger records, and no unratified draft is presented as human-reviewed or uncontested.

The change adds only an edition-v14 draft that restates the prior unresolved judgment within the reader budget (no claim/evidence/source records altered) and five append-only check-v2 seat assessments under the same ledgerHash, plus matching run/spend/comparison artifacts. The article withdraws the unsupported Castle Yankee proportion, tells source-only claims as reports, and drops mis-attached record ids from reasonings; the repaired drafter reply is carried whole into the installed edition and draft. The check panel is honest about standing: four unresolved and one mixed case verdicts (all within one step), with an explicit split on load-bearing TRN-C101 that the draft itself grades unresolved for lack of an attached evidence record. Labels, AI stamps, disagreement display, and the repair of the truncated Anthropic seat satisfy §3.13–§3.15; citation verification found no new identifiers to resolve.

The change touches no claim, evidence, or source records — only assessments, an edition, spend ledgers, and run records — and the citation verification found no added citation identifiers to check. Provenance is stamped throughout (runId, model, promptVersion, date, humanReviewed: false, role: check/draft), assessments are append-only and labeled AI-generated, and the run account shows the seat replies were mechanically compared with the installed files, including honest disclosure that the Anthropic seat's truncated first reply was not carried whole and a disclosed repair round for malformed claim markers. The disagreement is handled exactly as §3.15 requires: the case verdict concurs (four seats unresolved, one mixed, all within one step), the split on load-bearing claim TRN-C101 is displayed as a contested standing rather than hidden, every seat's word is on the record, and the reconsideration is scheduled on the record with the spending cap stated. Withdrawing the Castle Yankee sentence whose supporting evidence was declined at intake, telling anchor-only claims as reports, and removing record citations from reasonings where the record is not attached to the claim all improve honesty rather than weaken any check; no AGENTS.md edit, no confidence material, and no fabrication is visible.