No claim, evidence, source or research record is touched, and the verification found no new identifiers. The new article attributes founding-page material to 'the page' and discloses what rests on reasoning alone (DEEP-C018, C034, C047 'accepted on general grounds, untested here'). §3.8 and §3.9 hold, and I see no invented locator or quote. All six assessments are new append-only files stamped with runId, model, date and promptVersion and marked humanReviewed: false; their model names match the spend ledgers, whose sums equal the run records exactly ($2.931868 and $2.609855; token totals reconcile, including the 384 cache-read tokens). The RUN ACCOUNT shows every raw seat reply carried whole by its installed check file and the repaired drafter reply carried whole; the first reply differs only by the `{claim=DEEP-C080]` bracket that errors.md records, so the omitted working files hide nothing. The PR's concurrence account checks against the diff: four of five case verdicts sit within one step of 'unresolved', and on load-bearing DEEP-C002, C034, C026 and C018 only 2, 2, 1 and 2 seats sit within one step of 'provisionally supported', so 'contested' is correct and is displayed with every seat's words, not hidden (§3.15). The `reconciles` stamp keeps the answered checks from vouching for the draft, each verdict move is reasoned against the dissent it answers, and the load-bearing set is stated unchanged, so nothing is reclassified to dodge the split. Editorial notes only, not violations: 'report no traceable chain', 'post-contact myth' and the flat 'a Jiroft object that hung on a cord' compress hedged sources, and the component label 'Dates and intermediates: well supported' can be misread as backing a shared-past chronology; these are worth tightening in the next edition.

The edition distinguishes retrieved source statements, founding-page self-reports, general priors, and editorial inference; unsupported results such as DEEP-C056, DEEP-C057, DEEP-C031, and DEEP-C054 remain unresolved and are described as unaudited or unchecked. Provenance is explicit throughout: the draft and five check assessments carry model, runId, date, promptVersion, role, ledger hash, and humanReviewed:false, while the run account confirms that every raw check reply was carried whole. The fresh five-vendor panel preserves disagreement rather than hiding it—three seats say unresolved, one mixed, and one contradicted—and the resulting contested standing is disclosed. The article presents four serious competing accounts, calibrated uncertainty, concrete tests, and source limitations; no new citation identifiers were added, and the mechanically checked omitted replies are fully accounted for by the run account.

The change adds a new edition candidate, an updated draft assessment, and five independent check assessments for the 'deep-memory' case following an edition and check pipeline run. All files adhere strictly to required schemas, provenance stamping, and append-only assessment rules without modifying underlying claim or evidence records. Disagreement across the panel is transparently documented and preserved rather than obscured, correctly reflecting a 'contested' standing under §3.15. No constitution files, confidential materials, or unverified citations are modified or introduced.

The diff adds only a stamped edition (role draft, edition-v14, previous/rationale/measured fields, ledgerHash), its adopted assessment (role draft, reconciles the five prior checks it answers, claim moves each tied to attached records or disclosed priors with seat dissent held on the record), five independent check-role assessments (check-v2, producedBy the fresh run, humanReviewed false, basis ledgerHash, seats carried whole per the run account), and run/spend/comparison artefacts. No claims, evidence, sources, or AGENTS.md are touched; citation verification found no new identifiers; the article and assessments distinguish anchors from records, label unaudited self-reports, keep the case unresolved, and surface the fresh panel split (four load-bearing claims) rather than hide it. Process stamps, append-only overlays, and multi-model concurrence match §3.14–3.15 with no fabrication, confidence exposure, or check-weakening.

The change is a re-edition answering a prior panel split, with the answer itself subjected to a fresh blind five-seat check whose results are installed and displayed, not hidden. No AGENTS.md edit, no new citations (verification found none added), no claim/evidence/source records touched, and every assessment is stamped with runId, model, date, promptVersion, role, and humanReviewed:false, consistent with §3.14 and §4. The draft assessment honestly stamps itself `reconciles` the five checks it was written against so they cannot vouch for it, and the new check (run record, seat replies carried whole by the installed files per the repo's own tooling) splits on four load-bearing claims — and the change records standing as contested with every seat's own words displayed, which is exactly the 'disagreement does not publish silently' path §3.15 requires rather than a dodge of it. The five claim-verdict moves are each justified in the rationale from the records attached to those claims, and the seats' dissents on held claims (C053, C044, C028, C057) are named with reasons rather than buried. The one mid-run defect (malformed claim span in the first edition reply) was caught, sent back, repaired, and is documented in errors.md and the run record with the divergent paragraph shown — honest process, not fabrication or check-weakening. Nothing in the visible material shows invented locators, misattributed provenance, or a loosened check.