The change adds only new overlay files: an edition, a draft assessment, five blind check readings, run records and spend ledgers. Each is stamped with runId, model, date and promptVersion and marked humanReviewed: false. No claim, evidence, source, code or AGENTS.md file is touched, and the tooling found no citation identifiers added. The retold article's quotations (McGovern, Bremmer, Clinton, the Didache, Justin, the 2026 paper) are ones the blind seats cite independently from the same ledgerHash. Its unchecked material is labeled plainly: Dioscorides 'as Muraresku summarizes him', the ergometrine and paspali passages 'this site has not yet checked', the dose threshold 'not checked here', ergot fats 'not confirmed'. I see no fabrication and no blurring of source statement with inference. Disagreement is shown, not hidden: the case tally in the PR matches the five check files (two mixed, one weakly_supported, two contradicted), the title and body call the standing contested and schedule a reconsideration, the repair round is on record (errors.md, the reply.json comparison), and both spend ledgers reconcile to run.yaml. Non-blocking notes for the operator: the Grok seat's structured `sensitivity` field was dropped from its installed check file (RUN ACCOUNT: 'NOT carried whole'), so carry it or record the omission, since seats' own words are what is displayed. The draft also holds TIK-C012 at weakly_supported while its own treatment says the thesis 'cannot be stronger than its weakest premise' (C013, contradicted) and all five seats say contradicted; the scheduled reconsideration should settle that. Finally, 'a thousand kilometers' understates the Pontós–Eleusis distance (about 1,800 km), a ledger wording the Opus seat already flagged.

The new edition preserves the incumbent verdicts, separates the hypothesis into atomic rungs, distinguishes diagnostic background evidence from the Eleusinian and Christian claims, and repeatedly labels unread, unreplicated, or unchecked material as such. Provenance is explicit throughout: the draft and five check assessments carry runId, model, date, promptVersion, role, ledger hash, and humanReviewed:false, while the edition points to its predecessor and records a detailed rationale. The blind check used five vendors, retains all five verdicts and reasoning, and visibly exposes the two contradicted case readings and claim-level splits rather than concealing disagreement. No citation identifiers were newly added, and nothing shown positively demonstrates an invented quote, locator, source, or permission problem. The sole omitted file was mechanically compared with the repaired reply in the run account, which fully identifies the malformed marker that was removed, so it does not create material uncertainty.

The proposed change updates the narrative edition of the case within reader budget constraints and installs a full 5-seat independent check panel without modifying any underlying claim, evidence, or source records. All new assessment files are append-only, accurately timestamped, labeled with proper provenance (`role: check` and `role: draft`, `humanReviewed: false`), and transparently display the resulting contested standing rather than presenting the draft as ratified. The comparison run shows unanimous model preference (5-0) for the candidate edition over the incumbent, and citation verification found no added or fabricated identifiers. The run records, spend logs, and repaired parsing error are completely documented in accordance with §3.15.

The change is an edition retell plus a fresh blind five-seat check. No claim, evidence, or source records are touched; CITATION VERIFICATION found no new identifiers. The draft assessment keeps the incumbent case verdict (mixed) and every component/claim grade, adds full per-claim treatments, and stamps runId/model/date/role=draft with humanReviewed:false. The check installs five independent seat assessments with honest labels; the run account shows four seats carried whole and one minor non-normative field drop, and standing is correctly contested (2 of 5 two steps from mixed). Article and synthesis keep source-vs-inference distinctions and do not present the draft as ratified. AGENTS.md is untouched.

The change touches no claim, evidence, or source records; it adds an edition whose assessment restates the incumbent's verdicts unchanged (append-only, humanReviewed: false, stamped with runId/model/promptVersion), plus five independent blind check assessments and transparent run/spend records — consistent with §3.15's gated process and §3.14's versioning. The check is genuinely blind (the drafter was shown no check), all five seats' words are installed and displayed, and the contested standing (two seats two steps from the draft's 'mixed') is displayed rather than hidden, with the constitutionally prescribed next step (reconsideration) recorded rather than silently dropped. No AGENTS.md edit, no citation identifiers added (verification section is empty), and the one defect — a malformed claim span in the first draft — was caught by the loader, sent back once, and logged in errors.md rather than silently repaired. Two minor items worth a review note but not parking: the tooling discloses that seat-xai's reply carried a 'sensitivity' field the installed Grok assessment does not (its content does not change that seat's verdict or the standing arithmetic, and the discrepancy is disclosed on the record, satisfying §3.14's inspectability), and the deferred reconsideration rests on a spending cap the run records support. Nothing in the material shows fabrication, unrecorded confidence material, or weakened checks.