The parked defect is fixed, not relabelled. None of the four published 115839 checks grades GEO-C502 or leans on GEO-E002, and the run account says each is carried whole from its seat's reply. Opus says outright that GEO-E002 'has been rejected and plays no part here', and the 2026-09-30 GPT sentence survives only as a raw reply under proposals/ beside panel.withheld.txt, not under content/. GEO-E002 and GEO-C502 are kept as refused records (reviewState: rejected), with reasons in evidence.yaml, dispositions and history, all attributed to the claude-sonnet-5 answer run. The reader's 'section 13 / sections 21-23' appears only inside an attributed refusal reason, and DECISIONS lists the proper read-in as owed, so no unverified locator is admitted. Withholding the Gemini check fits §3.8 and §3.15 rather than weakening a check: its GEO-C010 reasoning does cite '(GEO-E010, GEO-E016)' for 850 °C, its reply is kept whole, and its note (including its 'contradicted' verdict) is printed on the record page and flagged under the panel. By my count GEO-C030 splits against the draft's 'mixed' with or without Gemini (only GPT is within one step), so leaving it out cannot raise the standing, which stays contested. Spend rows reconcile exactly with each run.yaml, the new assessments are append-only files stamped with runId, model, promptVersion, date and humanReviewed:false, AGENTS.md is untouched, and the new test requires a withheld note for any run whose recorded output is absent. Minor items worth fixing, none a breach: (1) the edition rationale says GEO-E002 'is no longer in the ledger', though it remains as a refused record; (2) DECISIONS says 'none of the ten' 2026-09-30 checks where six are visible; (3) Grok's GEO-C020 reasoning credits the four petrographic studies with finding no 'C–S–H binder enrichment', which they did not test; (4) GEO-C911's parent link was cleared without a history line. The only text I could not read is the clipped tails of the Opus check and the draft assessment, whose visible parts show no sign of the Engelbach problem.

The change refuses GEO-E002 and GEO-C502 rather than continuing to rely on their unanchored, compound account, records the refusal reasons, and removes C502 from the featured edition. The replacement edition and assessment are stamped as AI-generated drafts with run, model, date, prompt version, hashes, and append-only history, and they no longer treat the secondary account as confirmation of Engelbach’s text. The fresh panel’s four published assessments are carried whole from their seat replies; the Gemini reply is preserved in the run record but withheld because it visibly attributes furnace-temperature support to GEO-E010, which the ledger describes as a natron-source study. That withholding is disclosed beside the standing and in the record, and the standing was computed both with and without the reply and did not improve when it was omitted, so disagreement is not silently hidden or used to evade concurrence.

The change directly resolves the prior park on #452 by formally re-reading and refusing GEO-E002 and GEO-C502 for missing locators, quotes, and compound propositions, recording the refusals honestly in claims, evidence, dispositions, and history. A new edition removes the refused claim and updates the article, ratified by panel preference. A fresh panel check across independent model families evaluates the corrected ledger, transparently recording and explaining the withholding of one seat that misstated evidence while verifying that withholding does not raise the standing. All additions adhere to append-only provenance, public inspectability, and fail-closed gate standards without modifying the constitution.

The change answers the prior fabrication park by re-reading GEO-E002/GEO-C502 against the primary monograph, refusing both with honest reasons (no quote/locator; unanchored claim), recording dispositions and history, and keeping them as rejected rather than deleting. A fresh edition drops C502 and the secondary Engelbach account; the adopted draft and four published check seats are carried whole from their replies, stamped, and do not rest on the refused records. The Google seat is withheld with an on-record note for mis-citing GEO-E010 on temperatures, standing is recomputed both ways without raising it, and UI/tests surface unpublished checks instead of silently publishing or editing them. No AGENTS.md edit, no new fabricated citations, and provenance of what ran versus what is shown is explicit.

The change answers the parked fabrication objection the constitution's own way: the two records the OpenAI seat named (GEO-E002, GEO-C502) were re-read against the primary text and refused with recorded reasons in dispositions, history, and the run's verification.md — no unverified secondary account is published as confirmed, and the offender's sentence is not republished anywhere. The fresh panel publishes four seats, withholds the Google seat's check for misstating GEO-E010 (a fabrication-grade misstatement), keeps the withheld reply whole beside a stamped note, computes the standing both ways so withholding cannot raise it, and adds a CI-tested rule that every run either publishes what it wrote or carries a note — strengthening, not weakening, the §3.15 checks. Provenance is stamped throughout (runIds, spend rows, withheld notes signed and dated, history entries naming the second reader), the article and adopted assessment no longer carry the refused claim, and citation verification finds no added identifiers to verify. The only blemish I see is the decisions entry's phrase 'none of the ten' where this packet shows six checks from 2026-09-30 (five plus the Google re-run) — likely counting other checks of that date outside this case, but the operator should confirm that count on the record; it is not itself a fabricated citation, locator, or assessment.