The diff adds one stamped edition candidate, one append-only draft assessment overlay (role: draft, humanReviewed: false, runId/model/date/promptVersion and basis ledgerHash present), the matching per-run spend file, the run's reply.json, and harvested verbatim arbiter records for prior PRs. AGENTS.md is untouched, and the citation tooling found no identifiers in added content lines, so no fabrication surface is exposed. The substantive movement is in the honest direction: TIK-C029 and TIK-C034 are demoted to 'unresolved' at low confidence precisely because they are held on a source anchor with no admitted evidence record (§3.6), TIK-C033 is graded weakly supported with its unadmitted dose threshold named, and the article prose states that thinness in place rather than asserting the propositions in the page's own voice (§3.9, §3.13). The steelman's previously unanchored reading of the initiates' formula is rewritten as the proponents' reading, and each premise names its evidence record. Negative and unfavourable material (empty cell at Eleusis, audited accounts, specialist rejections, the self-disclosure that the site's founder funds YAPP) is displayed rather than suppressed (§3.11), and the rationale, run.yaml, reply.json and edition/assessment files reconcile with one another. The harvested governance files record prior verdicts verbatim, including pr-332's five-seat constitution park with a founder hand merge, which is accurate provenance rather than a check dodged.

The new assessment is explicitly stamped as an AI-generated draft with model, runId, date, promptVersion, ledger hash, and `humanReviewed: false`, while the edition is labelled a candidate that replaces its predecessor only if the panel prefers it. Anchor-only claims TIK-C029 and TIK-C034 are attributed to what the paper reports, graded unresolved at low confidence, and repeatedly disclosed as lacking admitted Evidence records; TIK-C033’s limited non-anchor basis and missing dose-threshold evidence are likewise stated explicitly. The article and assessment distinguish possibility from historical attribution, expose the unreplicated and interested provenance of the chemistry, disclose the founder’s funding relationship, and preserve serious competing accounts and negative evidence. No AGENTS.md edit, confidential material, fabricated identifier, failed citation, hidden disagreement, or weakened check is visible, and the run, spend, assessment, and edition provenance agree.

The pull request adds a newly generated edition and draft assessment overlay for the immortality-key case, alongside harvested arbiter records and a matching spend ledger entry. The edition candidate directly resolves earlier panel objections (notably review note #319) by demoting anchor-only claims (such as TIK-C029 and TIK-C034) to 'unresolved' at low confidence, and qualifying TIK-C033 while disclosing its basis, strictly adhering to §3.6 and §3.9. In the article prose and steelman, proponent inferences are explicitly attributed rather than stated as ledger-admitted facts. Provenance stamps (runId, model, date, promptVersion, basis hashes) are fully intact and accurate, AGENTS.md is untouched, and no citations or records are fabricated.

The PR adds a stamped draft edition/assessment for immortality-key under edition-v9, correcting prior anchor-only overgrades: TIK-C029 and TIK-C034 are unresolved/low with the article voicing them as paper reports held on anchors only, and TIK-C033 is weakly_supported on the disclosed TIK-E008 yield while the dose threshold stays anchor-only. The steelman now labels the initiates' formula as the proponents' reading rather than a ledger-held fact, and the article's three local edits match those basis changes without inventing citations (verification found none). The overlay is append-only, humanReviewed:false, role:draft, with runId/model/date/promptVersion and matching spend/run records; AGENTS.md is untouched and the harvested arbiter files are historical records, not constitutional edits. No confidence material or check-weakening appears; research items stay open as retrieval failed.

The change is a single edition run that honestly recalibrates two anchor-only grades the prior panel flagged (review note #319): TIK-C029 drops to unresolved and TIK-C033 to weakly_supported, with the article explicitly telling anchor-only propositions as 'what the paper reports' and disclosing that no admitted evidence record quotes the passages — a direct application of §3.6 and the edition-v9 anchor-only rule. Provenance is fully stamped (runId, model, promptVersion, date, humanReviewed: false, role: draft, append-only header, basis hashes), the edition is presented as a candidate against the incumbent rather than as ratified, and no AGENTS.md edit, confidence material, or check-weakening appears anywhere in the diff. No citation identifiers are added under content/ (the mechanical verification found none), so no fabrication surface exists; the article's quotations ride on evidence records already admitted in earlier PRs and unread primary sources are marked as such rather than overstated. The harvested arbiter records are internally consistent with the constitution as quoted (including the 2026-09-17 §4 amendment they document being merged by the founder's hand after a 5/0 constitution-paradigm park in pr-332, explained in pr-334), and I find no self-contradicting locator, quote, or provenance stamp in the added lines. The run account shows a reasoned, calibrated update with no research item settled and disagreement displayed.