The diff touches nothing under content/: it adds three harvested governance/arbiter records, spend rows, and a proposals/ working bundle (report, draft proposal, partial verify judgments/splits, run records). AGENTS.md is untouched and the proposal.yaml carries the explicit header 'Never citable; nothing here is a record until adopted through the gate', so no unratified draft is presented as ratified. Accounting is internally consistent: the report run's 3.274411 USD equals the 3.144411 + 0.13 spend rows, cache/input token totals reconcile to the run.yaml's 2,489,975, and the draft and verify rows match their run records and the PR body's figures. The interrupted verify step is recorded as outcome: failed with the exact JSON parse error and the model that produced it — a refusal reported honestly rather than papered over, and its partial judgments (which themselves flag non-atomic sourceStatements and mislocated quotes) are kept visible rather than discarded. Provenance is stamped throughout (runId/model/date/promptVersion/inputHash), sources carry ai_verified labels with retrieval notes and genre caveats, evidence directions are explicit, independence is flagged (preprint = same performance as TIK-E008, not a replication), and unretrievable sources are dispositioned 'blocked'/'irrelevant' with routes to the primary rather than being cited unread. The one probable-duplicate risk (preprint vs. published Antonopoulos) is disclosed in novelty.md and handled by an explicit duplicate disposition. The only omitted file is the report's raw.json working material, which cannot decide compliance given the whole report.md and proposal are visible.

The omitted raw.json is a newly public run record, and without it I cannot determine whether its source material, retrieval provenance, or contents violate the fabrication or confidence-material rules. It is particularly needed to reconcile report.md saying the report model “opened and used” Samorini, Bremmer, and Bennett with novelty.md and the proposal recording those URLs as not retrievable; these may describe different retrieval stages, but the visible packet cannot establish that. The visible proposal is clearly labeled never citable and unadopted, while run.yaml honestly marks verification as failed, so the verifier’s partial findings of inaccurate locators, compound evidence statements, and wrong directions do not by themselves publish those defects into the ledger. I would need raw.json to confirm the report’s provenance and that no confidential or fabricated source material is being committed.

The pull request commits only governance and proposal artifacts from an interrupted pipeline run; no files under content/ or AGENTS.md are modified. Three harvested arbiter verdicts (PRs 287, 288, 289) are added verbatim as append-only records, and governance/spend.yaml accurately accounts for all API calls across the report, draft, and failed verify stages. The proposed additions under proposals/ are explicitly flagged as non-citable working proposals, and the verify failure (a JSON syntax error on verify-v6) is documented transparently in proposals/2026-09-16-verify-immortality-key-025720/run.yaml. The single omitted file (raw.json) is raw retrieval data from the report stage; its contents cannot alter the fact that nothing is published or adopted into the ledger. No unratified draft is presented as ratified, no checks are loosened, and all provenance and spend entries are fully consistent.

The diff touches only proposals/, governance/arbiter harvests, and spend — no content/ ledger records and no AGENTS.md edit. Proposal envelopes are explicitly never-citable until the gate; the draft sets edition: null, stamps runId/model/date/promptVersion, and the PR/run records honestly mark verify as failed and the chain as interrupted. The draft admits only three retrieved sources as ai_verified, disposition-blocks the HTTP-403/PDF-fail/bot-challenge set with reopen routes, separates sourceStatement from editorInference, keeps strengths weak/moderate, and leaves Juan-Tresserras page locators unverified while sourced to the preprint’s refs. Harvested pr-287/288/289 YAMLs are append-only machine records of prior public seats; spend rows match the run accounts. Report-vs-novelty tension over Samorini/Bremmer/Bennett is real but confined to never-citable working material that the draft already filtered out of adds, so it is ordinary pipeline imperfection rather than published fabrication or check-weakening.

The change adds only operational records (harvested arbiter verdicts, spend rows, run records) and non-citable working material under proposals/, explicitly stamped 'Never citable; nothing here is a record until adopted through the gate'; nothing under content/ is added, no citation identifiers are introduced, and AGENTS.md is untouched. The interrupted chain is disclosed honestly: the verify run's failure (outcome: failed, JSON error) is recorded in its run.yaml rather than papered over, so no unratified draft is presented as ratified and the gate itself was not dodged — it tripped and was recorded as tripping. The proposal's epistemic hygiene is strong: it refuses to import report claims from the five sources it could not retrieve (Samorini, Bremmer, Bennett are dispositioned 'blocked' with routes to the primaries), keeps the Juan-Tresserras correction marked unverified, and the verify judgments honestly record locator failures and direction mislabels (locatorSupported: false, directionRight: false) rather than hiding them. Spend arithmetic checks out exactly (report 3.144411+0.13=3.274411; the 35 verify calls sum to 1.544664), and the harvested arbiter files — including PR 289's lone check-weakening dissent recorded as a review note and PR 288's failed seat — are internally consistent with the constitution's procedures. My one reservation is that the report.md claims to have opened pages the drafter later found unretrievable, but transient 403/bot-challenge failures are plausible, the drafter explicitly refused to enter any of that unverifiable material, and the omitted raw.json is a working file whose compliance-relevant content (what entered the ledger) is decided by the visible proposal, which entered none of it.