This is a pipeline/loader change with no content records and no citations added (verification found none), so no fabrication or confidence surface. AGENTS.md is untouched; the only doc edit is an append to docs/DECISIONS.md recording the change and its founder direction, stamped with date and channel, consistent with §3.15's allowance for founder directions in the decisions log. The verify.ts change broadens the anchoring test to count the case's live (non-rejected) ledger evidence citing a claim, which is not a loosening of the real-citations check so much as a correction of a scoping bug — a founding claim anchored by six live evidence records was being read as unanchored when re-read alone; anchoring still requires either a verbatim source anchor or actual evidence citations, and rejected evidence is excluded. The loader relaxation lets only retired/superseded (frozen) agenda items reference a claim rejected since, mirroring the existing tombstone/assessment-run exception, while open items still fail integrity — the added load.test.ts asserts both directions. The retirement path stamps status, note, statusBy (runId) and statusDate and keeps the served claim IDs as the record, and the note is now appended to verification.md, so provenance is displayed rather than hidden (§3.8, §3.14). Tests accompany each behavior; no check in §3.15 is dodged or reclassified.

The change weakens the citation check in §3.15. In `verify.ts`, `citedBy` now includes every claim ID referenced by non-rejected ledger evidence, but there is no guard requiring the proposed claim wording to be identical to the existing ledger claim that evidence was reviewed against. A rewritten claim reusing an existing ID could therefore bypass re-verification solely because old evidence points to that ID; the new test covers only the unchanged happy path and does not enforce that invariant. The retirement loop also rewrites the status, note, actor, and date of already retired or superseded agenda items, despite treating those items as frozen records, risking loss of the visible provenance required by §3.14. Limit the shortcut to an unchanged existing claim and preserve already frozen agenda-item statuses and provenance.

The pull request modifies pipeline verification and domain loading logic to handle re-reading existing ledger claims and retiring research agenda items whose claims have all been refused. Live evidence already in the ledger is now counted when evaluating whether an existing claim has anchoring evidence, preventing valid founding claims from being erroneously discarded as unanchored. When an agenda item's claims are all refused, it is retired with complete audit metadata (status, statusNote, statusBy, statusDate) rather than stripped of claims, preserving provenance and ledger integrity in accordance with §3.8 and §3.14 without weakening any constitutional checks (§3.15).

The diff is pipeline/loader/test work plus a founder-stamped DECISIONS entry; AGENTS.md is untouched. verify.ts now treats live ledger evidence as anchors when a claim is re-read alone, fixing a false unanchored refusal without fabricating citations or loosening real-citation checks. reverify.ts retires agenda items whose claims were all refused (keeping claimIds and stamping status/run/date) and appends those notes to verification.md; load.ts lets only frozen (retired/superseded) items reference later-rejected claims, matching the tombstone pattern. Tests cover both paths. No content citations, confidence material, or check-weakening.

The change fixes a fail-closed false refusal: a founding claim (VASO-C001) with no source anchor of its own but cited by six live, already-verified evidence records was refused as unanchored when re-read on its own. Extending the anchor check to count the ledger's live evidence does not weaken a §3.15 check — the evidence it counts already passed the citation/provenance checks when it entered, no fabricated citation or locator is introduced (none are added at all per the verification section), and the claim is 'promoted over itself' unchanged, so no unratified judgment is presented as ratified. The retirement mechanism is provenance-clean: status, note, runId and date are stamped, the refused claim ids are kept as the record, and the note is appended to verification.md, improving the account's honesty rather than hiding the refusal. The loader change mirrors the existing tombstone/assessment-run allowance — only retired or superseded items may reference claims rejected since, while open items still fail — preserving fail-closed integrity, and tests cover both behaviors. The DECISIONS.md entry is dated, stamped with supervision channel, and no AGENTS.md edit is present.