The code changes all tighten rather than loosen checks: the quote-edge rule now refuses to strip '?' and '!' so a misquote of tone fails (§3.8 fidelity), the per-claim bearing fails closed unless it names exactly the record's claims and otherwise records the reader's dissent aloud instead of silently rewriting direction (§3.5, §3.15), the correction writer's history line now serialises whole values so an entry can be reconstructed (§3.14), and the intake entry's reason now states the admitted set and labels the drafter's rationale as pre-verification argument rather than as what entered (§3.9, §3.14). Both history additions are appended, not edited, and each is stamped with actor, model, session, PR and date (§3.14, §3.15). AGENTS.md is untouched; docs/DECISIONS.md records the four answers. On citations: arXiv 1510.06537 resolves with exactly the title the history entry attributes to it ('Ring Type Structures in the Planck map of the CMB'), and arXiv 1207.6905 resolves to a concentric-rings CMB variance paper consistent with the CCC anchor's subject. The one FAILS — 'doi 10.1093/mnras/stx2299\' — carries a trailing backslash captured from the YAML escaping, and it is quoted as the *prior* value of an existing record rather than newly asserted provenance, so I read it as an extraction artifact, not fabrication; if the operator wants the record clean, I would want a re-resolution of 10.1093/mnras/stx2299 without the stray character confirmed on the record.

The added CCC history line repeats DOI 10.1093/mnras/stx2299, but citation verification reports `10.1093/mnras/stx2299\` as FAILS. The trailing backslash appears to come from the YAML escape before an embedded quotation mark, so this result does not establish whether the clean DOI resolves and is not conclusive evidence that the clean identifier was fabricated. Both added arXiv identifiers resolve, and the code changes strengthen quote matching, bearing validation, and append-only provenance rather than weakening checks. A resolver result for the clean DOI, without the YAML escape character, is needed to determine compliance with the real-citations and exact-provenance rules.

The diff directly addresses review notes from prior panel sittings by strengthening verification and provenance safeguards. In `src/pipeline/quotes.ts`, edge-punctuation trimming now preserves terminal '?' and '!' to prevent tone distortion in verified quotes. In `src/pipeline/verify.ts`, second-reader claim bearings fail closed if claim IDs are missing, repeated, or foreign, admitting the record with dissent recorded rather than silently misapplying partial bearings. In `src/pipeline/ledger-write.ts`, `shown()` serializes field transitions completely in JSON rather than truncating strings or printing '[object Object]', and the affected history files (`ccc` and `pre-columbian-amazon`) append clarifying entries to maintain full, auditable provenance without rewriting past entries. No citations are fabricated, and AGENTS.md is untouched.

The diff answers three review notes by strengthening checks and fixing provenance display: quote edge-stripping keeps ?/! (honest quotation), reader bearings fail closed unless they name exactly the record’s claims, history lines serialize full values via shown(), and intake reasons label the drafter’s pre-verification rationale beside the admitted set. Append-only housekeeping entries restate the two lossy CCC corrections and clarify the Amazon intake entry without rewriting prior lines; actor stamps and kinds are explicit. AGENTS.md is untouched; DECISIONS.md only records the operator answers. Citation tooling shows arXiv 1207.6905 and 1510.06537 resolve to titles matching the restated anchors; the DOI FAILS line looks like a locator artifact (trailing backslash) on an identifier already on the ledger, not a new fabrication. No confidence material, no check-weakening, no unratified draft presented as ratified.

This is a repair-and-transparency change: it tightens quote verification (a '?' or '!' the source lacks now counts as a mismatch), makes the second-reader bearing fail closed so a defective dissent is displayed on the record rather than silently applied, and fixes lossy history serialization — all strengthening, not weakening, the §3.15 checks. AGENTS.md is untouched; the append-only history is corrected by following restatement entries, exactly as §3.14 requires, and both entries carry date/actor/aiAssisted/kind stamps with honest disclosure of the prior report fault. On citations, the two arXiv identifiers the change relies on resolve and their titles match what the records claim ('Ring Type Structures in the Planck map of the CMB' for 1510.06537; a CMB ring-search paper for 1207.6905). The one FAILS result — DOI 10.1093/mnras/stx2299 — appears only as the quoted historical 'from' value in a restatement of a pre-existing ledger entry, not as a new citation introduced to carry an assessment, and the change's action there is to append the resolving arXiv identifier; I do not treat a pre-existing locator quoted as past record as fabrication by this PR, though the operator should confirm on the record whether that DOI's 404 is transient or the locator needs relabeling unverified in a follow-up correction.