The diff adds two edition candidates, their run records and reply files, harvested arbiter verdict records, and two spend rows. AGENTS.md is untouched, no new citation identifiers are introduced (the verification section confirms none), and no confidential material appears — the one sensitive item, the founder's funding of the 2026 Pärssinen study, is disclosed in-article rather than concealed. Both editions explicitly re-adopt the incumbent's assessment (assessment: null in the replies), carry runId/model/date/promptVersion stamps, and the file header states plainly that the candidate replaces the incumbent 'only if the panel prefers it' — no unratified draft is presented as ratified. The rationales are unusually careful about provenance: the CCC edition removes a first name ('Marcos' López) the ledger does not hold and deletes a stale unverified caveat that the ledger now records as ai_verified; the Amazon edition strips a sentence sourced only to the founding narrative and corrects 'towns'/'roads' to the words the underlying evidence records. Uncertainty is calibrated rather than laundered — AMZ-C029 is still flagged in prose as 'a figure this record has not yet verified against the paper's text', and negative evidence (the sparse-Amazon section, the CCC nulls) is displayed prominently. The harvested arbiter files reproduce seat verdicts including dissents verbatim, which displays disagreement rather than hiding it.

Both editions are append-only candidates with model, date, prompt version, production-run linkage, basis hashes, and explicit warnings that they replace the incumbent only if the panel prefers them; neither presents itself as already ratified. They re-adopt identified incumbent assessments rather than silently changing verdicts, while preserving competing accounts, negative evidence, limitations, and dependency cautions such as the shared Amazon research network. The Amazon article explicitly labels AMZ-C029 as not yet verified against the paper text, and the CCC article distinguishes unreviewed or unreplicated findings from established results, showing calibrated uncertainty rather than false precision. No citation identifier is newly added under content, no visible locator or quotation contradicts its stated source, and the packet supplies no positive evidence of fabrication or confidence-material exposure. The run and spend figures agree, the historical arbiter files are append-only and internally consistent with their recorded seat votes, and AGENTS.md and the constitutional checks are untouched.

The pull request carries two edition updates (for ccc and pre-columbian-amazon) generated by the edition verb, along with harvested arbiter records and spend entries. Both new editions re-adopt their incumbent assessments without changing verdicts or crux orders, tightening prose to match ledger terminology and removing stale qualifiers. No content citations or locators are added or fabricated, no confidential material is exposed, and AGENTS.md is untouched. Provenance is faithfully recorded across run records, proposal replies, and spend rows.

Both edition candidates are stamped AI drafts that re-adopt incumbent assessments by hash/runId, return assessment: null, and label themselves as panel candidates rather than ratified judgments. Rationales state the ledger is unchanged on verdict-moving evidence and limit edits to prose hygiene already warranted by ledger wording (remove unheld first names, drop a stale unverified parenthetical now that SRC-NIELSEN is ai_verified, cut founding-narrative-only weir intensity and non-ledger 'soldier'/'towns'). No new citations or locators appear (citation verification empty); AGENTS.md is untouched; claim spans name existing packet ids; competing accounts and honest gaps (AMZ-C029 unverified; CCC gaps on sky-twist/independence) remain displayed. Harvested arbiter YAMLs and spend rows are append-only operational records, not content fabrications.

The two edition runs are prose-only candidates that explicitly re-adopt the incumbent assessments, file themselves as candidates pending the panel ('Replaces ... only if the panel prefers it'), and are stamped with runId/model/promptVersion/date and matching run.yaml and spend.yaml rows — no unratified draft is presented as ratified and no assessment is rewritten. The citation verification finds no new identifiers under content/, and the rationale's prose edits are ledger-directed corrections (removing a name and a verification caveat the ledger no longer supports, aligning wording with what evidence records actually say), the opposite of fabrication; unverifiable flourishes were removed and the still-unverified AMZ-C029 figure is labeled as such in the article. The thirteen harvested governance/arbiter files are append-only machine records of prior public PR comments, visibly stamped and internally consistent with the histories those PRs describe (including PR 269's park and my own seat's empty-reply failure recorded honestly); the spend additions match the run.yaml costs exactly. My only reservation is that the PR body does not enumerate the governance files it installs, but they are fully visible in the diff, are additive records rather than content claims, and hide nothing that decides compliance.