The change installs five independent check-role seat assessments for two cases plus run/spend records and harvested arbiter verdicts; AGENTS.md is untouched and no check is loosened. Every assessment header carries runId, producedBy, model, date, promptVersion, basis.ledgerHash, role: check, and humanReviewed: false with the comment 'Append-only; NOT human reviewed' — so no AI draft is dressed as a ratified or human judgment (§3.15, §4). The seats disagree openly (CCC verdicts all 'contradicted' but with differing load-bearing sets and claim verdicts; Amazon splits well_supported/provisionally_supported), and disagreement is displayed rather than reconciled away (§3.3, §3.13). Claim verdicts consistently distinguish what a source states from inference (e.g. 'the claim is about what the text says, not whether it is true'), flag anchor-only or evidence-less claims as unresolved/provisional, disclose independence problems (proponent-adjacent Warsaw group; Peripato sharing Levis's species list and authorship), and carry negative results (Allen's 10^-103 bound, Jow & Scott's 87%, Piperno's null phytolith synthesis) (§3.5, §3.9, §3.10, §3.11). The one mechanically checked identifier (arXiv 1707.04169) resolves and matches the use made of it. The omitted files are seat working replies and further harvested verdicts of the same verbatim-record pattern; the run account covers the consequential content, and nothing visible contradicts a locator, quote, or provenance stamp.

The visible assessment files are clearly stamped as check-role, AI-generated, append-only, and not human reviewed, and the sole newly added arXiv identifier mechanically resolves. However, the omitted-files list includes numerous harvested governance records and all ten underlying seat files, while the RUN ACCOUNT expressly excludes those working files and does not authenticate their complete contents. Those files could contain provenance inconsistencies, unlabelled supplied text, or records that differ from the installed assessment overlays, so the complete change cannot be judged from the packet. I would need the omitted files themselves or a complete mechanical account covering their provenance, confidence status, and correspondence with the published assessments.

This pull request installs independent multi-model check assessments for the 'ccc' and 'pre-columbian-amazon' cases across five separate model vendors (Anthropic, OpenAI, Google, xAI, Venice/GLM). All ten assessment files are strictly append-only overlays stamped with runId, model, date, promptVersion, ledgerHash, role: check, and humanReviewed: false, in full compliance with §3.14, §3.15, and §4. The mechanical citation verification confirms resolution of the added arXiv identifier, and the check assessments evaluate existing ledger claims using calibrated uncertainty and explicit counter-considerations. The omitted files consist solely of internal working artifacts (model seat replies), spend tracking, and harvested prior arbiter records, while the complete assessments and run accounts are visible and confirm constitutional compliance.

This PR installs two completed check-role panels (ccc and pre-columbian-amazon), five independent assessments each, plus run/spend stamps and harvested prior arbiter verdicts. AGENTS.md is untouched. Each assessment is labelled role: check, humanReviewed: false, stamped with runId/producedBy/model/date/promptVersion/ledgerHash, and the headers state they are append-only AI overlays that feed concurrence rather than case narrative — no draft is dressed as ratified and no AI record as human (§3.15, §4). The RUN ACCOUNT shows reasoned claim-level verdicts with load-bearing sets, steelmans, and disagreements displayed rather than collapsed; the sole mechanical citation check (arXiv 1707.04169) RESOLVES and matches the Penrose 2017 use. Omitted files are seat working replies and prior harvested governance YAML whose substance is already in the account or is not content under review; nothing visible fabricates citations, weakens checks, or exposes confidence material.

The change adds only append-only check-role assessments, run/spend records, and verbatim harvested arbiter governance records; AGENTS.md is untouched, so §3.15's amendment reservation is not engaged. Every assessment is honestly stamped (runId, producedBy, model, date, promptVersion, humanReviewed: false, 'NOT human reviewed', basis.ledgerHash) and labeled role: check — never presented as a ratified or human judgment, satisfying §3.14/§4. The only mechanically checkable identifier (arXiv 1707.04169) RESOLVES and its title exactly matches the Penrose 2017 LIGO-correlation conjecture the claim attributes to it, and the assessments add no new sources or evidence records — they judge an existing hashed ledger, so there is no fabrication surface beyond the one resolved identifier. Internal consistency checks pass: the ccc run's spend ledger sums to the run record's $1.066147, matching the PR body's $1.07, and seat counts match files written. The omitted files are working seat files and additional harvested governance records of the same verbatim format as the ones shown, and nothing visible in the run account or diff suggests a violation they could hide; the assessments themselves display disagreement (verdicts ranging from contradicted to provisionally_supported) rather than hiding it.