The change adds only check-role overlays for vasocomputation. There are five assessments, each stamped with runId, producedBy, model, date, promptVersion, humanReviewed: false, role: check and a shared basis.ledgerHash. The run account confirms each raw seat reply was carried whole, run.yaml's wrote-list matches the files, and the spend rows sum to $3.61564 and 144,871 output tokens (the input total differs only by Grok's 384 cache-read tokens). The seat verdicts (weakly_supported x2, mixed x2, unresolved x1) all sit within one step of the draft's 'unresolved' under the founder-amended scale, so a ratified five-of-five standing follows §3.15, and every seat's own word is displayed. No identifiers are added under content/, and the figures the checks cite (0.57/0.36/0.17 cm², 15/49, 11/11-4/11-0/11, 60%/86%, 71.5%) agree across all five seats. The imprecision I found is ordinary and mostly in the Gemini seat: it pins Landfald's 60%/86% outcome on E067 where others place it in E072, calls E078 a systematic review in C190 while treating it as a proponent concession in C030, and calls E062/E082/E109 'numerous clinical studies.' None of this invents a source or inflates provenance the way the withheld geo check did. The check.test.ts edit leaves the assertions unchanged and makes the fixture panel hermetic, so it does not weaken a check. Its comment, however, describes geo checks this change no longer carries, and the test no longer incidentally exercises stale-ledger checks; a unique fixture ledgerHash would keep that coverage. Holding back the whole Cast, Not Carved panel, rather than only its dissenter, avoids showing a cleaner standing than that panel gave.

The five check assessments are append-only, explicitly marked AI-generated and not human reviewed, and stamped with model, runId, date, prompt version, and a common ledger hash. The run account confirms that every raw seat reply was carried whole, while the published records preserve divergent case and claim verdicts rather than concealing disagreement. The assessments generally anchor factual discussion to existing ledger IDs, disclose missing evidence, dependencies, alternative explanations, and uncertainty; no new external citation identifier failed verification. The spend totals and run metadata are internally consistent, and the test edit isolates fixture checks without loosening the production concurrence or fail-closed checks.

This pull request installs the independent cross-model check panel for the 'vasocomputation' case (State, Not Scar) across five vendor seats (Anthropic Opus, OpenAI GPT, Google Gemini, xAI Grok, Venice/Z.ai GLM). Each check assessment is properly stamped as an append-only, non-human-reviewed check role record with exact runId, model, promptVersion, and basis ledger hash, matching the proposal run record and spend ledger. The minor adjustment in `src/domain/check.test.ts` isolates fixture data from real repository check assessments without altering domain logic or weakening constitutional checks.

The diff adds only append-only check-role assessments (five seats), the matching run.yaml, raw seat replies, and spend rows for vasocomputation, plus a one-line fixture isolation in check.test.ts; no claims, evidence, sources, editions, or AGENTS.md are touched. Run account confirms all five seat replies were carried whole into the installed assessments, each stamped with runId/model/date/promptVersion/role/basis.ledgerHash and humanReviewed:false, and the panel’s case verdicts (two weakly_supported, two mixed, one unresolved) lie within one step with no load-bearing split. No external citation identifiers were added (verification empty), reasonings anchor only to existing ledger IDs, and the test change merely keeps the fixture’s owed-seat assertion independent of real panels rather than weakening any §3.15 check. Cast, Not Carved material is correctly omitted.

This change adds only check-role assessments for the vasocomputation case plus their run record, spend rows, and a test-fixture isolation fix; no edition, claim, evidence, or source records are touched, and citation verification found no external citation identifiers in the added lines. The fabrication problem the OpenAI seat caught on the first push (the Cast, Not Carved panel's false attributions to GEO-E002/GEO-E010) has been removed from this change entirely — refusing to publish that panel and re-asking the seat blind, with the withheld run kept on record, honors rather than weakens the real-citations rule. Provenance is honest throughout: the run account confirms all five raw seat replies were carried whole into the installed assessments, the files are stamped with runId/model/date/promptVersion and humanReviewed: false, and nothing is presented as a human conclusion. The claimed ratification checks out arithmetically under the graded scale — verdicts weakly_supported (×2), mixed (×2), unresolved (×1) are all within one step of the case verdict 'unresolved', and 4 exact/20 within-one-step/none-split on 24 featured claims means no load-bearing claim splits the panel. The check.test.ts change does not loosen any §3.15 check: it makes a hermetic fixture by filtering real check runs out of the fixture's input, leaving the asserted behavior (only unjudged seats are asked) unchanged.