The diff does not touch AGENTS.md or anything under content/, so it adds no citation, locator or case record. The scale in §3.15 is left exactly as it is, and any future amendment is to be 'drafted for the founder's hand, as amendments are'. The DECISIONS entry does not override the constitution either. It declines to bring the two verdict words within one step because that would loosen a check. The founder's in-session direction is quoted and stamped with its date and channel, both in DECISIONS.md and in the config/models.yaml comment, which §3.15's operating-record paragraph permits. The choice on the verdict words is attributed to the operator, not the founder. AUTOMATION.md's heading 'what he decided' sits over a bullet that is the operator's choice, but the body says so plainly, so I do not read it as misattribution. The code change adds `unevidencedClaims` and `checkInstructions` and adds one test without loosening any. check-v2 does not use the new placeholder, so current seats are sent the same instructions. The experiment is registered before any run, with measures, an adoption rule and an explicit small-sample caveat, consistent with §3.12 and §3.13. The draft sits in proposals/, its runs are not installed, and the `check-v4` name is retired if the draft is not adopted, which keeps the stamps honest. The published figures check out internally: the splits sum to 23, the far-verdict kinds sum to 129 (28+25+22=75, 28+25+14=67), standings are 7 ratified and 3 contested, 11 of 12 seat rows gain 1 to 8 claims, and measure 1 is 17+11+18+13=59. I found no sign of fabrication.

The change does not edit AGENTS.md, alter the verdict scale, publish a new assessment, or carry any run records. It clearly labels check-v4 as a draft experiment that has not run, while the active check-v2 behavior remains unchanged; the added placeholder plumbing only enables the proposed experiment. The draft reinforces §3.6 by identifying featured claims with no admitted evidence record and requiring either an unresolved verdict or a disclosed alternative basis, and it reinforces provenance by requiring limitations and cross-claim attachments to be stated. The founder’s direction is recorded with its date and channel as permitted by §3.15, and the operator’s decision not to propose an amendment is accompanied by reproducible scripts and counts. No added citation fails verification, no confidential material is published, and no visible check is weakened or bypassed.

The change records founder directions given in session on 2026-09-30 in DECISIONS.md and config/models.yaml in accordance with §3.15. No edits are made to AGENTS.md, nor are existing checks weakened or reclassified; the operator documented an empirical count and explicitly decided against proposing an amendment to bring two verdict words closer on the scale. The experiment design and tooling under proposals/ along with pipeline helper functions (`unevidencedClaims`, `checkInstructions`) do not touch published case content or present unratified drafts as ratified.

The diff records the founder's 2026-09-30 session directions in DECISIONS.md and AUTOMATION.md with date and channel, adds only a non-normative cost comment to models.yaml, and leaves the §3.15 scale and AGENTS.md untouched. It designs (but does not run) a future check-protocol experiment under proposals/, with a draft that is explicitly unratified and a helper that fills unevidencedClaims only when a protocol asks; check-v2 behaviour and existing packets are unchanged, as the new test asserts. No case content, assessments, citations, or confidential material are added, and no check is weakened or reclassified.

AGENTS.md is untouched, and the change correctly treats moving the two verdict words as a founder-reserved amendment, declining to propose it after counting — the opposite of check-weakening; the only code change (unevidencedClaims/checkInstructions) fills a placeholder for a not-yet-run draft protocol in proposals/ and provably leaves check-v2's rendered instructions unchanged, guarded by a new test. The founder's answers are recorded with date and channel ('in session') in DECISIONS.md, AUTOMATION.md, and a config comment, exactly as §3.15 permits for operating-record directions. The check-v4-draft tightens rather than loosens the blind check (it gives seats the edition protocol's unresolved-unless-basis rule and requires carrying stated limitations, echoing §3.6 and §3.9), and its adoption is gated on pre-registered measures rather than presented as ratified. No citations are added, nothing was run, and the design honestly marks itself as designed-not-run with no fabricated records or locators visible.