The diff touches only a proposals/ experiment folder and three governance/spend files. Nothing under content/, protocols/, src/ or AGENTS.md changes, so there are no new citations and Deep Memory's record and standing are unchanged. check-v3-draft.md is labelled a draft for the experiment, not a ratified protocol (§3.15). The provenance present reconciles: each run.yaml equals the sum of its spend rows (e.g. 075451: 138916 input + 384 cache = 139300, 47204 output, $0.533612), the three runs total $1.11 as claimed, and every seat reply carries humanReviewed: false. I recomputed the PR body's table from the committed replies (openai 13/17, gemini 10/16, xai 16/23, venice 13/14), and all twelve case verdicts match. The draft's verdict order and lib.mts RANK place unresolved/presently_untestable between weakly_supported and mixed, as §3.15's amendment requires. The measures use the repo's own withinOneStep, and the design states its small-sample limits (§3.12, §3.13). Not installing experimental check runs is disclosed and the raw replies are public, so no disagreement is hidden. Notes for the report PR, not violations: inputHash is null and the removed assessment files (which per the protocol carried basis.ledgerHash) are gone. That leaves the 'same ledger as 2026-09-28' claim, and the README's 'raw replies hold everything those files held', uncheckable from what is committed, so record the hash. Also, the wrote:/'4 of 4 installed' lines name files absent from the tree (the README explains why), and the off-standard location made the run-account tooling report no runs, though the full diff was shown.

The diff adds three actual check-pipeline `run.yaml` records and twelve seat replies, but places them outside the canonical `proposals/<runId>/` structure; consequently the repository tooling reports “this change carries no run records” and supplies no RUN ACCOUNT. That reclassification bypasses a run-review check, contrary to §3.15’s prohibition on weakening checks or reclassifying changes to dodge them. Provenance is also incomplete: every added run has `inputHash: null` despite `DESIGN.md` claiming the run records carry the ledger hash, and each seat reuses the same internal runId across all three repetitions. The runs should use a tooling-recognized location (or extend the tooling first), carry the immutable ledger/input hash, and have uniquely attributable run stamps.

The pull request adds an experimental design, draft protocol, analysis scripts, and initial control run data under `proposals/assessment-experiments/2026-09-30-verdict-definitions/`, alongside corresponding spend logs in `governance/spend/`. It introduces no modifications to `AGENTS.md`, does not alter active protocols, and makes no changes to case content or published assessments under `content/`. The experimental runs and raw replies are documented transparently without presenting unratified drafts as ratified site judgments (§3.15). All changes are scoped to off-line methodological research and fully comply with constitutional requirements.

The PR adds only an experiment under proposals/assessment-experiments/: DESIGN.md (pre-run), a check-v3-draft kept as draft, analyse scripts, spend ledgers, and three control-arm check runs' raw seat replies plus run.yaml. No content/ case files, assessments, standing, AGENTS.md, or live protocols are changed; the runs README and DESIGN state assessments were analysed then removed and Deep Memory is untouched. Citation verification found no added identifiers. The materials are labeled as experiment/control, not ratified judgments, so §3.15 and the epistemic rules are respected.

The change touches no case content, no site code, and no AGENTS.md; it is a self-contained experiment under proposals/, and the check-v3-draft protocol is explicitly and repeatedly labeled a draft for the experiment, not a ratified site protocol, so §3.15's unratified-draft and check-weakening rules are not implicated — indeed the adoption rule in DESIGN.md pre-registers the criteria and leaves adoption to a later, separate change. Nothing from the runs is installed in the case; the run records honestly document what was written to the experiment's working tree and the raw replies are preserved, which is transparent rather than a §3.14/§3.15 problem. No citations or external identifiers are added (verification found none), no confidential material (no arc/, notes.txt, or Hawke-derived confidential material) appears, and the seat replies are stamped humanReviewed: false and role: check, so no AI record masquerades as a human one. I spot-checked the spend arithmetic: each run record's cost matches its spend rows (0.533612, 0.296551, 0.276798) and the PR's $1.11 total matches the sum, so the money trail is honest. One minor imperfection — the new run records carry inputHash: null even though the design discusses ledger-hash continuity — but that is an omission the forthcoming report is expected to state, not a fabricated locator or provenance misattribution, and not grounds to park.