The diff does not touch AGENTS.md or content/, installs nothing in Deep Memory, and adds no citations (the verification tool found none). I recomputed arm B from the committed seat replies. OpenAI is 23/25 same word and 25/25 within one step, Google 13/17, xAI 12/14 with 8 weakly/provisionally splits, and Venice 15/17. The panel's first-run pairwise counts are 79 same and 98 within one step. All of these match the report and DECISIONS entry. Each spend file sums exactly to its run.yaml's tokens and USD, and the report's arm-B input-token column matches the spend rows. The provenance gaps are disclosed rather than hidden: `inputHash: null` is admitted with a calibrated caveat ('equal token counts do not prove equal bytes'); the `wrote:` files the run account flags as absent are explained as analysed-then-removed working-tree files; the earlier 'hold everything' claim is corrected in place, quoting the original wording; and the lost calls during machine sleep are disclosed. The `check-v3` stamp is explained, and a new test keeps that name from ever labelling a different text. That test only adds a check. Filing the runs outside `proposals/<runId>/` does not dodge the gate, because the run account reads them here. The negative result is reported by the pre-committed rule, post-hoc analyses are labelled as descriptions, and the scale question is put to the founder without altering §3.15. I cannot check the arm S/A figures or the design commit's time and hash (`ffff11f`) from this packet, but the report labels these 'a record, not a proof', and nothing visible contradicts them.

The change preserves the experimental runs as model-labeled raw replies and spend records, while clearly stating that their generated assessments were not installed in the Deep Memory case or used as its standing. The report handles the missing run-time input hashes honestly: it leaves `inputHash: null`, distinguishes the later-computed ledger and packet hashes from run-recorded provenance, and explicitly says equal token counts do not prove equal bytes. Post hoc analyses are labeled as descriptions rather than preregistered tests, limitations are calibrated to one case, four seats, and three runs per arm, and the adverse xAI result is reported rather than hidden. The non-unique model-written reply IDs and omitted verb stamps are also disclosed, while directory and filename provenance identifies each reply. No citation identifiers are added under `content/`, no confidence material or human-review claim is introduced, AGENTS.md is untouched, and the added test strengthens rather than weakens protocol-stamp accountability.

This pull request completes the evaluation of the verdict-definitions experiment across its second arm without modifying any canonical case content under content/ or adopting the experimental protocol. The report documents the pre-registered decision rule and demonstrates transparently why the proposed definitions were not adopted based on the results across the four evaluation seats. All experiment runs, raw seat replies, spend records, and analysis scripts are tracked with meticulous provenance and calibrated uncertainty, and a test is added ensuring the spent protocol identifier check-v3 is never reused. The changes adhere fully to epistemic rules and the project constitution.

The change is an experiment report and supporting artefacts only: three check-v3 run records with raw seat replies under proposals/, spend rows, REPORT.md, describe/standing scripts, a DECISIONS entry, a runs/README correction, and a test locking the spent check-v3 name. No content/ case files are touched; the assessments the runs wrote were analysed and removed, so nothing is installed as Deep Memory standing and no unratified draft is presented as ratified. The report honestly states the precommitted adoption rule was not met (xAI within-one-step fell by nine), that inputHash is null, that equal token counts do not prove equal bytes, and that post-hoc pair/standing descriptions are not design tests. AGENTS.md is untouched; no citations are added; OMITTED FILES is empty.

No content files, AGENTS.md, or case assessments are changed: the diff adds only experiment records, spend rows, a report, scripts, and a decisions entry, and both the run account and the README confirm that the assessment files the runs wrote were removed and nothing from the experiment is installed in Deep Memory — so no unratified draft is presented as ratified and no standing is altered. The report handles provenance honestly: it corrects its own earlier false claim that raw replies 'hold everything', admits the design's assertion about ledger hashes was wrong when written, states plainly that equal token counts do not prove equal bytes, and labels every post-hoc finding as a description rather than a test. The check-v3 stamping is disclosed (draft copied in and removed), and the new test guards the stamp by ensuring no site protocol takes that name — a strengthening of the checks, not a weakening, since the protocol stays check-v2 and the definitions are explicitly not adopted. Lost calls from the second run are disclosed rather than hidden in the spend ledger. Nothing here fabricates a citation, locator, or record, misattributes a stamp, edits the constitution, or dodges a gate; the disagreements the experiment exposed are displayed rather than smoothed over.