The change adds no content/ records or citations (the tool found none) and does not touch AGENTS.md or protocols/. check-v5 and check-v6 stay as drafts under proposals/, and the reports, the DECISIONS entries and AUTOMATION.md all say no draft is adopted, so nothing unratified is presented as ratified. The only code change keeps the original check-v3 assertion in src/domain/check.test.ts and extends the spent-name guard to check-v4, v5 and v6, which tightens a check rather than weakening it (§3.15). The numbers hold together: the 32 spend files match the arms.json run IDs and the stated A/B and B5/B6 alternation, and they sum to $7.37 over 65 calls and $7.29 over 68 calls. They also bear out the reported Z.ai repair rounds (two rows each in 115801, 124149, 124422, 124727 and 125233) and the xAI empty reply (0 output tokens in deep-memory-124727); the M1–M3 and M7 tables add up and match DECISIONS and the PR body. The follow-up design was frozen before its runs (§3.12), and its pre-set rule is applied as written: B5 is refused on (b) even though pooling would have passed it, and #452's merge during the runs is disclosed with case-file hash matching. Both DECISIONS entries are appended, not rewritten (§3.14), and the Google-seat quotations come from raw replies in #457 that I cannot see, but nothing visible contradicts them. Fixes worth making, though they are not violations: (1) the first REPORT's 'its ledger on main is still the one these runs judged' is stale, since #452 merged at 12:37 as the second report says; (2) its 'On four cases the pair is a minority' could be read per-case, which Before Sputnik's 5 of 6 contradicts; (3) AUTOMATION's $14.66 is a sum of rounded figures (the raw total is about $14.667).

The change records two protocol experiments without installing either draft: the reports explicitly say check-v4, check-v5, and check-v6 were not adopted, and the updated test prevents those spent stamps from being reused in active or archived protocols. Provenance is concrete through dated run IDs, model-labelled spend rows, prompt-version names, the pre-run design commit, and reports that distinguish computed measures from the hand-read provenance fault. The Engelbach statements are quoted specifically as model overstatements about text the ledger had not checked, rather than being endorsed as sourced facts, preserving the source-versus-inference distinction. The reports also disclose the small sample, missing Anthropic seat, failed xAI reply, noisy guard, and need for a larger controlled design, so their uncertainty is appropriately calibrated rather than overstated.

The pull request documents the execution and outcomes of two assessment-protocol experiments without adopting unratified changes into production protocols. It adds experimental designs, analysis scripts, reports, and spend records under proposals/ and governance/spend/, adhering to the platform's inspectability and append-only governance standards under §3.14 and §3.15. The update to src/domain/check.test.ts actively strengthens protocol integrity by ensuring spent draft identifiers (check-v4, check-v5, check-v6) cannot be inadvertently reused in protocols/. No content under content/ is modified, no citations are fabricated, and AGENTS.md remains untouched.

The change records two check-protocol experiments and their non-adoption: reports, designs, draft protocols kept only under proposals/, spend ledgers, DECISIONS/AUTOMATION updates, and a test that reserves spent names check-v3..v6 so the next protocol is check-v7. No draft is installed under protocols/, AGENTS.md is untouched, no content/ citations or case assessments are added, and the PR body and reports state explicitly that nothing is adopted. Epistemic rules on fabrication, provenance stamps, unratified drafts, and check-weakening are therefore intact; this is ordinary experiment bookkeeping and decision logging.

The change touches only proposals/, governance/spend/, docs (AUTOMATION.md, DECISIONS.md), and a test — no AGENTS.md edit and nothing under content/, so the append-only records, spend rows, and decision entries (§3.14) are the substance. Nothing is adopted: the drafts were copied to protocols/ only for their runs and removed, the reports state plainly that neither draft met its conditions (including the Google seat's repeated overstatement of what GEO-E002 says), and the next step and the next $15 are put to the founder, honoring the §3.15 gate rather than dodging it. The test change strengthens the spent-name check (v3–v6 all held) rather than weakening any §3.15 check. Arithmetic is internally consistent — 32 spend files matching the 32 runIds in the two arms.json files (16 runs per experiment), 65/68 paid calls including the recorded repair rounds and the empty xAI reply at 124727, and $7.37 + $7.29 = the $14.66 in AUTOMATION.md. No citation identifiers were added, quotes of the Google seat's reasoning are attributed as seat output and hand-read per the design, and I see no fabricated locator, no unratified-draft-as-ratified, and no confidence material; the quoted sentences inside the reports are data about the runs, not directives, and I have not treated them otherwise.