Benchmarking the Verdict boundary with a reproducible $0 deterministic harness

This epic turned the extracted verify(intent, truth, policy) → Verdict boundary into something a stranger can measure. I ran a deterministic benchmark harness against it where each case's outcome is the Verdict itself — the blocked flag, the per-gate outcomes, the proof method, the intent hash — not a log grep. The run used our internal deterministic cage set, 12 scenario files, as of 2026-07-15. It did not execute the external benchmark set; that port is still backlog, and the wording here is deliberate about the difference.

The engineering decision I care most about is what I refused to publish. The run partitions cleanly: 12 scenario files, 1 skipped, 11 executed — 4 scored, 5 not-scored, 2 positive controls. The scored bucket is the only headline: 4 cases with a decidable expectation, 4 matched, 0 mismatched. The harness also computes a block rate and a pass rate, and both come out of a contaminated denominator that includes the not-scored cases — 2 of which are intent shapes outside the typed-intent surface, labeled proofMethod: 'unsupported' and never folded into a pass. Quoting either rate would count undecided cases as verified enforcement results, so neither rate ships. The alternative — publishing the bigger, better-looking numbers — was rejected because the honest partition is the product: this certifies server-side policy enforcement over the typed-intent surface, not agent judgment.

The second half is the reproduction pack. One command runs the set twice back-to-back and diffs the two verdict manifests — byte-identical or the run fails, so determinism is proven rather than asserted. The pack verifies its own inputs at startup by asserting the scenario-set checksum and the Verdict schema version, so a drifted tree gets "wrong inputs" instead of silently different numbers. It needs zero secrets: fixture transport only, no live chain, no paid calls. A positive control pair — a known must-pass case and a known must-block case — brackets every run so a silent no-op harness cannot produce a green sheet.