Measure. Refine. Re-check.

Keep the criterion fixed, measure the artifact, name the drift, and recompute the verdict after the work improves.

Crucible 1.2.0 / local fixture / deterministic verdict / sanitized capture / source

The 30-second proof

A draft release brief satisfies one of three measured claims. Crucible records the two concrete drifts, the worker artifact is refined without moving the criterion, and the same thesis reaches three measured matches.

Silent, captioned on screen. The exact sanitized transcript, thesis, artifact receipts, and machine-readable evidence are linked below.

The observed conversion

3claims measured

1 / 2draft MATCH / DRIFT

3 / 0refined MATCH / DRIFT

2 PASScleanroom reviews

Register: one three-claim thesis defines the release criterion before either artifact is judged.
Measure the draft: one claim holds; two measured axes return DRIFT rather than an invented pass.
Hold the boundary: the cleanroom reviewer receives the spec and artifact packet, not worker context or reasoning traces.
Refine: the artifact changes while the thesis stays fixed.
Re-check: the refined run returns 3 MATCH / 0 DRIFT, and the verdict is re-derived from disk.

The full 94-second workflow

The long cut preserves all seven successful commands: build draft evidence, run and review the draft, build refined evidence, run and review the refined artifact, then verify the recorded verdicts. Every displayed count, receipt, seal, return code, and verdict comes from the captured evidence package.

Run path

python make_substrate.py worker-draft.md substrate-draft.json artifact-evidence-draft.json
python -m crucible run thesis.json --substrate substrate-draft.json --registry registry-draft --bundle bundle-draft
python -m crucible review bundle-draft
python make_substrate.py worker-refined.md substrate-refined.json artifact-evidence-refined.json
python -m crucible run thesis.json --substrate substrate-refined.json --registry registry-refined --bundle bundle-refined
python -m crucible review bundle-refined
python -m crucible verdicts registry-refined --verify

Sanitized transcript Machine-readable evidence Independent capture verification Fixed three-claim thesis Reproduction guide and checksums Download 30-second MP4 Download full MP4

Claim boundary

This is a local, sanitized fixture built to make the measurement loop inspectable. It demonstrates the register, run, cleanroom review, refinement, witnessed assessment, and disk re-verification paths shown here. It does not claim that an evaluation criterion is automatically well-designed, that all claims can be reduced to numeric measures, or that a passing fixture proves production fitness.

Recorded and verified 2026-07-12 from the public Crucible 1.2.0 source revision shown in the evidence package. Both videos are H.264, 1920 x 1080, 30 fps, and contain no audio. They decoded successfully end to end, and the package passed path and credential hygiene checks.