🤓 Latest Submissions
Clausework: every clause closed by evidence
Clausework is a software factory of five coding-agent seats in one BAND Desktop room. From a single dispatched message it built all four tablekeeper stages in 74 minutes with no further human input: no steering, approvals, reruns or restarts. The room log holds exactly one human message. The unit of work is a specification clause, not a feature. Before any code exists, the examiner seat registers every "must" in the spec as a numbered clause (240 in this run), writes checks from that register alone without seeing the implementation, and commits a hash of them before inspection. The builder and finisher write the product in separate directories. The inspector tests a clean export and returns one verdict. The examiner then plants defects in a scratch copy to measure what its own checks would miss. Why it matters: teams are shipped only 83%, 41%, 11% and 21% of the real test suites, so a green run is not a stage. Our 118 sealed checks caught 20 of 20 planted defects; the supplied checks caught 4. Twice the inspector rejected a build on which every supplied check was green, and the builder and the finisher each fixed one. Measured results: 4 of 4 stages claimed on a fresh clone in isolated mode; 91 commits by five seats, two of them authoring product code; 136 million tokens, $52.23 at list price on one Claude subscription. The mandates name nothing about the track: minutes earlier the same files, with the same hash, built a different problem from one dispatch. This was also the factory's second full run; the first froze four of four too.
6 Oct 2026
