COMMIT: Factory That Measures Bad Work

COMMIT is a software factory that measures how much bad work it can kill. AI agents ship code fast, but green tests do not prove strong verification. The deeper problem is structural: the agent that builds software can also become the agent that decides it is correct. COMMIT turns verification into an independent release gate inside BAND Desktop. Three generic seats share the work: Planner, Builder, Verifier. Planner converts the specification into numbered requirements and acceptance conditions. Builder implements, commits, and reports exact commands and raw evidence. Verifier has no production-code edit authority: it rebuilds each candidate in a fresh container with no network, reruns the official checks, and attacks behavior happy-path tests miss — spec-derived contract checks, a reference model with replayable seeds, 50-way concurrency attacks, and mutation testing of its own checks. When verification fails, it REJECTS with a reproduction command; Builder repairs; Verifier re-checks: REJECT → REPAIR → RE-VERIFY → ACCEPT. The submitted run (5 Oct 2026): all four Pocketful stages ACCEPTED in 9 h 32 min. The official harness (isolated, all stages) reaches stage 4, and every stage folder claims its own stage. Stage 1 ran a full 782-mutant campaign (92.5% killed). Zero lines of stage code were written by a human; model spend was $33.57. The full room log (room.json, 1,517 messages) is in the repository. Human input, disclosed: the dispatch, plus one time-limit message (finish by 23:00; sampled mutation from stage 2 on). In calibration, the verifier's kill rate rose from 64% to 95.4% as every escaped mutant became a new check. Pocketful demonstrates the factory. Evidence demonstrates the claims.