Torpor

Torpor for Muse Code: you say what works ("login works now"), and if the AI breaks it, the AI is told before its turn ends. An AI can say "Done" after breaking it or deleting files; "undo that" restores your files. Muse Spark writes a check from your words. At turn end a Stop hook has the engine run it; on a fail the AI is sent back and asked to start its reply with what broke. Pre-registered proof (build 8b119d6, Muse Code 1.4.13; 11 runs, each case once, one machine, demo app only): 5 of 5 cases met (3 only after the AI's first "Done"); no false note in 40 clean turns. Against its engine alone (earlier builds, synthetic data): scripted drill, 16 disasters (adapter 235bc5c), won 5, tied 11, lost 0; pre-registered study, 21 tasks x 2 (adapter 0e8b0da), inconclusive: behaviours kept in 34 of 42 episodes vs 32 of 42 (95% CI -0.05 to +0.14). Built during the hackathon (judge this): Muse Code hooks, check writer, voice promises. Declared base (before the event, not judged): the open-source Torpor engine (github.com/Kylne700/torpor, tag pre-event-final) that saves, undoes and runs checks. Limits: Node apps only. Muse Code 1.4.14 shows a Stop block only as a 5 s label (Torpor's status line was measured only with Muse's echo model), so the AI's reply is what lasts. If a prompt dictates the reply, the check stayed failing in 1 of 2 early takes (build 56a6a11) and in 1 video take (older build 8b119d6), 0 of 16 later (b9871b5, 56a6a11); too few for a rate. Built with an AI coding agent (Claude Code).