ProofRig

ProofRig is an evaluation harness for robot code, with a feedback loop that repairs what fails. AI now writes and patches robot controllers, and "fixed, it works" costs a model nothing to say. ProofRig makes the claim earn trust: Meta's Muse Spark repairs the controller's source code through tool calls, but deterministic code (a MuJoCo sandbox, an exact feasibility oracle, 400 hidden scenarios the model never sees) decides the result. Two verdicts: Verified (works on unseen cases) and Certified (and nothing that worked before regressed). Results are pre-registered and traceable: held-out success 76.9% to about 99% in 16 of 16 repairs, and 3-D arm 77% to 99.7–100%. At equal success, behaviour-aware repairs passed the behaviour contract 5/5 vs 0/5. A model-free CI gate blocks bad patches. Simulation only; limits are listed in the README.