LeoAtiS

Robots are trained in simulators whose physics is guessed, and controllers written by AI agents ship with no evidence they hold on the real system. TwinProof turns ten seconds of phone video, with a printed marker in frame, into a MuJoCo digital twin that knows how wrong it is. SAM 3.1 tracks the moving part in every frame; the marker turns pixels into metres; CMA-ES fits a grey-box twin with a Laplace posterior and an OU residual. A Muse Spark agent then proposes model fixes and controllers through tools, and a deterministic verifier runs seven gates (patch guard, admissibility, held-out prediction with whiteness, Lyapunov stability with a certified region of attraction, a torque bound, contact safety, 300 robustness rollouts). Each claim is ADMITTED with a witness or REFUSED with the number that failed. On five real rigs the twins predict held-out takes to 0.8% NRMSE on the best pendulum, a can's inertia ratio replicates across two ramps (0.78 vs 0.74), and the pipeline caught a ramp stated as 13° that the motion shows is 8°. The verifier still refused every real controller, because the twins' errors carry structure; its self-test catches 7/7 seeded bugs. A web console shows the ghost twin over the video, the phase portrait, the verdict, the claim stream and spoken what-ifs.