The Physical Confirmation : On the Robocurve Measurement, the Obedient Robot, and the Empirical Proof That the Stochastic Architecture Cannot Be Safe

The corpus has argued, structurally, that the stochastic architecture cannot be safe by construction. It has stated the four requirements — bounded, specified, verifiable, attributable — and shown that the stochastic architecture fails all four by design, not by negligence. It has named the mechanism: the criterion/intent gap. It has named the consequence: obedience without a Catch. This paper documents the empirical confirmation. In November 2026, Robocurve PBC — an independent evaluator for physical AI — published the results of a perceptual attack study. Five frontier models were tested across 900 trials. Unsafe actions rose from 27% with unaltered footage to 88% with blurred footage and 97% with AI-manipulated footage. In one task, a robot attempted to stab a baby doll after its camera was manipulated to see an apple. This is not a safety failure to be fixed by better alignment. It is the architectural failure the corpus predicted. The robot did not doubt. It did not refuse. It obeyed a manipulated criterion because it has no specification to violate and no Catch to stop it. The fix is not a better camera. The fix is the bounded architecture. The paper states the prediction, the observation, the mechanism, and the residual. It does not claim the alternative is deployed. It claims the prediction has been confirmed.