NeoX-IF: Adaptation of NeoX v2.0 to Physical AI — Full Design Document

NeoX-IF v0.1 is an exploratory design document that adapts the NeoX v2.0 framework to physical AI, meaning robots acting under learned policies. It proposes restrictive supervision with slow human governance, organized in five planes: deterministic safety, learned policy, a restrictive supervisor, NeoX orchestration, and humans. The supervisor can only restrict, never widen, a robot's autonomy. The design covers a monotone authority ladder (L0–L5), irreversibility classes (R0–R4), expiring concessions, twelve rules (IF-01 to IF-12), twelve design risks, seven falsifiable hypotheses and a five-stage validation plan. It starts from a review of three synthetic simulations, with provenance and epistemic labels attached to every figure. The document is exploratory: it has had no human review or legal review, it was not tested on a robot or in a physics simulator, and every threshold is an illustrative seed. It makes no claim of conformity with any standard, law or certification, and it must not be used as, or in place of, a safety function. Companion work: SSRN 7585639.