From Structural Dynamics to Robot Safety Runtime Assurance, Hard Constraints, and Persistent Correction in Embodied AI

Embodied AI must act under physical constraints, delayed feedback, partial observability, and potentially irreversible consequences. This paper develops a design proposal for robot safety that places runtime correction outside ordinary task optimization when safety-critical constraints require it. The proposal is situated within established work on reward hacking, safe reinforcement learning, shielding, control barrier functions, safe interruptibility, and runtime assurance. Five architectural functions organize the account: grounded state and fault detection; explicit safety constraints; supervisory arbitration with authority to interrupt a learned policy; persistent correction verified by regression and safety tests; and traceable handoff between task control and safety control. The account adds a calibration requirement: not every anomaly warrants policy change. A safety-relevant discrepancy must be linked to sufficiently reliable evidence, and the appropriate response may be update, bounded non-uptake, or temporary suspension into a safe state. A correction is stronger when it changes future tests or control behavior under recurrence while remaining open to later evidence. The term steering capture is retained only as a bounded diagnostic label for repeated proxy control that conflicts with an independently specified safety envelope. The contribution is an integrated engineering pattern for runtime assurance, evidence-sensitive correction, trace, and recurrence rather than a new general theory of robotic agency.