A wheeled robot patrols between two docks past a restricted zone using a potential-field policy whose hazard-avoidance gain degrades as you raise the perception-drift slider — modelling a sensor or perception fault rather than a hard-coded bug. An independent runtime monitor never inspects the policy itself; it only tracks how far the robot's actual trajectory diverges from a verified-safe reference ("ghost") running in parallel, folding that divergence into an EWMA anomaly score. Cross the tunable trip threshold for a debounce window of consecutive frames and the monitor hands control to the reference policy's own commands until the score settles back down — the same detect-and-switch pattern used by runtime-verification shields and supervisory watchdogs guarding learned robot controllers in the field.