E.23.CAE:5.3 - AI or robotic apparent forgetting

Tell. A continually updated model or robot previously expressed a task function or policy and later appears to forget it. A benchmark drop can reflect parameter change, task/context routing, function activation, interface state, sensing, actuation, support configuration, or a demand outside the tested envelope.

Show. Hold parameters fixed during the probe where feasible. Compare no task/context cue with a qualified task cue or routing intervention; vary cue reliability, state feedback, blocked versus interleaved order, and relevant transition statistics; then return to the earlier condition without further update. For a robot, keep sensing, controller state, actuation, calibration, tool, and environment conditions explicit. If the earlier function or policy reappears under a qualified activation condition, availability under that condition is supported and simple global loss is weakened. Function vectors, context inference, routing, controller state, continual-learning algorithms, and parameter overwrite remain model-specific explanations. If no response reappears and direct model or controller evidence supports change, capabilityClaimRevisionWarranted may be returned without claiming a universal memory mechanism.