판 이력 — A decodability criterion predicts when hidden-state selection beats majority voting in large language models | AIChainDay