판 이력 — Model Confidence Under Answer-Preserving Attacks: An Informativeness-Manipulability Frontier | AIChainDay