Outcome Calibration

A key property in AI user-simulator evaluation referring to the degree to which the failure modes and success-rate distributions produced by the simulator match those observed in actual human-user interactions with the agent.