When Is an Agent Evaluation Over? Outcome Finality and Cross-Unit Separation

Title: When Is an Agent Evaluation Over? Outcome Finality and Cross-Unit Separation Source: arxiv Current agent evaluations incorrectly treat the final state of a stopped run as a definitive score. This framework identifies two necessary conditions for valid labeling: outcome finality and cross-unit separation. Controlled replays show that terminal labels diverge from endpoint scores whenever delayed operations remain unresolved.