Counterfactual Intervention Standard for Agent Failure Attribution

This research proposes a standard for attributing agent failures by executing counterfactual interventions—freezing mid-trajectory states and altering history—rather than relying on post-hoc transcript labeling. It emphasizes the need for measured replay certificates to verify causal claims.