Systematically disabling individual attention heads to determine which ones are causally responsible for specific model behaviors.