A method to identify which attention heads in a transformer are causally responsible for specific model behaviors.