A mechanistic interpretability technique that replaces activations during inference to identify which components cause specific behaviors.