A measure of how much a model's predictions change when intermediate reasoning steps are modified, indicating whether steps actually drive the output.