A measure of how well an explanation captures the true reasoning of a model by testing prediction changes.