Determining whether a model's decision or correction is causally grounded in visual input rather than other factors.