formalization of "how well a linear classifier explains the performance of a model " has many advantages over prior