This also seems plausible for benign failures — the really bad failure cases are cases in which arriving at the right answer wouldn’t have involved much detailed knowledge or handling of uncertainty, and so it ought to be easier to learn.