[WithinReason]

I don't think that ever happens in training, at least in the image domain. The classifier can always can find some subtle clue.