What if the ML takes the conclusion exactly from the right pixels, but the cause is a rasterization issue.