Or you get the original output result was perfect and the adversarial "rethinking" switches to an incorrect result.