mattmacdermott comments on Validating against a misalignment detector is very different to training against one