A cheaper reviewer from a different model family beats self-review by 12 points.

On 100 olympiad math problems in an execute-review-revise pipeline, a cross-family mid-tier reviewer lifted final accuracy from 52% to 64% (p = 0.0005) with zero damaged answers (arXiv 2609.04270). Same-model self-review had the highest error-detection recall of any condition at 0.85 and still produced no significant gain, because it rejected 2.1x as often for a third the repair rate and falsely rejected 35% of its own correct answers against the cross-family reviewer's 2%. There's a floor too: the weakest reviewer changed zero of 100 answers while doubling token cost. Cheap insurance below the executor's level is just cost.