Review added 33 points. It still lost to one call.

Hand one model’s answer to another model for review and accuracy climbs — 33.2 points on one pairing — because you are measuring against the weaker model that wrote the answer. Measured against the stronger of the two, called once by itself, review won in 9 of 96 pairings, finished about 3.1 points below it on average, and cost 2.6 to 2.8 times as much. The reviewer is not verifying. It is handing over its own knowledge, and deferring to the answer already in front of it.

Zhang and Shrestha · 2026 · 12 open-weight models, 3B to 72B · 48 pairs × both directions · 4 domains, all graded by program