Depth is a pipeline, not a model: how Gauntlet splits peer review's work

Technical review has split in two. Depth moves with workflow: a structured pipeline with an adversarial synthesis stage consistently beats the same model run as a single agent. Trust does not move with workflow. It is the residue where human reviewers still win, and it is precisely the kind of error a synthesis pass cannot fix: confident-wrong claims that survive reconciliation, mechanisms described but not propagated, breadth no one prioritizes.

The Gauntlet paper makes the split concrete. In a 98-paper automated ablation, a multi-agent pipeline with adversarial synthesis beats the same model run as a single rich-persona agent on 96% of papers. In a smaller human-graded trial on 20 recent computer-architecture venue papers, evaluators preferred Gauntlet's analyses 15 times out of 20, with the per-analyst advantage significant at p<0.01. The depth lift is structural. The calibration gap is what remains for humans to own.

For the program committees now negotiating how AI enters peer review, the next move is concrete: deploy a multi-agent depth pipeline, gate its output on a human-stamped trust signal, and measure both. The 96% number is not a verdict on the models. It is a workflow diagram with the calibration step still blank.