An evidence-review system must decide when to spend additional model judgments and how to control their use. The authors of this arXiv preprint study those software decisions directly, using 6,164 frozen manufacturing source-task pairs drawn from supplier discovery. The dataset is fixed, so the comparison isolates the routing and revision logic rather than the underlying model.
The central proposal is benefit-targeted routing, which learns to route cases for extra judgment based on recovered agreement minus a cost term. The abstract is truncated mid-sentence, so the full objective and the numerical results are not available in the source. The paper also describes bounded revision, a mechanism that limits how and when model judgments are revised after routing.
Taken together, the work frames routing and revision as engineering decisions that can be learned and evaluated, rather than left to ad hoc prompting. With only the abstract available, the exact performance gains and baselines are not stated; the significance is in the framing and the scale of the evaluation.