The research presented Cros, a new approach to stopping clinical diagnosis agents. The system combines state-wise error ranking, policy design on disjoint development splits, and LTT-style tests. Evaluation on a 1,834-episode MIMIC-derived abdominal-pain benchmark demonstrated that the full ranker achieved an AUROC of 0.853, outperforming maximum class probability (0.715) and the native stop score (0.552).
Comparing Cros to native stopping on the 367-episode evaluation split, the system reduced selective error from 30.8% to 16.9% while maintaining 78.8% coverage, at a cost of 5.57 and 0.68 tests. Forced continuation resulted in non-monotone error rates, with 28.3% error using HPI alone and 34.3% after a full workup.
Key constraints for the system include freezing the candidate family, testing rule, and any randomization before accessing calibration labels. The uniform-weight mixture ablation performed cheaper on the viewed split despite missing locked development margins. The findings provide exploratory feasibility and audit evidence, not a confirmatory safety certificate.
Source: https://arxiv.org/abs/2609.09678