Safety Judges Are Part of the Clinical System

Reliable deployment in automated screening of medical-agent outputs depends on more than obtaining a strong benchmark result. This conceptual analysis uses evaluator robustness to study how claims travel from data to model output and then to action. Its central thesis is that the unit of assurance must be the complete chain from incoming signal or prompt to evidence retrieval, model reasoning, confidence, and human action. The reviewed evidence shows recurring risks from hidden distribution change, correlated evaluation error, missing provenance, and optimization objectives that omit downstream costs. In response, the article proposes a layered evaluation program combining controlled perturbations, subgroup and scenario analysis, repeated runs, calibration or selective prediction, and monitoring after release. It also asks who can inspect, override, and learn from failures. By integrating the assigned target papers with established scholarship, the synthesis clarifies which findings transfer across domains and which remain local to a benchmark, dataset, or experimental apparatus. The goal is a testable research program for bounded, traceable, and revisable systems.

1. Hu, Saisai. "Research on Security Enhancement Methods for Adversarial Robust Large Language Model Intelligent Agents for Medical Decision-Making Tasks." *arXiv preprint arXiv:2605.08257* (2026).
2. Li, Yuanhao, et al. "DRAFT-RL: Multi-Agent Chain-of-Draft Reasoning for Reinforcement Learning-Enhanced LLMs." *Proceedings of the AAAI Conference on Artificial Intelligence* 40.35 (2026): 29530-29537.
3. Sang, Yinghao. "Adaptive Quantization Strategies for Robust ML Inference Under Distribution Shift." *Proceedings of the 2026 5th International Conference on Cyber Security, Artificial Intelligence and Digital Economy* (2026): 362-368.
4. Zhou, Yongxi, et al. "Style Over Substance: Content-Invariant Wrappers Flip LLM Safety-Judge Verdicts." *arXiv preprint arXiv:2609.08236* (2026).
5. Goodfellow, Ian J., Jonathon Shlens, and Christian Szegedy. "Explaining and Harnessing Adversarial Examples." *International Conference on Learning Representations*, 2015.
6. Guo, Chuan, et al. "On Calibration of Modern Neural Networks." *Proceedings of the 34th International Conference on Machine Learning*, 2017, pp. 1321-1330.
7. Ovadia, Yaniv, et al. "Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty under Dataset Shift." *Advances in Neural Information Processing Systems*, vol. 32, 2019.
8. Geifman, Yonatan, and Ran El-Yaniv. "Selective Classification for Deep Neural Networks." *Advances in Neural Information Processing Systems*, vol. 30, 2017.
9. Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. "Why Should I Trust You?: Explaining the Predictions of Any Classifier." *Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining*, 2016, pp. 1135-1144.
10. Singhal, Karan, et al. "Large Language Models Encode Clinical Knowledge." *Nature*, vol. 620, 2023, pp. 172-180.
11. Wiens, Jenna, et al. "Do No Harm: A Roadmap for Responsible Machine Learning for Health Care." *Nature Medicine*, vol. 25, 2019, pp. 1337-1340.
12. Rajkomar, Alvin, Jeffrey Dean, and Isaac Kohane. "Machine Learning in Medicine." *New England Journal of Medicine*, vol. 380, 2019, pp. 1347-1358.
13. World Health Organization. *Ethics and Governance of Artificial Intelligence for Health*. World Health Organization, 2021.
14. National Institute of Standards and Technology. *Artificial Intelligence Risk Management Framework (AI RMF 1.0)*. U.S. Department of Commerce, 2023.