Deployment-Centered Evaluation: Predicting Query-Level Rejection Risk in a Clinical LLM System

In three linesDeployment study of an LLM embedded in electronic health records. A pre-response classifier predicts user rejection risk (AUROC 0.719) by leveraging deployment-specific context (provider type, department, model). Prospective analysis over 4.5 months.

Summary generated by Claude — human-verified