GLEAN: A guideline-grounded evidence accumulation framework for high-risk agent validation

GLEAN is a verification framework for high-risk agents that compiles expert-developed domain guidelines into trajectory-aware, calibrated correctness signals and accumulates them into a correctness probability using Bayesian logistic regression. On agentic clinical diagnosis tasks for three diseases in MIMIC-IV, its AUROC exceeds the best baseline by 12% and its Brier score is 50% lower; uncertainty-triggered active verification expands guideline coverage and performs differentiated checks. A clinician expert study recognized its practical utility.