Dynamic Evidential Routing: An Auditable Neurosymbolic Interface for Clinical Transformers Under Extreme Missingness

Dynamic evidential routing adds an auditable expert-intervention channel to transformer inference while preserving comparable held-out predictive behavior. The interface maps Monte Carlo dropout uncertainty into Non-Axiomatic Logic (NAL) truth values, revises those values with explicit clinical rules, and uses revised confidence to gate attention. Every intervention exports its trigger, neural and symbolic truth values, revision, and attention effect. We evaluate TCGA-THCA lymph node metastasis prediction as a clinical-genomic feasibility check and WiDS ICU hospital-mortality prediction as the primary high-missingness scale test. The symbolic path is active at scale, firing 13,031 times across 8,551 of 13,757 held-out ICU stays. On WiDS, paired Brier and ECE differences favor NARS-gated routing over the ungated transformer, but comparisons with flat-confidence and MC-confidence-only gating include zero. The results therefore support confidence-gated, auditable intervention without isolating symbolic revision as the cause of the calibration difference. The implementation is a NAL truth-value interface, not a full Non-Axiomatic Reasoning System (NARS) cognitive architecture or a clinically validated decision device.