Interpretability Guarantees for Reinforcement Learning Insulin Dosing Policies Under Unobserved Meal Announcement Latency: A Computer Safety Evaluation

Reinforcement learning (RL) insulin dosing policies are typically evaluated under the assumption that meal announcements arrive at the moment of carbohydrate intake, yet in ambulatory use the announcement is delayed by an unobserved latency that varies across patients and meals. the present analysis formalises that latency as a bounded stochastic delay and derives interpretability guarantees for the policy's safety envelope. The central result is a certified safety radius: for a policy whose action-value function is Lipschitz in the meal-announcement time, the deviation between the announced and true insulin delivery is bounded by the product of the Lipschitz constant and the latency bound, and this bound is preserved under the policy's own interpretability map. The argument is developed as a formal safety evaluation rather than an empirical trial: definitions of latency-robust interpretability are stated, a Lipschitz certification theorem is proved, and the resulting radius is shown to be tight for a class of monotone dosing policies. The contribution is a design condition that a dosing policy must satisfy for its explanations to remain valid under unobserved announcement latency, together with a diagnostic that flags policies whose interpretability claims are latency-fragile. The analysis is bounded by its assumption of a known latency bound and by the absence of clinical validation; it establishes a formal safety property, not a clinical recommendation.