Title: Zero-shot automated insulin delivery for type 1 diabetes via dynamic physiology-aware reinforcement learning

Insulin therapy for type 1 diabetes requires continual dose adjustment to meals, activity, stress, illness, and changing insulin sensitivity, creating a substantial self-management burden and increasing the risk of dosing errors. We developed the Dynamic Physiology-Aware Reinforcement learning Controller (DPARC), a zero-shot automated insulin delivery algorithm that infers latent physiological dynamics from recent continuous glucose monitoring and insulin-delivery history without subject-specific tuning, carbohydrate announcements, or preset clinical parameters. DPARC uses a rolling 24-hour history window, but closed-loop control can begin after 1 hour of observed data by initializing the unobserved history with neutral normalized padding and progressively replacing it with real observations. In silico, a single frozen DPARC policy adapted within 1 hour, improved time in range compared with a total daily insulin-conditioned reinforcement learning baseline, and approached the performance of a fully personalized upper-bound model across stochastic unannounced meals with randomized timing, carbohydrate amounts, absorption variability, and meal skipping. Additional perturbation studies showed adaptive insulin delivery during changes in insulin sensitivity and carbohydrate load. In rodent experiments, DPARC was evaluated as a rapid-physiology stress test, revealing feasible zero-shot deployment while highlighting the need for further safety tuning. In supervised porcine studies with unannounced meals, the same frozen policy maintained high time in range without manual configuration, supporting large-animal feasibility. Learned latent representations aligned with physiological markers including insulin sensitivity and plasma insulin, providing post-hoc explanatory anchors for the policy. These findings establish DPARC as a preclinical proof-of-concept zero-shot AID framework and motivate supervised prospective human evaluation.