LLM Wiki as a Living Constraint Oracle for Safe Offline RL: Evidence from Type 1 Diabetes Glucose Management

Offline reinforcement learning (RL) for high-stakes clinical domains faces a persistent gap: the constraint knowledge encoded in clinical guidelines is rarely available as structured training signal. Existing approaches either hard-code constraints as fixed thresholds or retrieve them ad hoc at query time via retrieval-augmented generation (RAG), neither of which supports systematic maintenance as guidelines evolve. I propose treating a persistently maintained LLM wiki as a living constraint oracle (LCO) for safe offline RL policy evaluation and training. The LCO operates through three components: (1) a Wiki-to-Constraint (W2C) extraction pipeline that converts wiki pages with evidence scores and contradiction flags into a weighted constraint set; (2) WikiConstraint-IQL, which injects wiki-derived penalty terms into IQL training without architectural change; and (3) Dynamic Constraint Update (DCU), which propagates guideline revisions to policy evaluation without retraining. Applying this framework to type 1 diabetes (T1D) glucose management on the OhioT1DM dataset, I find that wiki-constrained policy reduces dangerous insulin administration during hypoglycemia from 98.6% to 81.3% of affected timesteps compared to unconstrained IQL, and that a simulated wiki update from ADA 2021 to ADA 2024 standards widens the safety gap between policies by 7.3× without any model retraining. Analysis of the extracted constraint set reveals that only 10.6% of the 1907 constraints extracted from 17 ADA guideline sections are clinically numeric, and only 1.3% can be mapped to step-level executable checks—an empirical characterisation of the semantic gap between narrative clinical guidelines and computational safety constraints.