Verifiable AI Abstention Boosts Water Leak Diagnosis Accountability

Water utilities face significant losses due to leakage, yet often hesitate to trust AI systems for dispatching repair crews because AI predictions lack accountability—they can't explain when they *shouldn't* act. This research redefines leak localization as a decision-making process that includes "verifiable abstention," ensuring AI only acts when confident and can justify its decisions. The proposed system features two main agents: a physics-grounded executor agent that tests hypotheses (e.g., leak, demand, sensor issues) against a digital twin of the water network, and an independent supervisor agent. The supervisor, aided by a large language model (LLM) auditor, verifies evidence against a code-verifiable contract. Based on this audit, it either certifies a dispatch for excavation, requests more evidence, or abstains from making a decision. Under realistic field noise conditions, the system dramatically improved decision precision on acted events from a 32% forced baseline to 96%. On an independently generated benchmark, it correctly identified all 4 of 33 leaks it chose to act upon. A real-world dataset of 194 audited leak events, simulated with twin pressures and flows, resulted in five excavation dispatches, three of which were correct, achieving 44% survey recovery at full district precision. This accountable abstention mechanism offers a robust and defensible pathway towards autonomous operation of critical water infrastructure.