Preprint tests abstaining AI workflow for water-leak dispatches

A preprint describes an AI leak-localization workflow for water networks that can certify a dispatch, request more evidence or abstain. The system tests leak, demand, sensor and valve hypotheses against a digital twin, then has an independent supervisor and LLM auditor check evidence against a code-verifiable contract. In a 194-event register of audited leak locations with simulated pressures and flows, it made five excavation dispatches, three of them correct. On an independently generated benchmark, it acted on four of 33 leaks, all correctly.