Abstract
High-frequency monitoring networks raise an alarm whenever a measured value crosses a regulatory threshold, and each alarm consumes scarce investigative capacity; unreliable alarms spend that capacity chasing non-events, yet how reliably these alarms indicate actionable pollution is rarely audited against what field investigators actually find. We audit an operational alarm stream against its own verification records to (i) quantify how many alarms reflect an actionable point-source discharge, (ii) identify the mechanism behind the rest, and (iii) determine how much investigative effort can be recovered. Pairing 281,039 station-hours from a 25-station Chinese river network with investigators’ conclusions for 702 of 1049 alarms, two coders independently coded the narratives (Cohen’s kappa 0.807); only 7.7% of the 428 station-matched modeling alarms (20 of 101 verification-narrative groups) recorded an actionable discharge. Dissolved-oxygen exceedances cluster before dawn (Rayleigh R = 0.311) and scale with temperature, consistent with a strong diel metabolic contribution, though hydrological and anthropogenic oxygen demand cannot be excluded. Under leakage-free validation, no signal-based model beat a station base-rate prior that uses no monitoring data; a simulated triage nonetheless de-prioritizes about two-thirds of narrative groups while retaining 18 of 20 confirmed discharges. We conclude that alarm reliability, not monitoring volume or model complexity, governs the sustainability value of smart-water systems.