Twelve days after the catastrophic flooding in Nepal, we are still learning why warnings failed and what actions helped get some people to safety. We know this: On 26 August, the warnings that reached people in time to act were human—phone calls, word-of-mouth, and a school bell. 08:40 — flood crosses into Nepal. 08:50 — upstream gauge destroyed. 08:56 — downstream gauge operator calls it in. 09:16 — 679,295 SMS alerts sent. 11:20 — first automatic warning. At 09:10, school principal Rajendra Dawadi was teaching when his accountant warned him by phone about the flood. Dawadi rang the bell and evacuated about 900 students and 16 staff. Fourteen minutes later, the flood destroyed the school. Nepal's official system sent 679,295 SMS alerts at 09:16. But the school was already being evacuated through a human warning network. The event was a cascading mountain hazard: glacial rock collapse, mass flow, then a devastating flood debris cascade. That matters because a warning system designed around conventional river flooding can miss the hazard entirely or lose its sensors before they trigger. This is not simply a story about missing technology. It is a story about systems failing at the moment the hazard becomes most dangerous. Nepal's flood warning thresholds were based on water level. But a rapidly rising, debris-laden cascade can destroy the gauge before the water reaches the threshold. A sensor that sits in the path of the hazard can become the first casualty of the warning system. Nepal is now planning a new architecture: seismic sensors, cameras and satellite/VSAT connectivity, plus better real-time information from upstream in China. The proposed system is designed to withstand sensor and communication failures. But there is another failure mode: The warning reaches people, but does not tell them what to do. Reporting from Nepal found that the official SMS alert was late, in Nepali only, and lacked actionable, location-specific instructions. Meanwhile, community networks warned people directly. The community warning worked because it was trusted, specific and connected to someone who could act. It also exposed a critical communication gap. People outside those local networks—including tourists and construction workers—could be left out. They may not have the relatives, neighbors, local language, WhatsApp groups or institutional relationships that made the informal warning chain work. The real metric is not: “Can the system send an alert?” It is: “Can people still receive and act on a warning when the system is failing or responding to a cascading event?” The Nepal tragedy points to the need for investing in a few low-cost system-wide decisions and actions with immediate effect. For national agencies: • Trigger on rate of rise, not only absolute water level. • Recalibrate rating curves after major channel changes. • Build independent communication paths. • Establish named duty officers and alert-to-action protocols. • Define what happens when the automated system goes offline. For local authorities and operators: • Assume sensors will fail. Build redundancy. • Pre-authorize action when decision-makers cannot be reached. • Publish evacuation destinations and routes, not just alerts. • Drill through schools, hospitals and communities. • Build independent egress and refuge capacity for tunnels and other high-risk worksites. • Treat community networks as communication infrastructure—not as an informal backup to the “real” system. They are part of the system. For funders: Stop measuring success by equipment installed. Instead, report on installed vs operational and tested. And let's fund the operational life of the system. Responsible capital commitments must come with years of: • maintenance • calibration • communications • spare parts • staffing & training • drills/exercises • independent testing An early warning system is not simply live when the sensor is installed. It has to work at 3 a.m., during a power failure, after a landslide, when the network is down and the sensor has disappeared. Finally, early warnings must be treated as a transboundary issue, especially for fragile mountain ecosystems experiencing cascading hazards. #EarlyWarningsForAll needs a cross-border data-sharing architecture for cascading cryosphere hazards. This requires stress-testing the entire warning chain across cascading scenarios: Assume the sensor fails. Assume the power fails. Assume the network fails. Assume the decision-maker cannot be reached. Then measure whether people can still be warned and know where to go for safety in a complex cascading event. The lesson from Nepal is not that we need more flood gauges. It is that a multi-hazard warning system must detect the cascade upstream, survive the cascade, and still deliver an actionable warning downstream, with people at the center of its design and delivery. #EarlyWarningsForAll #EW4All #DisasterRiskReduction