Every alert HazardNet has published, district by district, with the evidence and the policy behind it. A watch is the highest level the engine can publish on its own; anything above a watch waits for a named duty officer to approve it. The list is rendered from live data, so this page also works from the snapshot the deployment carries when the API cannot be reached.
Four levels are published: no alert, watch, warning and severe. The level comes from the alert policy in force, not from the model alone. A watch is issued when the model score crosses the watch threshold or when its severity band does; a warning additionally requires a calibrated probability, and no calibration map is fitted for the current model version, so warnings are not currently issuable from model evidence.
Each alert states the hazard class, the district, the 7- or 15-day horizon, the forecast issue time, the data cutoff and the model version it came from.
Every alert carries a model severity, an independently computed physics severity and the divergence between the two tracks. When the tracks disagree beyond the policy threshold the alert is flagged rather than silently averaged.
The model score is a softmax output used as a relative priority signal. It is not a probability that the hazard will occur, and no calibration map is fitted yet — the alert page says this next to every number.
The engine publishes at or below the watch level automatically. A warning or severe alert requires a named duty officer, a review timestamp and the model version and data cutoff recorded against the decision, so a published alert can always be traced back to the run that produced it.
Alerts are also delivered by SMS and Telegram to subscribers who have opted in. Both channels carry the same disclaimer as this page, and a single SMS is capped at 70 Bengali characters per segment, which is why the message body is short and links back to the evidence card.
Loading the interactive HazardNet application…