Record
Every call, and what followed
This page lists each forecast we served for a 1 km cell. Beside each call it shows what a satellite pass, a verified report or a news report later showed for the same cell. A call that was wrong appears here in the same size and weight as a call that was right.
- A satellite passes every 6 to 12 days. An event between passes is unobserved, not disproved.
- Early entries come from the baseline method, not from a trained model, and the record says so.
- Nothing is removed from this record.
This record is not complete. From 14 October 2026 the Delhi forecast covers about 2,900 cells a week instead of 18. Keeping one entry per call for a forecast that size would add about 6 MB to this file every week, so those calls are not recorded until the record has a smaller format. The record stops at its newest call.
The file is data/record/calls.json. It only grows. Each entry
carries a hash of the entry before it, so a removed or edited entry breaks the chain.
Summary
How far ahead our skill holds
For each lead day, how much better our judged calls scored than the base rate alone (the Brier score, lower is better). The forecast range in the map panel is grey beyond the last lead day where the whole interval is above zero. This is recomputed from the record each time the record is.
- our forecast, skill established (dot, with its 95% interval)
- our forecast, skill not established
- climatology, the base rate (the 0 line)
- lead days where the forecast range is grey
Skill is measured against climatology on the record so far, which is thin early on.
A greyed horizon is not a claim that nothing happens there. It is a statement that we cannot show we beat the base rate.
The file is data/record/skill.json.
Map
Each pin is the centre of one 1 km cell. The shape and the colour both give the result, so the map does not depend on colour alone.
Forest-loss detections
Every forest-loss detection we serve is in this record on the same terms as a flood call. Each is checked against Hansen, the annual forest-loss product, once Hansen has published the year it falls in. Every detection stays a candidate.
- A detection awaiting Hansen is not a failure. Counting it as one would understate the detector.
Confirmation rate by month of detection
Each row is the detections whose backscatter step was first seen in that month. A month with nothing checked yet shows no rate, not zero.
The file is data/record/forest.json.
It only grows, with the same hash chain as the flood calls.
Against official warnings
IMD and the state agencies are the official warning authority. We are not. This page does not yet say how often a call of ours disagreed with an official warning. Showing 0 would be wrong, because 0 would read as "never", when in fact it cannot be worked out. That needs a dated copy of the official warnings (from India's NDMA Sachet feed) taken while our calls were live, and no such copy exists for the calls above. We have set up a record of those warnings, taken every 12 hours from the day its first run completes, so that the comparison can be made later. It can only cover calls issued after that day, and it is a sample taken twelve hours apart, so a short warning between two samples will not be in it.
All calls
Every call is listed, oldest first. There is no filter that hides a result.