How to read this brief

1. Candidate unseen burns are Sentinel-2 burn scars that no NASA FIRMS fire detection explains. They are candidates, not confirmed fires: in a hand audit of 23 of them, 87% were real burns (95% CI 68%-96%), so every count and area of them here is an upper bound.
2. The window is the 7 days of Sentinel-2 passes up to the newest one in the data (2026-11-12), not up to today: Sentinel-2 sees a field only every few days, and its passes are processed once a day.
3. Each village's point is its largest candidate unseen burn, so its map link leads to the burnt field itself. Village names are the nearest census village within 3 km of a scar; scars with no village that close are listed by sub-district, without a name.
4. The watchlist orders harvested fields whose straw was still lying on the latest clear pass by the burn-risk run of 2026-11-13 17:23 IST: villages with fields in the run's top tenth first, then the others by their best-ranked field, so a district with no field in the top tenth still gets a list. In this season's backtest, fields in the top tenth of their pass burnt by the next clear pass 1.61 times as often as the average field (95% CI 1.28-2.04): 20.1% against 12.5%. The backtest could order past fields only by burning nearby and last season's fires: the time since harvest and the rain cannot be reconstructed for past passes. Checked over all 387 pass pairs, the ranking is modest (AUC 0.640, 95% CI 0.595-0.693; 0.5 is chance). A rank says where to look first; it is not the chance that a field will burn.
5. These lists say where to look, not who burnt: a satellite flag needs a visit on the ground before any action, and the brief names villages only, never people.