In the August 2026 north Odisha replay we ran the unmodified rule engine on the forecast exactly as it was issued that week. At forty-eight hours it caught all six location onsets. At twenty-four hours, four of six. At seventy-two hours, one of six.
A seventy-two hour warning is a much better thing to sell than a forty-eight hour one. It is most of an extra day. It is the difference between moving livestock at leisure and moving them at night.
One of six is not a lead time. It is a coincidence with good manners.
So seventy-two hours does not appear anywhere on this site. Not in the deck, not on the product pages, not in a footnote with an asterisk. The number we publish is forty-eight hours, labelled BACKTEST, because it was measured on history and has never been quoted as live accuracy.
The same rule cost us more than that
Through the recession that followed that same event, fourteen of eighteen location-days scored LOW while 13.44 lakh people were still displaced. The rain had stopped; the water had not. The engine reads rainfall, not flood state, and the active-flood hold we had built for exactly this case keys off river level, which made it dead code for ungauged locations.
That is the single largest known gap in the system and it is written up on the product page it belongs to, with a MISS label, rather than in a place where nobody would look for it.
Backtest and Miss. Both figures appear on the Flood Ops page and in docs/ModelEarth_NorthOdisha_Aug2026_Replay.md.