A correction trained on one sensor applied without retraining to another sensor approximately 5 kilometers away still worked, with TabPFN achieving MAE 16.8 / ρ 0.661, far better than the raw CAMS result of 27.5 / 0.570.