Climate emulators often pair a neural network with a hard constraint that corrects each predicted state to satisfy physical laws. A new arXiv preprint (2610.23561) argues that this creates a blind spot in evaluation: if researchers look only at the final corrected forecast, they cannot separate the network's own skill from the effect of the correction.
The authors say this conflation is a problem for understanding where an emulator's accuracy comes from. The abstract announces a new approach to address it, but the provided text cuts off mid-sentence, so the details of the proposed method are not available in this source.
Because the source is truncated, this summary covers only the problem statement. Readers should consult the full paper for the method and results.