In this work , Fabrizio Falasca and Laure Zanna introduce a flexible framework that combines response theory and score matching to eliminate spurious, noncausal dependencies in reduced-order neural emulators of turbulent systems. Using the stochastic Charney–DeVore model as a proof of concept for low-frequency atmospheric variability, they demonstrate that enforcing causal constraints significantly improves emulator responses to both weak and strong external forcings, even when trained solely on unforced data. The framework is broadly applicable to complex turbulent systems and can be seamlessly integrated into standard neural network architectures, offering a principled path toward more reliable climate emulators.