Hybrid climate modeling has emerged as an effective way to reduce the computational costs associated with cloud-resolving models while retaining their accuracy. The approach retains physics-based models to simulate large-scale atmospheric dynamics, while harnessing deep learning to emulate cloud and convection processes that are too small to be resolved directly. In practice, however, many hybrid AI-physics models are unreliable. When simulations extend over months or years, small errors can accumulate and cause the model to become unstable.