Improving dynamical climate predictions with machine learning: insights from a twin experiment framework | Synapse