Due to computational constraints, running global climate models (GCMs) for more than a few years requires a spatial grid (≳50 km) too coarse to resolve two key atmospheric physical processes: cumulonimbus convection and airflow over orography, coastlines and other land-surface heterogeneities. The subgrid variability of these processes are approximately represented in GCMs via expert-designed physical parameterizations. Subjec tive choices made within these parameterizations contribute significantly to uncertainty in GCM predictions of precipitation, cloud cover, etc.