This paper develops a generative emulator for monthly climate fields used in impact assessments, including temperature, precipitation, humidity, and wind. A score-based diffusion model on a spherical mesh learns their joint distribution from Earth system model output. The evaluation checks spatial and cross-variable relationships, distribution tails, and the emergence of forced changes across three climate models. The results reproduce many key statistics but also expose failures when seasonal behavior changes strongly between climate regimes. Readers get both an approach to generating inexpensive climate ensembles and a detailed set of diagnostics for judging whether an emulator is useful for impacts research.