The most extensive and most frequently used dataset for climate model hindcasts and projections comes from the internationally organized CMIP process (notably, CMIP5 and CMIP6). These data have been used to evaluate models against past data and constrain projections going forward. However, the models participating in these exercises have many assumptions in common (related to ocean resolution, or interaction for ice sheets, for instance), and the parametric uncertainty and even the forcing uncertainty is woefully under-sampled. Credibility in model projections is built through success in hindcasts, but there is also a danger in overfitting imperfect models to imperfect observations, hence the use of true out-of-sample paleo-climate tests can be very useful, although again, this is not done universally. I will discuss some of the new techniques/processes and results from the GISS suite of models that attempt to address at least some of these problems.