The problem of quantifying uncertainties in climate is scientifically ill-posed. One of the largest hurdles is the absence of a performance metric that would be predictive of the scatter in model predictions that come about from feedbacks that emerge only after the system has been hit with a sizeable forcing. Model biases are large enough and processes controlling cloud feedbacks are subtitle enough within simulations of modern climate such that relatively small differences in how a model is tuned make a big difference in how a model responds to greenhouse gas forcing. I envision a two-stage process that within a Bayesian framework sequentially uses general and specific performance metrics for selecting a set of models and outcomes that represent observational constraints on predicted quantities of interest. The first stage would use a general performance metric (i.e. a metric that uses observations to estimate the likelihood of the observations given a particular configuration of the model) to estimate a distribution of alternate models that are consistent with the usual Top 10 observational constraints that are commonly used as tuning targets. Our project team can report progress in the design of a performance metric based on multivariate normal statistics that is able to represent observed space and field dependencies using Gaussian Markov Random Fields. The point of such a metric is to down-weight the selection of models that mimic observed climate for the wrong reasons. A limitation of the first stage selection process is that not all model biases are important to predictions of quantities of interest. A second stage may be needed that places additional observational constraints on the selection of models that represent uncertainties in particular processes important for predictions. As an example of this, the project team has identified structures within CAM that are associated its cloud feedbacks. We then can use these model structures against observations to generate a probabilistic assessment of climate change and its uncertainty. The second stage involves the development of a statistical model that uses observations to constrain model feedbacks based on what we have learned in the first stage experiments.