Due to factors such as spatial discretization and the parameterization of certain processes, the presence of bias in models of the Earth's atmosphere is unavoidable. Whether we are selecting a model to explain past phenomenon, forecast weather patterns, or make inferences about the future, the target of any selection process is to minimize the discrepancies between model output and observations. Some discrepancies have a greater effect on the scatter of model predictions though. We exemplify this in the case of CO2 forced warming using multivariate empirical orthogonal functions (EOF), created using an ensemble of plausible parameter configurations of CAM3.1. When subjecting this ensemble to a doubling of atmospheric CO2, some EOFs exhibit significantly higher correlation than others with the resulting increase in mean global surface temperature. Therefore, there are discernible bias patterns that effect its predictive scatter. By targeting these patterns in the model evaluation process, it is plausible to use this information to constrain the resulting range of predictions. We take a first step towards showing this by creating a metric to evaluate model skill based on these EOFs and their correlation to a model's sensitivity to CO2 forcing. Using model output, for which we know the resulting temperature increase, as a surrogate for observations in this metric, the resulting distribution of skill scores indeed agreement in sensitivity to CO2 forcing. / text