The distinction between a model's ability to produce novel outputs versus its ability to evaluate existing candidates.