In statistical accounts of conditioning, subjects' behavior should reflect their prior assumptions about reinforcer contingency. In particular, behavior should reveal their generative model of stimulus and reinforcer delivery. I discuss two examples. First, attentional and surprise effects in conditioning may reveal subjects' assumptions about how cues compete or cooperate in mixture generative models to predict reinforcement. Second, I discuss how subjects' inferences about generative model structure may explain additional cue-combination phenomena in so-called "configural" conditioning experiments.