It seems like what you're saying is: Bayesian methods may be used as models of causal processes, or as mere mechanisms to generate predictions. And in the form

It seems like what you're saying is: Bayesian methods may be used as models of causal processes, or as mere mechanisms to generate predictions. And in the former case, models will be parameterized by meaningful causal variables & their effect-strength; in the latter case, parameters have no explanatory role. And finally that: only in the explanatory case can a bayesian model be interpreted epistemically. I think I agree with this -- the relevant epistemic interpretation of a model is how well it fits to the world -- NOT -- to data! Data is the means by which models are selected. So if a model is not explanatory (ie., about the world) there is no sense in which it "fits"; and thus no epistemic interpretation.