I suppose that makes sense, for some reason I was under the impression that the models need to be aligned / have the same tuning or they'd have different probability distributions and would reject the draft model really often.