Using average of the posterior draws for posterior predictive checking?

Maybe I am misreading your post but it sounds like posterior_predict vs posterior_epred (Expected Value of Posterior vs. Posterior of Expected Value with epred - #2 by scholz).