Reviews: AIDE: An algorithm for measuring the accuracy of probabilistic inference algorithms

I do still have some concerns regarding applicability, since a gold standard "ground truth" of inference is required. I do appreciate the situation described in the feedback, where one is trying to decide on some approximate algorithm to "deploy" in the wild, is actually fairly common. That is, the sort of setting where very slow MCMC can be run for a long time on training data, but on new data, where there is e.g. a real-time requirement, a faster approximate inference algorithm will be used instead. The approach is based on constructing an estimator of the "symmetric" KL divergence (i.e., the sum of the forward and reverse KL) between an approximation to the target distribution and a representation of the "true" exact target distribution. The overall approach considered is interesting, and for the most part clearly presented.