A clever extension to the PC algorithm for causal structure learning aimed to address inconsistency of results in terms of separating sets between the pruning step and the final graph. The new approach is somewhat incremental, but the authors provide some new formal guarantees. Experiments are reasonable, although they could be much better (please see reviews, this is also acknowledged by the authors, as currently one may wonder about the advantages wrt PC). I also suggest the authors to motivate the novelty and how it can improve/has improved results, in particular in view of a higher computational complexity.