The paper initially received two strong reviews and one very negative review. I solicited the feedback from an additional reviewer who read the paper thoroughly and agreed with the first two reviewers. I believe this paper makes a valuable contribution to the study of generalization in modern machine learning settings. In particular, they appear to have isolated ways in which the learned classifier inherits microscopic structure from the data that causes it to misclassify a ghost data set distributed identically to the training data. The result is catastrophic for two-sided generalization error. I suspect this paper will generate much downstream insight.