We have violations after CI since we do early stopping - satisfying them till end can sometimes hurt overall

Thank you for your detailed comments. We will make all clarifications below in the next version. We note that our formulation in Sec 3.2 can handle any Learning the constraints automatically is a direction for future work. There are important differences compared to Diligenti et al. We said that Diligenti and Mehta are task specific since they only experiment on a single task. Jin et al. assume arbitrary non-convex non-concave form for a twice We will add this important theorem in the paper.