This problem, where a model fails on unseen (out-of-distribution) data, is best solved by designing a system that uses model uncertainty and user feedback to mine and label these "hard examples."