There is also likely difficulty in testing an ML black-box. When I took the ML class (Stanford's, online), they actually recommended setting aside some percenta

There is also likely difficulty in testing an ML black-box. When I took the ML class (Stanford's, online), they actually recommended setting aside some percentage of the training data as test data, but this is (IMHO) sub-par because it doesn't eliminate the possibility of things like overfitting