[bryanlarsen]

1000s of exceptional circumstances, but once added to training data, is solved forever.

[bcrl]

You say that with the air of confidence of someone that believes their test cases cover all possible circumstances. I know for a fact that none of the large, complex real world systems I've worked on have ever had complete test coverage.

[lores]

Only if they are discretely identifiable, in a wide range of conditions, in a short enough time, and do not conflict with other circumstances...