Comment by blindriver

The same way you validate scores given by NLPs I assume. You run various tests and look at the results and see if they match what you would expect.