Can you pick some good tests ("good" meaning they involve edge cases) and then just run each 100 times and see if the percentages come out OK. This seems more reliable than randomly generating 100 tests and running each of them.