I agree that the discussion is interesting and I also agree that generating once a file of pseudo-random numbers is probably the best solution for garanteeing reproductible results in tests.