It is important to remember that you are changing the code (and probably the tests) at the same time, so your coverage data is getting stale. So you can only "mostly" predict which tests to run, which is somewhat useful but can also make these cut down runs misleading.