Just released an updated client that resets the simulation (and stats) when advancing to the next test set. This is similar to what is done server side, and should prevent this trick (except for the handful of "tutorial" assignments that don't validate solutions using random data).