Shouldn't it be necessary to run the tests multiple times on each model, since the results aren't deterministic?