Exactly. And presumably had a more sophisticated harness around the model, longer reasoning chains, best of N, self judging, etc