Rate 8 pairs of answers. Your picks train a reward model. Then an optimizer writes for it, and you decide how hard to let it push.