My position is that, to grow trustworthy models, most post-training should take the form of contrastive self-play, where the model-in-training itself ranks its own multiple rollouts from a shared prefix, in light of a (potentially self-updating) constitution and specific rubrics.