[swordsmith]

Seems like he thinks RLVR == learning from binary reward for the whole chain, completely discounting techniques to provide denser rewards like process reward supervision?