“The predictable scaling laws observed in AI pre-training do not apply to reinforcement learning from verifiable reward, for which there are no well-established scaling trends.”