“The use of specialist models in post-training is organizationally scalable because training a specialist with SFT and RL on a single domain is a well-understood and parallelizable task.”