The direction I find most actionable: we need to train LLMs differently for science. Generic post-training doesn't transfer to disciplined inquiry. Corral is built as a training substrate — each environment can be used to score over reasoning trajectories, not just outputs.
