Submit RL environment collections; a fixed recipe post-trains a fixed model on them and scores sealed held-out tasks.