Local causal-LM view of the model owned by a running TRLTrainer session.

It exists so Imp's ordinary program rollouts and the subsequent TRL update use the same pinned policy process. It cannot start a worker or download a model on its own. Training LMs sample by default; artifact deployments are constructed in greedy mode. A capable single-field enum adapter may bind an exact finite choice set during greedy deployment, which the causal policy scores and selects using its own logits rather than post-hoc output repair. Sampled training rejects that constraint because a choice-normalized policy requires a different objective than TRL's ordinary token-policy GRPO.