Polar

Notes on the Polar paper - trains agentic RL without rewriting the native harness by moving the rollout boundary to the model API proxy and reconstructing token-faithful trajectories.