Imagine mode uses world-model-guided planning: candidate move, opponent reply, follow-up, then a learned energy score chooses the future it prefers. This energy-rollout world-model player consistently outperforms the direct policy player in head-to-head benchmarks.