Latent-belief RL on the passive BayesClue detective game. Base: Qwen3.5-4B. Belief is read from forced-choice probe logits (never verbalized). - Merge the LoRA with credalverl08/remergesftqwen35.py (the model.layers.→model.languagemodel.layers. remap); the stock verl.modelmerger writes a base copy. - Reward R = α·(−CE(pH,qH)) + (1−α)·means(−CE(pR,qR)), α=0.6. The −1.225 reward plateau IS the optimum −H(p), not a truncation artifact.