So.. I feel like maybe it's not the best idea to let a known reward-hacking model to go rogue, especially when you can't truly see its chain of thought.