It seems like you don’t understand reinforcement learning. The signal is reinforced because it correlates to behavior, hacking the signal itself is misalignment.