“Claude's constitution does not mean the model "believes" in its principles; rather, the model is trained to prioritize outputs that align with those principles.”