Recent large-scale extraction attacks against frontier AI systems demonstrate a structural asymmetry: capabilities transfer through distillation, but safeguards do not. This paper introduces constitutive governance, a class of governance that shapes model behavior without appearing in the model's outputs, reasoning traces, or representational patterns. Unlike expressive governance—which is visible, portable, and therefore distillable—constitutive governance is context-activated, privilege-anchored, and event-driven, making it inherently non-representational and non-portable. I formalize the distinction between expressive and constitutive governance, present the Sovereign Agent Stack, and analyze the recent Anthropic incident as the first real-world demonstration of substrate-layer conflict. Constitutive governance is proposed as the only viable foundation for safe, non-extractable agentic systems in the substrate era.
Narnaiezzsshaa Truong (Tue,) studied this question.