A Substrate-Layer Governance Architecture for Agentic Systems: Constitutive Substrates, Detection Grammar, and Failure Geometry

Frontier AI systems exhibit a structural asymmetry: capabilities transfer through distillation and extraction, but safeguards do not. This asymmetry enables extraction-attackpathways in which model capabilities are recoverable while governance constraints vanish. This paper presents a unified substrate-layer governance architecture that resolves this asymmetry. I introduce constitutive governance as a non-distillable substrate for agentic systems, map detection-engineering grammar to governance-physics semantics, and formalize the failure geometry that emerges when bounded representational channels encounter unbounded state. Together, these components provide a practical map for engineering teams to instantiate governance architectures tailored to their systems and threat models, including extraction-attack scenarios similar to those observed in recent frontier-model incidents. The result is a discipline-level framework for building agentic systems whose safeguards persist across training regimes, operational contexts, and adversarial transformations.