Abstract This document defines a missing architectural layer in contemporary AI systems: the governance substrate required to preserve meaning, authority, provenance, and constraint over time. It establishes that instability in LLM systems is not incidental but structural, arising from the absence of governed state. Governed state is not stored, inferred, or approximate state. It is state constrained by explicit invariants, authority boundaries, provenance, and temporal coherence. This is not a critique of model capability. It is an analysis of why current systems cannot maintain admissible state in high-accountability environments, and what is required to make them deployable in such contexts. Cognitive Memoisation (CM-2) is introduced as a complete, normative architecture that enforces these constraints by validating admissible state entering inference and restoring required context when invariant violation occurs. This document is intended for readers evaluating the deployability of AI systems in domains where correctness, auditability, and authority must be preserved.