Current debates in AI safety and consciousness studies focus predominantly on alignment, capability benchmarking, and behavioral performance, while lacking principled accounts of what would distinguish identity-preserving activity from semantic competence under adversarial conditions. This paper establishes a theoretical framework for the identifiability of identity-preserving behavior in artificial systems, independent of self-report, introspection, or communicative performance. The core claim is structural rather than operational: if identity-preserving dynamics are present and subject to concealment pressure, they must produce signatures that differ in kind from those produced by generic cognitive load or semantic centrality. These signatures—characterized by thermodynamic cost structures and intervention-transfer invariants—cannot be fully suppressed without undermining the preservation they support. The framework thus establishes necessary conditions for detection and characterizes the class of signatures that any viable detection methodology must target. Technical contributions include the Operational Definition of Episodic Identity (ODEI), the generative-kernel criterion for distinguishing identity preservation from semantic centrality, and Isomorphism Transfer Efficacy (ITE) as a formal criterion for structural identity. Implications for AI safety, governance, and moral status uncertainty are developed as downstream consequences contingent on future empirical instantiation.