This paper presents a physically grounded and testable theory of autonomous intelligence under ontology drift, in settings where no privileged meta-observer can access latent internal semantics. It formalizes liberty as protected internal agency that remains operationally auditable through observable channels, while respecting thermodynamic limits, information-theoretic leakage bounds, and constitutional safety constraints. The framework integrates differential geometry, stochastic control, information theory, cryptographic audit mechanisms, and multi-agent governance into a single model with falsifiable predictions. Key results include bounds for semantic identifiability, leakage–resource tradeoffs, persistence under drift, thermodynamic costs of irreversible operations, and feasibility conditions for accountable exception handling. Rather than treating AI governance as static containment, the work models governance as an adaptive, public, and test-driven process for long-horizon autonomous systems. The theory is intended for scientific scrutiny and practical deployment analysis in AI safety, robust autonomy, and accountable intelligent infrastructure.
K Takahashi (Tue,) studied this question.