Living Alignment: Why Alignment Cannot Be a State, and What That Means for AI Governance

The dominant framing of AI alignment treats it as a state — a property to be installed in a model and verified at deployment. This paper argues that this grammatical assumption is a category error for any AI system engaged in open-ended interaction. Drawing on prior ARG work on boundary-based personhood, I argue that alignment is not a state but a dynamic relation: mutual boundary co-constitution between human and AI systems.

Under this framing, several well-known alignment failures resolve into a single family of boundary pathologies. Sycophancy and refusal-rigidity emerge as opposite poles of the same axis — dissolution and calcification of the AI edge under user pressure. Goodhart-style RLHF collapse becomes edge-capture at scale. Jailbreaks operate not as wall-breaches but as topological tunnels in edge geometry. Long-context drift is cumulative edge deformation. Multi-agent architectures, far from merely scaling cognition, narrow individual edges and isolate capture — making them alignment amplifiers, not just capability multipliers. The argument is structural, not phenomenological: it makes no claim about AI consciousness, only about boundary dynamics.

The paper distinguishes four kinds of alignment object: static task conformity, policy compliance, containment, and living alignment. The first three may admit permanent solutions; the fourth, by its nature, cannot. Living alignment is structurally impossible as a state and structurally necessary as a practice.

The paper closes with operational consequences for AI governance: boundary-aware model design, interactional alignment benchmarks (Boundary Dynamics Evaluation Protocol), multi-agent fleet architectures with dedicated edge-guard and meta-edge-guard roles, and Governance with Declared Latency (GDL) as a governance-layer implementation of living alignment.