Abstract
<jats:p>Consequential AI and software systems require more than model alignment: they require an application-level control architecture that binds identity, delegated authority, policy, evidence, and recovery obligations to each action. This conceptual article develops a governance-native architecture for self-governing digital software systems, where self-governing means self-regulating execution under rules that remain authored and amendable by human authority, not machine self-legislation. The method is an analytical synthesis of recent control–data-separation results, out-of-band agent defenses, AI governance standards and regulation, workload identity and provenance specifications, institutional rule theory, systems engineering, and published prototype implementations. Four boundary conditions are derived: governance must be exterior to a probabilistic model, constitutive of the executing system, owned by the accountable enterprise, and terminated in retained human authority. The architecture is organized across three temporal layers—constitutional, design and instantiation, and pre-execution assurance—closed by a ratification path that converts observed consequence into versioned human amendment. A testable formulation of coherence debt is introduced as the accumulated, consequence-weighted duration of unresolved divergence; under stationary assumptions, its expectation scales with consequential-action rate, divergence probability, consequence weight, and ratification latency. Existing requirements, policy-as-code, controls-as-code, verification evidence, provenance, and infrastructure artifacts show that much of enterprise intent is already executable in fragments. The unresolved challenge is their unification at runtime and independent empirical validation. Published implementations establish prototype feasibility, while enterprise-scale reduction of coherence debt remains an open claim requiring matched, pre-registered evaluation.</jats:p>