BoundaryG: A Decision-Centric Runtime Governance OS for Enterprise AI

Abstract Autonomous AI systems increasingly participate in enterprise decision‑making, yet existing governance approaches remain centered on pre‑execution controls (policies, risk assessments) and post‑execution mechanisms (audits, logging, output filtering). These mechanisms provide no structural guarantees at the moment where decisions become actions. Recent work in runtime governance—such as decision modeling (DMN [1]), integrated runtime telemetry and containment (MI9 [2]), cryptographic policy enforcement (Aegis [3]), kernel‑level governed execution (Cosmocrat [4]), path‑based policy formalization (Kaptein et al. [5]), formal behavioral contracts (ABC [6]), and machine‑readable governance standards (Policy Cards [7])—addresses isolated components of runtime behavior, but lacks an integrated operational structure capable of ensuring accountable, predictable, and compliant execution within business processes. This paper identifies the structural gap created by this fragmentation and argues that enterprise‑scale autonomy requires a unified governance layer that operates at runtime. We introduce BoundaryG, a Decision‑Centric Runtime Governance OS, which provides a minimal but coherent set of architectural primitives: Decision Unit Architecture, Responsibility Boundary Architecture, Fail‑Closed Execution Boundary, Drift Boundary Control, Escalation Topology, Regulatory‑to‑Process Synchronization, and a Process OS that binds these elements into a unified operational layer above workflow engines. BoundaryG does not prescribe model‑level accuracy, legal interpretation, or kernel‑level sandboxing; instead, it defines the structural conditions under which autonomous systems may act within enterprise workflows. By formalizing these primitives, BoundaryG offers a previously unaddressed integration of runtime governance primitives, filling the conceptual space between policy governance and technical enforcement.