Deterministic Engines and Deterministic Governance: Why Correctness Is Not Permission

This paper clarifies a growing source of confusion in contemporary AI discourse: the conflation of deterministic computation with authorized decision-making. As physics-based simulators, constrained optimization systems, and other deterministic engines gain prominence, claims increasingly imply that deterministic outputs are sufficient for safe, compliant, or permissible action. This paper explains why that assumption is incorrect. The paper distinguishes deterministic engines, which generate reproducible outputs under declared assumptions, from deterministic governance, which authorizes actions by binding decisions to policy, evidence, semantic definitions, authority, and time in a replayable and independently verifiable manner. It argues that determinism at the model or simulation layer, while valuable, does not resolve questions of compliance, liability, or accountability. Through a systems-level analysis, the paper examines why methods used to achieve determinism in engines (e.g., constraints, solvers, proofs, and canonical execution) do not automatically transfer to governance, which must operate at the authorization boundary where legal and organizational responsibility attaches. It introduces the concept of proof-carrying authorization artifacts and illustrates the minimal elements required for third-party verification. The paper is vendor-neutral and intended for regulators, enterprises, investors, and technical leaders operating in high-stakes or regulated environments. It positions deterministic governance as a distinct architectural layer, model-agnostic by design, necessary to ensure that decisions informed by both deterministic and probabilistic systems can be independently audited, replayed, and defended under external scrutiny. Version update: Added Section 7.6 (“Physical AI: When Outputs Become Effects”) to explicitly address governance requirements for AI systems that drive real-world actuation, clarifying how the authorization boundary applies to safety-critical and embodied systems.

Перевод пока недоступен