Symbolic Governance for Probabilistic Intelligence: Why Governable AI Requires Symbolic Structure Without Reverting to Symbolic AI
Annotatsiya
This publication presents a canonical architectural framework for governing probabilistic AI systems in regulated environments: runtime authorization infrastructure, the enforcement layer of AI governance. Modern AI systems, including large language models, autonomous agents, and tool-calling workflows, operate through statistical inference and cannot provide the deterministic guarantees required for authorization, auditability, and legal accountability. Existing approaches based on alignment, guardrails, human-in-the-loop review, or lineage tracking therefore fail to provide enforceable control over consequential actions. This is not only a practical observation: a formal impossibility result establishes that observability-based architectures cannot satisfy ex-ante authorization requirements, on the independent grounds of causal posteriority and independent-verifiability failure. The paper argues that governable AI requires a strict architectural separation between intelligence and governance. Neural systems generate proposals (recommendations, hypotheses, candidate actions), while authorization must occur through an independent, deterministic verification layer that is structurally independent of the governed system. This layer evaluates proposed actions against explicit policy constraints before execution and renders one of exactly three verdicts: ALLOW, DENY, or ABSTAIN, where ABSTAIN blocks execution pending authorized human override. Probabilistic compliance is rejected as a governance strategy; authorization is a deterministic, fail-closed decision rendered at a non-bypassable runtime boundary. The core contribution is the tamper-evident authorization artifact, implemented as a Proof-Carrying Decision (PCD) under the Five Tests Standard (5TS): a structured artifact that carries all evidence necessary to independently reconstruct the verdict rendered at execution time. Authorization artifacts enable replayable verification, allowing auditors, regulators, and investigators to mechanically re-verify authorization decisions without relying on post-hoc narratives or probabilistic model behavior. The paper defines the deterministic envelope that enforces “no artifact, no action,” specifies decision compilation requirements, establishes a replay contract, and formalizes ABSTAIN resolution and bounded human discretion. The work clarifies the relationship between governance and provenance by distinguishing two senses of the term. Provenance as lineage tracking provides evidentiary inputs but cannot substitute for authorization: lineage is not authorization evidence; authorization evidence is produced by deterministic verification at the boundary. Provenance as a test, the fifth test of 5TS alongside Stop, Ownership, Replay, and Escalation, operates inside authorization: it requires established origin of the inputs grounding the verdict. It is origin, not truth. The dependency direction runs through the authorization boundary: input origin flows into evidence, evidence feeds constraint evaluation, and constraint evaluation produces the verdict that gates the action. This publication is not a return to symbolic AI as a theory of intelligence. Instead, it demonstrates that symbolic structure (determinism, inspectability, verifiability) is essential for governance even as probabilistic models remain indispensable for intelligence. The framework is model-agnostic, vendor-neutral, and applicable across regulated domains including healthcare, finance, and public-sector systems. This document is intended as a stable, citable technical publication and canonical reference for deterministic AI governance architectures. Version 2.0.0 (July 2026) supersedes version 1.2.0 and aligns the paper with the Five Tests Standard, the three-verdict space, and the authorization-artifact terminology. The authoritative version is maintained by FERZ, Inc. at ferz.ai.
Hali tarjima qilinmagan