Deterministic AI Governance: An Executive Guide to Runtime Authorization and Spotting the Fake

Series Introduction to the AI Governance Executive Guide Series from FERZ, Inc.

AI systems in production already produce consequential actions: who is flagged, who is denied, who is escalated, who is restricted. Accountability for those actions attaches at deployment, not when a governance program is stood up. This guide is written for executives and directors who must decide whether an AI system is fit to touch consequential work, and who need a way to test vendor and internal claims without a technical background.

Governance is enforced at the authorization boundary: the point where an output becomes an effect-bearing action, and where evidence, not explanations, must exist.

What this guide provides

• A working definition of runtime authorization and the authorization boundary, including the three-verdict model of ALLOW, DENY, and ABSTAIN. The boundary is the runtime control that makes execution depend on a valid, action-bound ALLOW and its required authorization artifact across every execution path within the declared scope; placing a check before an action does not, by itself, establish it.
• Five tests that separate enforced governance from theater, each with the shape of a passing answer and the deflection patterns that indicate an answer is avoiding the question.
• A separate assessment question the five tests presuppose but do not settle: which conditions determine permission, which the system evaluates, and which remain assumptions. A system may enforce its verdict reliably while evaluating an inadequate condition.
• Five recurring moves by which observation is presented as control, including monitoring-as-governance, policy without runtime enforcement, and synthetic authority.
• A minimum bar of five deliverables to require in writing, with named owners, before deployment or purchase.
• A one-page intake script and a glossary of the terms used.

The five tests

• Stop. An authorized human can halt the system before an action produces an effect, through a control enforced at runtime.
• Ownership. Each class of consequential action is assigned to a named accountable role, and that assignment is recorded and enforced.
• Replay. The authorization verdict for each action can be independently reconstructed, under the declared replay mode, from the authorization artifact and its preserved bound materials: inputs, applicable policy and version state, authority chain, and proposed action.
• Escalation. When permission cannot be determined on the available grounds, or policy requires an authorized human resolution, the system returns ABSTAIN and execution remains blocked pending an authorized human override. The override is a separate authorization event with its own artifact. If evaluation cannot run, no verdict is issued and execution remains blocked.
• Provenance. The origin of the inputs grounding the verdict is established. Origin does not establish truth, and whether reliance on an input is permitted is determined under the applicable policy.

Screening versus conformance

The five questions screen a claim; an unanswered question leaves the corresponding requirement unestablished. Formal conformance is assessed against the Five Tests Standard and its conformance materials, which the guide cites by specification, bundle, and repository version. Provenance is normative in the standard, but machine-checkable conformance is deferred pending a defined method for binding input origin to the decision: a serious supplier should answer the Provenance question substantively, and no claim of mechanically verified 5TS Provenance conformance should be accepted under the current specification.

Scope

This is an introduction, not a procurement standard and not a technical specification. It describes requirements, failure modes, and evaluation methods. It does not describe the capabilities of any named product, and no vendor is identified or assessed. Scenarios are illustrative composites. Readers who need procurement depth should proceed to The Enterprise AI Governance Buyer's Guide; readers who need controls expressed in a sector's native language should proceed to the corresponding volume in this series.

Related publications

• Five Tests Standard (5TS), the published standard defining the five tests and their conformance bundle: 10.5281/zenodo.21040295
• The Enterprise AI Governance Buyer's Guide: 10.5281/zenodo.18002693
• Authorization Boundary Integrity Model: 10.5281/zenodo.20929115
• Deterministic Governance Is Multi-Dimensional, on the assurance functions that must compose: 10.5281/zenodo.18902166
• What Deterministic Authorization Requires, on distinguishing authorization depth and assessing incomplete vendor answers: 10.5281/zenodo.22749462

Version

Version 2.1, September 2026. This edition aligns the guide with the framework published in What Deterministic Authorization Requires: it corrects the definition of the authorization boundary, completes the treatment of ABSTAIN and of evaluation that cannot run, separates the executive questions from formal conformance, defines the authorization artifact as an evidence class without product-status claims, tightens the reconstruction requirement, and adds the assessment question on whether the evaluation addressed the right conditions. Length, executive questions, scenarios, and tone are unchanged. It supersedes Version 2.0 (August 2026, 10.5281/zenodo.21864147); the earlier 2025 edition, written before the Five Tests Standard was published, is not deposited.

Published by FERZ, Inc. under the Creative Commons Attribution 4.0 International License. The authoritative version is maintained at ferz.ai.