Decision Admissibility & Governance Gates: Runtime Governance for Agentic AI

Decision Admissibility & Governance Gates: Runtime Governance for Agentic AI

This working paper introduces Decision Admissibility and Governance Gates as core constructs within Decision Engineering Science™ (DES), an emerging research program developed by the REGEN AI Institute for the systematic design, evaluation, governance, and control of consequential decisions in AI-enabled and agentic systems.

As artificial intelligence evolves from systems that generate predictions and recommendations toward systems capable of planning, tool use, delegation, resource access, and autonomous execution, a critical governance problem emerges: a technically valid or high-confidence decision is not necessarily a decision that should be permitted to become action.

Existing approaches to AI governance frequently concentrate on models, organizational controls, risk management, compliance processes, or post-hoc accountability. These remain necessary, but they do not fully resolve the runtime question faced by increasingly autonomous systems:

Under what conditions should a proposed AI-generated or AI-mediated decision be permitted to proceed to execution?

This paper develops a formal framework for addressing that question at the level of the Decision Object.

Decision Admissibility

Decision Admissibility is defined as a runtime governance determination concerning whether a proposed Decision Object satisfies the conditions required to progress toward execution under a particular governance environment.

The paper distinguishes four principal admissibility states:

ADMISSIBLE — RESTRICTED — ESCALATION REQUIRED — PROHIBITED

This distinction separates several concepts that are frequently conflated in AI systems:

decision validity ≠ decision quality ≠ decision admissibility ≠ authorization ≠ execution

A decision may therefore be technically coherent, supported by strong evidence, or generated with high model confidence while still being restricted, escalated, or prohibited because of constraints, risk, policy, authority, or other governance conditions.

Decision Admissibility Function

The paper introduces the Decision Admissibility Function (DAF) as a formal runtime governance operator:

DAF(Dᵢ, Γᵢ(t)) → sᵢ(t)

where the governance environment Γᵢ(t) captures the evidence, context, constraints, risk, authority, policy, and other conditions relevant to Decision Object Dᵢ at time t.

The framework explicitly adopts non-compensatory governance logic. Strong performance on one dimension cannot automatically compensate for failure of a mandatory governance condition. High confidence, expected utility, or evidence strength cannot manufacture authority or override a binding prohibition.

Governance Gate Architecture

To operationalize Decision Admissibility, the paper develops a modular Governance Gate Architecture positioned on the runtime path between decision formation and consequential action.

The proposed canonical architecture distinguishes:

• Evidence Gate — whether the decision is sufficiently evidenced for its consequence;

• Risk Gate — whether residual decision risk remains within the permitted envelope;

• Constraint Gate — whether mandatory boundaries and hard constraints are respected;

• Policy Gate — whether applicable legal, regulatory, institutional, contractual, and governance requirements are satisfied;

• Authority Gate — whether the relevant actor possesses legitimate authority to progress the decision;

• Authorization Gate — whether this specific Decision Object may cross the decision-to-action boundary under current conditions.

Gate outcomes are represented through:

PASS — CONDITIONAL PASS — ESCALATE — FAIL

and resolved through explicit precedence and enforcement semantics rather than an aggregate governance score.

From Governance Assessment to Runtime Enforcement

A central contribution of the paper is the distinction between governance observation and governance enforcement.

A gate becomes a governance mechanism only when its outcome changes what the system is legitimately permitted to do next.

The framework therefore connects gate evaluation to execution capability through explicit restriction contracts, escalation pathways, authorization checks, temporal validity, re-evaluation triggers, hard stops, and runtime enforcement.

The paper further develops the concept of Governance Integrity as an end-to-end property of the decision-to-action path.

Governance integrity requires that consequential execution remain traceable to valid gate evaluation, correct state resolution, applicable restrictions, valid authorization, temporal validity, and execution-path integrity.

This extends the analysis beyond the nominal execution path to include potential governance bypass through alternative tools, retries, APIs, delegated agents, cached permissions, indirect execution routes, and multi-agent architectures.

Runtime Governance for Agentic AI

The proposed architecture is particularly relevant to agentic AI systems, where decision formation and execution may occur within the same computational process and where human review may no longer exist between every model output and consequential action.

The paper therefore argues for a transition from predominantly model-centric governance toward complementary decision-level runtime governance.

The proposed control sequence can be summarized as:

Observe → Reason → Proposed Decision → Governance Gates → Decision Admissibility → Authorization → Governed Action

Within this architecture, autonomy is not treated as an unrestricted property of an AI agent. It exists within an explicitly governed decision space and terminates, narrows, or escalates when the system reaches the boundary of legitimate autonomous authority.

Contribution to Decision Engineering Science™

This working paper extends the emerging DES research program by providing a formal bridge between Decision Objects, decision evidence, governance conditions, admissibility states, authorization, and execution control.

Its principal proposed contributions include:

1. the formalization of Decision Admissibility as a distinct decision-level governance construct;

2. the Decision Admissibility Function (DAF) for runtime state determination;

3. a multidimensional representation of admissibility conditions and decision boundaries;

4. a modular Governance Gate Architecture;

5. explicit non-compensatory governance and enforcement precedence;

6. formal separation between admissibility, authority, authorization, and execution;

7. runtime mechanisms for restriction, escalation, prohibition, re-evaluation, and override governance;

8. the concept of Governance Integrity across the complete decision-to-action path;

9. formal properties and invariants that can support future implementation testing, runtime assurance, and formal verification;

10. a foundation for pre-action governance of autonomous and agentic AI systems.

The constructs presented in this working paper should be understood as a proposed conceptual and formal architecture. They are not presented as universally empirically validated governance mechanisms. The framework is intended to establish a basis for subsequent implementation, falsification, comparative evaluation, formal verification, domain-specific validation, and empirical research.

Research Program

This publication forms part of the:

REGEN AI Institute — Decision Engineering Science™ Foundations Series

and contributes to a broader research agenda examining the decision layer between AI inference and consequential action.

The central proposition of the paper is:

The ability of an AI system to produce a decision does not establish its permission to act.

Runtime governance therefore requires an explicit mechanism for determining not only whether a decision can be generated, but whether that decision is admissible, authorized, and governable into action.

Keywords: Decision Engineering Science; Decision Admissibility; Governance Gates; Agentic AI; AI Governance; Runtime Governance; Decision Governance; Decision Objects; Autonomous Agents; AI Safety; AI Assurance; Decision Authorization; Governance Integrity; Pre-Action Governance; Responsible AI; Decision Control