Governance / Comparisons

FERZ is not Human-in-the-Loop

FERZ is designed as a runtime AI-action authorization architecture: a non-bypassable, fail-closed boundary evaluates each proposed action against the applicable policy and governed state before execution and emits ALLOW, DENY, or ABSTAIN. The resulting verdict is represented in an independently reconstructable authorization artifact. Review-only human-in-the-loop is an arrangement in which a human reviews a flagged action and the determination is recorded as a review event. A human determination becomes material to authorization only when the runtime authorization boundary materially consumes it as authority-bound input to emit the separate action-bound verdict on which execution depends, and that verdict is represented in an independently reconstructable authorization artifact.

This comparison addresses a confusion that surfaces in regulatory and procurement conversations: because FERZ's architecture supports authorized human resolution of ABSTAIN verdicts, observers sometimes collapse it into a review-only human-in-the-loop architecture. That classification is wrong. The distinction is developed in The Override Asymmetry: Why ABSTAIN with Authorized Resolution Is Not Guardrails with Review-Only Human-in-the-Loop (Meyman, 2026; concept DOI 10.5281/zenodo.19772248, Version 2.2), which states the evaluation rule this page applies. The architectural-class threshold is stated in The Authorization Artifact Test (Meyman, 2026; concept DOI 10.5281/zenodo.20013582). Both arrangements involve humans. They involve them at different points, consume their input differently, and produce different evidence.

The comparator is review-only human-in-the-loop: an arrangement in which a human reviews an action that a guardrail or workflow has flagged, and the determination is consumed by the immediate workflow and recorded as a review event, but is not materially consumed as an authority-bound input by a runtime authorization boundary. Human review incorporated into a non-bypassable, fail-closed pre-execution authorization procedure is not review-only, and this comparison does not describe it. The test is the rule, not the label: an arrangement that calls itself human-in-the-loop is evaluated on what the boundary consumes and what execution depends on.

Where the categories diverge

The two arrangements diverge along three axes. The second is load-bearing; the other two follow from it.

Trigger meaning
FERZ

ABSTAIN is emitted by the runtime authorization boundary when the applicable policy does not resolve the proposed action, or when the policy reserves the action class for human decision. The question that invokes the human: does the applicable policy resolve this action?

Review-only human-in-the-loop

A workflow rule sends the action to human review. The rule may use a classifier, threshold, sample, action class, or other condition. The trigger establishes that review is requested; it does not establish what authorizes execution.

How the human's determination enters the architecture
FERZ

The human supplies authority-bound input concerning the held action. The boundary materially consumes that input and emits a separate resulting ALLOW or DENY verdict. Execution of the held action depends on that verdict, and the verdict with its authority chain is represented in the authorization artifact. The human supplies the input; the boundary emits the verdict.

Review-only human-in-the-loop

The determination is consumed by the immediate workflow. The flagged action proceeds or does not. A review event is recorded. No runtime authorization boundary is defined over the determination.

Record and release
FERZ

Fail-closed by design: ABSTAIN blocks execution unless and until authorized resolution produces a separate resulting action-bound verdict through the boundary; the original verdict does not convert. Each verdict is represented in an authorization artifact that binds the proposed action, the bound inputs, the applicable policy and version state, and the authority chain, so that the verdict is independently reconstructable under a declared replay mode.

Review-only human-in-the-loop

Workflow-dependent: if the reviewer is unavailable, the workflow may stall, fall back to an automated decision, or fail open. The record is evidence that a workflow determination occurred. It may be detailed and may cite policy; it need not bind the elements from which an authorization verdict could be reconstructed, because no boundary verdict was produced.

Why the distinction is structural

A human determination becomes material to authorization only when the runtime authorization boundary materially consumes it as authority-bound input to emit the separate action-bound verdict on which execution depends, and that verdict is represented in an independently reconstructable authorization artifact. Three conditions are packed into that rule: the input is materially consumed by the boundary; execution is unavailable without the resulting verdict; the verdict is represented in the artifact. Participation that fails any of the three may support authorization. It is not authorization.

ABSTAIN arises in two cases. In the first, the policy does not resolve the proposed action. In the second, the policy reserves the action class for human decision regardless of coverage, so that every action in the class returns to a human by design. Actions the policy resolves are decided without human involvement and produce policy-versioned authorization artifacts on their own. How many actions return ABSTAIN depends on policy maturity and on which classes the policy reserves. The volume is a measurement of the authorization state against the actions proposed, and each ABSTAIN is attributable in an artifact that records what was not resolved and under which policy.

The evidence differs in kind. A review record attributes a determination to a person: "reviewer R approved action A at time T." It describes who said yes. An authorization artifact represents the verdict the boundary emitted for the proposed action under the applicable policy and authority: "policy P-17 evaluated action A-921 in governed state G-558 and emitted ALLOW." When the boundary emitted ABSTAIN and an authorized human resolved it, the artifact records the resolution as a boundary event: "authority R supplied resolution input on held ABSTAIN V for action A-921; the boundary consumed it under policy P-17 and emitted ALLOW, represented with the authority chain." The original ABSTAIN does not convert. A separate resulting verdict exists, and execution depended on it.

Resolving a held action and amending the policy that governs later actions are separate governed acts. The first supplies authority-bound input for action-specific authorization. The second changes the policy governing later actions. Where a separately authorized policy-amending pathway exists, it may later consume the resulting authorization artifact. That later act is optional and is not what authorized the held action. A review record may likewise inform a later policy change; that later use does not make the earlier review event the authorization of the held action.

The structural claim follows. Review-only human-in-the-loop and runtime authorization are distinct architectural classes even though both place human judgment in the decision path. The two compose: human review incorporated into a fail-closed pre-execution authorization procedure, whose input the boundary consumes and on whose resulting verdict execution depends, is no longer review-only. Adding a human to a workflow does not produce authorization, and adding an authorization boundary does not remove the role of human judgment where policy reserves it.

What buyers often confuse

The common conflations place a human-review mechanism where an authorization boundary belongs. Each is evaluated by the rule above, not by its label.

Review queues branded as AI governance.Manual queues that route AI proposals to human reviewers, with dashboards labeled governance. The determination is consumed by the workflow; the arrangement produces no pre-execution verdict at a boundary and no artifact from which one could be reconstructed.
Approval features in agent frameworks.Framework features that require human approval for certain tool calls, presented as built-in governance. The framework intercepts and queues. The question is whether execution depends on a resulting boundary verdict that consumed the approval, or whether the approval is a workflow determination that some path can bypass.
"AI with oversight" positioned as governance.Products that combine LLM workflows with human oversight queues, marketed as enterprise AI governance. Oversight roles and documentation establish who is answerable. They do not establish that execution depended on a verdict that consumed the authorized person's input.
Sampling-based human review.Systems that route a sampled subset of AI actions to reviewers for spot-checking. The system samples observations; it does not authorize the population of actions against policy. Sampled review is a quality-control mechanism, not an authorization architecture.

Architectural diagnostic

Any arrangement that places a human in the decision path, whatever it is called, is evaluated against four questions. They are the questions The Override Asymmetry applies to guardrail-labeled "abstain," anchored to The Authorization Artifact Test (Meyman, 2026; concept DOI 10.5281/zenodo.20013582):

1
Verdict before execution

Is a verdict on the proposed action produced before the action executes?

2
Boundary properties

Is it produced at a non-bypassable, fail-closed runtime authorization boundary, so that absence of a verdict implies absence of permission?

3
Material consumption and execution dependency

Does the boundary materially consume the human's authority-bound input to emit a separate resulting action-bound verdict, and does execution of the held action depend on that verdict, such that no path in the declared scope, including retry, administrative, recovery, and human-resolution paths, releases the action without it?

4
Reconstructable artifact

Is the verdict represented in an authorization artifact from which an independent party can reconstruct it, under a declared replay mode, from the bound inputs, the applicable policy and version state, the authority chain, and the proposed action?

If all four answers are yes, the human's participation is within the authorization architecture, whatever the arrangement is called. If any answer is no, the arrangement may support authorization but does not itself constitute it. ABSTAIN is how this architecture keeps an unresolved case from collapsing into a refusal; it is not itself the test. FERZ's architecture is designed against these four conditions. Conformance to any of them is a matter of demonstration under a declared replay mode and a declared execution-path scope, not of description.

References

Meyman, E. (2026). The Override Asymmetry: Why ABSTAIN with Authorized Resolution Is Not Guardrails with Review-Only Human-in-the-Loop. Version 2.2, with companion Questions About Human Review and Authorized Resolution, Version 1.0. FERZ, Inc. https://doi.org/10.5281/zenodo.19772248 (concept DOI; version 2.2 DOI https://doi.org/10.5281/zenodo.22259718)
Meyman, E. (2026). The Authorization Artifact Test: Applying the Impossibility Result to Ex-Ante Authorization Requirements. Version 1.2. FERZ, Inc. https://doi.org/10.5281/zenodo.20013582 (concept DOI; version 1.2 DOI https://doi.org/10.5281/zenodo.22128176)
Meyman, E. (2026). The Authorization Non-Substitution Principle. Version 1.0. FERZ, Inc. https://doi.org/10.5281/zenodo.22017004 (concept DOI; version 1.0 DOI https://doi.org/10.5281/zenodo.22017005)
Meyman, E. (2026). Authority versus Authorization. Version 1.0. FERZ, Inc. https://doi.org/10.5281/zenodo.21341907 (concept DOI; version 1.0 DOI https://doi.org/10.5281/zenodo.21341908)
Meyman, E. (2026). The Authorization Boundary Integrity Model. Version 1.2. FERZ, Inc. https://doi.org/10.5281/zenodo.20929115 (concept DOI; version 1.2 DOI https://doi.org/10.5281/zenodo.22262364)
Meyman, E. (2026). Five Tests Standard (5TS). Specification v1.2.0. FERZ, Inc. https://doi.org/10.5281/zenodo.21040295 (concept DOI; specification v1.2.0 DOI https://doi.org/10.5281/zenodo.21040296)
Meyman, E. (2026). A Taxonomy of AI Governance Approaches: Distinguishing Visibility, Alignment, and Authorization. Version 1.7.1. FERZ, Inc. https://doi.org/10.5281/zenodo.18275969 (concept DOI; version 1.7.1 DOI https://doi.org/10.5281/zenodo.22037345)
Meyman, E. (2026). From Monitoring to Authorization: The Structural Shift Emerging in Agentic AI Governance. FERZ, Inc. https://doi.org/10.5281/zenodo.18743974 (concept DOI)
Meyman, E. (2026). On the Impossibility of Observability-Based Authorization. Version 1.4.0. FERZ, Inc. https://doi.org/10.5281/zenodo.19647542 (concept DOI)

Related doctrine and concepts

Frequently asked questions

FERZ has ABSTAIN with human resolution. Isn't that just human-in-the-loop?

No. In review-only human-in-the-loop, the human's determination is consumed by the workflow and recorded as a review event. Under ABSTAIN with authorized resolution, the human supplies authority-bound input, the boundary materially consumes it and emits a separate resulting ALLOW or DENY verdict, execution depends on that verdict, and the verdict with its authority chain is represented in the authorization artifact. The human supplies the input; the boundary emits the verdict. The label is the same; the architecture is not.

For regulated decisions, isn't human-in-the-loop required by law?

Whether a given obligation calls for human review before a defined action is for the applicable authority to determine. Where it does, policy can reserve that action class so that every such action returns ABSTAIN, the authorized person's input is consumed at the boundary, and the resolution is represented in the artifact. Role assignment and documentation, standing alone, do not establish that execution depended on a verdict that consumed the authorized person's input. Nothing on this page interprets regulatory text or constitutes legal advice.

Can human-in-the-loop workflows and FERZ compose?

Yes. A workflow can still sequence which humans see which classes of decision and when. When the human's determination is supplied as authority-bound input that the boundary consumes, and execution depends on the resulting verdict, the review is incorporated into a fail-closed pre-execution authorization procedure and is no longer review-only. The workflow may implement routing, but the applicable policy and current authority state determine whose input is eligible. The boundary emits the verdict on which execution depends.

Does resolving an ABSTAIN change the policy?

No. Resolution is scoped to the held action and does not author policy. Where a separately authorized policy-amending pathway exists, it may later consume the resulting authorization artifact to change the policy for a covered class of later actions. That second act is optional, changes future policy, and is not what authorized the held action.

Does a valid resolution establish that the human decided well?

No. A valid resolution establishes that the boundary consumed an authority-bound input, emitted the verdict, and that execution depended on it. It does not establish that the resolver considered the evidence with adequate attention and judgment. The artifact can bind what was presented to the resolver and the input the resolver supplied; it cannot represent a distinction the boundary did not bind. The Override Asymmetry states this limitation at Section 8.

Why is review-only human-in-the-loop not sufficient for AI-action authorization?

Because the determination is not consumed by a runtime authorization boundary and execution does not depend on a resulting boundary verdict. The record says "a human approved this," not "this action was evaluated under policy version P-17 against governed state G-558 and the boundary emitted ALLOW." Authorization requires the latter. A human's presence in the path does not supply it.

Cite this page

FERZ, Inc. (2026). FERZ is not Human-in-the-Loop. Revised September 2026. https://ferz.ai/governance/comparisons/ferz-is-not-human-in-the-loop

BibTeX
@misc{ferz2026nothitl,
  author = {{FERZ, Inc.}},
  title  = {FERZ is not Human-in-the-Loop},
  year   = {2026},
  month  = {9},
  url    = {https://ferz.ai/governance/comparisons/ferz-is-not-human-in-the-loop},
  note   = {FERZ Architectural Comparison, revised September 2026}
}