The Artifact Gap: Why Observability-Based AI Governance Cannot Prove Authorization

FERZ Technical Advisory TA-2026-01 places a structural limitation of monitoring-based AI oversight on the public record.

Term introduced by FERZ, Inc. in Technical Advisory TA-2026-01 (June 11, 2026), grounded in the formal result of On the Impossibility of Observability-Based Authorization (DOI: 10.5281/zenodo.19647542).

The presumption most AI governance frameworks rest on

AI governance frameworks, across statute, regulation, supervisory guidance, audit standards, and acquisition practice, frequently operationalize control through monitoring, logging, incident reporting, post-deployment evaluation, and human review of AI system behavior. Where those mechanisms are treated as sufficient to establish that consequential AI actions were authorized before execution, the framework rests on a false presumption.

The presumption appears wherever a framework requires human oversight, continuous monitoring, or incident response as the mechanism by which an institution demonstrates control over AI systems acting in high-consequence settings: benefits adjudication, cybersecurity response, critical infrastructure operation, credit and underwriting decisions, claims adjudication, and clinical decision support. It also appears in the frameworks used to examine those systems. Supervisory programs, model risk management, compliance attestations, and audit practice routinely accept monitoring records as evidence that AI-affected decisions were subject to control. Where that evidence is accepted as establishing pre-execution authorization, the supervisory or audit conclusion inherits the same false presumption.

The formal result behind the gap

The presumption is false for a structural reason, proven formally in On the Impossibility of Observability-Based Authorization. An observability architecture is any system whose verdicts derive from properties of the governed AI system: its outputs, traces, internal states, or self-evaluations. Monitoring stacks, logging and audit instrumentation, guardrails, and human-in-the-loop approval arrangements that review what a system has produced all fall within this class.

The theorem establishes that no output of an observability architecture can constitute the authorization artifact an ex-ante requirement demands, on two independent grounds. First, observational signals exist only after the governed system has begun to generate the action they would authorize; observation of an event cannot precede the event. Second, even where such signals could in principle yield a pre-execution verdict, the resulting artifact could not be verified independently of the governed system, because reconstructing the verdict requires access to the system itself.

Two corollaries close the common objections. The result holds regardless of observational latency: zero-latency observation remains observation. And it holds regardless of intervention capability: a monitoring system that halts execution upon detecting a violation has prevented an action, but a halt is not a verdict, and the record it produces is a report that an unauthorized action was attempted, not a judgment that an action was authorized.

What counts as observability

Classification turns on input provenance, not labels. A verdict derived from the governed system's outputs, traces, internal states, or self-evaluations remains observational however it is packaged. A system may still include observability components and satisfy the authorization requirement, provided the authorization verdict itself is produced from policy, context, and the proposed action specification independently of the governed system. Timing alone does not change classification: review before dispatch remains observational if the judgment derives from the governed system's generative process.

Why the gap stays invisible until it is tested

Three properties distinguish the artifact gap from ordinary oversight shortfalls. It is universal across the architecture class: no observability-based arrangement avoids it, regardless of sophistication or expenditure. It is invisible until tested: systems governed this way operate normally until the moment an authorization record is demanded. And it is provable in advance: an institution can determine today whether its AI systems can produce pre-execution authorization artifacts, without waiting for an incident to reveal the answer.

The operational test is one question: can the system produce, for each consequential action, an independently verifiable artifact showing that the action was authorized before execution? If the answer is no, the institution has implemented observation, review, response, or intervention. It has not implemented authorization governance for that action.

What this means for procurement, supervision, and audit

The remedy is a requirement, not a product. Acquisition, supervisory, and audit language for AI systems in high-consequence settings should require that each consequential action is evaluated against explicit policy before execution, producing a verdict in a defined decision space (ALLOW, DENY, or ABSTAIN, where ABSTAIN blocks execution pending an authorized human decision); that the verdict is represented in an artifact an independent reviewer can verify without access to the governed AI system; that the absence of a verdict blocks execution; and that no execution path exists outside the authorization boundary.

For examiners and auditors, the requirement is evidentiary: request the pre-execution authorization artifact for sampled consequential actions, and treat its absence as a control deficiency rather than accepting observational records in its place. Observational records, however complete, evidence behavior, not authorization.

These conditions define the enforcement layer of AI governance: the architectural class for which FERZ builds runtime authorization infrastructure. Observability explains what happened. Enforcement determines what is allowed to happen. The two are complementary layers of a compliant governance architecture; neither replaces the other. The artifact gap exists precisely where the first is asked to do the work of the second.

Frequently asked questions

References

• Meyman, E. (2026). On the Impossibility of Observability-Based Authorization: A Formal Impossibility Result for Ex-Ante AI Governance. FERZ, Inc. https://doi.org/10.5281/zenodo.19647542
• Meyman, E. (2026). Observability Is Not Enforcement: A Doctrinal Framework for Distinguishing Compliance Instrumentation from Runtime Authorization in AI Governance Architectures. FERZ, Inc. https://doi.org/10.5281/zenodo.18663864
• Meyman, E. (2026). A Taxonomy of AI Governance Approaches: Distinguishing Visibility, Alignment, and Authorization. FERZ, Inc. https://doi.org/10.5281/zenodo.18275969
• FERZ, Inc. (2026). The Authorization Artifact Gap in AI Governance, Regulatory, and Audit Frameworks: Formal Notice of a Known Structural Limitation in Observability-Based AI Oversight. FERZ Technical Advisory TA-2026-01. https://doi.org/10.5281/zenodo.20646404

Related concepts

@misc{ferz2026artifactgap,
  author = {{FERZ, Inc.}},
  title = {The Artifact Gap: Why Observability-Based AI Governance Cannot Prove Authorization},
  year = {2026},
  url = {https://ferz.ai/research/artifact-gap},
  note = {FERZ Technical Advisory TA-2026-01, DOI: 10.5281/zenodo.20646404}
}