Fail-Closed Design
Fail-closed design is the structural property that requires a system to block an action when authorization cannot be confidently granted. Under uncertainty, error, or evaluation failure, the system defaults to blocking rather than permitting. In the FERZ runtime authorization framework, fail-closed is expressed through the ABSTAIN verdict, which blocks execution pending authorized human override.
A FERZ-defined structural property in the runtime authorization framework, articulated in On the Impossibility of Observability-Based Authorization (Meyman, 2026) and developed across the FERZ corpus.
What fail-closed means
Fail-closed is a binary property of the failure path. When the authorization gate cannot return a definite ALLOW, the action is blocked. The condition that prevents an ALLOW may be an internal evaluation error, a policy state that cannot be resolved, an input that falls outside the encoded constraints, or any other governed condition in which the gate cannot confidently authorize. In every such case, the action does not cross the runtime authorization boundary. The default state is closed. The gate must affirmatively open the path; absence of permission is denial of effect.
Fail-closed versus fail-open
Fail-open systems default to permitting when something goes wrong. A firewall that passes traffic through when its rules engine crashes is fail-open. A policy engine that returns ALLOW on internal error is fail-open. Fail-open is acceptable in some availability-critical contexts where blocking is more harmful than permitting. It is not acceptable in authorization, because the cost asymmetry runs the other way: an unauthorized action that takes effect cannot be retracted, while a blocked action can be re-evaluated. AI governance is an enforcement problem, not a monitoring problem. Fail-open enforcement is, structurally, no enforcement at all.
The ABSTAIN verdict expresses fail-closed
Fail-closed is the design property. The ABSTAIN verdict is its expression in the canonical verdict space. When the authorization gate cannot confidently issue ALLOW or DENY, it issues ABSTAIN, which blocks execution pending authorized human override. ABSTAIN is fail-closed. A Proof-Carrying Decision is emitted for the ABSTAIN verdict with the same cryptographic guarantees as for ALLOW or DENY. The block is recorded, signed, and independently verifiable.
What fail-closed is not
Fail-closed is not a circuit breaker. A circuit breaker is an availability pattern that halts calls to a degrading downstream service to prevent cascade failure; it operates on error rate and call volume, not on authorization. Fail-closed is not graceful degradation, which reduces service quality when components fail; degradation continues to permit reduced-fidelity action, while fail-closed blocks. Fail-closed is not rate limiting, which constrains throughput without evaluating individual authorization. Fail-closed is also not a security guardrail in the advisory sense: guardrails warn or filter within the governed system and can be modified by it, whereas a fail-closed authorization gate operates from outside the governed system and cannot be reconfigured by it.
Where it sits in the framework
Fail-closed design is the failure-mode property paired with non-bypassable AI governance. Non-bypassable governance ensures that every action passes through the gate; fail-closed design ensures that the gate's default behavior under any condition short of a definite ALLOW is to block. Together with execution-time authorization as the temporal property and the runtime authorization boundary as the structural location, fail-closed completes the architectural conditions under which a runtime authorization framework is enforcing rather than advisory.
References
- Meyman, E. (2026). On the Impossibility of Observability-Based Authorization, v1.3.2. FERZ, Inc. https://doi.org/10.5281/zenodo.19647542
- Meyman, E. (2026). The Override Asymmetry: Why ABSTAIN-Plus-Human-Override Is Not Guardrails-Plus-Human-in-the-Loop. FERZ, Inc. https://doi.org/10.5281/zenodo.19772248
- Meyman, E. (2026). Execution-Time Authorization for AI Agents: A Formal Framework for Deterministic Governance Boundaries. FERZ, Inc. https://doi.org/10.5281/zenodo.18764562
The full FERZ corpus is available at /research/papers, on Zenodo, and on SSRN.
Related concepts
Frequently Asked Questions
FERZ, Inc. (2026). Fail-Closed Design: Definition and Architecture. https://ferz.ai/governance/concepts/fail-closed-design
BibTeX
@misc{ferz2026fcd,
author = {{FERZ, Inc.}},
title = {Fail-Closed Design: Definition and Architecture},
year = {2026},
month = {May},
publisher = {FERZ, Inc.},
url = {https://ferz.ai/governance/concepts/fail-closed-design}
}