Technical Report SSL-TR-2026-002
No Perfect Guardrail & No Perfect Guard
On the Decidability of Output Policies for Large Language Models
PDF · BibTeX · All reports
Abstract
Expecting guardrails to provide a general guarantee of safety fails to recognize the limits of both computational enforcement and policy specification. We argue that promising such perfection is an ethical failure of precision, and that invoking it to abdicate responsibility is a moral failure. Guardrails can support responsible action, but cannot relieve researchers, developers, policymakers, or users of their obligations to examine what they claim, permit, and prevent.
We establish a precise computational boundary. If every finite string is a possible output of a language model, a computable guard that blocks every forbidden output and passes every permitted one exists if and only if the policy is decidable. If the policy is undecidable, every guard misclassifies infinitely many strings, and every guard that blocks all forbidden outputs excludes infinitely many permitted ones. We apply this characterization to arithmetic truth, program behavior, and harmful instructions to an idealized recipient. We distinguish guards from execution monitors, identify bounded and interceptable cases admitting exact enforcement, and explain why the results establish no positive lower bound on average error.
These distinctions matter to the practical and ethical interpretation of the proof. Exact enforcement, where possible, does not establish that a policy adequately captures safety. Conservative exclusion cannot be presumed harmless: exclusions may withhold warnings, remedies, or challenges needed to recognize the policy's own defects. Human oversight likewise cannot substitute for the exercise of judgment: an apparent controller can become an instrument of purposes they have ceased to examine. Responsible governance therefore requires continued critical challenge and the intellectual and moral humility to recognize possible flaws in our machines, rules, policies, and ourselves. The impossibility of a general assurance of perfection is a condition responsible practice must confront, not a responsibility that another guardrail—or another guard—can remove.
Keywords: decidability, output policies, large language models, guardrails, responsibility
Cite as
Ahmed Amer, Loren Kinczel, and Darrell D. E. Long. No Perfect Guardrail & No Perfect Guard: On the Decidability of Output Policies for Large Language Models. SSL Technical Report Series, SSL-TR-2026-002. Sustainable Systems Laboratory, Santa Clara University, November 5, 2026. Preprint v0.1. sustainable-systems-lab.org/reports/SSL-TR-2026-002/
Version history
-
Initial version. AI tools were used in the preparation of this version (v0.1) of the manuscript. Author details corrected November 19, 2026; version number and report date unchanged.