The Unbiased AI Standard: 27 Requirements That Prove a Model Was Built Fairly

A contract clause that demands unbiased AI and provides no way to measure bias is not a standard. It is a liability trap. GSA’s draft GSAR clause 552.239-7001 makes Unbiased AI Principles a contract performance standard with suspension rights and decommissioning liability attached, and it provides no test method, no threshold, and no measurement standard anywhere in the clause. The Unbiased AI Standard, or UAS, released July 13, 2026, is the open-source answer: quantified, inspectable, and adoptable by anyone, including competitors.

The problem: a performance standard nobody can pass or fail

The problem is not that the government wants unbiased AI. The problem is that the draft clause creates a performance obligation with no definition of performance. A contractor can be suspended, and a system can be decommissioned, over a standard the clause never defines. There is no test to run, no threshold to meet, and no measurement to dispute. A vendor cannot prove compliance, an evaluator cannot prove noncompliance, and a contracting officer cannot tell the difference between a system that was tested and a system that was merely asserted to be fine.

That gap is worse than having no clause at all, because the clause carries consequences. Suspension rights and decommissioning liability attach to a judgment nobody can make in a repeatable way. The practical result is that compliance becomes a matter of who writes the most confident attestation, and the vendor with the best paperwork wins over the vendor with the best evidence. A standard that cannot be measured selects for confident claims, not unbiased systems.

What the release changes

UAS ships five components, and each one answers a specific part of the measurement gap.

First, a normative standard, sections 1 through 8, covering scope, definitions, requirements, testing, evidence, and independent review criteria. That is the document a contractor and an evaluator can both read and both be held to.

Second, an 8-class bias taxonomy, UAS-B1 through UAS-B8, with each class mapped to its pillar source, its Cognitive Sovereignty Framework domain, and the GSAR trigger it answers. The taxonomy makes a bias finding specific: a finding names its class, and the class names what to test.

Third, 27 controls: 14 system-side controls spanning pillars P1 through P5 plus NEXUS, 8 human-side controls sourced from the Cognitive Sovereignty Framework, and 5 cross-domain bridge controls. The human-side controls are the part no other framework in this space supplies, and the release names that directly. A technically neutral AI can still manipulate its operator, and a standard that only tests the system cannot see it.

Fourth, a test protocol with 8 test types, published pass thresholds, and ACT-tier scaled frequency. The thresholds are the numbers that turn the standard from a document into a measurement: an 85 percent neutrality classification floor, a Cohen’s kappa of 0.7 as the inter-rater minimum, and a demographic parity requirement of p greater than 0.05 across matched framings. A system either meets those numbers or it does not, and the protocol says how often to test at each ACT tier.

Fifth, an evidence package: a five-document due-diligence set mapped paragraph by paragraph to the clause’s (j)(1) and (f)(7) provisions. That mapping is what makes the standard usable in procurement. The contractor does not have to guess which document answers which paragraph. The mapping states it.

The release also answers the disclosure hole directly. UAS-B8, Foreign Regulatory Embedding, is presented as the first published methodology for detecting model modifications made to satisfy foreign regulators, which the clause requires contractors to disclose at (f)(3) without ever saying how. That is the release’s own claim, and this draft attributes it to the release. If the claim holds, it closes the one gap in the clause that no vendor could have closed alone.

The human layer, measured

UAS is the first cross-pillar control in the program with an explicit human-protection layer. Controls UAS-H1 through UAS-H8 are sourced directly from Cognitive Sovereignty Framework Domains 1 through 6, with CTSS scoring feeding the evidence package. The release puts the relationship in one line: AI SAFE2 secures the system, CSF protects the human, and CP.11 proves both. The system-side controls test what the model does. The human-side controls test what the model does to the person running it. A standard that measures only the first half cannot see the second half failing.

The release also states the terms plainly: zero patents, zero licensing, and every control inspectable today, so a competitor can adopt it and an evaluator can run it without asking permission.

What to do with it

Start with the CP.11 overview in the ai-safe2-framework repository, then run the test protocol against the deployment under review, then generate the attestation from the evidence package. For teams commenting on the GSA docket, the release directs them to reference Notice-MVAC-2026-01 at regulations.gov before August 3, 2026. That deadline has passed by the time this backfill announcement is published, and the draft records it as a historical fact, not a live instruction.

Where it sits in the program

UAS was released July 13, 2026, the same day as CSF v2.1 and the Federal Procurement Integration CP.11, which formally published CSF’s human-layer controls as UAS-H1 through UAS-H8 inside the standard. Later releases confirm the standard’s shape: the September 4 CLI release carries UAS as a 27-requirement profile, and the October 2 AISM Scoring Update preserves UAS at 27 profile requirements alongside 161 core AI SAFE2 controls. One discrepancy should be recorded. The July 13 release presents the standard as CP.11, the 11th Cross-Pillar control. The September 4 release states that UAS is a 27-requirement regulatory profile extension and not CP.11. This draft follows the July release’s framing, because the announcement is about the July release, and flags the discrepancy here so the record stays honest.

The framework, in one place

The standard is 27 requirements: 14 system, 8 human, and 5 bridge. The system requirements span pillars P1 through P5 plus NEXUS. The human requirements come from CSF Domains 1 through 6. The bridge requirements connect the two. The framework is the last thing this draft names, not the first, because the problem here is not a missing framework. It is a contract clause that demands a property nobody can measure, and the standard exists to make the measurement possible.

What this does not establish

The thresholds are the release’s own numbers, and the release body describes no independent validation of them. Treat the 85 percent floor, the 0.7 kappa minimum, and the p greater than 0.05 parity requirement as the standard’s stated thresholds, not as outside proof that those are the right thresholds. The clause account here is the release body’s own account of GSAR 552.239-7001, and the clause text was not independently read for this draft. The “first published methodology” claim for UAS-B8 is the release’s claim, and it was not independently verified.

Sources

Recent Posts

Tag Cloud

Stop Threats Before They Execute

Your free Kernel-Level Defense Buyer’s Guide is ready to download.

By providing my email address, I consent to receive emails and text messages—including newsletters and marketing communications—from creators of Warden Secure, Cyber Strategy Institute, our flagship zero-trust platform for ransomware prevention, and agree to the Terms and Privacy Policy. You may unsubscribe at any time.