The US Government’s AI Contract Clause Tests Everything but Value

← Back · ACT III · Chapter 16: MAVEN: Controlled Evaluation

For three years the standard objection to buyer-governed value has been that no buyer would ever get a seller to accept it. Sellers write the terms, define the metrics, and ship the dashboard that grades their own work. A buyer who insisted on setting the standard would be told the request was not commercially reasonable.

On 3 August 2026 the comment period closed on a draft federal contract clause that ends that objection. GSAR 552.239-7001 would govern how the General Services Administration buys large language model systems that process government data. Most of the coverage treats it as a data-safeguarding rule. Read paragraph (j) of the June draft and it is the most complete buyer-held verification apparatus yet drafted for a standard-form contract.

It is also aimed at the wrong question.

What The Government Reserved For Itself

Four things, in sequence.

The Government reserves the right to run its own tests. It may “conduct automated assessments of the LLM, as deployed and configured for government users, at any time using its own benchmarks.”

The seller must build the port. “The Contractor must provide tools and interfaces that enable the Government to run its benchmarks in an automated fashion to test the production LLM and identify any material gaps.” The cost of being measured is transferred to the party being measured.

The methodology stays with the buyer, with one deliberate exception. Benchmarks, test data and methodologies are Government Data, and the Government “is under no obligation to disclose or provide access to the underlying data, methodologies, or systems, with the exception of data or methodologies used as the basis for an adverse action under paragraph (j)(3)(ii).” The seller cannot tune to a test it cannot see, and where the buyer terminates for cause under that paragraph, it has to show the basis. Keeping the method closed is what protects the instrument, and opening it at the point of an adverse action is what protects due process.

The remedy has teeth. The Government may suspend use until performance issues are addressed, and on termination for cause, the contractor is liable for reasonable decommissioning costs up to a percentage of contract value.

A buyer defined the evidence, retained the method, compelled the instrumentation, and priced the failure, and the whole of it was published in the Federal Register.

What Carries A Consequence

Now read what the Government attached the consequence to.

The assessments “may assess bias, truthfulness, safety, unsolicited ideological content, and other factors determined by the Government.” That last phrase is open, so nothing stops the Government from testing whether the funded result arrived. Assessment breadth is not the constraint.

The constraint is the remedy. In the June draft the decommissioning liability applied to failure to remediate “the Unbiased AI Principles described in paragraph (j)(1),” and paragraph (j)(1) was three principles: the model must be truthful, it must be neutral and nonpartisan, and the contractor must run continuous improvement on bias detection. Every one of them describes the model, and none describes the result.

So an LLM could satisfy paragraph (j)(1) completely and produce nothing the agency funded it to produce. It could be truthful, unbiased, safe and ideologically neutral while the case backlog it was bought to clear stays where it was. The draft priced the embedding of a political judgment and put no price on a capability that did not work. The word “outcome” does not appear in the draft’s paragraph (j) at all.

That is the pattern the verification failure keeps producing. When a buyer finally builds the instrument, the enforceable part gets pointed at the supplier’s conduct rather than at the buyer’s result, because supplier conduct is what the contracting discipline knows how to write, and nobody has written the clauses for delivered value yet.

GSA has since issued the clause as a deviation, effective 19 October 2026. On FedScoop’s reading of the final text, a duty to use “reasonable efforts” to make the model respond factually has mostly replaced the “unbiased AI principles,” the Government keeps its right to run automated assessments, and decommissioning liability after a for-cause termination is capped at 25 percent of the affected order. That duty is still about how the model behaves, and none of the reported changes ties a consequence to whether the agency got what it paid for.

Why The Sellers Objected Anyway

The reaction tells you what the mechanism is worth even aimed imperfectly.

At GSA’s 14 July listening session, Palantir’s head of US government legal and contracting called the terms “not consistent with customary commercial practice,” and said that if the major LLM developers will not accept GSA’s terms, agencies will have to turn to non-GSA vehicles for the most advanced LLM-powered solutions.

The objection is that the standard is unusual, and that the sellers are large enough to route around it. Custom is being offered as a defense against a buyer’s terms. That is what happens the first time a buyer writes the standard instead of receiving it. A term one buyer invents can be refused. A term every buyer holds cannot.

The Correction Is One Sentence

The clause does not need rebuilding. It needs the consequence extended.

Everything required to verify delivered value was already granted in the draft: buyer-defined tests, buyer-retained methodology, mandatory instrumentation, production access, suspension, and cost recovery. Name the funded outcome alongside the duties the clause already places on the model, and the same machinery that prices a partisan model prices a capability that did not deliver.

The buyer’s leverage was never the missing piece. GSA proved a buyer can compel a vendor to build the interface for a test the vendor does not control. What is missing is the definition of what failure means. A buyer who sets terms without setting the standard has built an instrument and left the dial on the wrong setting.

The strongest verification right yet drafted into a federal contract can ask whether the machine is honest, and only punishes the answer to that question. It never prices whether the money worked.

A test the buyer owns can still miss what a capability displaces, which is the subject of Wendell Berry’s Nine Tests for Any New Tool.

References: GSAR 552.239-7001 draft clause, published for comment as Notice-MVAC-2026-01, Federal Register 17 June 2026 (FR Doc. 2026-12205), comments due 3 August 2026; clause text read in full at federalregister.gov/documents/2026/06/17/2026-12205. Palantir remarks at the GSA listening session of 14 July 2026, as reported by FedScoop, 15 July 2026. Final clause: GSA, RGO-2026-01 (section 539.71 and clause 552.239-7001, effective 19 October 2026); FedScoop, GSA issued an AI acquisition policy for new contracts, 1 October 2026.

Keep reading: ACT III · Chapter 16: MAVEN: Controlled Evaluation