A Judge Forced the Army's AI Evaluations Into the Record. Then It Made DOJ Counsel Show Cause.

The contract was large and unremarkable: $449,372,096 for White Sands Missile Range mission support, awarded to Southwest Range Services, a joint venture of SA-TECH, Amentum and EWA Warrior Services. Trax International had bid $419,992,258, about 6.5 percent lower, and lost on technical superiority. It filed a GAO protest, lost that, and sued the United States in the Court of Federal Claims.

Then it sued about something else entirely.

The Government's position was tidy. The Army had used a tool called FAST TRACK, described in the opinion only as "an AI platform." A procurement analyst with the Mission and Installation Contracting Command Headquarters Theater Support Center had run it, to assess whether it "could be used in future source selections." He created outputs for all three offerors.

The only output the source selection evaluation board ever saw, the Government said, was a single document for a third bidder, displayed on a screen. The board "did not receive digital or physical copies." The board concluded the document was "[n]ot useable and could not be used as a reliable tool." It was never entered into the contract file and never used. Nothing else existed, and nothing in the record had been contaminated.

On 22 September 2026, Judge William E. Lerner found that the record was not tidy at all.

The procedural run-up#

The path there took five months and it is worth tracing, because the case would not exist otherwise.

The Army issued the solicitation on 22 May 2025, RFP number W51EW7-25-R-A003, for experimentation, test, research, assessment, development and training support at White Sands. Southwest Range Services won it on technical superiority, at a price of $449,372,096 against Trax's $419,992,258. Both proposals were rated Outstanding overall on mission capability, and both received Substantial Confidence on past performance. The price premium was justified by technical superiority.

A GAO protest followed on 17 February 2026, denied on 14 May 2026. The suit in the Court of Federal Claims came on 29 May. SRS intervened on 4 June. The Government submitted the administrative record on 15 June, and a motion to complete that record was filed on 22 June. Cross-motions were filed on 9 July. The hearing was 19 August. The opinion came on 22 September.

That motion, on 22 June, is the hinge of the case, and it was granted.

The Government's account, and the record#

Trax moved to complete the administrative record. That motion succeeded, and with the AI outputs in it, the shape of the thing changed.

The court held, on the third bidder: "the entire SSEB convened to review the FAST TRACK output for [***] proposal. Whether the SSEB ultimately found that output 'useable' or 'reliable' has no bearing on whether it was considered."

And on the overlap: "But had the SSEB found the AI output useful, it could have revised its evaluation. Indeed, the overlap between the AI evaluation for [***] and the SSEB's final evaluation suggests that it may have done just that."

For the other two proposals, the Government conceded the opposite of what it had argued for the third. The opinion puts it plainly: "The Government does not contend, as it did for [***], that these were generated only after the corresponding SSEB evaluations had been finalized. The record therefore reflects that generation and review of the AI evaluations occurred during the agency's decision-making process."

The standard the court applied is the reusable part. Under it, information is considered if it was reviewed, regardless of whether it was relied upon. "Under either formulation of the standard, TRAX satisfied its burden as to all three AI evaluations."

The legal standard is old and the reading of it is not. The court took the rule from Joint Venture of Comint Systems: an agency "may not exclude information merely on the grounds that it did not rely upon the excluded information when reaching a final decision when there was evidence that the information was, in fact, reviewed." The opinion notes the precedent language in Smith is conjunctive, where the disjunction is what the court applied here, and that under either formulation Trax won, so the choice did not change the outcome.

The reason the standard matters more than the holding is Westwind Partners, the rule that a reviewing court must "have before it neither more nor less information than did the agency when it made its decision." Under that rule, an agency that reviewed something and then kept it out of the file has handed the reviewing court a record that is not the record the agency actually had. That is the whole mechanism, and it is a rule from before generative AI was a term of art.

The sentence-for-sentence overlap, and where it actually is#

Once the record was complete, the comparison the court made is the one that has since been repeated in every write-up of this case, usually in a way that gets one detail catastrophically wrong.

The court found that "the SSEB's evaluation of [***] contained multiple passages nearly identical, sentence for sentence, to the AI-generated evaluation of the proposal." The pinpoint citations run from the human evaluation at AR Tab 261, page 5059, to the AI output at AR Tab 434, pages 9089 and 9090.

That third bidder is redacted. They are not a party to the case, their identity is withheld, and their evaluation is not in dispute. The court immediately qualified the finding in the sentence that follows: "But any substantive or textual overlap in TRAX's and SRS's evaluations is minimal."

So the answer to the question most coverage asked is no. The AI output was not copied into the winning bid's evaluation. It was not copied into Trax's evaluation either, to any extent the court could find. What the court found was near-identical wording between a human evaluation and an AI evaluation of the same proposal, for a bidder nobody was fighting about, in an opinion that contains no discussion whatsoever of FAR 15.308 or of agency practice in source selection generally.

The belated admission#

Then the Government changed its story under cross-examination, and the court noticed.

"In his supplemental declaration, Mr. [***] acknowledged for the first time that he had shared the AI summaries with three individuals." The court called it a "belated admission."

The order to show cause, reissued publicly on 5 October 2026, is blunt about what that means. Twice, the court says, the Government filed declarations which omitted the fact that sitting SSEB members and the contracting officer had received the AI-generated evaluations the Government was representing no decisionmaker had seen.

"That omitted half of the answer is clearly material to the issue before the court."

And on the account as a whole: the Court "made a finding that the Government's account was 'a shifting and, at times, inaccurate account.'"

The order opens: "This Court has determined that Counsel for the Department of Justice ('DOJ') and the Department of the Army must SHOW CAUSE why they should not be sanctioned for factual misrepresentations in the Government's briefs and declarations."

It quotes Rule 11, and makes two points that matter for how much this is worth. "A court need not find bad faith to impose sanctions." And: "the Court is not required to find that any single individual acted with subjective intent to mislead."

The order also puts the overlap itself in issue. Question nine asks the Government to "Explain the overlap in language between the SSEB's evaluation of [***] proposal and the FAST TRACK summary for [***], and address whether the SSEB considered that output in conducting its technical evaluation of [***] proposal."

And it says this: "It strains credulity that this degree of overlap could be coincidental."

Government counsel apologised at the hearing: "I sincerely apologize. Obviously, had I been aware, I would not have allowed those representations to go forward."

Trax still lost#

None of this produced relief, and that is where most of the coverage of the case goes wrong.

TRAX identifies no actual errors or inaccuracies in the challenged evaluation findings that it can trace to AI. Even if the Court were to find AI was relied upon in some way, TRAX has not demonstrated why that reliance independently undermines the validity of the Army's findings.

And: "TRAX has not explained what remedy or instruction TRAX seeks from this Court related to the Army's use of AI."

On the one error the Army did concede, the court sided with the GAO, which "found no prejudice because of this error. Id. The Court agrees." De minimis errors do not justify relief. As the opinion says, "As TRAX fails on the merits, it cannot obtain a permanent injunction."

So the scoreboard reads: no sanction imposed, no damages, no injunction, no finding of hallucination, no error attributable to AI, and no general remarks about AI in procurement from either document.

What exists is a federal court that put the AI outputs into the administrative record, held that the board had considered them, found that the Government's own account of that had shifted and been inaccurate, and ordered its lawyers to show cause.

What the coverage got wrong#

Three errors, repeated widely enough to be worth naming.

One outlet headlined it as an "AI Hallucination Case Opens New Era for Procurement Litigation." The court made no hallucination finding. The word appears in the opinion only in the description of Trax's theory and in a note that the GAO did not address whether evaluation errors resulted from AI hallucinations.

The same coverage did not say which evaluation contained the overlap, which is the detail that changes everything about how much weight the story can bear. A third bidder nobody was arguing about is a different fact from the winner.

Another described the Court of Federal Claims as "a federal appellate court." It is the trial court. The opinion is the first instance.

And the coverage attached FAR 15.308, the provision on independent judgment in source selection, as the legal hook. The opinion never mentions it. The hook the court actually used is Rule 11, and the hook the judge used to order the record completed in the first place is a line from the 1960s and 70s case law: the reviewing court must "have before it neither more nor less information than did the agency when it made its decision."

The one piece nobody outside the docket had#

There is a document in the record that is not the opinion and not the order, and it is the only place the tool itself is described at all.

The July declaration of the procurement analyst, filed as an exhibit, says the tool was run on 1 May 2025, on White Sands and Dugway Proving Ground, and that it was "initially tested using fictitious acquisition documents," with clear instructions that none of the results generated from the tool would be used to inform the source selection evaluation boards for any real-world acquisition.

None of that appears in the opinion. The opinion contains no vendor, no model, no version, no configuration and no methodology, and the tool's name is FAST TRACK. It is a black box by the time it reaches the page, and the only description of what it was and how it was tested is a declaration that the judge did not need to reach.

That matters for a reason beyond the case. The Government told the court the AI was a single unusable screen view. The record shows it was run on all three proposals, on a live procurement, on real proposals, and distributed. The only thing that stops this from being a story about a broken evaluation is that the two proposals anyone was fighting over came out nearly identical, and that one of them turned out to be the redacted third bidder.

That is luck, not design, and nothing in the opinion or the order suggests the court understood it that way.

Why the mechanism travels#

The reusable finding survives contact with the next case, and it is not about hallucinations.

An agency's record of its own AI use is not self-authenticating. In this case the account was that the AI was a one-off screen view, unusable, never filed. The record showed otherwise for two of three proposals, and surfacing that took a motion to complete the administrative record, a court-ordered supplemental declaration, and cross-examination. The information was in the Government's files the whole time. It was in the files of the party that made the claim, and the party that made the claim controlled what the record said until a judge made it produce.

The corollary is a specific, court-tested move available to any protestor. Before this case, an agency could say we looked at AI output and did not use it, and the record would often support it. Now, "we looked at it" is discoverable through the record, and the overlap between the AI output and the final evaluation is the evidence. Demand the outputs, then compare them line by line.

As of today no follow-on protest has adopted that move, and the Government's response to the show cause order was filed under seal on 5 October. A joint status report and a redacted document went in on 19 October, also sealed. No sanction decision has issued.

The most likely next place this shows up is not in a sanctions ruling. It is in the next protest where a bidder asks a court to make an agency put its AI outputs in the file, and someone in a government counsel office realises that the file already contains an answer nobody had asked for.