Part of the AI Regulation News hub.
The Court of Federal Claims has held that AI-generated proposal evaluations the Army produced and showed or sent to its evaluators belong in the bid protest record, while rejecting TRAX's claim that AI tainted the award because it traced no error to the AI
TRAX got the Army's AI outputs into the record and still lost the protest. Its AI taint argument failed because it tied no specific evaluation error to the AI tool.
Bottom line: Binding on the parties; a trial-level ruling in one protest. The opinion was filed under seal on 3 September 2026 and reissued publicly with agreed redactions on 22 September 2026. TRAX lost: its motion was denied and the government's and SRS's cross-motions were granted.
Who this affects: Bid protest and government contracts counsel, contracting officers and source selection evaluation boards trialling AI tools, and defense contractors competing for Army services awards.
Issue date: Filed under seal 3 September 2026 (ECF No. 70); reissued with agreed redactions 22 September 2026. The court said a separate Order to Show Cause to the Government would follow.
What changed: The opinion gives the reasons for the court's orders of 7 and 10 July 2026 adding three FAST TRACK AI evaluations to the administrative record. On this record, showing or distributing them to evaluators counted as consideration, whether or not the evaluators relied on them.
Analysis: An agency saying an AI output was unusable did not keep it out of the record. But TRAX identified no evaluation error it could trace to AI, so the court found its request to factor AI use into deference unavailing.
Primary sources: TRAX International Corp. v. United States, No. 26-796, Opinion and Order, public version reissued 22 September 2026 (Court of Federal Claims)
- Instrument (EN)
- TRAX International Corp. v. United States, Opinion and Order
- Authority
- United States Court of Federal Claims (Judge Carolyn N. Lerner)
- Jurisdiction
- United States, federal. Post-award bid protest of an Army Mission Support Services award at White Sands Missile Range
- Status
- Decided. TRAX's motion for judgment on the administrative record denied; cross-motions of the United States and Southwest Range Services granted; clerk directed to enter judgment
- Bindingness
- Binds the parties to this protest. A single trial-court opinion; we make no claim about how other judges of the court will treat it
- Issue date / next deadline
- Filed under seal 3 September 2026; reissued publicly 22 September 2026. No deadline stated
- Legal basis
- Administrative Procedure Act standard of review applied through 28 U.S.C. 1491(b)(1) and (4)
- Document
- No. 26-796, 22 page public version with redactions marked [***]
- Primary source
- https://ecf.cofc.uscourts.gov/cgi-bin/show_public_doc?2026cv0796-76-0
What did the court decide, and when?
Judge Carolyn N. Lerner denied TRAX International's post-award protest of an Army contract for Mission Support Services at White Sands Missile Range, awarded to Southwest Range Services at a price of $449,372,096. The court found TRAX had failed to show the Army acted irrationally or contrary to law in its evaluation, and held that, having failed on the merits, TRAX could not obtain a permanent injunction.
The Opinion and Order was filed under seal on 3 September 2026 as ECF No. 70. The parties jointly proposed redactions (ECF No. 75), and the court reissued it with those redactions on 22 September 2026. That redacted reissue is the event we report; the reasoning itself dates from 3 September, and the record-completion orders it explains date from July.
Most of the opinion deals with conventional protest grounds. The AI element sits in Part III, which explains the court's orders of 7 July 2026 (ECF No. 38, granting the motion as to the third offeror) and 10 July 2026 (granting the rest) on TRAX's motion to complete the administrative record and then addresses TRAX's argument that reliance on AI tainted the evaluation.
How did the FAST TRACK tool enter the procurement?
According to the government's declarations as the court summarises them, a procurement analyst was tasked with assessing FAST TRACK, which the opinion calls "an AI platform", "to determine whether it could be used in future source selections." He generated AI outputs for all three offerors and displayed the one for the third offeror, not a party to the suit, on a screen for the Source Selection Evaluation Board. The board concluded it was not usable.
The government first said that was the only AI output the board saw. In a supplemental declaration the analyst acknowledged, for the first time, that he had shared the AI summaries with three people. The court found that those who received the evaluations of TRAX's and SRS's proposals included the contracting officer and members of the board.
The court called this a "belated admission" and said it would issue a separate Order to Show Cause "requiring the Government to explain the errors and misrepresentations in its briefing" on the motion. That matter is not decided in this opinion.
Why did the AI outputs belong in the record?
The test is what was "actually before the agency". The court noted the "generated and considered" standard is commonly applied disjunctively and held TRAX met it under either formulation for all three AI evaluations.
The outputs were relevant because they were generated to reflect the comparative merits of the proposals. They were generated during the procurement; for TRAX and SRS the government did not contend they were produced only after the board's evaluations were finalized. And they were considered. Whether the board found the third offeror's output "useable" or "reliable" had no bearing on whether it was considered, and for TRAX and SRS, distribution to the contracting officer and board members was sufficient "regardless of whether the SSEB relied upon or used the outputs further."
Why did the AI taint argument fail?
Once the record was complete, the court observed that the board's evaluation of the third offeror contained "multiple passages nearly identical, sentence for sentence, to the AI-generated evaluation". TRAX used that to argue that over-reliance on AI tainted the whole technical evaluation.
It did not carry the argument. The court found any overlap in the TRAX and SRS evaluations "minimal" and said that "TRAX identifies no actual errors or inaccuracies in the challenged evaluation findings that it can trace to AI." Even if AI had been relied upon in some way, TRAX had not shown why that reliance independently undermined the Army's findings, and its request that the court weigh AI use in deciding what deference the findings deserve was "unavailing".
The court did not clear the Army's process either. It said it "cannot entirely rule out the use of AI in assessing the proposals".
What remains open?
The court said a separate Order to Show Cause on the government's briefing would follow. We have not seen it and do not know whether it has issued.
GAO, which denied TRAX's earlier protest on 14 May 2026, did not discuss whether any evaluation errors resulted from AI hallucinations, according to the opinion. This opinion does not decide whether using an AI tool in a source selection is itself improper.
What we did not verify
What we opened: the 22 page public Opinion and Order, retrieved from the Court of Federal Claims' own public document link cited above and read as extracted text. We confirmed the file is byte-identical (same SHA-256) to the copy our research desk retrieved, and re-fetched it from the court link on 24 September 2026 with the same result. The public version does not print its own docket entry number: footnote 1 identifies the sealed opinion as ECF No. 70 and the proposed redactions as ECF No. 75, and we cite the document by its case number, No. 26-796. We read the introduction, background, Part III on the AI evaluations in full, and the conclusion.
What we did not open: the sealed version (ECF No. 70), the three FAST TRACK outputs, the declarations, the parties' briefs, the court's July 7 and July 10, 2026 orders, the GAO decision, the Order to Show Cause, and the docket. We did not examine the opinion's treatment of TRAX's non-AI grounds beyond its conclusion.
What we refuse to claim: we do not say the Army relied on AI to make its award, that FAST TRACK produced errors, or that the Army or its counsel misrepresented anything, since the court has so far only announced a show cause order. We do not identify FAST TRACK's developer, which the opinion does not name. We do not say this ruling binds other judges or sets a government-wide rule, and we do not predict an appeal.
Quotations are reproduced with ASCII punctuation in place of the typographic quotation marks and apostrophes in the opinion, which is a house typesetting convention and not a change to any word.
Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
If an agency runs an AI tool over proposals and shows or sends the output to members of the evaluation team, this court treated that output as part of the record, useful or not. For protesters, access is only the first step: TRAX got the AI evaluations, but its AI taint argument failed because it could not point to a finding the AI got wrong, and it lost on its other grounds too.
Source File
https://ecf.cofc.uscourts.gov/cgi-bin/show_public_doc?2026cv0796-76-0
Open the public Opinion and Order and check the caption and footnote 1 for the sealed and reissue dates, Part III.B on pages 5 to 7 for how the FAST TRACK outputs were shared, Part III.C on pages 7 and 8 for the record holding and the AI taint analysis, and Part VI on page 22 for the disposition.
Because all three documents were relevant to the Army's decision-making process, and because each was generated and considered by the Agency during the evaluation process, they should have been included in the AR. ยท TRAX International Corp. v. United States, No. 26-796, Opinion and Order, Part III.C, page 8, reissued 22 September 2026
FAQ
Did the court find that AI caused errors in the Army's evaluation?
No. It found TRAX identified no actual errors or inaccuracies in the challenged findings that it could trace to AI, while saying it could not entirely rule out the use of AI in assessing the proposals.
Why does an unusable AI output still go in the record?
Because the court treated consideration, not reliance, as the test. Once the output was shown to the evaluation board or distributed to the contracting officer and board members, it was considered, whether or not anyone found it useful.
Why are there two dates on the opinion?
It was filed under seal on 3 September 2026. After the parties jointly proposed redactions, the court reissued it publicly on 22 September 2026 with the redactions marked.
What is the Order to Show Cause about?
The court said it would separately require the Government to explain errors and misrepresentations in its briefing on the motion to complete the record, including the late disclosure that the AI summaries had been shared. That matter is not addressed in this opinion.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.