Part of the AI Regulation News hub.
The Court of Federal Claims upheld the National Geospatial-Intelligence Agency's award of its SEQUOIA AI data-labeling contract to a small business, rejecting every ground of the incumbent's protest
The dispute is about who labels the data that trains government computer-vision models. The incumbent was rated Unacceptable on a single subfactor, which made its proposal unawardable, and the contracting officer kept it in the competition anyway. It still lost, and then lost the protest.
Bottom line: Judgment for the government and the intervenor on every count. Scale's motion for judgment on the administrative record was denied, its request for injunctive relief was denied, and the cross-motions of the United States and Enabled Intelligence were granted.
Who this affects: Government contracts and national security counsel, AI data-labeling and data-services vendors bidding federal work, contracting officers running AI training-data procurements, and incumbents on AI programs facing recompete.
Issue date: Filed under seal 31 July 2026. Reissued publicly with the parties' redactions on 18 August 2026.
What changed: Nothing in the law. What the opinion supplies is a worked example of how a court reviews an AI training-data source selection: security-requirement failures, a single Unacceptable subfactor that renders a proposal unawardable, past-performance confidence ratings, risk ratings, and the decision not to run a price realism analysis.
Analysis: The opinion opens with three pages on advocacy, not procurement. Judge Somers holds that Scale devoted substantial portions of its briefs to arguments for which it lacked standing while confining potentially dispositive theories to a sentence or two. The court granted Scale fifty pages and Scale used just over forty-two. That framing is unusual enough to be the part practitioners circulate.
Primary sources: Opinion and Order, Court of Federal Claims (PDF)
- Case
- Scale AI, Inc. v. United States, et al.
- Docket
- No. 26-155
- Court
- United States Court of Federal Claims
- Judge
- Zachary N. Somers
- Jurisdiction
- United States, federal
- Status
- Reported opinion. Judgment entered for the United States and defendant-intervenor Enabled Intelligence.
- Bindingness
- A decision of the Court of Federal Claims. Binding between the parties; persuasive, not binding precedent, in later protests.
- Issue date / next deadline
- Sealed 31 July 2026, reissued 18 August 2026. No deadline runs from this opinion; any appeal lies to the Federal Circuit.
- Procurement at issue
- National Geospatial-Intelligence Agency, SEQUOIA, services to create labeled geospatial data in accordance with the Government's standards
- Editorial Note
- Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
- Primary source
- https://ecf.cofc.uscourts.gov/cgi-bin/show_public_doc?2026cv0155-69-0
What the contract is for
SEQUOIA is the National Geospatial-Intelligence Agency's procurement for services that, in the solicitation's own words, create labeled geospatial data in accordance with the Government's standards. That is the human-and-tooling layer that turns raw imagery into training data for government computer-vision models.
The competitive significance is in the lineage. Scale AI bid as the incumbent contractor on NGA's predecessor Maven program. Defendant-intervenor Enabled Intelligence, a small business, bid against it. Both advanced past the first phase, and NGA told them and two other offerors that they were viable competitors and advised them to proceed to Phase 2.
This is worth stating plainly because it is the part most likely to be summarised badly: the case is not about whether AI may be used in intelligence work, and it decides no question about autonomy, targeting or model behaviour. It is a source-selection dispute about who supplies labelled data.
How the incumbent became unawardable
After Phase 2 evaluation, Scale was rated Unacceptable under Subfactor 2.1, and Scale alone failed the Factor 4 security requirement. On the record as it then stood, Enabled Intelligence's was the only awardable proposal.
The contracting officer nonetheless exercised her discretion to recommend Scale's inclusion in the competitive range. The stated reasoning was that although Enabled Intelligence had the most highly rated proposal, there was also a consideration of the extent to which other offerors could, through discussions, reasonably address weaknesses, deficiencies, or other issues in the proposal in such a way that would afford a proposal to be among the most highly rated.
NGA opened discussions on 25 June 2025 by issuing Items for Discussion to both offerors, allowed revised proposals, and evaluated them. Scale ultimately passed Factor 4. Its proposal nevertheless remained unawardable because the Unacceptable rating under Subfactor 2.1 stood.
On the adjectival ratings the two were close where it did not matter and apart where it did. Scale took Outstanding on Subfactors 1.1 and 1.2 and Good on 1.3; Enabled Intelligence took Outstanding on 1.1 and 1.3 and Good on 1.2. Enabled Intelligence also received Substantial Confidence under Factor 3 and Low Risk under Factor 5.
NGA announced the award to Enabled Intelligence on 25 September 2025.
The stay override, and why it matters
Scale first protested to the Government Accountability Office on 29 December 2025. After GAO partially dismissed its challenges, NGA notified GAO that it was overriding the automatic stay of performance under the Competition in Contracting Act.
The stated basis is the detail worth keeping. NGA said the override was in the government's best interest and cited an urgent and compelling need to continue data labeling services to prevent mission failure.
That is an agency putting on the record that a lapse in AI training-data labelling is a mission-failure risk, in the document it uses to justify continuing performance over a protest. For anyone assessing how defence and intelligence agencies now characterise data-services contracts, this is a more useful data point than the ratings tables.
Scale filed in the Court of Federal Claims on 30 January 2026 and amended its complaint on 27 March 2026.
The four challenges, and what happened to them
Scale's amended complaint alleged that NGA's source selection decision was arbitrary and capricious because it was infected by a series of evaluation errors. Four were pressed.
First, its own Unacceptable rating under Factor 2, specifically the Deficiency assigned under Subfactor 2.1. Second, Enabled Intelligence's favourable Substantial Confidence rating under Factor 3. Third, Enabled Intelligence's favourable Low Risk rating under Factor 5. Fourth, NGA's decision not to perform a price realism analysis on Enabled Intelligence's proposed price.
None succeeded. The court held that Scale failed to meet its high burden of establishing that NGA's award decision was arbitrary, capricious, an abuse of discretion, or otherwise contrary to law.
A structural point runs underneath the merits. Because Scale's proposal was unawardable on its own subfactor rating, several of its attacks on the awardee's ratings could not have produced a contract for Scale even if they had succeeded, and the court treats a number of them as outside Scale's standing rather than as losers on the merits. Protest counsel should read the standing discussion before the evaluation discussion.
The advocacy holding
The opinion's first three pages are about brief-writing, and they are not decorative. Judge Somers opens with Jones v. Barnes on winnowing out weaker arguments, then collects Seventh Circuit authority condemning scattershot, kitchen sink and blunderbuss briefing as the antithesis of sound advocacy.
Applied to this case, the finding is specific. Scale sought and obtained leave to file an oversized brief, then devoted substantial portions of it to arguments for which it lacked standing while leaving several potentially dispositive theories confined to only a sentence or two, often with sparse citation to the administrative record. Having asked for fifty pages, it used just over forty-two.
The court is careful about the consequence: this is not a case in which every underdeveloped argument is deemed waived, and Scale generally presented enough analysis for the court to reach the merits. But, it adds, scattershot advocacy exacts a cost even when it avoids forfeiture.
The closing line returns to the same image. Rather than concentrating its fire on the bullseye, Scale put its efforts toward issues lying in the outer rings.
Two things travel beyond this case. First, an Unacceptable rating on a single evaluation subfactor made the incumbent's proposal unawardable and then narrowed what it could even argue, because challenges to a rival's ratings do not help a protester who could not have received the award. Check awardability before building a protest theory around the awardee. Second, NGA justified overriding the CICA stay on an urgent and compelling need to continue data labeling services to prevent mission failure, which is a useful record of how an intelligence agency now characterises continuity of AI training-data supply. Any appeal lies to the Federal Circuit.
Source File
https://ecf.cofc.uscourts.gov/cgi-bin/show_public_doc?2026cv0155-69-0
Open the court's PDF and confirm four things: the caption naming Scale AI, Inc. against the United States with docket No. 26-155 on page 1, the seal and reissue dates of 31 July 2026 and 18 August 2026 immediately beneath it, the BACKGROUND section identifying the award as NGA's SEQUOIA procurement, and the CONCLUSION denying Scale's motion and granting the cross-motions of the government and Enabled Intelligence.
Scale failed to meet its high burden of establishing that NGA's award decision was arbitrary, capricious, an abuse of discretion, or otherwise contrary to law. Scale AI, Inc. v. United States, No. 26-155 (Fed. Cl., reissued 18 August 2026), Conclusion
FAQ
What was the contract?
NGA's SEQUOIA procurement, for services to create labeled geospatial data in accordance with the Government's standards. Scale AI bid as the incumbent contractor on NGA's predecessor Maven program; the award went to defendant-intervenor Enabled Intelligence, a small business.
Why was Scale's proposal unawardable?
It received an Unacceptable rating under Subfactor 2.1, and at the Phase 2 evaluation stage it was also the only offeror to fail the Factor 4 security requirement. It later passed Factor 4, but the Unacceptable rating under Subfactor 2.1 stood, and that alone kept the proposal unawardable.
If Scale was unawardable, why was it still in the competition?
The contracting officer exercised her discretion to include Scale in the competitive range, reasoning that beyond which proposal was most highly rated there was a consideration of the extent to which other offerors could, through discussions, reasonably address weaknesses or deficiencies so as to become among the most highly rated.
What did the court actually decide?
That Scale did not carry its burden. Scale's motion for judgment on the administrative record was denied, its request for injunctive relief was denied, and the cross-motions of the United States and Enabled Intelligence were granted, with judgment entered accordingly.
What was the CICA stay override?
After Scale protested to GAO on 29 December 2025 and GAO partially dismissed its challenges, NGA notified GAO that it was overriding the automatic stay of performance, stating the override was in the government's best interest and citing an urgent and compelling need to continue data labeling services to prevent mission failure.
Does this decide anything about AI regulation?
No. It is a bid protest decided on administrative-record review of a source selection. It sets no standard for AI systems, and its transferable content is procurement law plus the court's treatment of standing and of underdeveloped briefing.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.