The Weakness No One Checked

A weakness filed against Trax International’s proposal on the Army’s $449 million White Sands contract rested on references the company says don’t exist. The Army conceded it was “errantly assigned” — and the award stood anyway. A federal court is now weighing Trax’s allegation that the finding began inside an Army AI platform and entered the record unchecked.

Somewhere in the evaluation record of a $449 million mission-support contract at White Sands Missile Range, a weakness was filed against Trax International’s proposal. The company, the finding said, proposed to “mirror” the installation’s own organizational structure in realigning incumbent personnel.

Trax says it proposed no such thing: its complaint calls the references behind the finding “made-up.” And on this point, the government did not argue. Before the Government Accountability Office, the Army conceded the weakness was “errantly assigned,” while contending the mistake was “harmless error” that had not prejudiced Trax’s chance at the award.

Then the award stood anyway. GAO denied the protest because the record showed the source selection authority’s decision “did not rely upon the erroneously‑assigned weakness.” The contract went to Southwest Range Services at $449,372,096 — $29.4 million above Trax’s $419,992,258. Whether the errant finding shaped that outcome is precisely what remains unresolved; no proven loss attaches to these numbers, and this record implies none.

That no one checked is Trax’s charge, and a contested one. That the weakness was errant is the Army’s concession. This record rests its weight on the concession.

The theory now before a federal judge

What happened next moved the case out of ordinary procurement law. In its complaint before the Court of Federal Claims, Trax points at the source of the finding: an internal Army AI platform known as FAST TRACK. “The weakness appears to be a classic AI hallucination, with made-up references to TRAX’s proposal, that no one on the SSEB checked before sending the SSEB Report to the SSA,” the complaint states — the SSEB is the Source Selection Evaluation Board, the human review layer; the SSA, the deciding official it reports to. The government’s sworn declarations in the record maintain that the experimental tool did not influence the award; the Army itself has declined public comment. GAO’s decision never mentions AI at all.

The causation fight is fully briefed before Judge Carolyn Lerner — Trax arguing “there is a reasonable possibility that the removal of this Weakness could have changed the award decision,” the government holding to harmless error. A ruling could land any day.

The part already conceded

The record does not need to wait for the ruling, because the part already conceded is the part that matters. Set the machine aside entirely. Accept, for argument’s sake, every government position: the tool did not influence the award; the deciding official never relied on the errant weakness; the error was harmless. What survives untouched is narrower than the accusation, and still consequential: a finding the agency itself calls errant entered the evaluation record of a nine-figure decision. Whatever review stood between that finding and the award did not remove it. The error surfaced only because the losing bidder protested. The unresolved failure here is not proven causation. It is that an unsupported finding survived every stage. Litigation became the only instrument left to expose where it may have come from.

The court has since added one fact that outlives the causation fight. The FAST TRACK analyses were not part of the administrative record until Judge Lerner ordered the government to produce them, for all three proposals. The tool’s work product entered the case record under compulsion, placed before the court for examination after the award. After the protest. After the lawsuit.

The record has a clinical name for the pattern. Decision Exclusion — the fifth characteristic of Digital Derangement Syndrome™ — is usually read as the machine leaving you out of its answer. If Trax’s theory holds, this case files the harder version: the exclusion of the entity’s actual record from the decision made about it. And either way, it sharpens the doctrine to its working edge: an invented finding does not need to be the reason you lost. It only needs to enter the record where no one is positioned to check it. From that moment, every downstream decision inherits a burden: detect the invention, or carry it forward silently.

The question that travels

If this were only federal procurement, it would be a specialist’s story. It is not. The same evaluation layer is moving into the rooms where businesses are judged at volume: lending files, insurance underwriting, vendor qualification, enterprise procurement, hiring. The structural question in the Trax record is portable to every one of them: when the machine’s account of you diverges from your actual record, who in the chain is positioned to notice — before the decision, not after?

Most businesses cannot answer it. They do not know what the machines currently say about them, which means a divergence would surface the way this one did: after the award, in the record of a decision already final — or under court order, filings later. The defense is not better prose in the next submission. It is a record structured to leave an evaluator the least possible room to misread, and monitoring close enough that an invented claim is caught while it is still a sentence, not a verdict. The question is not whether the machine was wrong. The question is who was positioned to know.