The claim file was built by someone else's AI. Verify it before you act on it.
A referral narrative, a claim-file summary, a provider statement, an attorney demand package. The documents that land on an SIU desk are increasingly drafted or summarized by AI that belongs to someone else, and every party in the chain had a reason to make the story read cleanly.
You work between two expensive errors. Act on a misidentified provider, a stale exclusion, a billing code that does not mean what the narrative says, or a pattern that turns out to be coincidence, and the case comes back, or exposes the carrier to a bad-faith or wrongful-denial action. Miss a real scheme, and the carrier absorbs the loss. The file arrives fluent. On a read, a fabricated or stale specific is indistinguishable from a correct one. Checking the file with another AI does not fix it: a second model shares the same blind spots and confirms the error with equal confidence.
The AI is the witness. The math is the judge. GauntletScore verifies each claim against primary sources, then computes the trust score with deterministic math. The agents gather the evidence; no model sits in the verdict. The same evidence always produces the same score.
GauntletScore runs a dedicated analysis of cause-and-effect claims, testing whether the timing, proportionality, and logic of an alleged scheme actually hold together. A pattern that fails the test counts heavily against the document. The engine is built to catch the story that reads like fraud and does not hold up, which is the story you do not want to act on.
Upload the document at gauntletscore.com. Minutes later you are reading the Gauntlet Report: a trust score with a 95% interval, every flagged claim with its verdict and the primary source behind it, and a cryptographically signed, tamper-evident certificate. For each specific in the file, you have the authoritative source that confirms or contradicts it. Ten minutes decides which files deserve the unit's time, and the ones you escalate carry the record that they were checked. In a live production run, the engine examined nineteen claims in one fluent, credible document and returned two debunked against the court record, each with the source that contradicts it. Every voting agent independently recommended against proceeding.
LIMITSIt does not decide the case, it does not replace your investigative judgment, and it does not verify what no public source can confirm. A claim it cannot check is reported as unverifiable, not as fraud. The tool tells you which facts hold; the investigation and the call are yours.
Run the file the day it lands, and put the unit's time on the claims whose specifics hold.
Confirm every named provider, credential, exclusion, and billing specific against the primary record before the referral goes to the carrier or the bureau.
Keep the signed certificate as the record that the basis for a denial or referral was independently verified.
There is a file in the stack right now that you are half-trusting. Run it. Either every specific holds and you move it with the record to prove it was checked, or one does not, and you found it before you acted on it.