A real Supreme Court case, cited for the exact opposite of its holding. Our verifier refused to pass it and sent it to a human. The brief cited Erie Railroad v. Tompkins for the idea of a uniform federal common law. Erie held the opposite. The case is real, and a shallow check that only asks "does this citation point to a genuine case?" waves it straight through. When we built CiteGuard, we split that question in two. A deterministic check grounds every citation against real case law on CourtListener and proves whether the case exists. Then a separate support-check weighs whether the cited opinion actually supports the claim it was attached to. On Erie, it could not establish that from the opinion, so it did the honest thing: it abstained and routed the citation to an attorney with the contradicting context, rather than bluff a green verdict. That abstention has teeth. The policy gate is deterministic: a brief is filing-ready only when zero citations are fabricated and zero are left unresolved. One unresolved citation is enough to hold the entire filing at Do Not File. And every decision lands in a signed audit certificate, each cite mapped to its real CourtListener opinion, its verdict, and whether the call was deterministic or an abstention. The uncomfortable part is that a better model does not retire this. A court does not care how confident the AI was; it cares whether the citation says what you claimed. So the most valuable thing our verifier does is know when to say "I am not sure, a human needs to look." 🔍 If your firm is working out where AI drafting ends and human verification has to begin, we would genuinely like to hear how you are drawing that line. #LegalTech #AIGovernance #LegalOps #Litigation #AICompliance