FAQ
The questions lawyers and firm administrators ask before piloting CitationShield. Don't see yours? Email us.
What does CitationShield actually do?+
It reads a legal brief (.docx or PDF), extracts every citation, and verifies each one against authoritative sources — primarily CourtListener (Free Law Project). For each citation it returns a status: verified, partial match, or not found. The output is a color-coded report with one-line explanations.
Is this a Westlaw or Lexis replacement?+
No. Westlaw and Lexis are research platforms — you go there to find cases. CitationShield is a verification platform — you bring it cases you've already cited and ask 'are these real and correctly cited?' Different jobs. Most firms will use both.
Doesn't Westlaw / Lexis already check citations?+
They have citation tools (KeyCite, Shepard's) that tell you a case's treatment — affirmed, overruled, distinguished. They are excellent at that. They were designed before LLMs were generating fully-fabricated case names with plausible-looking reporters and pages. CitationShield is purpose-built for the post-LLM threat: catching citations that look correct but reference cases that do not exist anywhere.
What citation formats do you support?+
Federal: U.S. Reports, F.2d/F.3d/F.4d (where applicable), F. Supp./F. Supp. 2d/F. Supp. 3d, Federal Reporter, Federal Rules Decisions. Regional reporters (A., A.2d, A.3d, P., P.2d, P.3d, NE, NW, SE, SW, So.). State reporters for the top 20 states by filing volume. Statutes (U.S.C., C.F.R., state codes). We're expanding coverage every release.
What's your accuracy?+
On a benchmark of 200 real briefs from federal litigation, the regex extractor catches 82% of standard Bluebook citations. The GPT-4o-mini fallback brings extraction to ~96%. CourtListener verification is exact-match by reporter+volume+page; partial matches surface as yellow flags. The remaining ~4% are short-cite chains where context is required — flagged for human review.
What happens if a real case isn't in CourtListener?+
Rare but possible — most often very recent unpublished decisions or some state appellate filings. We mark those 'unverified — not found in CourtListener' rather than 'fabricated.' The yellow status. You then check Westlaw or PACER directly. Enterprise customers can configure secondary sources.
Will this catch every AI hallucination?+
Anything cited with a fake reporter+volume+page combination — yes, very reliably. AI hallucinations that invent a real-sounding case name AND a coincidentally-real reporter+volume+page combination that maps to a different real case will surface as a yellow 'partial match' for the lawyer to review. It is a substantial improvement over the unaided eye and is not a substitute for it.
What about ABA Model Rule 1.1 (competence) and 3.3 (candor to the tribunal)?+
CitationShield is a tool that supports your professional duties. It does not discharge them. Rule 1.1's competence requirement and Rule 3.3's duty of candor to the tribunal remain with counsel. We provide an additional verification layer — and the audit-trail receipt some carriers and clients are starting to require.
Do you train models on our briefs?+
No. Customer briefs are never used for model training, fine-tuning, or any analytics beyond serving your own verification request. By default the brief is deleted within minutes of the run.
Where is data stored?+
U.S. cloud regions only (currently AWS us-east-1, with us-west-2 failover). Enterprise customers can request data residency in a specific U.S. region or an on-prem deployment.
Do you have an API?+
Yes — included on Firm and Enterprise tiers. Most firms wire it into their document management system (iManage, NetDocuments) so verification runs automatically before any brief moves to the partner-review folder.
Who's behind it?+
Chris Swofford and Marlin Swofford lead the company. The product was built in close collaboration with practicing litigators and an advisory board of malpractice-defense counsel. We are not anonymous — see the About page.