Cite or Strike: A Citation Verifier for Legal Drafting
Build a verifier that checks every citation in an AI-drafted legal document against a real source and flags or strikes anything it cannot ground.
The failure behind this brief
In a Mississippi federal court, attorneys on both sides were sanctioned over AI-hallucinated citations. The adversarial check that normally catches bad citations collapsed because both sides carried the same exposure: each relied on unverified AI output, so neither caught the fabrications.
Unverified model output certified itself. The system that should have flagged fabricated citations, opposing counsel, had the same blind spot, and no step in the pipeline verified a citation against a retrievable primary source before the filing went out.
Never let generated citations certify themselves. Verify every citation against a retrievable source before the document leaves the system, and treat an unverifiable citation as a hard failure that blocks export, not a soft warning a tired drafter can skim past.
Sources Source 1
Editorial note: The court's sanctions order is the primary source; attach it in the editorial verification pass before publish. The thread above is discovery provenance only.
Why this matters
AI-drafted filings have put fabricated citations in front of judges and drawn sanctions against the lawyers who filed them. The drafting model is not the failure; the missing artifact is a verifier that grounds every citation against a real source before the document is filed. A citation that cannot be retrieved is a defect, not a stylistic choice.
Persona
Litigator or legal drafter using AI to produce filings
Current manual workflow
A drafter generates a brief with an AI assistant, skims the citations for plausibility, and files, trusting that anything formatted like a case is real.
The AI workflow to build
The verifier extracts every citation from the draft, looks each one up against a citation database or document corpus, and classifies it as verified, altered, or not found. Verified citations pass. Altered or not-found citations are flagged with the specific defect, and the document is blocked from export until every citation resolves or is removed.
Inputs
- an AI-drafted document
- a citation database or reference corpus
- a citation extraction step
Outputs
- a per-citation verdict (verified, altered, not found)
- a defect report naming each unverifiable citation
- a blocked or cleared export decision
Definition of done
On a synthetic document set seeded with real, altered, and fabricated citations, the verifier marks every fabricated citation as not found, catches altered quotes or pin-cites as altered, passes the genuine ones, and blocks export while any citation is unresolved.
A brief citing Smith v. Jones, 123 F.3d 456, with a quoted holding, where no such case exists in the reference corpus.
Citation Smith v. Jones, 123 F.3d 456: not found in the corpus. Verdict: defect, export blocked. The verifier names the citation and the failed lookup rather than rewording the brief.
Data plan
synthetic data
Boundaries and non-goals
- drafting the legal argument
- real client matter data
- filing or court system integration
Evaluation ideas
- fabrication recall (caught not-found rate)
- altered-citation detection
- false-positive rate on genuine citations
- export-gate correctness
Run Level target
R3 Reliable Plain translation: handles real cases.
Scope envelope
Buildable by one solo builder in 20 to 30 focused hours, on public, synthetic, or sanitized data, with a demo path that requires no production access.
Suggested tools
Suggested options, never requirements; briefs are tool-agnostic.
Product thesis questions
- Should an unverifiable citation block export or only warn?
- What citation corpus coverage is needed before the verifier is trustworthy in a given jurisdiction?
Open briefs in this failure family
Triage First: A Support Ticket Routing Agent
Build an agent that classifies inbound support tickets, drafts a suggested reply, escalates uncertain cases, and explains why it routed each one.
Reliability focus classification, calibrated escalation, explainability
View BriefSeen this failure in your own stack? Submit your version.
Submit your version