Check every AI-drafted citation before it reaches the court.

How does the citation checker work?

The instant check answers the single most important question about an AI-assisted brief: is this a real case? It is the primary anti-hallucination signal, because fabricated cases do not resolve against a live corpus. The tool below runs that existence check against a public case-law source and returns a plain verdict — resolved, or not found. There is no dashboard to learn and no account to create to try it. You paste a citation, and the resolver tells you whether an opinion with that reporter citation actually exists in the corpus it can reach.

For a filing you are about to sign, the full certification goes further: it also matches quoted passages against the published opinion, overlays good-law standing from your firm's citator, and seals the result to a tamper-evident log. The instant checker is the front door; certification is the checkpoint between your draft and your signature. One is a fast sanity check you can run on a single string of text; the other is a durable record of what you verified, when, and against what. You can read the deeper mechanics on how it works.

Both start from the same discipline. The resolver never reasons about whether a case sounds plausible. It looks the citation up. A model can generate a citation that reads exactly like a Ninth Circuit opinion — correct reporter, believable volume and page, a party name that fits the doctrine — and be entirely invented. The only reliable way to separate the real from the fabricated is to check the string against a source of truth, which is precisely what the resolver does and what a human skim, however experienced, cannot do at speed.

What does the free existence check actually confirm?

The free check confirms one thing precisely, and it is careful not to overstate it: an opinion matching the reporter citation you entered exists in the public case-law corpus the resolver reaches. The existence check runs against CourtListener and public case-law sources. That scope is a feature and a boundary at the same time. When a citation resolves, you have strong evidence the case is real. When it does not resolve, you have a clear signal to stop and investigate before that authority goes anywhere near a filing.

It is worth being exact about the limit, because honesty here is the entire value of the tool. The resolver is only as complete as the case-law corpus it reaches. A very recent opinion, a sealed matter, an unpublished disposition, or a citation to a source outside the public corpus may not resolve even though the case is genuine. A non-resolution is a prompt to verify, not a verdict of forgery. What the free check reliably catches is the failure mode that is sinking firms today: a confidently formatted citation to a case that was never decided by any court, because a language model produced it to fill a gap in an argument.

That is the difference between the free existence check and the good-law question people often conflate with it. Existence tells you the case is real. It does not tell you the case is still binding, has not been overruled, or supports the proposition you cited it for. Those are separate dimensions, and full certification handles them separately — because collapsing them into a single green checkmark is exactly the kind of overclaim that gets a tool trusted for something it never measured.

What does the three-dimension certification cover?

These three dimensions fail independently, which is why certifying only one of them leaves real exposure. A case can exist and be good law but be quoted for a holding it does not contain. A case can exist and be quoted accurately but have been overruled two years ago. A case can be quoted perfectly from an opinion that was never written. Certifying existence alone — the thing most free checkers stop at — closes the loudest failure but leaves the quieter, arguably more damaging ones open. Full citation certification addresses all three and records the result of each, so the certificate reflects what was actually checked rather than a single blended score.

• Existence — the citation is matched exactly against live case-law, so a fabricated case is flagged before you sign. This is the same resolution the free tool performs, now recorded as part of a signed result.
• Quotation accuracy — quoted passages are matched against the text of the published opinion, catching misquotes, paraphrases dressed up as direct quotes, and quotations that appear nowhere in the cited case.
• Good-law standing — overruled, superseded, or questioned authority is flagged, sourced from your firm's citator (KeyCite or Shepard's). We never guess good law; the standing signal comes from the citator you already trust, and we record what it returned.

Why does AI fabricate citations that look perfect?

A large language model does not retrieve cases. It predicts text. When it writes a citation, it is generating the most statistically likely sequence of characters that looks like a citation in that context — the right reporter for the circuit, a plausible volume, a party name that fits the doctrine, a pin cite in the expected range. Correct format is the easy part for a model to reproduce, because format is the most regular, most learnable pattern in the entire training corpus. That is the trap. The very thing a hurried reviewer uses as a proxy for legitimacy — clean Bluebook form — is the thing an AI reproduces most faithfully whether or not the underlying case is real.

This is why formatting is worthless as an existence signal and why the checker never treats it as one. Two citations can be byte-for-byte indistinguishable in form; one resolves to a real opinion and one resolves to nothing. Only a lookup separates them. It also explains why the problem hits busy small and midsize firms hardest. The fabricated citation is engineered, by the nature of the model, to survive exactly the review a solo or small team has time for: a glance at the format, a nod at the plausible name, and into the brief it goes. The deeper mechanics of how these fabrications enter filings, and what they cost, are covered on AI hallucination in legal filings.

How large is the fabricated-citation problem?

The numbers are not fringe. A Stanford RegLab study, published in the Journal of Empirical Legal Studies in 2025, found that leading purpose-built legal AI research tools still hallucinate on a meaningful share of queries — roughly 17 percent for Lexis+ and around 33 percent for Westlaw AI — with general-purpose chatbots performing far worse. These are the tools marketed specifically to lawyers, with retrieval built in. The takeaway is not that the tools are useless; it is that a nonzero, meaningful hallucination rate is the baseline you are working with, and that unverified output cannot be trusted at the point of filing.

Public trackers such as the widely cited AI Hallucination Cases database maintained by Damien Charlotin now catalog well over a thousand filings worldwide involving AI-fabricated or misused citations, and the count keeps climbing. The landmark matter, Mata v. Avianca, drew a 5,000-dollar sanction for citations to cases that did not exist; some later matters have reached six figures. The reputational cost — a published opinion naming you, an adversary who now reads every one of your citations with suspicion — is harder to price and often larger than the sanction itself.

Why does a signed certificate beat a dashboard flag?

Most citation tools flag a suspected problem and produce nothing durable. A red badge appears in a dashboard, you resolve it, and the badge disappears. When a judge's standing order later asks what you did to verify your authorities, "our tool flagged nothing" is an assertion — a claim about a screen no one else saw, in software the court cannot inspect, at a moment that has already passed. It asks the reader to trust you, and increasingly, courts are not inclined to take that on faith.

A certificate is evidence instead of an assertion. Each result is signed with post-quantum cryptography and sealed to an RFC 6962 transparency log, so a court, opposing counsel, or your insurer can confirm the check ran — what was checked, and that the record has not been altered since — without trusting you or us. The distinction is the whole point: a flag is a transient state inside a vendor's UI, while a certificate is a portable, independently verifiable artifact that outlives the session and travels with the matter. You can see how anyone verifies one on the transparency page.

Flag versus certificate — what is the real difference?

The table is not a knock on flags as a working aid — a flag is a fine way to catch a problem in the moment. It is a statement about what evidence survives the moment. When the question is no longer "did the tool notice" but "can you prove to a third party what you verified," only the right-hand column answers it.

Where does the checker fit in your workflow?

Nothing here asks you to abandon the tools you already use or to route privileged work through a black box. The checker sits between drafting and signature as a verification step, not a research replacement. Because it operates on citations rather than on your work product, it fits a small firm's workflow without a litigation-support team to run it. How the underlying material is kept separated from the verification path is described on privilege isolation.

1. Draft with any toolUse whatever research or drafting assistant you prefer. The checker is model-agnostic; it does not care how the citation was produced, only whether it resolves.
2. Run the free existence checkPaste a citation, or a set of them, into the tool below. Each is resolved against public case-law by its exact reporter citation. Non-resolutions are your stop-and-verify list.
3. Certify before you signFor a filing, run full certification: existence, quotation accuracy against the published opinion, and good-law standing pulled from your citator. Each dimension is recorded separately.
4. Seal and keep the recordThe result is signed and sealed to a transparency log. If a standing order or an insurer later asks, you produce a certificate — not a recollection.

How is the record sealed and independently verified?

Every certificate is signed with composite post-quantum signatures — ML-DSA and SLH-DSA, the algorithms standardized as NIST FIPS 204 and 205 — and sealed to an RFC 6962 transparency log. Composite signing means the record carries two independent post-quantum signatures, so it stays verifiable even if weaknesses are later found in one scheme. The transparency log gives each certificate an inclusion proof: anyone can confirm the record was entered at a specific point and has not been altered since, without contacting us and without trusting us.

One deliberate design choice matters for any lawyer weighing this: the records store digests, never filing substance. A digest is a cryptographic fingerprint of what was checked, not the brief, the quotation, or the client's matter. That is what lets a court or insurer verify that a check occurred and what its result was, while the underlying work product stays with you. Verifiability and confidentiality are not in tension here, because the thing that gets published is a fingerprint, not the document. More on the cryptographic posture is on the security page.

What is enterprise-grade protection, firm-sized?

Large firms manage citation risk with dedicated litigation-support staff, internal review protocols, and the budget to build accountability into their process. A solo practitioner or a five-lawyer firm carries the same duty of candor to the court and the same malpractice exposure, without any of that infrastructure. The gap is not one of care — small-firm lawyers are not less diligent — it is one of tooling and time. That gap is exactly where the fabricated citation slips through.

The checker is meant to close that gap without adding headcount. The same verifiable accountability a large firm would build in-house is delivered as a checkpoint you switch on: run the existence check for free, certify a filing when it counts, and keep a record you can produce on demand. Enterprise-grade protection, firm-sized, means the control is proportionate to the practice — no team to staff, no platform to administer, just the verification step at the moment it matters most.

What does this tool not do, and not claim?

• MythIt guarantees an AI that never hallucinates.

  TruthIt never claims to be hallucination-free. It certifies which of your citations are real, accurately quoted, and good law. The model is not the thing we vouch for; the verified citations are.

• MythThe existence check confirms a case is good law.

  TruthExistence and good-law are different dimensions. The instant checker confirms a case exists; good-law standing comes from your citator (KeyCite or Shepard's) during full certification. We never guess good law.

• MythIt resolves every real case, so a non-resolution means forgery.

  TruthThe resolver is only as complete as the case-law corpus it reaches. A very recent, sealed, or non-public opinion may not resolve although it is genuine. A non-resolution is a prompt to verify, not a verdict.

• MythIt reads your brief and stores your filing.

  TruthThe certified record stores digests — cryptographic fingerprints — never filing substance. Verification confirms what was checked without exposing the underlying work product.

• MythIt is legal advice.

  TruthRankShield is a security vendor. The output is verifiable technical evidence, not a legal opinion, and it does not replace your professional judgment or your independent duty to verify authorities.