A fabricated citation is not a formatting error; it is a Rule 11 problem waiting for a show-cause order. Legal citation certification moves the catch to before your signature, and it produces something a flag never does: a durable, independently checkable record of exactly what was verified. This page explains what the three dimensions cover, why a certificate is evidence rather than an assertion, how it connects to the rules that now govern AI-assisted authorities, and — just as important — the honest limits of what any resolver can promise.
What does legal citation certification check?
Every authority in your filing is resolved on three separate dimensions, and a citation passes only when all three hold. The dimensions are deliberately distinct because they fail in different ways: a case can exist but be misquoted, or be quoted perfectly yet no longer be good law. Treating them as one check is how firms convince themselves a brief is clean when it is not.
The output of that resolution is not a flag on a dashboard. It is a certificate — a signed, independently verifiable record of exactly what was checked, against what source, and when. The distinction between a flag and a certificate is the whole point of the product, and the sections below return to it repeatedly.
- Existence — the citation is matched exactly against live case-law by its reporter citation, not just the case name, so a fabricated case is flagged before you sign. This is the primary anti-hallucination signal.
- Quotation accuracy — quoted passages are matched against the published opinion, catching misquotes, paraphrases dressed as quotations, and language that was never in the case at all.
- Good-law standing — overruled, vacated, or superseded authority is flagged, sourced from your firm's citator such as KeyCite or Shepard's. We never guess good law.
Why is a certificate different from a flag?
Most citation tools flag a suspected problem and produce nothing durable. The flag lives inside the tool, disappears when you close it, and cannot be handed to anyone as proof. When a judge's standing order asks what you did to verify your authorities, a screenshot of a green checkmark is worth exactly what any screenshot is worth: it can be edited, staged, or fabricated, and everyone in the room knows it.
Certification produces a different artifact. Each result is signed with post-quantum cryptography and sealed to an RFC 6962 transparency log, so a court, opposing counsel, or your malpractice insurer can confirm the checks ran and have not been altered after the fact. The certificate does not ask anyone to take your word for it. That is the difference between saying "we were careful" and being able to demonstrate it on the record.
For a firm without a dedicated litigation-support desk, this is the practical meaning of enterprise-grade, firm-sized: the same verifiable accountability a large firm would build in-house, delivered as a checkpoint you switch on rather than a department you staff.
What is the difference between existence and quotation accuracy?
Existence and quotation are the two dimensions people most often collapse into one, and the collapse is dangerous. Existence answers a binary question: does this reporter citation resolve to a real, published opinion? A generative model that invents a case will also invent a citation that looks correct, because correct Bluebook formatting is a language pattern the model has learned. Perfect format is therefore never treated as evidence a case is real. It is often the opposite — flawless formatting on a case no one can find is a classic signature of fabrication.
Quotation accuracy is a separate problem that begins only after existence is confirmed. A real case can be quoted for a proposition it never stated, or a passage can be stitched together from fragments that never sat next to each other in the opinion. AI drafting tools are especially prone to this: they reproduce the holding they expect a case to contain rather than the words it actually contains. Certification matches your quoted passages against the published text of the opinion, so a quotation that drifts from the source is caught even when the case itself is genuine. A citation that exists but misquotes still fails certification, because a court does not distinguish between a fake case and a real case bent to say something it did not.
Where does good-law standing come from, and why don't you determine it yourselves?
Good-law standing is the third dimension, and it is the one where honesty about sourcing matters most. Whether a case has been overruled, vacated, superseded, or limited is an editorial determination made by the citator services — KeyCite from Westlaw and Shepard's from LexisNexis — that maintain the treatment histories lawyers already rely on. RankShield does not re-derive those histories and does not guess. We read the signal your firm's citator provides and bind it into the certificate alongside the existence and quotation results.
This boundary is deliberate. A security vendor inventing its own view of whether a case is still authoritative would be worse than useless — it would be a second, unaccountable opinion competing with the tools courts and firms already treat as canonical. Instead, certification records which citator was consulted and what it returned, so the good-law dimension on your certificate is traceable to a source a court recognizes. If a citator result is unavailable for a given authority, the certificate marks that dimension incomplete rather than passing it silently, so you always know exactly which of the three checks was actually verified.
How does FRCP Rule 11 make certification a forcing function?
Federal Rule of Civil Procedure 11 requires that every filing be the product of a reasonable inquiry, and that the legal contentions in it be warranted by existing law. When a lawyer signs a brief, that signature is a representation to the court. A fabricated or misquoted citation is not a stylistic slip; it is a breach of the duty of candor and a failure of the reasonable inquiry the rule demands. That is why the sanctions in these matters attach to the signer, not the software.
Certification is built around this obligation. It converts "reasonable inquiry" from a vague standard you hope you met into a concrete, timestamped step you can show you performed. The certificate is not a defense that the argument was correct — Rule 11 is about the inquiry, and so is certification. What it establishes is that before you signed, every authority was resolved against live case-law, every quotation was matched to its opinion, and every good-law signal was pulled from your citator and recorded. When a court asks what your inquiry consisted of, that record is the answer.
A proposed amendment to Rule 11 addressing AI-generated authorities is currently before the Advisory Committee on Civil Rules. It is a proposal, not yet a rule, and should be treated as one. But the direction is unmistakable: the obligation to verify authorities before filing is being made explicit rather than left to inference, and firms that already produce verifiable evidence of that verification are positioned for whatever the final language requires.
What do court AI standing orders now require?
Beyond the federal rules, individual judges have begun issuing standing orders that govern AI-assisted filings directly. These orders are the sharpest forcing function firms face today, because they carry immediate, court-specific consequences and they name the exact representation a lawyer must make.
A representative example is the standing order issued by Judge Nina Wang of the District of Colorado, effective December 1, 2025, which requires counsel to certify that any AI-assisted authorities in a filing are non-fictitious and have been reviewed by a human. Read that requirement closely: it is not enough for a tool to have run. Counsel must be able to attest that the authorities are real and that a person checked them. A certificate that records the existence, quotation, and good-law results for every authority is precisely the artifact that lets a lawyer make that attestation from evidence rather than hope.
| What the obligation asks | What certification supplies |
|---|---|
| Rule 11 reasonable inquiry | A timestamped record that every authority was resolved before signing |
| Standing-order "non-fictitious" attestation | An existence result for each citation, matched against live case-law |
| Standing-order human-review attestation | A certificate the reviewing lawyer can bind their review to |
| Good-law duty of candor | Citator-sourced standing recorded per authority, never guessed |
| Proving the inquiry after the fact | A signed, sealed record a court can verify without trusting you |
What does a certification run actually look like, step by step?
- Ingest the filing's authoritiesEvery citation in the document is extracted, along with the quoted passages attributed to each authority. Nothing about the substance of your argument is required for this — certification operates on citations and quotations, not on your theory of the case.
- Resolve existence against live case-lawEach citation is matched by its exact reporter citation against a live case-law corpus. A citation that resolves to no real opinion is flagged as fabricated. The completeness of the corpus reached is recorded on the result, so coverage is never hidden.
- Match every quotation to its opinionQuoted passages are compared against the published text of the resolved opinion. Misquotes, invented quotations, and paraphrases presented as direct quotes fail this dimension even when the underlying case is genuine.
- Overlay good-law standing from your citatorThe treatment history from KeyCite or Shepard's is read and bound to each authority. Where no citator signal is available, the good-law dimension is marked incomplete rather than passed.
- Sign and seal the certificateThe combined result is signed with composite post-quantum cryptography and sealed to a transparency log, producing a certificate whose inclusion proof anyone can verify independently.
How bad is the AI citation problem really?
The scale of this is not speculative, and it is not confined to careless outliers. The failure mode is structural: modern drafting tools produce citations that read as authoritative whether or not the underlying case exists, and a human skim cannot reliably tell the difference.
Measured directly, a Stanford RegLab study published in the Journal of Empirical Legal Studies in 2025 found that leading legal AI research tools still hallucinated on a meaningful share of queries — from roughly 17 percent for one major tool to roughly 33 percent for another. These are purpose-built legal products, not general chatbots, and they were still inventing or misstating authority on between one in six and one in three queries. The takeaway is not that the tools are useless; it is that trusting their output unverified is a bet against odds no litigator should accept.
The downstream consequences are already on the public record. In Mata v. Avianca, the matter that put this problem in front of the profession, the court imposed a landmark $5,000 sanction after fabricated cases reached a federal filing; some later matters have drawn penalties into the six figures. Public trackers now catalog well over a thousand filings worldwide that cited authority which does not exist or was materially misused. Certification exists because this is a recurring, systemic gap — one that a verifiable checkpoint before signature is designed to close.
How is the certificate signed and sealed, and what do you store?
The certificate is signed with a composite post-quantum scheme: ML-DSA and SLH-DSA, the algorithms standardized as NIST FIPS 204 and FIPS 205. Using both a lattice-based and a hash-based signature together means the certificate stays verifiable even if one algorithm family is later weakened — a conservative choice for evidence that may need to be checked years after it was created. The signed result is then sealed to an RFC 6962 transparency log, the same append-only, tamper-evident structure used for certificate transparency across the web. Anyone holding the certificate can check its inclusion proof against the log independently, which is what turns "we checked the citations" into something a third party can confirm.
What the log holds is deliberately narrow. Certification stores cryptographic digests and public citation metadata — the citations themselves and the results of the three checks — and never the substance of your filing. Your argument, your client's facts, and your work product do not enter the log and are not needed to verify a certificate. This keeps the accountability record public-checkable without turning your evidence trail into a disclosure risk. The same discipline underlies privilege isolation across the platform, and you can read more about the signing and sealing model on the security page and in how it works.
What does legal citation certification not do?
-
MythIt guarantees an AI that never hallucinates.
TruthThere is no such guarantee, and we do not make one. Certification does not vouch for the model that drafted your brief. It certifies which of your citations are real, accurately quoted, and good law — the results, not the generator.
-
MythIt resolves against the entire universe of case-law.
TruthIts resolution is only as complete as the case-law corpus it can reach. Coverage is recorded on every result, so an authority in a jurisdiction or source outside the corpus is shown as such rather than silently passed.
-
MythIt tells you whether your argument is correct.
TruthIt certifies process and results — existence, quotation, good-law standing — not the merits. A brief can pass certification and still lose on the law. Certification is about the integrity of your authorities, not the strength of your position.
-
MythIt is legal advice.
TruthRankShield is a security vendor. The certificate is verifiable technical evidence, not a legal opinion, and it does not replace the professional judgment or the human review the rules require of you.
Who is this built for?
Legal citation certification is built for the firms that are actually getting caught: solo practitioners, small firms, and midsize practices that have adopted AI drafting to keep pace but do not have a litigation-support team to independently verify every authority before it goes out the door. The large firms most exposed to this problem can staff around it. The firms most likely to appear in a sanctions order cannot — which is exactly the gap certification is meant to close.
The design goal is enterprise-grade accountability at a scale a small firm can actually operate: a checkpoint that sits between your draft and your signature, produces a verifiable record every time, and asks nothing of you beyond running it. If you want to try the existence dimension on a single citation before adopting the full workflow, the AI legal citation checker runs that resolution instantly, and the broader context on why these failures keep happening is covered in AI hallucinations in legal filings.
