Every authority, with its citation attached.

The problem we designed around

Large language models produce citations that do not exist. They are fluent, plausible, correctly formatted, and entirely invented. This is not a rare failure — it is a well-documented property of the technology, and it has already reached the courts.

In Mata v. Avianca, the Southern District of New York imposed Rule 11 sanctions on lawyers who submitted a brief containing non-existent judicial opinions produced by ChatGPT, complete with fabricated quotations and fabricated internal citations.[1] That case is now the standard cautionary tale, and it involved trained lawyers.

Self-represented litigants are more exposed, not less. They are less likely to recognize a fabricated citation on sight, and less likely to have anyone check the filing before it goes in. Any honest account of AI in the hands of a pro se filer has to start here.

A pleading must be signed. By signing, an unrepresented party certifies to the court that the legal contentions are warranted by existing law and that the factual contentions have evidentiary support. Fed. R. Civ. P. 11(a)–(b) (paraphrased)[2]

That certification is made by you, personally. It cannot be delegated to a vendor. So the only defensible way to build this software is to make verification easy, obvious, and unavoidable — rather than to ask you to trust a machine you have no way to audit.

How it is built

Retrieval over the published record, not recall from memory

The engine does not answer from a model's parameters. It searches a corpus of published material — court opinions, statutes, and the rules of civil procedure — retrieves the passages that match, and puts those passages in front of the model as the material to work from. The answer is grounded in text that exists, because the text is fetched before the answer is written.

Every authority resolves to a public source

Each citation the engine surfaces links to the public record — for federal case law, to the docket and opinion as published. If an authority cannot be resolved to a source, it does not get shown to you as an authority. A citation you cannot click is a citation you cannot check.

The verification gate

Before a document is finalized, the engine re-checks every citation in it against the record and flags anything that does not resolve. This is a floor, not a guarantee. It catches citations that do not exist. It does not and cannot tell you that a real case means what you hope it means — that is legal judgment, and it is exactly what we do not do.

What we ask of you, without apology

• Read the source before you rely on it. Every authority is one click from the record. Open it.
• Verify all citations independently before filing. Do not take ours, or any tool's, on faith.
• Check your judge's standing order on AI disclosure. A number of judges now require filers to disclose the use of generative AI, and the requirements differ from judge to judge and change often. This is on you to check for your court.
• If it matters, consult a licensed attorney. Software has a ceiling. Yours may be a case that needs a lawyer, and there is no shame in that being the answer.

What "inference" means here — and what it does not

Our name is a description of a method, not a claim of expertise. An inference is a conclusion drawn from premises: the facts, the authorities, a line of reasoning, a conclusion. The mark is a set of scales for the same reason — a judgment is only ever as good as what gets placed in the pans.

Applied to law, the premises are the facts you supply and the authorities the engine retrieves. The conclusion is the one you draw. We do not draw it for you, we do not tell you it is a good one, and we do not warrant that any document is valid, sufficient, or will be accepted by any court.