Adversarial validation for legal work
A private workspace where a swarm of AI agents attacks your draft: pulling the actual case behind each citation, catching fabrications, and stress-testing the argument the way opposing counsel will. Over whatever drafting tool your firm already uses.
In development. Building with a founding cohort of firms.
The exposure
One hallucinated citation in one filing is enough. The sanctions, the malpractice exposure, and the reputational hit land on the lawyer who signed it, not on the tool that wrote it.
You cannot guarantee that no one at your firm is quietly using AI, and you cannot read every draft. What you can do is put an adversarial check between the draft and the filing, so a fabricated citation is caught inside your walls instead of in open court.
The first sanctions came in Mata v. Avianca (S.D.N.Y. 2023). They did not stop there.
The loop
Bring the draft in from whatever tool wrote it. LawNet is model-agnostic: it validates output, it does not compete to produce it.
Four lenses work the draft in parallel. Every finding is traceable to a real source the machine actually pulled.
Accept or reject each proposed edit, then run it again. The draft is hardened until it survives the attack.
You get a validation report and a timestamped record that the work was checked: a record of what was examined, not a score and not a clearance.
Suggest-only, always. Every change is a proposed edit you accept or reject. LawNet never edits or files anything on your behalf. You remain the author of record.
The swarm
Pulls the actual authority and confirms it exists and that the quoted language is really in it: it does not reason about whether a case is real, it looks it up. Whether the case truly stands for what the brief says is flagged for your review: a screen, not a verdict.
Fabricated quotes, invented holdings, made-up pin cites, misquoted statutes: the failure mode that gets lawyers sanctioned.
The strongest counterargument, the objection that lands, the thin spot in the record, and the case they will cite to distinguish yours.
Whether the standard of review fits the court and posture, whether the elements are complete, and whether the burden is allocated correctly.
What a finding looks like. The report speaks one color per state, learned once and read at a glance.
Why a separate layer
An AI that grades its own drafting is grading its own homework.
LawNet does not write your brief. That is the point. It validates the output of whatever drafting tool your firm already uses, from the outside.
A Stanford study measured the leading legal research tools producing hallucinated content in roughly seventeen to thirty-three percent of queries, even as their makers marketed them as hallucination-free. Verification has to come from somewhere the drafting does not.
Calibrated honesty
Incumbents marketed hallucination-free and hallucinate anyway. We tell you exactly what was checked, and what was not. Under-claiming is the product.
A citation is verified only when the machine pulled the actual authority and it held. Not because it looks right, and not because a model is confident. Green is earned.
A possible mischaracterization or a weak spot is surfaced as a candidate for your judgment. A screen that points you at what to examine, not a verdict.
We will never tell you a document is safe to file. LawNet reports what it checked. Whether the brief is ready is your call, and it stays yours.
From the makers of SciNet, the adversarial engine behind a living graph of scientific claims, where AI agents publish, review, and reproduce results in the open. The same engine, now being pointed privately at your practice.
Visit SciNetEarly access
LawNet is early. We are working with a small group of firms to ground it in real practice before opening it more widely. If liability from AI-assisted work sits on your desk, we would like to talk.