Ask a chatbot a legal question and it writes first, then decorates the answer with citations — real or not. JuriScripta is built the other way round. It finds the cases, confirms each one in the court record, and reads the full opinions. Only then is the AI allowed to write, and only from that text.
Law first, AI second · patent pending
The draft is assembled from the text of opinions it has already confirmed and fetched. A case it could not find in the court record is kept out of the drafting material, every citation in the finished draft is checked against the confirmed list, and you are shown the cases it wanted to cite and was refused.
Every AI model stops learning on a fixed date, and knows nothing decided since. JuriScripta follows the citation network forward: it finds opinions that cite your authorities and were filed after that date, reads them, and marks them as new. Recent decisions, not the model's memory of older ones.
A second model, built by a different company, reads each opinion again and tests every claim the draft makes about it. Two models rarely share a blind spot. What fails is flagged in red at the top of your brief, not buried.
The problem
A language model asked a legal question will produce a confident answer with citations attached. Some of those cases are real. Some are invented, and some are real cases that say something else. Courts have sanctioned lawyers for filing them, and a person representing themselves has even less chance of catching one.
Checking a finished draft catches the invented case. It does not catch the real case that was misread — and that is the more common error. Catching it takes the opinion's text, and something willing to read it twice.
One run, in order
What comes back
Every brief arrives with the list of cases it was allowed to use, the cases it proposed and was not allowed to use, and a table of each claim with a verdict. Unsupported claims sort to the top. Nobody should have to scroll to find the one the system itself flagged.
Illustration
3 claims checked · 2 grounded · 1 not supported
| Claim | Authority | Verdict | Why |
|---|---|---|---|
| A notice to appear that omits the hearing time deprives the immigration court of jurisdiction. | Pereira v. Sessions, 585 U.S. 198 (2018) | Not supported | The opinion decides the stop-time rule and says its holding is narrow. It does not address jurisdiction. |
| A notice to appear that omits the time or place of the hearing does not trigger the stop-time rule. | Pereira v. Sessions, 585 U.S. 198 (2018) | Grounded | Stated as the holding. |
| The required information must arrive in a single document, not across several. | Niz-Chavez v. Garland, 593 U.S. 155 (2021) | Grounded | Stated as the holding. |
Who it is for
Ask in plain language. The answer comes with the actual opinions linked, so you can read what a court said rather than take a summary on faith.
A first research pass on more matters than there are hours for, with the weak points marked before an attorney picks it up.
Federal court and Board of Immigration Appeals decisions, and a search for opinions newer than any model's training data.
The limits
It is a research and drafting tool, not a lawyer. A brief takes about fifteen minutes, because it reads the opinions.
Limited beta
Each brief costs real compute, so access opens in small groups. Tell us who you are and what you are working on, and we will be in touch.