How they have reasoned
What has mattered to this judge on your issue, quoted from their own opinions and tentative rulings — never paraphrased into a claim you cannot check.
Every court has the same reporters. Only one of them has the judge you drew on Tuesday. Judge Predict reads what they have actually written — and lands every claim on the words.
Not what the law says. What this judge has said, in the posture you are about to be in, and whether your file can carry it.
What has mattered to this judge on your issue, quoted from their own opinions and tentative rulings — never paraphrased into a claim you cannot check.
The elements you must establish, preferring how this judge states the standard, each checked against your actual file. Gaps are named, with the declaration or deposition topic that closes them.
Written in this judge’s doctrinal language, citing the authorities they rely on. Assertions the record cannot support are marked in place, never written around.
This is output recorded from an actual run against a fictional judge and a corpus of her rulings — replayed here, not regenerated. Press run, then select any citation marker to see the source it resolved to.
Recorded against a local model on a fictional corpus, so it shows the mechanism rather than the ceiling — the product runs on Claude Opus 5, and the judge, the parties and every ruling here are invented.
It will never say 73% likely to grant.
Published opinions over-represent contested, appeal-worthy matters, and most trial-court rulings are never published. A rate computed over what you can download is not the judge’s rate.
France criminalised judge-outcome analytics in 2019. Other bars have taken a dim view. A number is a liability in a way a quotation is not.
“Lead with the fee-splitting term — this judge has twice called a one-sided carve-out the clearest indicator” goes in a brief. A number goes nowhere.
An opinion the judge merely joined is weak evidence of their thinking. One they wrote is strong. A dissent they wrote is stronger still as to their own view, and is not law. Every passage is labelled so the three can never be conflated.
A ruling about “a one-sided termination provision” answers a question asked about “the right to cancel”. Keyword search finds neither in the other. Both retrievers run, and are fused before anything is weighed.
Offsets are preserved against the stored text, so a citation resolves to a precise span of the real document rather than a gesture at a case.
Judicial passages are marked as authority. Client documents are marked as evidence of a fact and never as precedent — and the caveat sits inside the text the model reads, so it cannot be separated from what it qualifies.
A trial judge’s tentatives are the best guide there is to how they decide the motions you are about to file — and they are not precedent. Used as evidence of reasoning, never cited as authority.
Your firm runs the server. Client documents never leave it except as retrieved passages sent to the backend you choose — and with a local model, not even then.
| Capability | Desktop | Web | Android |
|---|---|---|---|
| Analysis, motions, drafting | Yes | Yes | Yes |
| Case-file upload | Yes | Yes | Yes |
| Scans read by OCR | Yes | Yes | Yes |
| Ruling-corpus import | Yes | — | — |
| Fully offline model | Yes | — | — |
The corpus importer reads a server-side directory, so it is desktop-only by design: exposing it over HTTP would be arbitrary file disclosure.
The useful demonstration is not ours. Send a judge you appear before and the motion you are working on, and we will load their rulings and go through it live.
Request a walkthrough