How gno.law works, in plain words

Draft policy, 2026-10-12. Questions and corrections: legal@gno.law.

What this site is

gno.law is a public record of court opinions, searchable, with a profile for every judge built from that record: how many of their decisions went up on appeal, how those appeals came out, and in which areas of law. It starts with San Diego — the California Court of Appeal, Fourth District, Division One, and the trial judges whose work it reviews, family court included.

Where the cases come from

Every opinion comes from CourtListener, the Free Law Project's open archive of court records, published under the CC Public Domain Mark. gno.law copies the public bulk files on a schedule and keeps them as they are; it does not edit an opinion. The trial judge, the outcome and the area of law are read from the opinion's own first lines by a parser, and the parser's coverage is measured on every run and shown on the status page.

How an analysis is made

A language model reads one opinion at a time and writes down, in a fixed form: a two-sentence summary; each trial-court ruling the appellate court reviewed, the standard of review, the court's view of it and the model's own view with its reasons and a quote from the opinion; the authorities the court relied on, any it misapplied, and any binding published California authority decided earlier that the model thinks the court should have addressed; a view of the appellate opinion itself; a confidence; and flags.

• Names are removed before the model reads. The trial judge's and the justices' names are replaced with their roles, so the model judges the reasoning on the page and not the person.
• Every quote is checked against the opinion's text. A quote that is not in the opinion is marked as such.
• Every authority is checked against the reporter citations: it must exist, be decided before the opinion, and be published (an unpublished opinion is no authority under rule 8.1115). A "missed" authority must not already be cited by the opinion. Anything that fails is dropped from the page and from every count.
• The court's view is always shown beside the model's, never replaced by it. An analysis is one reading of a public opinion; the opinion is the record.
• The model and the prompt are named on every analysis, with the date.

Drafts and review

A new analysis is a draft. Drafts are not shown to readers. The first analyses are read by a person before the public switch is turned on; after that, analyses are published as they are made and the review continues on the ones readers report.

What a judge's profile shows

Counts from the record: cases reviewed, authored or sat on; outcomes on appeal (affirmed, reversed, modified, dismissed); areas of law; years. A reversal rate appears only from ten decisions on the merits, with the number of decisions beside it. Analysis counts (rulings the model agreed with, rulings it did not, authorities it found missed) appear only from ten analysed cases.

The grade

A judge's profile carries a letter, A to F with ++ / + / − / −− within each band, per model — the model is named beside it and a switch shows another model's grade when one exists. It is made on one stated assumption: that the model's reading is right. A trial judge's grade is the share of their reviewed rulings the model finds they got right: a ruling the Court of Appeal upheld and the model agrees with, or one the court faulted and the model defends. A justice's grade is made from the opinions they authored: sound counts one, sound with gaps a half, questionable nothing. Ten scored rulings or opinions are needed for a standing letter; with fewer the letter is provisional — shown with a dashed mark and the count, ranked against the judges who clear the bar, and never counted among them. The letter is on a curve — the judge's rank among all the judges graded under that model: A is the top fifth, B the next, C, D, F the bottom fifth, with ++ / + / − / −− within each band — because the raw shares cluster high (appeals mostly affirm and the model mostly agrees). The raw share, the counts and the rank sit beside every letter.

How often the model reads the court right

The grade assumes the model is right. This is how often it is, measured against the Court of Appeal's own dispositions over the analyses published here: a reading agrees when the model finds no error where the court affirmed, or finds one where the court reversed (dismissals carry no signal). Authorities are verified against the reporter citations — real, decided in time, published.

Recomputed hourly from the published analyses. A model that reads the disposition right less often grades judges less reliably; the profile's switch shows each model's grade.

Disputes and corrections

Anyone can report an analysis from the page it is on. A reported analysis goes to a review queue; a disputed one is hidden while it is checked. A judge, a party or counsel who believes an analysis is wrong can write to legal@gno.law. A judge's own response to an analysis, when one is sent, is shown with the analysis as it was written, unedited.

Takedown

The opinions are public records and stay. An analysis is taken down when it is shown to be wrong, or when a court orders it. Write to legal@gno.law with the case and the reason.

Privacy

gno.law keeps no accounts and sets no tracking cookies. Reports are kept with the analysis they concern. Nothing on this site is legal advice.