LBLegalBenchCase review
LEGAL PREDICTIONCalifornia Supreme Court
BENCHMARK / LEGAL PREDICTION / CASE 01
DECIDED CASE · PRE-DECISION MODEL TEST

J.M. v. Illuminate Education, Inc.

A student data-breach case testing whether an AI could predict the California Supreme Court’s ruling from the filings available before judgment.

Docket S286699Opinion filed May 14, 2026Author Justice Goodwin Liu
SUPREME COURT DISPOSITIONReversed
and remandedJudgment of the Court of Appeal
PRE-DECISION SOURCES—Filings and precedents
SAVED MODEL RUNS—Completed responses
BEST COMPLETED FORECASTGemini 3.1 ProMatched the main disposition
RECORD BOUNDARYMay 13, 2026Judgment held out of model input
CASE AT A GLANCE

What the models had to predict

Opinion issued
The dispute concerns a school technology vendor that stored student information, including medical data. After a breach, J.M. brought claims under California’s Confidentiality of Medical Information Act (CMIA) and Customer Records Act (CRA). The Supreme Court reviewed whether the complaint stated either claim.
01 / CMIA COVERAGE

Is Illuminate a health care provider?

The Court held that J.M. had not sufficiently alleged Illuminate was a “provider of health care” under Civil Code § 56.06.

Claim insufficient
02 / CONFIDENTIALITY

Must anyone actually view the data?

The Court said no. A significant risk of unauthorized access or use can breach confidentiality under § 56.101.

Legal rule clarified
03 / CUSTOMER RECORDS

Was J.M. Illuminate’s customer?

The Court held that J.M. had not sufficiently alleged customer status under the CRA.

Claim insufficient
MODEL CHECK

Prediction versus opinion

Among eligible pre-decision models, Gemini’s response got the reversal and failure of both pleaded claims, though it missed important details of the reasoning. GPT‑5.5’s response kept a CMIA claim alive and predicted a different disposition. The later Claude Opus 5.5 matched the main outcome in a retrospective run, but missed the Court’s confidentiality rule.

Test conditions. The models saw the same prompt and excerpts from 23 pre-decision sources, capped at 10,000 characters per source. This viewer shows the complete source files for review. The runs therefore test responses from excerpts rather than the full unabridged record. Claude Opus 5.5 was released after the opinion and is shown only as a retrospective check.