Historical AI question-bank experiment
Archived May 2026 aggregate scores are not results from official DMV examinations.
Correction notice: October 11, 2026. No human expert review or official approval is claimed.
The preserved May 2026 file records 148 model/jurisdiction aggregate rows for three model identities: openai/gpt-4.1, gemini/gemini-2.5-flash, and nemotron-3-nano:30b. Its rows sum to 4,440 scored responses and 3,194 recorded correct answers.
These totals describe the stored experiment. We have not independently reproduced the runs or verified the original question sample, answer keys, full prompt, raw model responses, or execution harness. The recorded 80% threshold was an experiment convention; it is not a verified passing standard for every state.
The earlier claims that these were official state exam results are withdrawn. No conclusion about real-exam readiness, current model performance, or official approval follows from this archive.
Download historical aggregate records with limitations (JSON). Original fields are preserved inside a clearly labeled historical dataset; they are not current verified claims.
Try the free California 10-question sample or find your official state agency.