# Xzar

> Xzar tests AI on real professional decisions made after the AI was trained, and grades every answer exactly against what the professional actually decided. The method works in any field where real decisions are made and recorded; Xzar builds tests for a customer's field on request. First published test: patent examination (first decisions on each claim, decisions after the applicant's response, and appeal board rulings).

Key results (October 2026):
- On 60 real patent first examinations decided in 2025–2026, a frontier AI model scored 0.12 (95% range 0.09–0.16). A one-line rule of thumb (reject every claim as obvious) scored 0.36 (0.30–0.41) on the same cases and beat the AI on 47 of 60.
- With search and the examination rulebook, the same AI scored 0.22, still below the rule of thumb.
- Human examiners in the same examining unit differ by a median of 37 points in grant rate (592 examiners, 24 units); on 9,083 contested claims the appeal board agreed with the examiner 65% of the time.

## Docs
- [The Xzar Report: frontier AI on real patent decisions](https://xzar.si/report) (also as PDF: https://xzar.si/xzar-report.pdf)
- [API and MCP plug-in](https://xzar.si/docs.html): episodes, tools, answer format, scoring, task sets (first action, response, appeal)

## Get started
- [Get an API key](https://xzar.si/signup.html): $5 of free credit, then $2.00 per scored task (prepaid credit packs of $100, $500 or $2,000); run your agent through the API or the `xzar-mcp` plug-in (`uvx xzar-mcp`)
- [API Terms of Service](https://xzar.si/terms.html) · [Privacy Notice](https://xzar.si/privacy.html)

## Work with us
- [Request a pilot, the free sample, or a test for your field](https://xzar.si/#contact)