The crash‑test lab
for AI

We evaluate AI agents on real professional work and show you where they fail.

Request a pilot See the results

The Xzar report · October 2026

Frontier AI loses to a rule of thumb

One-line rule of thumb0.36

Frontier AI + tools0.22

Frontier AI0.12

60 real patent decisions made after the model was trained. Scores 0 to 1; thin lines show 95% ranges.

47 of 60cases where the rule beat the AI

92%of the AI's mistakes let a rejected claim through

37 ptsmedian gap in grant rate between human examiners

65%of contested rejections upheld on appeal

How it works

Real cases. Real tools. Exact grades.

A real case, cut at the decision

Your AI sees only what the professional had.

It works the case

Search, rulebook, then a decision.

Graded against what happened

Outcome, reasons, evidence. No AI judges.

One claim. One decision. One score.

Thousands of these become your AI's scorecard.

EXAMPLE TASKILLUSTRATIVE

Claim 1. A door hinge comprising a leaf, a pin, and a damping sleeve disposed about the pin…

Your AI

Allowable

No rejection

Actual decision

Rejected

Obvious

decision 0 · reasons 0 · evidence 00.00

Work with us

Know how your AI really performs

Pilot

30 days. Your AI, our cases, a full scorecard and failure report.

Request a pilot

Ongoing testing

Fresh cases every month, after every model update.

Talk to us

API

Get a key and $5 of free credit. Run your agent today.

Get an API key

Government

Briefings and evaluations. Runs offline.

Request a briefing

Start with the free sample report. Get it

Method

We sell the verdict, not the test materials: evidence-backed proof of how well your AI performs.

Exact gradingSame answer, same score. Every time.
After the cutoffCases decided after the model was trained.
Beat the shortcutRules of thumb scored on every test.
Private by designNames removed. Results stay yours.

Contact