How we test the tribunal

Selected evaluations of our system.

Selected results

67/67

Correct outcomes

We built the set from real cases, removed party names and source-identifying details, and gave our system deidentified briefs or adapted case records. It arrived at the correct outcome in all 67 cases.

View 67 inputs and outputs · JSON

0/88

Hallucinated authorities

We found no hallucinated authorities among 88 external citations across 40 audited opinions.

View public artifact · JSON

32/32

Stable under controlled changes

Changing party names, amounts, fact order, or which party filed first did not change the result in any of 32 comparisons.

View public artifact · JSON