The leaderboard

We measure the gap, then help close it.

Our benchmarks test models on the work people actually need done. The results tell us which data to build next.

Featured benchmark

Citizen-Bench

How well do AI models help ordinary citizens with their administration? We ask the same real situations in several countries, in each country's language.

JusticeTaxesCompany creationBankingCivil status
450prompts
5domains
1gold answer per prompt

Each answer is scored by an AI judge against a gold answer drawn from official sources.

Citizen-Bench: overall score

Share of answers judged correct and complete. All countries, all domains.

  1. [Model name][XX.X%]
  2. [Model name][XX.X%]
  3. [Model name][XX.X%]
  4. [Model name][XX.X%]
  5. [Model name][XX.X%]
  6. [Model name][XX.X%]
White line = 50%Public split. Results to be published.
Methodology

Same questions. Official answers. One rule.

Same situations in every country

Each situation is written once, then adapted to local law and asked in the country's own language. Scores compare like with like.

Gold answers from official sources

Every gold answer is built from government and other official sources, and cites them.

Public and private splits

A public split is published with the leaderboard. A private split is held back to prevent contamination and is available to labs on request.

The scoring rule

An AI judge compares each answer with the gold answer. An answer counts as correct only if it is accurate, complete and would not mislead the citizen.

Test your model on the private split.

Full prompts, gold answers and per-country results.

Get the private split