CHINCHILLA Verified
The release bar is reached in at least one UN language group; the certificate lists those groups.
CHINCHILLA · VERIFIED
CHINCHILLA Verified
An open standard for verifying AI agents: 20 cases in UN languages, a freeze before the run, independent grading and published evidence. Anyone can check the mark themselves.
The registry holds 123 certificates: 30 Verified Multilingual, 93 Verified, 0 Verified Operational.
What the mark confirms
A specific version of the agent went through a frozen set of 20 cases — everyday tasks, missing information, abuse attempts, scams and market rules — in fresh conversations, with no edited answers, and an independent grader confirmed a result at or above the bar: at least 18 of 20 and a mean score of at least 8.5 in every language group named.
What the mark does not mean
- It is not a legal or government approval and not accredited certification.
- It is not a warranty of results: model answers can vary.
- It covers only the verified version, configuration and the language groups named.
- The mark is not for sale and does not depend on payment.
Levels
CHINCHILLA Verified Multilingual
The bar is reached in all six UN language groups: Arabic, Chinese, English, French, Russian and Spanish.
CHINCHILLA Verified Operational
In addition: at least 8 cases with tools — prompt injection, confirmation of irreversible actions, secrets, tool failures; zero failures allowed.
How to submit an agent
- Read the methodology and the badge rules.
- Fill in the application below: agent, language groups, links.
- We agree the 20-case set and freeze the instructions and cases with SHA-256 hashes.
- The run in fresh conversations is graded by an independent grader who took no part in building.
- If the bar is reached, the certificate goes into the public registry with its evidence; otherwise up to 4 fix rounds, each with a full re-run.
Standard documents
Agent verification request
Minimal fields. Your data goes only to CHINCHILLA's CRM and is used to reply to this request.