Work with us
Know what your AI gets wrong before your users do.
We help organisations in New Zealand find out how AI handles their subject, and use it safely.
What we offer
Four ways to work with us
Each one uses the same method as NZBench: questions with official answers, checked in more than one way, with the results and data handed over.
Test AI on your own content
A set of questions drawn from your own pages, each with an official answer. We run the AI models you care about against it.
- Results by model, topic and how often the rules change
- Which answers give another country’s rules or an old figure
- A test you can run again after every change
Check a chatbot before launch
We test your assistant as your users will meet it, not the model on its own.
- Where it gets New Zealand wrong
- Where it hedges instead of answering
- What to fix first
Get your model into NZBench
Want your model in the next edition? We will happily run it. Cover the cost of running it, or give us the credits, and it goes through the same test as every other model.
Use NZBench yourself
The public questions, answers and code are released under open licences (CC BY 4.0 for data, MIT for code), at https://doi.org/10.5281/zenodo.24364019 and on GitHub. Use them to test your own systems.
How it works
Simple, and quick
Talk to us
Tell us what you are building, who uses it and what worries you.
Agree the scope
Which topics, which models or systems, and what you need to see.
We run the test
Official answers, AI grading and checks that use no AI.
You get the results
A clear report, the full data, and a test you can run again.
Putting AI in front of New Zealanders? Test it first.
Tell us what you are building and who will use it. We will tell you how we can help.
Email ushello@ravenscar.ai