Evaluation, red-teaming and audit of deployed AI. The work is to demonstrate and measure, not to advocate for a vendor or a ban.
The same questions, whether the system is a credit score, a claims model, or a language model answering in Arabic.
Disparate approval, score or error rates across groups that matter in the book: nationality, gender, language, neighbourhood, employer type. We report the gap and the test, not a slogan.
Performance when the population drifts, a feature disappears, or the model is applied in a second country. A model that only works on last year's sample is not assured.
Whether a decision can be reconstructed for a credit committee, a regulator, or a customer complaint. Black-box accuracy without a reconstructable reason fails the test.
Behaviour under messy input, prompt attack, translation, and missing fields. For language models this includes Arabic and African-language red-teaming.
People who have to show that an AI system works as claimed, under local rules and local data.
Credit scoring, collections, fraud, pricing and claims models. The first product is a credit-model audit, because that is where the lab already has the quantitative background.
Model-risk evidence that can be read beside existing credit-risk and conduct files, not a generic AI ethics memo.
Procurement and deployment checks: does the system work in the languages and populations it will actually see?
An outside test before a launch. Later: an open Arabic and African-language eval suite they can run themselves.
We start from the claim the system makes (who it scores, what it decides, what "good" means). We rebuild the test on data that matches that claim, or we say the claim cannot be tested. Results that cannot be reproduced by a second reader do not go out.
Finance, health, governance, climate and geopolitics appear as applications of these tests, not as five other institutes inside the lab.