A sequence the lab uses. It is a method note, not a certificate and not a client case study.
Most "AI audits" start with a framework and never reach the book. A credit-model audit starts with a claim: who is scored, what decision is taken, and what "good" means in that market.
Name the population (new-to-credit retail in Kenya; salaried UAE residents; SME overdraft). Name the decision (approve, price, limit). Name the threshold. If the owner cannot state this, the audit stops at a model card and says so.
The test set has to look like the book the model will see. Western bureau samples do not stand in for thin-file African or GCC books. If the right data cannot be shared, say what cannot be tested.
Compare approval or score distributions across groups that exist in that market: nationality, gender, language, neighbourhood, employer type. Report the gap, the sample size, and the test. Do not collapse this into a single "fairness score".
Split by time window and by a second population slice. A model that only holds on last year's originations is not ready for a second country or a downturn.
Pick a handful of real decisions and reconstruct them in terms a credit committee can read. If the story depends on a feature the bank is not allowed to use, that is the finding.
Code, seeds, and the exact slices. SAFE Labs will not put its name on a result a second reader has not reproduced. See independence.