About

The systems are already here. The testing is not.

Loan decisions, fraud holds, clinical triage, identity verification, benefit eligibility, and hazard warnings across Africa are increasingly produced by models. Very few have been evaluated by anyone without a commercial interest in the result, and almost none have been evaluated on data that resembles the conditions they operate in.

The gap is not a shortage of models. It is the absence of an instrument between the people building these systems and the people they are used on. Materials get tested by a materials laboratory. Drugs get tested by someone other than the manufacturer. Deployed AI, in most of the markets we work in, does not.

We are a testing laboratory, not a consultancy and not an AI company. We do not build the systems we evaluate, we do not sell tools to the organisations we evaluate, and the value of everything we produce rests on our measurements being reproducible by someone who does not trust us.

Across Africa, with the ground truth held locally.

We take engagements across the continent. Africa is not one market and we do not treat it as one: a credit book in Lagos, a clinic network in Nairobi, and a registration programme in Accra fail in different ways, and a test set built for one establishes nothing about the others.

Field collection, annotation, and domain expertise are held locally to the system under test, because a test set that resolves local conditions cannot be assembled at a distance from them. Where we cannot build that ground truth in a given market, we say so and decline the work rather than substituting data from somewhere else.

What holds regardless of who is paying

• Every completed evaluation is published, whatever it concludes.
• No operator funds its own evaluation. Ever.
• The harm threshold is defined before analysis begins, never after.
• The protocol is pre-registered and timestamped before the study runs.
• Every figure carries an interval, a denominator, and the version and date it was measured under.
• The access tier we worked under is stated on the front of every report.
• The subject sees the finding first and has at least twenty-one days to reply. It may not block publication, and its reply is published unedited.
• The notebook, the results file, the harness, and the test set are published wherever law and licensing allow.
• Corrections are published as prominently as the original.
• Every funder is named on every finding, with amounts.
• If anyone reproduces our work and gets a different answer, we publish that too.

These are set out in detail on Independence.