Engagements
Organisations retain us for independent AI assurance, which means establishing, in advance of a decision, what a system does on the populations it will actually meet.
Who this work is for
- Operators putting a system into service, and re-checking it at intervals.
- Organisations procuring AI whose performance figures were established on populations other than their own.
- Supervisors and public bodies needing evidence the operator did not give them.
- Investors, acquirers, and insurers conducting diligence on a model.
How an engagement is arranged
Scope, method, harm threshold, and terms are agreed at the outset and recorded, and none of them moves once the measurement has begun. Access is settled at scoping, because it bounds what can be claimed. The tiers.
What a client receives
- A report with uncertainty on every figure and the access tier on its front.
- The results file, the notebook, and the harness, so your engineers can reproduce the measurement.
- The conditions under which the system may be relied upon, and the conditions the evidence does not reach.
- Re-measurement as the system, the population, or the requirement changes.
An engagement establishes how the system we were given behaves under the conditions we could construct. It is not a certification, an accreditation, or a mark of approval, none of which we issue. A client pays for the measurement and never for the result. See Independence.
Beginning a conversation
Tell us the system, who it decides about, and the access likely to be obtainable, and we will tell you what an evaluation could establish and what it could not.