Technology · AI and automation · Component

AI evaluation record

The test set, the measured quality, the thresholds accepted, and how performance is watched once the system is live.

The deliverable

What it is

AI systems are evaluated as other measurements are: against cases with known answers, with the types of error kept apart because they do not cost the same. A summarisation tool that omits a clause and one that invents a clause have different consequences, and a single accuracy figure conceals which is happening.

Evaluation continues after launch. Providers update models, inputs drift, and a prompt changed to fix one case can break others; re-running the same test set after each change is the practical safeguard.

One level in

What it is made of

Each element is a constituent part of the component. Follow one to see the attributes it carries.

  1. Test set

    A fixed collection of real cases with agreed good answers, including the difficult and adversarial ones.

    3 attributes: Cases · Drawn from · Version

    Learn
  2. Quality thresholds

    The error rates accepted before launch, set separately for each kind of error according to what it would cost.

    4 attributes: Error type · Accepted rate · Human baseline · Basis

    Learn
  3. Live monitoring

    How performance is watched in use: sampled review, user reports, and a re-run of the test set after each change.

    3 attributes: Review cadence · Current error rate · Within threshold

    Learn

Keep the test set fixed and re-run it after every change. Without that, nobody can say whether an update improved anything.