How do you write an AI quality requirement?

Name the characteristic, state the task and the context, give a measure and a method, set a threshold, state the conditions the result holds under, name the evidence and give an owner. A requirement that repeats a definition without a measure cannot be evaluated, and a single accuracy figure is not a correctness requirement.

Start with the definition

Take the characteristic from the model and read its definition. The definition tells you what the quality is; it does not tell you what good looks like for your system. That is the work.

Add the five parts

A task, a measure, a method, a threshold and the conditions. For functional correctness that means the error rate and the evaluation set as well as the success rate. For robustness it means a labelled case in each named category. For transparency it means the information and the audience.

Name the evidence

Before the requirement is approved, name the artefact that will show it is met and where it is kept. A requirement whose evidence cannot be named is not ready.

Give it an owner and a date

A requirement with no owner is not controlled, and one with no review date drifts. Put both on the row, and add the trigger for an earlier review, such as a model change.

Use the model for completeness

When the set is written, compare it against the characteristics in the model and record the include or exclude decision for each. The free check does a first pass.

Where to go next

• Run the free readiness check on your own requirement set.
• Read the model element the answer rests on.
• See the kit and the watch.