Clause 5.4: Functional correctness
The characteristic describes how far the system gives the correct results with the needed precision.
What clause 5.4 is
The characteristic describes how far the system gives the correct results with the needed precision. For AI systems, and especially probabilistic machine learning, correct results are not guaranteed in every circumstance, so both correct and incorrect outcomes have to be measured. The error rate and the conditions the result holds under are part of the requirement. Correctness also trades against other qualities such as performance efficiency and robustness.
What the 2023 edition said
The first edition amended the definition inherited from the base model to add the note that AI systems do not usually provide correctness in all observed circumstances, and pointed to the evaluation methods for machine learning.
What the second edition changes
The definition is restated against the 2023 base model. The substance is unchanged: measure correctness and incorrectness carefully, and do not treat a single accuracy figure as the whole story.
The FDIS text is not public. The change line above is read from the public draft, which is not the FDIS and may differ. No clause level diff is invented here. The ISO stage record is the authority for the dates.
Why it matters to the buyer
An accuracy number without a denominator, a test set and a condition is a marketing claim. The clause asks for the error side as well, which is the part a reviewer or a regulator will test.
What to write in the requirement
Write the task, the measure, the evaluation set, the threshold and the conditions in one requirement. Report the error rate with the same prominence as the success rate. Record the trade-offs you accepted against performance efficiency and robustness.
Evidence a reviewer asks for
A reviewer will ask what the accuracy figure is measured on and what it does not cover. A single number with no error analysis is the common finding.
- Correctness measureThe task definition, the measure and the threshold, with the error rate stated alongside the success rate
- Evaluation recordThe dataset, its provenance and the conditions under which the result was obtained
- Trade-off recordThe accepted trade-offs against performance efficiency and robustness, with the reason
Where it commonly fails
- One accuracy figure with no error analysis
- The evaluation set no longer matches production
- The result is quoted without its conditions
What to do before publication
Re-baseline the correctness requirement on the current model and the current data. A result from an older model is not evidence for this one.
Read across
Source for clause 5.4: ISO/IEC 25059, Clause 5.4, retrieved 20 October 2026. The change line rests on the public draft of the second edition and ISO page 88234. The requirement, evidence and gaps are written from our record. Every source.