Choose metrics that reflect the task
For issue detection, define what counts as a distinct finding and how it matches a reference issue. Precision asks how many proposed findings are correct. Recall asks how many reference issues were found. Neither measure alone establishes safe or useful deployment.
Synthetic worked example
There are 40 reference issues. The assistant returns 38 findings: 32 correct and 6 false. It misses 8 reference issues.
- Precision: 32 ÷ 38 = 84.2%.
- Recall: 32 ÷ 40 = 80.0%.
- If 10 reference issues are critical and 2 are missed, critical recall is 8 ÷ 10 = 80.0%.
These teaching results use issue-level denominators. Do not call them document-level accuracy. A document may contain several issues, and findings within a document may be related.
Results by issue severity
The fictional set contains 12 protocol documents with 40 distinct reference issues. Document count and issue count are separate units. Ten issues are critical and 30 are major under the fictional project's predefined severity rubric. A qualified reference panel establishes and adjudicates the findings before the protected evaluation. No real company or product results are represented.
Six additional AI findings have no matching reference issue and are false findings. They belong in the precision denominator, not the recall denominator. The scoring rubric handles duplicates, partial matches and disputed findings before calculating results.
An overall score can conceal the two critical misses. Suppressing uncertain findings might remove false findings and improve precision while also removing true findings and reducing recall. Examine both effects by severity and by relevant document type.
Representativeness and uncertainty
Select test material to reflect the permitted languages, formats, requirement types and process conditions. Deliberately challenging cases reveal particular weaknesses; they do not, by themselves, estimate their frequency in routine use. Record coverage gaps and avoid extrapolating to unsupported populations.
Several issues in one document share context, parser behavior and source material. Repeated runs also share the same underlying case. Neither creates new independent documents. Report the number of documents, distinct issues and repeat runs separately; use an analysis that respects those dependencies before making a wider performance claim.
The fictional acceptance basis excludes any observed critical miss from supporting reduced independent review. This is a local criterion tied to the consequence of inadequate protocol coverage; it is not a regulatory threshold or a guarantee of zero future error. The observed failures mean the proposed reduction in review is unsupported, even if a general average looks attractive.
Set acceptance rules before examining final results
Define the permitted use, required evidence, severity treatment, uncertainty, and unacceptable failure conditions. Any numerical threshold must be justified for the use; there is no universal “95% means validated” rule.
Zero observed critical misses in a small test set does not establish that critical misses cannot occur. Report sample size, coverage, limitations, and an appropriate uncertainty analysis for any performance claim.
What changes for this type of AI?
Related work: testing, supplier assessment, records, change control, monitoring.
Exercise
A team improves precision by suppressing difficult findings. What else should it check?
Show suggested answer
Recall may fall, including recall for consequential issues. Examine the tradeoff and the combined workflow. A cleaner-looking list can hide more missed problems.
Expected records
An issue-level performance results sheet with requirement/risk reference, category, expected and AI results, severity, disposition, deviation and conclusion. Adapt record names and detail to the applicable procedures; these are practitioner examples, not a universal mandatory document list.