A school-lunch allergy data class has both valid and invalid partitions, but a GenAI-generated pack covers only the valid band and misses the invalid one. Which metric best names this gap?

When a school-lunch allergy data class has valid and invalid partitions but a GenAI pack covers only the valid band and misses the invalid one, the metric that best names the gap is recall—the ability to identify all relevant instances, such as covering valid and invalid partitions. Civic meal safety depends on exercising invalid servings and allergen edges, not only happy-path portions. Calling the gap execution success based on stylish Markdown fails because render polish is not partition coverage. Calling it time efficiency as model training cost fails by confusing economics with detection completeness. Defining accuracy solely as the model’s self-reported confidence fails because confidence is not measured coverage of partitions. Exam caveat: recall emphasizes finding all relevant classes of cases the basis requires. Operational check: map allergy partitions to generated cases, flag missing invalid bands, prompt GenAI to add those civic instances, and re-score recall before the district lunch release ships. Retain the municipal audit note so later testers can see why the decision protects public service quality and fairness. Retain the municipal audit note so later testers can see why the decision protects public service quality and fairness.