Artificial Intelligence Benchmarking of Sparse-Sensing Glucose Forecasters Under Unmeasured Meal-Announcement Noncompliance in Insulin-Treated Cohorts: A Healthcare Safety-Regret Evaluation

Sparse-sensing continuous glucose monitors supply the input stream for most deployed insulin-dosing forecasters, yet the meal-announcement channel those forecasters rely on is systematically incomplete in free-living use. the present analysis treats unmeasured meal-announcement noncompliance as a design property of the evaluation, not a nuisance to be imputed away, and asks what a safety-regret benchmark can identify when the announcement indicator is missing not at random. A formal framework is developed in which the forecaster's action set, the announcement-compliance process and the hypoglycaemia loss are defined jointly, and a regret functional is derived that bounds the safety cost of benchmarking on the observed announcement channel alone. The framework is characterised analytically under a before-after-control-impact design with an intraclass correlation of 0.64, a between-cluster standard deviation of 6.0 and a residual standard deviation of 4.5; under that variance structure, 54 clusters are required to detect a 5-percentage-point effect at 80% power, and power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis yields three results: an identification bound showing that observed-channel benchmarking understates regret by a factor that grows with the noncompliance rate; a design theorem giving the cluster count at which the bound is dominated by sampling error rather than by noncompliance bias; and a reporting protocol that makes the noncompliance assumption auditable. The contribution is a benchmark design that treats announcement noncompliance as a first-class safety parameter rather than a preprocessing step.