Artificial Intelligence Governance of Antimicrobial Stewardship Recommendation Models Under Unmeasured Pharmacy Dispensing Substitution Latitude: A Healthcare Political Safety Evaluation

Antimicrobial stewardship recommendation models are typically evaluated on their ability to predict resistance or to align with prescribing guidelines, but their governance is complicated by a political fact that prediction metrics cannot capture: pharmacists hold discretionary latitude to substitute, delay or decline recommended antimicrobials, and this latitude is rarely measured. The argument is conceptual and methodological rather than empirical: it specifies how substitution latitude operates as a governance variable, distinguishes it from adherence and from prescribing behaviour, and shows why model safety claims that ignore it are systematically overconfident. A before-after-control-impact design is characterised as the evaluation architecture that would make substitution latitude observable at the cluster level, with a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) and power estimated across 2000 replications per design cell. The design analysis indicates that 54 clusters are required to detect a 5 percentage-point effect at 80% power.