Hospitals could learn from one another’s medication-related risks without transferring patient-level records into a common repository. However, keeping records local does not by itself establish privacy, security, interoperability, fairness, clinical usefulness, or accountable decision use. This article develops an original, non-empirical federated medication-safety architecture for collaboration among heterogeneous hospitals. It defines a medication-safety lesson as a proposed, versioned package linking a bounded clinical task with locally derived model information, provenance, uncertainty, validation evidence, workflow purpose, and explicit limits on use. The architecture separates six functions: local data stewardship, local training, protected update exchange, shared candidate-model construction, site-specific qualification, and consortium governance. Raw records remain within participating institutions; only authorized computational updates and contextual metadata enter the shared process. A resulting shared model remains a candidate until each hospital assesses local calibration, medication-safety relevance, subgroup performance, workflow burden, technical compatibility, and professional oversight requirements. Separate controls address statistical and semantic heterogeneity, privacy leakage, malicious or unreliable updates, inequitable performance, institutional burden, and unequal distribution of benefits. Evaluation must extend beyond aggregate predictive performance to credible local comparators, site-separated results, workflow consequences, human–AI interaction, attack resistance, equity, change control, and prospective validation. The architecture does not establish that federated learning is inherently safer, more private, more equitable, or more clinically effective than centralized or local alternatives. Its original contribution is a proposed boundary-governed approach in which collaboration creates locally contestable medication-safety candidates rather than transferable clinical authority.