Synthetic pharmacy records may widen access to data for software development, education, interoperability testing, and algorithm research, but plausible records can create false confidence when realism is treated as evidence of privacy, fairness, medication safety, or decision readiness. This article develops an original, non-empirical governance architecture for synthetic pharmacy records. The problem is defined as synthetic certainty: the unwarranted conversion of generated plausibility into claims about observed patients, valid clinical relationships, safe workflow use, or authorized downstream decisions. The proposed architecture begins with an intended-use and consequence specification, then connects the source-data envelope, generator configuration and versioning, provenance, separate privacy–fidelity–bias–disclosure controls, independent use-conditioned evaluation, bounded release states, and continuing monitoring, expiry, withdrawal, and supersession. The architecture separates model output from medication decisions, technical performance from clinical usefulness, association from causation, automation from professional authority, and implementation from validated benefit. Evaluation therefore requires structural, temporal, relational, task-specific, privacy, subgroup, human-factors, medication-safety, and organizational evidence selected for the declared use. The contribution is not a validated standard, regulatory pathway, or deployment manual. Its release states and decision gates require empirical calibration across pharmacy settings, populations, data environments, and jurisdictions. The central proposition is that synthetic records should be governed as versioned, purpose-bounded, auditable artifacts whose uncertainty remains attached throughout generation, release, reuse, and withdrawal.