Generative artificial intelligence can produce fluent medicines information while obscuring whether individual claims are supported, qualified, contradicted, or unsupported by current evidence. Retrieval augmentation may improve access to external information, but retrieval alone does not establish source eligibility, claim faithfulness, clinical applicability, or safe decision use. This article develops an original, non-empirical Retrieval–Evidence–Generation Architecture for evidence-bound medicines information. Evidence-bound generation is defined as a proposed form of generation in which externally verifiable claims are constrained by eligible evidence, linked to inspectable evidence units, assigned explicit support states, and accompanied by visible uncertainty, conflict, or evidence absence. The architecture separates task specification, governed retrieval, evidence construction, claim planning, bounded generation, claim-level verification, professional review, and lifecycle monitoring. It further proposes a provenance chain connecting each displayed claim to its supporting passage, source identity, version, retrieval context, and transformation history. Five qualitative claim states—supported, supported with qualification, conflicting evidence, unsupported, and evidence absent—are proposed to prevent citations or model confidence from functioning as substitutes for verification. Validation would require technical assessment of retrieval and evidence binding, pharmacist assessment of medication-content correctness and harmful omission, human-factors testing of reliance and verification burden, equity assessment, and prospective workflow evaluation. The architecture does not establish clinical benefit, medication-safety improvement, regulatory acceptance, or deployment readiness. Its original contribution is to make evidence control, uncertainty, professional authority, and governance integral to medicines-information generation rather than optional post-generation checks.