Foundation models are increasingly used to encode chemical structures, protein sequences, three-dimensional conformations, cellular states, biomedical knowledge, and clinical narratives. Their shared premise is that large-scale pretraining can produce reusable representations that reduce task-specific data requirements and support adaptation to multiple pharmaceutical problems. The unresolved issue is not whether such representations can improve selected benchmarks, but what information transfers across pharmaceutical scales and which evidence is required before transfer can influence medicine-development decisions. This systematic review evaluates foundation-model evidence using two complementary dimensions: scale distance, extending from molecules through proteins, structures, cells, omics, disease contexts, and development decisions; and evidence distance, extending from internal benchmark capability through external evaluation, experimental confirmation, prospective decision impact, and bounded operational readiness. Protocol-defined searching, eligibility assessment, qualitative risk-of-bias appraisal, structured evidence extraction, and narrative synthesis were used to distinguish demonstrated capability from plausible utility and translational value. The synthesis indicates that transfer is most convincingly demonstrated within individual modalities or between closely adjacent biological scales. Cross-scale claims become less secure when representations must preserve mechanistic relevance across changing assays, biological contexts, populations, or decision environments. Adaptation, grounding, calibration, leakage control, external validation, and experimental confirmation therefore function as evidence gates rather than optional performance refinements. A scale-aware interpretation of foundation models is proposed in which transfer is treated as a validated relationship among a source representation, target task, changed context, and decision consequence. The available evidence remains heterogeneous, frequently benchmark-centred, and insufficient to establish routine pharmaceutical readiness. Progress will require evaluation designs that connect representation reuse to reproducible, externally valid, and experimentally consequential decisions.