Pharmaceutical foundation models are increasingly expected to learn across chemical structures, biological sequences, molecular profiles, experimental images, and scientific language. Although joint representation learning may increase task coverage, naive multimodal fusion can erase the differences that determine what each source actually measures, predicts, or claims. Molecular structures encode chemical identity through representation-dependent abstractions; sequences reflect evolutionary and functional regularities; omics measurements describe context-sensitive biological states; images capture assay-conditioned phenotypes; and scientific text contains reported interpretations shaped by publication, terminology, and evidentiary quality. This article develops a proposed evidence-preserving multimodal pharmaceutical foundation-model architecture for organizing these non-equivalent evidence classes without reducing them to anonymous latent vectors. The architecture combines modality-specific evidence contracts, specialized encoders, typed alignment interfaces, conditional fusion, missing-modality management, disagreement preservation, uncertainty calibration, abstention, and bounded scientific-use controls. Its central principle is that fusion should be authorized by the scientific task, relation type, provenance, biological context, modality availability, uncertainty state, and intended decision rather than applied automatically whenever multiple inputs are available. Evaluation is correspondingly separated into within-modality representation validity, cross-modal alignment, transfer, grounding, calibration, disagreement handling, missingness robustness, and prospective scientific usefulness. The proposed architecture remains conceptual: it does not establish empirical superiority, mechanistic validity, clinical utility, regulatory acceptability, or deployment readiness. Its contribution is a methodological basis for designing pharmaceutical foundation models that preserve evidentiary distinctions, expose unsupported inferential transitions, and support more disciplined evaluation of when multimodal learning may contribute to drug discovery and development.