Pharmaceutical knowledge is distributed across chemical databases, bioassay repositories, target resources, pathway models, disease ontologies, experimental studies, computational predictions, and curated therapeutic claims. Although knowledge graphs can connect these resources, conventional graph integration may obscure how an assertion was generated, which source supplied it, whether apparently supporting records are independent, and under which molecular, biological, experimental, or temporal conditions it applies. This article develops the Therapeutic Evidence Graph as an original pharmaceutical knowledge architecture for connecting molecular data, biological mechanisms, disease knowledge, studies, and experimental claims while retaining their evidential identities. The proposed architecture distinguishes therapeutic entities from computational representations, scientific assertions from their supporting evidence, and source agreement from evidential independence. Its central unit is a qualified therapeutic assertion accompanied by provenance, context, polarity, temporality, uncertainty, and applicability information. Contradictions remain represented as inspectable relationships rather than being removed through premature harmonization, while source lineage and version history permit conclusions to be examined under alternative evidence configurations. The architecture may support traceable reasoning for target discovery, drug repurposing, mechanism assessment, and evidence governance, but graph connectivity or predictive ranking cannot independently establish causality, pharmacological efficacy, developability, safety, clinical utility, or regulatory acceptability. The contribution is conceptual and requires implementation-specific evaluation of semantic fidelity, provenance completeness, evidence independence, temporal reconstruction, uncertainty calibration, interoperability, and decision usefulness. By treating provenance as part of therapeutic meaning rather than peripheral metadata, the Therapeutic Evidence Graph offers a bounded foundation for pharmaceutical evidence integration without converting heterogeneous observations and claims into an artificial consensus.