Computational drug discovery commonly represents a molecule through a fixed chemical graph, descriptor vector, three-dimensional conformer, or learned embedding and then evaluates that representation against one or more isolated prediction tasks. Although such representations are useful, they cannot by themselves maintain an evolving account of how molecular identity, conformational behavior, target engagement, metabolism, toxicity, and off-target interactions change with biological context and accumulating evidence. This article proposes the digital molecular twin as a versioned, molecule-specific computational architecture rather than as a single predictive model. The proposed construct comprises linked state variables for molecular identity, structural ensembles, dynamics, binding, disposition, metabolic transformation, toxicity, and off-target risk; an evidence ledger that preserves source conditions and provenance; update operators that integrate simulation and experimental observations without silently overwriting prior states; a counterfactual query layer for examining possible lead-optimization changes; and an uncertainty layer that records model, data, sampling, and context limitations. The central contribution is an architectural distinction between predicting an endpoint and maintaining a living, auditable account of the evidence relevant to a molecule. The framework further separates association from mechanism, prediction from experimental confirmation, benchmark performance from pharmaceutical usefulness, and local technical validity from decision fitness. It is presented as an original conceptual and methodological synthesis that requires prospective evaluation, domain-specific calibration, and human oversight. It does not establish universal state definitions, validated causal inference, clinical utility, regulatory acceptability, autonomous optimization, or operational readiness. By organizing heterogeneous computational and experimental evidence around explicit states, versions, uncertainties, and decision boundaries, the digital molecular twin may provide a disciplined basis for developing more traceable and scientifically bounded molecular reasoning systems.