Artificial intelligence in pharmaceutical science increasingly generates molecules, predicts reactions, integrates biological measurements, and coordinates scientific tools. These capabilities are valuable, but they do not by themselves constitute a world model capable of representing what happens to a biological system after a pharmaceutical intervention. The unresolved problem is that models optimized to continue statistically probable sequences may produce chemically or biologically persuasive outputs without encoding intervention-specific state transitions, temporal adaptation, causal structure, uncertainty, or multiscale consequences. This article proposes a consequence-aware pharmaceutical world-model architecture that treats biological simulation as a structured relationship among an initial biological state, an explicitly specified intervention, context-dependent transition dynamics, time, observations, uncertainty, and counterfactual alternatives. The proposed construct distinguishes latent biological state from assay output, molecular action from therapeutic consequence, predictive association from mechanism, and model confidence from calibrated uncertainty. It further links learned cellular representations with mechanistic and quantitative pharmacological constraints, multiscale coupling, provenance, applicability boundaries, and experimental feedback. Validation is framed as a graduated evidentiary problem requiring more than reconstruction or aggregate predictive performance: models should be challenged through held-out perturbations, dose and temporal variation, context transfer, interaction recovery, mechanism-directed experiments, uncertainty assessment, and prospective falsification. The architecture is conceptual rather than empirically validated and cannot establish clinical utility, regulatory acceptability, or universal biological completeness. Its principal implication is that pharmaceutical world models should be judged by whether they support bounded, testable, intervention-conditioned claims about biological consequences—not by whether they generate fluent, realistic, or statistically familiar continuations.