Artificial intelligence models for drug discovery are commonly optimized within individual datasets, yet pharmaceutical prediction frequently involves assays with few observations, targets absent from model development, and compounds that occupy unfamiliar regions of chemical space. Under these conditions, a strong retrospective score does not establish that a model has learned information that can be transferred safely or usefully to a new task. This Original Pharmaceutical Meta-Learning Theory Article develops a conceptual account of how pharmaceutical models should learn across tasks while preserving the distinctions among assay similarity, target relatedness, chemical-space coverage, uncertainty, and experimental usefulness. The proposed contribution is a pharmaceutical meta-generalization framework in which transferable knowledge is conditioned on task provenance, an explicitly defined adaptation operation, a multidimensional description of distribution shift, uncertainty-aware deferral, leakage prevention, and progressively stronger evaluation. The framework treats assay sparsity, target novelty, and chemical-space shift as interacting but non-equivalent sources of difficulty. It further argues that task construction is part of the scientific hypothesis because decisions about episode composition, support examples, endpoint harmonization, and test partitions determine what a reported transfer result can mean. No single accuracy, ranking, or calibration measure is sufficient to demonstrate successful pharmaceutical meta-learning. Evaluation must instead test whether adaptation survives entity-disjoint, assay-disjoint, target-disjoint, temporal, and prospective conditions while remaining interpretable within a defined decision context. The framework is conceptual rather than empirically validated and does not establish mechanistic correctness, experimental success, clinical utility, regulatory acceptability, or deployment readiness. Its value lies in organizing the evidence and validation requirements needed to distinguish genuine learning across pharmaceutical tasks from memorization, inappropriate pooling, or misleading transfer.