Machine-learning models can rank molecules, estimate molecular properties, and predict compound–target relations with increasing computational accuracy. Nevertheless, an accurate prediction does not necessarily reveal why a molecule behaves as predicted, whether the model has learned pharmacologically relevant information, or whether its explanation can support an experimental decision. This conceptual gap is especially consequential in early drug discovery, where model outputs are frequently interpreted across virtual screening, target assessment, lead optimization, and experimental prioritization. This Original Conceptual Theory Article develops evidence-constrained interpretability as a proposed theory for distinguishing molecular explanations that merely describe predictive behavior from interpretations that may support bounded pharmacological reasoning. The theory treats representation traceability, attribution faithfulness, independent evidence linkage, mechanistic-status qualification, uncertainty characterization, and decision-context alignment as interdependent conditions rather than interchangeable indicators of explanation quality. It further distinguishes benchmark performance from prospective usefulness, model attribution from biological mechanism, molecular plausibility from developability, and predictive confidence from calibrated uncertainty. Under the proposed theory, a molecular feature acquires pharmacological meaning only provisionally and only when the inferential step connecting that feature to a target, mechanism, or discovery decision is supported by evidence appropriate to the claim. The contribution is conceptual rather than empirically validated and does not establish causal explanation, clinical utility, regulatory acceptability, or operational readiness. Its principal implication is that interpretable molecular prediction should be evaluated as an evidence-governed scientific practice rather than as the production of visually plausible feature maps. This perspective may support more disciplined use of prediction and explanation in early discovery while identifying the experiments required to test, revise, or reject pharmacological interpretations.