Computational pharmacology models are increasingly judged through predictive metrics that permit rapid comparison among algorithms but offer only partial evidence about pharmaceutical usefulness. A model may reproduce held-out observations while relying on unstable data associations, biologically implausible relationships, undocumented implementation choices, or causal interpretations that its design cannot support. This creates an unresolved methodological problem: predictive performance is often treated as the principal threshold of model quality even when the intended application requires explanation, mechanistic extrapolation, intervention reasoning, or consequential pharmaceutical decision support. This Original Methodological Position Article proposes a mechanism-first ordering of computational pharmacology evidence. The proposed position distinguishes structural coherence, biological coherence, and causal coherence from predictive adequacy and argues that these dimensions should be examined before performance results are granted substantive pharmaceutical meaning. Predictive evaluation remains necessary, but it is repositioned as a later test of an already scrutinized model rather than a substitute for plausibility, reproducibility, applicability, calibration, uncertainty characterization, or context-of-use specification. The article also introduces a bounded qualification logic in which the permissible influence of a model depends on the strength and relevance of its supporting evidence, the consequences of error, and the reversibility of the decision. The proposal is conceptual rather than empirically validated and does not establish a universal validation sequence, regulatory standard, clinical recommendation, or deployment criterion. Its purpose is to organize more defensible judgments about computational models and to clarify why benchmark success, mechanistic explanation, causal inference, experimental confirmation, and pharmaceutical utility must remain distinct evidentiary claims.