Foundation models can produce scientifically fluent accounts of targets, pathways, compounds, and disease mechanisms, but linguistic plausibility does not establish that an output constitutes a pharmacological hypothesis. A decision-worthy hypothesis must specify a directional intervention–mechanism–outcome relationship, identify the biological and experimental context in which the claim applies, expose assumptions and contradictory evidence, and state observations that could count against it. This Original Hypothesis-Generation Theory Article develops a conceptual framework for evaluating whether foundation-model outputs are mechanistically coherent, falsifiable, novel, experimentally discriminating, and consequential for pharmaceutical decision-making. The proposed contribution rejects the use of any single performance measure as a sufficient indicator of hypothesis quality. Instead, it treats quality as a structured profile comprising evidentiary grounding, mechanistic specification, falsifiability, scientific novelty, experimental feasibility, uncertainty, expected information gain, and decision consequence. The framework distinguishes prediction from explanation, association from mechanism, model confidence from calibrated uncertainty, and molecular plausibility from synthesizability, developability, safety, and therapeutic value. It further proposes that minimum evidentiary and falsifiability conditions should operate as non-compensatory gates: high fluency, novelty, or predicted activity cannot offset the absence of a traceable mechanism or an experiment capable of challenging the claim. The article outlines how molecular structures, biological networks, literature evidence, alternative hypotheses, and experimental constraints can be integrated into an auditable hypothesis object and evaluated prospectively. The framework remains theoretical and requires domain-specific, leakage-resistant, blinded, and experimentally grounded validation. Its purpose is not to authorize autonomous discovery decisions, but to define the conditions under which machine-generated propositions may become scientifically interpretable candidates for testing.