Artificial intelligence increasingly supports molecular generation, structural prediction, affinity estimation and property prediction in drug discovery, yet a favorable computational score does not itself establish that a candidate should advance. The central problem is therefore not whether AI can predict binding or related molecular properties, but how such predictions should be interpreted alongside mechanistic plausibility, translational evidence, uncertainty and the consequences of an advancement decision. This conceptual article develops an evidence-bounded candidate-advancement account that separates prediction quality from candidate-worthiness. It first identifies the justificatory limits of prediction-centred selection, including state dependence, benchmark leakage, out-of-distribution failure, calibration limitations and the conditional relevance of biological data. It then examines evidence that can strengthen or weaken an advancement case, including target engagement, pharmacological characterization, target assessment, orthogonal experimental confirmation and human translation. The resulting argument treats evidence layers as heterogeneous rather than automatically commensurable and proposes that uncertainty changes the burden of evidence required for progression Favorable AI predictions can therefore remain insufficient when mechanistic incoherence, unresolved distribution shift, weak target biology or adverse translational evidence persists. Conversely, a candidate can become more defensible as distinct uncertainties are retired through evidence appropriate to their level. Candidate advancement should thus be framed as a state-dependent decision under uncertainty, not as the maximization of a single predictive score.