The distinction between discriminative and generative models is well established in machine learning as an architectural classification. This paper argues that the distinction carries unrecognized epistemic weight: discriminative architectures compress uncertainty toward checkable answers, while generative architectures expand learned structure into probability spaces of possible continuations. These architectural differences produce categorically different relationships to truth, calibration, and epistemic accountability. The paper formalizes this claim through the *flexibly deterministic* and *structured probabilistic* categories, showing that they correlate strongly with, but are not reducible to, the discriminative/generative architectural divide. Where discriminative models estimate conditional distributions over bounded labels evaluable against external targets, generative models estimate distributions over data itself, optimized for distributional similarity rather than correspondence with fact. The paper specifies the boundary conditions under which a generative system can achieve task-level epistemic accountability, demonstrating that every such crossing requires importing discriminative or rule-based governance from outside the generative mechanism. These findings have direct implications for enterprise AI governance, where the failure to distinguish between epistemic categories contributes to the high abandonment and low-ROI rates documented across multiple enterprise AI surveys. The paper concludes that epistemic accountability emerges when probability is subordinated to an external evaluator, and that this subordination must be architecturally enforced rather than behaviorally trained.