Large language models achieve remarkable performance on language tasks but exhibit systematic failures in factual reliability, reasoning stability, and epistemic transparency. We identify the root cause: modern architectures treat facts and inferences as statistically equivalent tokens, leading to hallucination, fact mutation, and unwarranted confidence. We propose the Fact-Constant / Narrative-Variable (FC/NV) architecture, a hybrid epistemic framework that enforces strict separation between immutable factual constants and mutable explanatory hypotheses. Under FC/NV, reasoning becomes constrained symbolic transformation over verified facts, while language generation remains probabilistic but subordinate to factual integrity. We provide: (1) formal definitions of epistemic categories, (2) architectural specifications, (3) algorithmic implementations, (4) evaluation protocols, (5) theoretical analysis, and (6) empirical validation strategies. The framework is falsifiable, implementable, and compatible with existing neural systems.