Probability Comes After Identity_ Hallucination as Invariant Drift

Abstract This paper proposes a structural reordering of explanatory priority in bounded epistemic systems: identity → invariant → observable → probability. The central claim is that probability is not rejected, but repositioned. Probability governs uncertainty over observables, while observables presuppose invariant structure, and invariant structure presupposes identity-persistence conditions sufficient for same/not-same continuity. On this view, hallucination in generative AI is not primarily random output or stochastic misgeneration. It is symbolic continuation that has drifted beyond lawful invariant anchoring while still preserving local probabilistic plausibility. The probabilistic surface remains smooth even after identity continuity has broken. The paper develops this reordering as a structural and epistemic argument, not an ontological claim. It distinguishes invariant governance from probabilistic estimation, explains why scale alone does not resolve hallucination, and shows why retrieval, replay, verification, audit trails, and deterministic artifacts help: they supply invariant anchoring that probability alone cannot provide. The result is a methodological reframe for AI systems and bounded reasoning contexts. Hallucination is treated as a persistence problem visible through the probabilistic surface, and verification infrastructure is interpreted as one engineered way of preserving identity-faithful continuation under transformation.