Epistemic Field Theory: Predicting Hallucination in Large Language Models via Multi-Model Consensus

Large language models exhibit hallucination—generating confident but incorrect outputs—at rates that undermine their reliability in high-stakes applications. We introduce Epistemic Field Theory (EFT), a formal framework that predicts hallucination probability from multi-model consensus. EFT defines a consensus field σ ∈ [0, 1] over query space and derives the hallucination predictor P(H) = (1 − σ) · η, where η is a model-specific noise coefficient. We establish theoretical conditions under which consensus bounds error probability and prove that independent model errors yield superlinear consensus-reliability scaling. Empirical validation across 13,728 responses from four models in three domains confirms the core prediction: consensus and hallucination correlate at r = −0.38 (p < 0.001), with hallucination rates dropping from 51.9% (σ < 0.2) to 5.9% (σ = 1.0). The framework provides a principled, model-agnostic mechanism for uncertainty-aware decision gating in automated pipelines.