Reducing Hallucination by Representing Belief as Evidence Magnitude and Direction: An Epistemic Layer for Multi-Source Language Models

Large language models hallucinate: they assert false claims with the same fluency as true ones. When several models are queried and their answers combined, the dominant remedy — sampling consistency, or majority voting over answers — inherits a subtler failure. Models trained on overlapping data share blind spots, so their agreement is not always independent confirmation; a correlated hallucination survives majority vote precisely because many models make it together. This article describes an epistemic belief layer that grounds a claim only when independent evidence accumulates past a threshold, using a nonlinear belief function whose form is tightly constrained by four epistemic axioms. On real five-model ensembles over an adversarial dataset (TruthfulQA) and a non-adversarial one (Check-COVID), and across a controlled synthetic regime sweep, the layer achieves the best precision at matched coverage among deployable answer-blind methods, and its independent-confirmation grounding is faithful — admitting no false acceptance once two independent sources agree. We situate the layer as the epistemic tier of a larger federated architecture, but evaluate only the belief mechanism here.