VKB-Training: Epistemically Categorized Training Data for Hallucination Reduction in Large Language Models

NOTE: ILLUSTRATIVE NUMBERS - WIPLarge language models hallucinate because their training data carries no epistemic metadata: facts, hypotheses, value judgments, and acknowledged unknowns occupy the same embedding space with identical weight. We propose VKB-Training (Verified Knowledge Base Training), a data-centric approach that assigns each training sample a five-category epistemic tag (Fact, Model, Value, Hypothesis, BlindSpot), a calibrated confidence score, and a provenance chain. We introduce a four-stage hybrid annotation pipeline:(1) AI triangulation — multiple LLMs classify independently; inter-model disagreement signals normative content (the "Caesar/God boundary")(2) Human sampling with axiom extraction — domain annotators resolve high-disagreement cases; recurrent decision principles are extracted as reusable rules(3) Expert calibration with reputation weighting — formalized Galton's ox-weighing insight (per S.V.E. XI, DOI: 10.5281/zenodo.18109198)(4) Logical consistency filters — contradiction detection and symmetry verification via the CGS Method (DOI: 10.5281/zenodo.18776172) Three training mechanisms are proposed: confidence-weighted loss, provenance-aware attention, and a BlindSpot training objective that maximizes output entropy at known knowledge gaps. VKB-Training was first described as part of the CogOS framework (DOI: 10.5281/zenodo.18109244). This paper extracts and formalizes the VKB component as a standalone, empirically testable proposal with a falsifiable experimental protocol and pre-specified success thresholds. Section 7 (Ethical Data Sourcing: Author Revenue Sharing, 10-50%) is included in the preprint but will be omitted from the workshop submission. Prepared for submission to NeurIPS 2026 Workshop.