Disagreement Entropy in Multi-LLM Knowledge Graph Verification: When Ensemble Conflict Signals Epistemic Uncertainty

Knowledge graph (KG) verification increasingly relies on large language models (LLMs) as automated evaluators, yet single-model pipelines inherit systematic biases that compromise reliability. In an earlier experimental phase on the same infrastructure, we identified a cascade of three such biases—Verbosity Bias, Acquiescence Bias, and Negation Asymmetry—in single-model KG verification. This paper extends that work with a new ensemble-based phase (N = 1,125) by introducing Disagreement Entropy (DE), an information-theoretic metric that quantifies inter-model divergence in multi-LLM verification ensembles, and by presenting a self-contained account of both the bias-cascade findings and the ensemble analysis. Evaluating 1,125 Wikidata-sourced triples across three locally deployed models (Qwen2.5:7b, Gemma2:9b, Llama3.1:8b) on the OmniCore v17 knowledge graph system, we report three principal findings. First, the empirical population of the three structurally determined DE zones shows 54.8% of triples at consensus (DE = 0), 44.1% at majority split (DE ≈ 0.58), and 1.2% at full disagreement (DE = 1.0). Second, we measure a 28.4 percentage-point (pp) Acquiescence Gradient (from 57.8% to 86.2%) across models evaluated on identical triples, providing a controlled quantitative characterisation of model-specific response tendencies. Third, all 13 triples reaching maximum entropy (DE = 1.0) resulted in SPLIT verdicts with zero achievable consensus, suggesting that DE = 1.0 serves as a candidate heuristic for flagging content where automated consensus fails and human review is warranted. We do not assume access to ground-truth labels; therefore, we characterise disagreement patterns but leave the link between DE and factual error rates to future work. These results suggest that ensemble disagreement may provide a promising diagnostic signal for identifying potentially unreliable KG content and prioritising human-in-the-loop review. All experiments were conducted on the PARAM BILIM supercomputer cluster at IKCEICT, L.N. Gumilyov Eurasian National University. Keywords: knowledge graph verification, LLM ensemble, disagreement entropy, acquiescence bias, multi-model consensus, epistemic uncertainty, evaluation bias cascade, OmniCore