NeutroStats-AI: A Neutrosophic Statistical Framework forEvaluating Uncertainty in Large Language Model Outputs

Resumen

The deployment of Large Language Models (LLMs) in high-stakes domains demands statistical frameworks capable of quantifying epistemic uncertainty beyond binary accuracy metrics. This paper introduces NeutroStats-AI, a Python-based framework implementing neutrosophic statistical inference for LLM evaluation. The framework provides neutrosophic confidence intervals, hypothesis tests, and a novel Epistemic Calibration Score (ECS) that penalizes indeterminacy suppression. Validated against six state-of-the-art LLMs across four benchmarks, NeutroStats-AI demonstrates that classical evaluation systematically underestimates model uncertainty by 34-67% relative to neutrosophic evaluation.