CITATION.cff

cff-version: 1.2.0
message: "If you use this software, please cite both the software and the companion paper."
type: software
title: "snc-core: Behavioral Trust Clustering for production LLMs"
abstract: >-
  A model-agnostic governance layer that wraps any decoder-only LLM and reduces
  hallucination rate at inference time without retraining. Composes self-reported
  confidence, behavioral clustering of K stochastic samples, and a closed-form
  thermodynamic trust score. Reduces hallucination rate by 52% on the official
  HumanEval benchmark with Qwen2.5-Coder-7B at the conservative threshold.
authors:
  - family-names: Culotta
    given-names: Daniel
    email: daniel.culotta@gmail.com
    country: IT
version: 0.4.0
date-released: "2026-05-04"
license: MIT
repository-code: "https://github.com/dculotta/snc-core"
url: "https://github.com/dculotta/snc-core"
keywords:
  - LLM
  - hallucination mitigation
  - trust calibration
  - selective prediction
  - abstention
  - behavioral clustering
  - HumanEval
  - regulated AI
preferred-citation:
  type: article
  title: "Behavioral Trust Clustering: A Thermodynamic Governance Layer for Production LLMs"
  authors:
    - family-names: Culotta
      given-names: Daniel
  year: 2026
  doi: "10.5281/zenodo.20028123"
  url: "https://doi.org/10.5281/zenodo.20028123"