snc-core reduces hallucination rate by 52% on HumanEval with Qwen2.5-Coder-7B via inference-time governance layer

snc-core wraps any decoder-only LLM with a governance layer that reduces hallucination rate from 16.5% to 7.8% on HumanEval with Qwen2.5-Coder-7B (z = 2.12, p < 0.05), requiring no model retraining.

The layer combines three signals: confidence elicitation (instructing the model to emit a self-confidence score), behavioral clustering (sampling K = 5 candidates at temperature 0.8 and grouping by output equivalence), and trust thermodynamics (a closed-form score that discounts as candidates diverge, compared against a user-supplied threshold θ).

At conservative threshold θ = 0.65, the system achieves 92.17% precision while recovering five previously failing test cases, though nine residual failures stem from adversarial mode collapse where multiple candidates make the same systematic error.

Available via pip install snc-core for Python 3.9+ with no mandatory dependencies beyond the standard library; supports Ollama-served local models and OpenAI-compatible APIs.