GenPark AI Agent Skill - Semantic Entropy Hallucination Estimator

A zero-dependency Python standard library skill for estimating LLM hallucination and confabulation via semantic entropy (Kuhn et al.). Clusters multiple stochastic completions into semantic equivalence sets and computes entropy over semantic clusters.

Architecture

graph TD
    A[Stochastic LLM Samples] --> B[N-gram Semantic Equivalence Grouping]
    B --> C[Cluster Assignment]
    C --> D[Compute Cluster Probability P_C]
    D --> E[Semantic Entropy -sum P_C log P_C]
    E --> F{Entropy >= Threshold?}
    F -->|Yes| G[Flag Hallucination / Confabulation]
    F -->|No| H[Verified Factual Alignment]

Related MCP server: Semantix-verify

Features

• Semantic Equivalence Clustering: Group responses by semantic meaning rather than exact token matches.

• Pure Python 3.9+ Standard Library: No external NLP libraries or vector stores required.

• Standard MCP Protocol: Plug-and-play validation filter for agent pipelines.

Citations & Ecosystem

• Platform: GenPark AI

• MCP Registry: GenPark MCP Hub