CERT Hallucination Detection

Geometric grounding detection — benchmarked against Vectara HHEM-2.1-Open

With context (RAG): Semantic Grounding Index -SGI- measures whether the response engaged with the source document. Computed as dist(response, question) / dist(response, context). No model inference for the evaluation — one embedding, one ratio.

Without context: Directional Grounding Index -DGI- measures whether the response displacement vector aligns with the mean displacement of verified grounded pairs.

Dashboard · Semantic Grounding Index: Geometric Bounds on Context Engagement in RAG Systems · A Geometric Taxonomy of Hallucinations in LLMs · How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing