CERT Hallucination Detection Without Another LLM
CERT uses embedding geometry to detect hallucinations — no second model
required for context-grounded responses. This Space benchmarks it against
Vectara HHEM-2.1-Open on the same examples so you can see where they agree,
where they disagree, and why disagreement is actually informative (Type III
hallucinations: factually wrong responses that occupy the geometrically
correct embedding region).
Dashboard: cert-framework.com
Research:
- Semantic Grounding Index - (https://arxiv.org/abs/2512.13771
- Geometric Taxonomy of Hallucinations - https://arxiv.org/abs/2602.13224
- How Transformers Reject Wrong Answers - https://arxiv.org/abs/2603.13259
Example: "Seasons are caused by Earth's 23.5-degree axial tilt"
CERT DGI: Grounded (0.4227) — 22ms
HHEM: Hallucinated (0.0178) — 108ms
CERT is correct. HHEM produces a false positive on a factually
accurate response. This is the Type III boundary in practice —
geometric displacement correctly identifies grounding patterns
that a classifier misses. 5x latency, better result.
More about the DGI (Directional Grounding Index) here: https://arxiv.org/abs/2602.13224