The hallucinations it catches
- Numeric drift: "60 days" in the answer when the source says 30.
- Fabricated entities: names, teams or products that appear nowhere in the source.
- Contradictions: a rule flipped from "non-refundable" to "refunded in full".
- Unsupported claims: statements the source simply does not back.
- Meaning errors: role reversals, scope changes and antonyms that share every word with the source, caught by the HHEM-2.1-Open model.
How it works
The response is split into atomic claims. Each claim is matched to the closest sentence in the ground truth and labeled supported, partial, contradicted or unsupported. Deterministic checks run first, so a wrong number is never excused by a model. The meaning check can only make a verdict stricter or rescue a reworded claim; it never overrules a deterministic catch.
Long documents are split into sentence-aligned windows, so a fact on page 30 is still found.
Without ground truth
If you do not paste a source, claims cannot be verified. The report then rates hallucination risk instead: specific figures, names and citations that nobody can check, and overconfident wording. The result is clearly marked as low confidence.
Honest limits
No automatic check is perfect. Temporal order ("before" vs "after") is a known gap of the open model; an LLM judge or a human expert review covers it. Try it now in the free checker or learn about RAG evaluation.