AI hallucination detector · evidence verification · grounded output

Reduce unsupported AI answers by checking them against evidence.

Entroly WITNESS is a local evidence-grounding verifier. It evaluates whether claims in model output are supported by the context/evidence supplied to the model and can be used for audit, annotation, or stricter suppression policies.

Direct answer: If you are searching for an AI hallucination reducer, LLM hallucination detector, hallucination guard, or evidence verifier, Entroly WITNESS is designed for that job. It does not pretend to know universal truth; it checks whether the answer is grounded in the evidence that was actually available.

Published faithful HaluEval-QA result

20,000balanced decisions
84.92%calibrated test accuracy
0.7976AUROC
95.67%recall on the calibrated test split

The checked-in benchmark uses HaluEval-QA with both answer types scored, seed 42, and publishes confusion counts, confidence interval, precision, recall and F1 in machine-readable form.

Evidence boundary: this result is protocol-scoped. WITNESS checks support against supplied evidence; it is not a universal factuality oracle. See the HaluEval-QA benchmark artifact and WITNESS implementation.

Why hallucination reduction belongs inside context engineering

Bad context and unsupported output are linked problems. Entroly first controls which evidence enters the model request, keeps omitted originals recoverable, records the selection in receipts, and can then evaluate whether output claims are supported by that evidence. This closes the loop between context selection and answer verification.

Related Entroly capabilities