Hallucination Detector — Finance

Three checks documented on the methodology page: (1) entity grounding — does every named entity in the output appear in the source context? (2) numerical grounding — do numbers in the output match values in the source within tolerance? (3) self-consistency — does the same prompt produce stable outputs across N samples? Any failure flags the response.

Around 8-15% on the benchmark suite documented on the methodology page. A 'false positive' here is the detector flagging a correct answer as hallucinated — usually because the entity matches a known synonym or the number was rounded differently. The methodology page lists the benchmark mix and per-category accuracy.

Entity and number checks work for any language with named-entity tooling (English, Spanish, German, French, Mandarin all have decent NER). Self-consistency is language-agnostic. Languages with weaker NER tooling get a higher false-negative rate.

Partially. Entity checks catch 'invented company names'. Number checks catch 'made-up percentages'. The hardest case — a plausible but wrong claim that doesn't contradict anything in the source — usually slips through unless it triggers self-inconsistency across samples. Multi-sample voting helps but isn't bulletproof.

5 samples is the sweet spot per the methodology page — more catches subtler inconsistencies but multiplies cost. At 5 samples, a 3/5 agreement is typically enough to ignore minor wording variation; below 3/5, the response is flagged as inconsistent (and probably less reliable).