The grounding scorer that's right when it says yes — proven on public data.
LLM-as-judge evaluators wave through most hallucinations at the threshold you actually deploy. We measured the gap on public, labeled datasets — same data into every scorer, reproducible scripts, and we publish the false-positive rate nobody else does. Pick a benchmark to see the full head-to-head.
How many hallucinations does each tool catch?
Healthcare RAG answers. How many hallucinated answers does each tool catch? · 2,000 evaluations
“Caught” = the tool correctly flagged a hallucinated answer as unsupported.
Clinical · PubMedQA
variA/Bly vs DeepEval
On 2,000 clinical Q&A evaluations, variA/Bly catches 94.5% of hallucinations — vs 22% for the LLM-as-judge — while scoring the highest AUC in the set.
General RAG · RGB
variA/Bly vs RAGAS
On 592 general-domain RAG evaluations, variA/Bly matches LLM-judge accuracy out of the box — deterministic, with no LLM judge and a verdict for every case.
Reproducible by design
Every number on these pages is recomputable from committed raw scores in the public benchmark repo — no account, no cost — and each scorer can be re-run from scratch. Full commands are on each benchmark page and in github.com/varia-bly/variably-benchmark.