Scope
What we catch, and what we do not.
General-purpose hallucination detection, meaning a reliable answer to “is this arbitrary statement true”, is not a solved problem and is not something anyone can currently sell you. Any product claiming otherwise is describing something narrower than it sounds. Here is the narrower thing, stated precisely.
Four failure modes we detect
A claim your own sources do not support
GroundednessPass retrieved documents, a database record or an API response as context, and every factual claim in the answer gets held against it. A claim the context does not support fails the check and the request escalates.
Context says refunds take five business days. The answer says two. That fails, and you never see the two.
A second model materially disagreeing
Cross-modelA different model family answers the same question independently, and the two answers are compared for material conflict in fact, number or conclusion.
Two models from different labs, trained on different data, rarely invent the same false specific. Agreement is weak evidence; disagreement is strong evidence.
Malformed output
Schema and rulesJSON that does not parse when JSON was requested, URLs that are not URLs, an empty response, a response that ignores a length or format constraint.
Deterministic, free, instant, and never wrong about the thing it checked.
An answer that will not hold still
Self-consistencyThe same model is resampled at temperature and the spread across runs is scored. Instability correlates with the model being unsure even where it sounds certain.
Reserved for requests that are both critical and genuinely uncertain, because three extra completions is the most expensive thing we can do.
What gets past us
Free-floating factual error
Ask a general-knowledge question with no context attached, and get back something false but internally consistent and plausible. Groundedness has nothing to check against. Cross-model only helps if the second model happens to know better, and two models trained on overlapping data often share the same wrong belief. Nothing here catches that, and we will not market this product as if it did.
Wrong because your sources were wrong
Groundedness checks the answer against what you supplied. If what you supplied was out of date or incorrect, a faithful answer is a wrong answer, and it passes. This check is a retrieval-quality amplifier, not a substitute for retrieval quality.
Tool and function calls
Passed through, but no rung inspects them yet. The checks operate on text answers. An agent that calls the wrong tool with the right-looking arguments is not something this catches today.
Anything inside a stream
Streaming is not implemented, precisely because verification needs a finished answer. Shipping it would mean either holding the stream until the checks clear, which defeats the point of streaming, or streaming tokens we have not checked, which defeats the point of us.
Why this page exists
Every competitor in this category quotes a benchmark number, and most of those numbers were produced by the company quoting them, on a suite they designed. We would rather be the one you can check. A verdict from llm11 is evidence about a specific failure mode, produced by a named check whose logic is described above, at a cost printed on the receipt. It is not a guarantee of correctness, and treating it as one would be a misreading we would rather prevent now than argue about later.
If you find a failure mode we describe here but miss in practice, that is a bug and we want the request id.