What we actually detect
General-purpose hallucination detection - “is this arbitrary claim true” - doesn’t exist yet, and any product that claims to have solved it is overselling. What follows is the honest, specific list of what llm11’s verification ladder checks, and what it doesn’t.
We detect
Claims unsupported by the source you gave it
When you pass context- retrieved documents, a record, an API response - the groundedness rung checks every factual claim in the answer against that context and fails it if the model said something the context didn’t. This is real, checkable groundedness, not a vibe.
A second independent model materially disagreeing
The cross-model rung asks a different model family the same question and flags a different fact, number or conclusion. Two unrelated models rarely invent the same false detail, so disagreement is a strong signal something went wrong.
Schema, format and reference violations
Malformed JSON when JSON was requested, malformed URLs, an empty response - deterministic, free, and never wrong about what it checked.
An answer that isn't stable
The self-consistency rung resamples the same model and flags disagreement between runs, which correlates with the model being unsure even when it sounds confident. Reserved for genuinely critical, genuinely uncertain requests - see how it works.
We do not detect
Free-floating factual error with no source to check the claim against. If you ask a model a general knowledge question with no supplied context, and it states something false but internally consistent, no rung here catches that - nothing on the market reliably does. We will not claim otherwise, and we will not market this product as “eliminating hallucination.”