Tessary groundedness

Groundedness is one of Tessary's built-in classifiers. For each answer built from retrieved documents, it checks whether what the agent said is supported by those documents, sentence by sentence.

It catches the shape most production hallucinations take: an answer that asserts things its documents don't support while reading as fluent and confident. A sentence counts as unsupported when the documents contradict it or never say it at all.

The classifier is a small open model, not an LLM judge. It reads the retrieved documents and the whole answer together in one pass, so a claim that draws on two documents is judged against both. It scores each sentence for how likely it is to be unsupported, and an answer is flagged when any sentence crosses a high bar.

It judges the answer against the documents, not the documents against the world. If retrieval returns a stale or wrong document, an answer that faithfully repeats it reads as grounded. An answer that is true but came from the model's own knowledge can still be flagged.