How Evidence-Grounded Answers Work: Chunks, Pinned Results, and No Absence Claims
In short
A grounded answer is one where every claim maps to a retrieved chunk, a pinned tool result, or a cited media region. Inteleto runs acceptance checks before returning: unsupported claims are stripped, and claims about what does not exist are blocked outright.
Three classes of evidence
Retrieved chunks are passages pulled from your documents by hybrid search — vector similarity and BM25 keyword matching, fused into one ranking. Pinned tool results are outputs from an external call, recorded verbatim at run time so the answer can be re-audited later. Media regions are specific areas of a page or image that a claim points at.
All three are recorded on the Run. An answer is only as good as the evidence attached to it, and the evidence is inspectable after the fact rather than reconstructed from a log.
Why 'no absence claims' matters
Retrieval returns the top matching passages, not every passage. That makes any statement of the form "there is no clause covering X" unsupportable: the system searched a ranked subset, not the whole corpus.
So that check is mandatory on retrieval tasks. An agent may say what the documents state; it may not say what they fail to state. Exhaustive coverage is a different mode of operation, and until it runs, nothing is allowed to sound like it did.
The run is the audit trail
Every execution — chat message or scheduled task — produces the same Run shape: the question, the resolved scope, the evidence retrieved, the model call, the checks that ran, and the final answer with its citations.
There is no separate logging layer to keep in sync, because the record of what happened is the same object that produced the answer.