If the brain can't cite it, don't trust it

Language models have a well-documented flaw: they answer with the same confidence whether they know or not. For a chatbot writing birthday poems this is charming. For a system that answers 'what did we promise this client?' it is disqualifying.

The standard fix is retrieval: instead of asking the model what it remembers, you hand it the relevant documents and ask it to answer from those. This works, but only if you enforce it. A retrieval system that falls back to the model's general knowledge when it finds nothing has the same flaw with extra steps.

We enforce two rules. Every claim in an answer links to the document it came from, at the passage level, so a human can check in one click. And when the index has nothing, the system says it found nothing. A refusal is an answer. It is often the most useful answer, because it tells you the knowledge genuinely is not written down anywhere.

There is a cultural effect we did not fully expect. Once people can check the citation, they do, and then they start trusting the system more, not less. The teams that get the most value from IINDBRAIN are the ones that treat a missing citation as a bug report. Trust in an AI brain is not built by fluent answers. It is built by checkable ones.