Most assistants give you a paragraph. If the paragraph matters — if it goes into something a client, a regulator, or a court will read — you then spend the next twenty minutes finding out whether it is true. The assistant has moved the work, not done it.
Axon is built so the answer arrives with its evidence attached. Every response carries numbered citations, and each one opens the source: the file, and the page, the sheet-and-row, or the section it came from. In a PDF the passage is highlighted where it sits. You are not being asked to trust the summary; you are being handed the thing the summary is a summary of.
A confident wrong answer is the expensive failure
When retrieval finds nothing close enough to the question, the safe move is to say so. Axon does — and it shows the nearest passage it found, so you can see for yourself that it looked and that the material genuinely is not there. A model that fills the gap with a plausible sentence is worse than useless in reviewed work, because the cost of catching it later is far higher than the cost of not having an answer now.
The record is part of the answer
Every turn is written to an audit record: the question, the answer, and each passage behind it, including the ones that were retrieved and rejected as too weak. Each passage carries the hash its file had at the time you asked. Come back a month later, and the record tells you which sources have changed underneath you — so an answer that was right when it was given, and is now being read against a revised document, doesn't quietly become a mistake with your name on it.
The record exports as Word, Markdown or JSON through the same writer that produces the deliverable, so the working file and the evidence file look like the same document.