I asked my preferred LLM and (Urs) the same question, on the same document. The model invented numbers. Confidently. Convincingly. When I challenged it, it admitted:
“I invented these figures. It's an error on my part.”
(Urs)“I can't find this data: it does not exist in the document.”
The entire difference between extraction and preservation lives in those two sentences. Language models synthesize brilliantly — but when they don't know, they fill the gap. A cognitive infrastructure doesn't guess. It says “I don't know”, because provenance matters more than the answer.
Note the exact wording. (Urs) does not say “this data does not exist” — a statement about the world, unverifiable. It says: it does not exist in the document — a statement about a defined perimeter, verifiable, replayable. That is our entire verdict doctrine: never an absolute negation, always a verdict scoped to the sources mobilized, with the fingerprint to prove it.
A trust system is not recognized by what it knows. It is recognized by the way it says what it does not know.
That is what we are building. Not instead of the models — on top of them. Every claim typed, sourced, scored; every “I don't know” documented and handed to the expert. See Evidence →