Trust
What "traceable to the record it came from" actually means
Every tool in this category promises traceable answers. Most of them mean something modest by it: the answer includes some references, and the references look like the kind of thing a source system would produce.
That is not nothing, but it is not the property you need when you are about to put a figure in front of a board. The property you need is that a reference which does not correspond to a real record cannot appear in the answer — and to have that, the checking has to be done by something other than the model.
Why the model can't check its own work
A language model asked to cite its sources will produce citations. If it found a real record, it cites the real record. If it did not, it produces a citation in the same shape — a plausible id, formatted correctly, referring to nothing. In a chat bubble the two are indistinguishable, and asking the model whether it was careful gets you a confident yes either way.
So the citation has to be resolved outside the model's reach. In Synaptic Intelligence, every cited source id in a finished answer is passed back to the connector that supposedly surfaced it and asked: did you actually return this record? Only ids that come back yes are shown as clean. The rest are marked, and the whole answer is flagged as not auditable end to end.
A broken trail does not mean the number is wrong. It means you should not forward it anywhere until it resolves — which is a far more useful thing to be told than nothing.
The receipt id
There is a subtle failure mode this creates. If a tool searches for something and finds no matching records, it has an honest answer — "nothing matched" — but nothing to cite. A model handed an empty result and a requirement to cite something will invent an id to satisfy the requirement. This genuinely happened during development.
The fix is that every tool call returns a receipt id even when the result set is empty. An empty answer then has a legitimate thing to point at: the receipt for the search that returned nothing. The model never has a reason to fabricate.
What you can do with it
Open any figure in an answer and you see the method, the inputs and the record ids. For a looked-up value, the id resolves in the source system and you can check it by hand. For a calculated value, the arithmetic is shown in full with every input record, so you can reproduce it — and people do. For a projection, it is marked as a projection for its whole life, including in whatever document it ends up in, and it carries the assumptions it rests on.
A number you cannot trace is a rumour with a decimal point.
The security page covers how this sits alongside read-only connections, per-request access checks and tenant isolation. Or ask us to walk the provenance of one of your own figures on a call.