In systems Rottawhite builds, every extracted field carries a source citation: the document, page, and passage it was read from. It is the question that ends most AI pilots in regulated work. If the honest answer is that the model said so, the system does not survive its first review.
A reviewer checking an extracted figure, an auditor sampling processed documents, a client querying a number on an invoice — each of them needs to get from an output back to the document it came from, in seconds. If that path does not exist, every output has to be re-verified by hand, which removes the entire point of the system.
This is usually what people mean when they say an AI pilot “did not get adopted”. The extraction worked. The trust did not.
Document, page, and the passage the value was read from. A reviewer clicks the field and lands on the highlighted line rather than opening a sixty-page PDF and searching.
Recorded alongside the value, so a low-confidence field that later turned out wrong is visible as a known risk rather than an unexplained failure.
What the agent wrote, where it wrote it, and the inputs it acted on. Reconstructing what happened does not depend on anyone remembering.
The original output, the correction, who made it, and when. This is the audit record and the feedback loop in one artifact.
Illustrative. Every field the system produces carries a record of this shape.
field: order_total value: 4,250,000 confidence: 0.91 source: document: PO_AcmeLogistics_2026.xlsx sheet: Line items cell: G47 reconciled: cover sheet stated 4,180,000 status: exception raised resolution: routed to a reviewer reviewed_by: s.mehta@example-client.com reviewed_at: 2026-09-12T14:22:07Z final_value: 4,250,000 (line items accepted)
Note what the record contains beyond the answer: the disagreement between two documents, the fact that a human resolved it, and which source won. Six months later, that is the difference between explaining a decision and guessing at one.
Five questions worth putting to any AI vendor, including us. A system with a real audit trail answers all five without a meeting.
Document, page, and passage, on demand. If the answer involves the word "usually", the trail has holes.
Not just "the agent updated the record" but what it read, what it wrote, and when.
A correction that overwrites the original destroys the evidence of what the system actually did.
An auditor asks about a number from last quarter. The answer must come from the log, not from memory.
Confidence recomputed after the fact is a rationalisation. It has to be captured when the value was produced.
It is the practical half of it. We are not claiming to explain a model's internal reasoning. We are claiming that for every value the system produced, you can see the document, the page, and the passage it came from, plus the confidence it carried. That is what a reviewer actually needs.
As long as your retention policy requires. The log is structured data in your own storage, not a vendor dashboard you lose access to when a contract ends.
It helps with the evidence side — showing that processing is logged, traceable, and reviewable. We are engineers rather than auditors, so we build to what your auditor asks for rather than claiming a certification on your behalf.
Those are the most valuable records in the log. The original output, the correction, and who made it are all retained — which is both the audit record and the raw material for improving the test set.
The audit trail answers how did it get here. The eval harness answers is it right, still. Both ship with every production build, and you can see the shape of both in a two-week Workflow Pilot.
Next step
The scanned copy, the one-off format, the packet of six attachments that disagree with each other. That is where the difference between a demo and a system shows up.