AI that shows its work
An AI answer you cannot check is an answer you cannot use for anything that matters. Every answer, and every element an AI system generates, should point back to exactly where it came from.
Most AI tools answer confidently. Few of them tell you why. For someone asking for a recipe, that is fine. For an engineer deciding whether a connector fits, an analyst briefing a decision, or a compliance officer answering an auditor, a confident answer with no source is a liability. The person still has to check it, and without a source they have to check it from scratch.
Our rule is simple: every output carries its provenance. An answer cites the document, page and line it relied on. A generated model element carries the document, page and row it came from, plus a confidence score. When a system cannot point to a source, it should say so instead of guessing.
Why provenance matters as much as accuracy
An accuracy number describes a system on average. Provenance lets a person check one specific answer in seconds. Those are different things, and in practice the second one decides whether people trust the tool enough to use it.
- Review gets faster. A reviewer who can jump from an answer to its source page confirms it in seconds instead of searching for it.
- Errors become visible. When an answer is wrong, the source shows why: a misread table, an outdated revision, two documents that disagree. A bad answer turns into a defect someone can fix.
- Accountability survives. In regulated work, “the system said so” does not hold up. “The system read this revision of this document, on this page, in this row” is a record an auditor can follow.
What we measured
We applied this rule to our own document-to-model pipeline. It reads the kinds of documents engineering teams live with (interface documents, bills of material, cable run sheets and port tables) and builds a populated engineering model from them. Every element in the model links back to its source.
We built and measured the pipeline at our own expense, on a synthetic test corpus written for the purpose. It contains no controlled or classified material. The corpus was deliberately messy: 18 documents in 7 formats, with 40 defects planted on purpose to see whether the pipeline would catch them or be fooled by them.
F1 is a standard score that balances two questions: did the system find everything it should have, and was everything it found correct? A score of 1.000 is perfect on both. The pipeline also reported zero false conflicts, meaning it never flagged two documents as disagreeing when they did not.
The provenance score is the one we care about most. A 0.974 on connectors means a person still needs to review some of them. A provenance score of 1.000 means that person can go straight to the source of every one of them.
Two caveats. First, this is a synthetic corpus. Real document sets are larger and stranger, and we expect the numbers to move when they meet one. Second, the pipeline is at technology readiness level 3: the method works in a controlled test and has not yet been proven in production. We publish the numbers because they were measured, and we will publish the next ones the same way.
Four questions to ask of any AI system
Whether you build or buy, ask these before you let a system near a decision:
- Does every answer cite a specific source, down to the page and line or row?
- Can a reviewer open that source in one step?
- Does the system say when it has no source, instead of filling the gap?
- Has anyone measured accuracy on documents like yours, with the test method written down before the test was run?
If the answer to any of these is no, the system may still help with drafting. It is not ready to support decisions.
The same rule for companies and agencies
This applies equally to a manufacturer reconciling supplier drawings, a utility reading decades of maintenance records, and an agency answering questions from its own policy library. In every case, someone has to be able to check the answer.
It also shapes where the system runs. Provenance is easiest to trust when the documents never leave your control, so we design assistants to run inside your own boundary, including disconnected environments, and to log every action they take.