Source trail
- Industry term
A source trail links a claim back to where it came from, through the original record and its version, the exact passage, and each step that changed or checked it along the way.
Each step between a record and a claim can change something. A search keeps some records and skips others, a model restates a passage in its own words, and another step labels or adds things up. An error can enter at any of them, and the finished claim reads the same either way.
Where you’ll meet it
Your assistant's monthly budget summary says eating out cost $410, which seems high. Under the total, it lists your bank statement as the source and nothing more.
What to do
Ask it to list every transaction it counted toward the $410 and why each one counted as eating out. Then read down that list against your statement. A grocery run or a gas station snack filed under restaurants shows you which step slipped.
Easy to mix up
| Source trail | Personal retrieval |
|---|---|
| A finished claim, like a total in a report | A question, put to your own email, notes, and files |
Check yourself
An AI summary written last month quotes a line from your team's project plan, a shared document, and links to it. You open the link today and can't find the line anywhere. What can you conclude?
Go deeperA trail is recorded as sources are read, and finer trails take more upkeep
Data management calls this provenance, and the W3C, the body that sets web standards, published a model for recording it, called “PROV,” in 2013. An AI tool that answers from documents has a trail only if it records where each passage came from at the moment it fetches it. A model asked afterward to cite its sources may cite a plausible one it never drew on.
A sentence-level citation lets a reader check a number in seconds, and a link to a whole folder leaves the search to them. Finer links take more work to store and maintain, so they suit figures someone will act on, while a rough draft may need only a list of its sources.