Every article we publish carries a record of how it was made: which steps it passed through, which sources it drew on, and, for every claim in it, the sentence in the source that supports it. As of this week that ledger holds 5,185 sourced claims across 2,577 production runs. It is queryable. Nothing about it depends on anyone's memory.
This is the piece about provenance, which I think matters more than anything else a news organization can own in 2026, and which most newsrooms cannot show.
What the ledger is
When a story is produced on our platform, three records are written to the database. A run: this story, this publication, this date, opened and closed. The steps: each stage the story passed through, from loading the writer's voice to fact-checking to the cold read, each with its result, pass or fail, honestly. And the facts: one row per claim, with where it appears in the piece, which source it came from, and the verbatim excerpt from that source that backs it.
The newsroom desk, our internal tool, shows all of this on one screen per article. Click a claim, see its source. Click the source, see every article that cites it.
Why it exists
Because in May a draft with fabricated details reached our review gate, and the review caught it, and I asked how we would ever prove afterward what had been checked. The answer at the time was: we could not. The checks had happened in a conversation that scrolled away. So we made the checks leave a record, and then we made the record a requirement. Since June, a story cannot be declared published on our platform without its run and its ledger. If the ledger is missing, the publish is not complete, and the system says so.
There is one more step that I think matters more than any of the automated ones. Before a story is presented for approval, it is read cold by a reader who has seen none of the sources and none of the drafting: what is the news here, what reads as overstated, what did you not understand, what is buried. The findings are written into the ledger verbatim. A cold read with no recorded findings counts as a cold read that did not happen.
What provenance does not do
I want to be honest about the limits, because the word gets used as a magic one. A ledger proves that a claim was checked against a source. It does not prove the source was right. It does not catch a claim that was never made, or a story that is accurate and about nothing. A wrong fact with a receipt is still a wrong fact.
What it does do is make every error findable and every correction specific. When we get something wrong, and we do, we can say exactly which claim, which source, and which step should have caught it. That is the difference between a correction and an apology.
Why a small newsroom needs this more than a large one
A big newsroom has institutional memory in the form of people. Ours has it in the form of a table. When a reader asks how we know something, the answer should not be "trust us." It should be a record. When an AI agent, reading our coverage on behalf of a reader, asks the same question, the answer has to be a record, because it will not accept anything else.
We are building for that reader. Every article carries its receipts because in a year or two, a source that cannot show them will be arguing for its credibility from a position it cannot win.
This is No. 07 in We Built Our Own. Previous: Analytics without a black box. Next: The people, places and bodies behind the news.