What it was for
Nothing with this shape, because nothing needed it. When a person wrote every sentence, the facts lived in their head and the byline carried the accountability. You could ask the author where a number came from and they would remember, or admit they didn't. Provenance was a property of the writer, not of the writing.
Style guides, fact-checking desks and source files are the nearest ancestors, and they are all built around the same assumption: a small number of humans producing a small number of artifacts slowly enough to remember what went into them.
The verdict
EMERGES. Once a machine can assert things at volume, the scarce artifact is a store of claims a human has actually adjudicated — with a source, a date, an approver, and a list of everything built on top of it.
Two properties make it new rather than a tidier source file:
It gates. The machine may only assert what has been approved. This inverts the usual order, where generation comes first and verification chases it. A claim that has not been ruled on is not available to the writer, so the expensive failure — plausible, fluent, wrong, already published — cannot occur by omission.
It heals. Because each artifact records which claims it used, retiring a claim produces the list of everything now stale. The refresh queue is derived rather than remembered.
That second property is the one worth building for, and it is the one most systems skip. Gating is a discipline you can approximate with care. Knowing what breaks when a fact stops being true is not something care produces; it requires the link to have been written down at the time.
What changes
The unit of review moves from the artifact to the claim. Reviewing a page asks whether this page is right. Reviewing a claim asks whether this is true, once, and then every page inherits the answer. The first scales with output. The second scales with how much you actually believe.
The approval step is an answerable call in the sense this playbook means it: you can say what the claim is, why it was safe to publish, and what happens if it turns out to be wrong. A store with no named approver per claim is a store that has recorded facts without anyone standing behind them.
And the failure it prevents is specific. Published work revises. Verdicts get re-stamped, numbers get superseded, a source retracts. Without dependency links, the revision is invisible downstream: the essay, the deck and the eleven posts that cited the old version stay live and stay wrong, and nobody is lying — nobody knows.
The caution
The thing most likely to go wrong here is measuring the store instead of the work. Facts counted, coverage percentage, pages published: these are inventory metrics, and inventory is exactly what got cheap. A graph with 500 approved claims and no argument is a well-organised warehouse.
The number that means something is how often a retirement fires and how much it catches. If nothing has ever been retired, the store is not being used as a store. It is being used as a filing cabinet, and it will quietly rot in the same way the writing it replaced did.