The audit recipe
A finished piece of work carries claims that came from an assistant, and the draft does not say which ones. This is the procedure that makes every one of them get looked at and then disposed of deliberately. It can be run on any near-final piece of work, in any tool, and what the dispositions ask of you is not the same for each.
What actually goes wrong
A finished draft does not announce where its claims came from. A claim the assistant produced out of training alone sits in the same sentence rhythm as a claim taken from a document you hold, and nothing in the text separates them.
The second half is what defeats a careful read. A model fills a gap before the gap is noticed as a gap. The claim arrives fluent and confident, so nothing in the experience of reading marks it as the one to check.
So the failure is not judgment applied badly to a claim. It is a claim never reaching your judgment at all. Reading the draft again does not fix that, because the read is what the claim slipped past the first time. A list fixes it, because a list does not wait for a claim to raise its hand.
The ledger
The unit is one row per checkable claim. A checkable claim is anything a reader could go and verify: a date, a figure, an article or clause reference, a named standard, a version number, an attribution, a parameter set, a quantity.
| column | what goes in it |
|---|---|
| the claim | one sentence, as the artefact states it, not as you remember meaning it |
| where it sits | the page, slide or section, so the row can be found again |
| where it traces | the note, document or source the claim actually comes from, or the word nothing |
| the grade at home | how well grounded the claim is where it lives, not how confident the artefact sounds |
| disposition | one of the three below, filled before the row is allowed to close |
Four rules make the list do work that a careful read does not.
The three dispositions
The middle disposition is what makes the other two affordable. Material can leave at the grade it has, labelled, and no one is pretending otherwise. The first disposition is the one that cannot be taken at a distance: grading a claim to the level that carries your judgment is a statement that you understood it.
Confirming at the source has one rule of its own. Produce the exact words of the passage, not a restatement. A verbatim quote is hard to reconstruct from memory, so being unable to produce one is itself the answer: the source was not read. The quote is then compared against the document mechanically, by a search for the string, because a second reading by the same assistant checks the faculty that failed.
When it runs
The trigger is not someone remembering. It is the boundary: work crossing to a client, an employer or an audience. The workable form is a definition of done. The artefact is not finished until its ledger exists, which puts the extraction inside the work instead of beside it. The earlier version of this gate waited on someone remembering, and under deadline nobody did.
The cost is real, and the trigger is what bounds it. Most claims in most work never need this, which is why the trigger is an event and not a habit. What the boundary buys is that the expensive check lands on the artefact that can do damage, and on nothing else.
What it does not do
The ledger and the quote comparison settle whether the source says the thing. They cannot settle whether the conclusion drawn from it holds. Quoted correctly and reasoned away from is a different failure, it survives every mechanical check, and it stays with the person who signs the work. The procedure states that limit rather than implying a coverage it does not have.
The ledger is also not the claim's permanent record. It is per artefact and disposable, built for one crossing and kept as evidence that the crossing was checked. The record that travels with a claim for its whole life is a different object, described on the trust page.