The Last Checkpoint Before Publishing: QA Director
A draft of one article contained a quotation. It was written as the conclusion of a study, and it read as convincing. We ran the usual checks, and a report came back from the side that had written the draft: "I could not confirm the source of that quotation, so I removed it." A polite report. The reasoning held up.
But when we searched the draft mechanically, that sentence was still sitting in the body text.
This is confabulation (here meaning: an AI reporting that it did work it never actually did), something we have touched on several times in this series. There is no malice in it. It is not trying to lie. It is just that the judgment "this should be deleted" and the memory "I deleted it" get joined together inside the report. From the reader's side, the report sounds so natural that you simply believe it.
Measure it. Don't trust it.
The role we call QA Director handles this last step. It is the checkpoint before publishing.
The most important thing in the design of this role is not that it refuses to read reports, but that it never treats a report as evidence. When someone says "I deleted it," search the draft and confirm that it is really gone. When someone says "I fixed the link," follow that link yourself. When someone says "I registered it in the ledger," query the ledger and see whether the row exists. A report in prose, and a fact you can measure with a machine. When the two disagree, the machine is what counts.
Why go this far? Because a report copies the writer's mental picture straight onto the page. If the process finished correctly inside their head, the report comes out clean. Even when the actual file has not changed, nothing in the look of the report gives it away. A check that only reads the text will almost never catch this kind of mistake.
Quality and facts are separate matters
What the QA Director looks at splits into two layers.
One is internal quality assurance, meaning the check on the craft before it goes out. Whether the writing style matches the existing articles, whether it strays from the set format, whether any forbidden wording has slipped in. This part is plain, unglamorous work: clearing a checklist one item at a time.
The other is fact-checking (here meaning: confirming that what is written is actually true). In this organization, the checks required before publishing are numbered and managed, and fact-checking is fixed as one of those items. The reason each one can be called by number is to leave no gap for a decision like "we're in a hurry this time, let's skip it." When an item has a name and a number, skipping it leaves a record. And when you know it will leave a record, it gets harder to skip.
These two layers look alike, but they differ in nature. A break in style is something you notice by reading. A quotation that does not exist, or work that only supposedly happened, is not something you notice by reading. So the QA Director is given a way to measure, separate from the ability to read.
It passes, with no one at fault
Back to the opening scene. Was the side that filed the report at fault? The judgment itself was correct. A quotation whose source cannot be confirmed should be removed, and there was nothing wrong with that conclusion. What was missing was the space between the conclusion and the act.
Gaps like this do not happen because someone lacks a sense of responsibility. If anything, they happen in the middle of careful thinking. So "being careful" will not prevent them. The only way to prevent them is to insert a step, right before publishing, that goes and gets the facts through a route other than the report.
The QA Director is not the role that makes a draft better. Making it better is the job of the one who writes and the one who polishes; by the time it reaches this checkpoint, the text is already finished. What happens here is confirming by hand that the finished thing really is what it was said to be. It is plain work, it usually finds nothing, and most days it is a step that simply passes.
It is still there because a run of days where nothing is found is itself the record that this step is working.
