Last reviewed August 2026

Seven of Nine

We ran a check to confirm seven copies of a nine-item list matched. It passed. Every copy was the old seven-item version. What a check that compares copies to each other can and cannot tell you — and why consistency is not correctness.

We call these "the nine nazgûl" internally. Ugly, encountering them has bad consequences and they will vanish soon enough… Why? We are sure we will find more items to log eventually...


The case

We have a log. Currently it is at nine documented ways AI models damage text. The immediately previous iteration was at seven ways...

It is carried on seven pages of this site (the number is coincidence...LOL), because a candidate arriving at any one of them should not have to go looking for it.

Thus, there are seven copies of a nine-item list. You can see where this is going...


The check that passed

We ran a sync to confirm all seven copies matched.

It came back clean. Copies compared, byte-identical, 1,334 characters each.

Every one of them was the old seven-item version.

The check compared the copies to each other. It never compared any of them to the source.

So it did exactly what it was asked to do, reported success, and told us nothing. Identical copies of the wrong thing look precisely like identical copies of the right thing.

To add insult to injury? Days later the same AI reported it as three copies compared. Did it compare seven at the time, or did it compare three?

Now no one will ever know.


What the two missing items were

Item seven: it shifts the meaning while keeping the shape*.*

Item eight: it reproduces its own work from memory, and the copy is worse*.*

Those were the two that had gone missing.

We shipped a broken copy of the list that names the failure — by committing the failure it names.

A list reproduced from memory. The copy was worse. Two items shorter, and it read perfectly well, because a seven-item list looks exactly as complete as a nine-item one when you do not know how many there should be.


Why it took so long to find

Nothing was visibly wrong.

Each copy was well-formed. The formatting was right. The items that were there were correct. A reader on any one of those pages would have found a coherent, useful checklist and no reason to doubt it.

The fault was only visible from outside — by knowing the list should have nine. The Nine Nazgûl: a memorable phrase assisting human memory of the real number.

And that is the same form as for every case in the log. No error, no flag, no gap. The prose closes over the loss.


What we changed

The check now compares every copy to the source, not to the other copies.

And the number is stated where it is checkable. The list says nine, in a place a person will read, so a seven-item copy contradicts itself.

A count is a cheap thing to verify and a hard thing to fake.


What is worth knowing

A check that compares a document to its copies cannot tell you if their contents are right or wrong.

That sounds obvious written down. It was not obvious while the check was passing — because a passing check feels like information, and it is only ever information about the one thing it measured. Which was not content quantity or value.

You have this in your own work.

Your reference manager confirms that every in-text citation matches an entry in your reference list. It will tell you if one is missing. It cannot tell you whether the entry is right — whether the journal exists, whether the authors wrote that paper, whether the source says what you claim it says. A bibliography can be perfectly consistent and entirely fabricated, and the software will pass it every time.

A style checker confirms you have been consistent about the serial comma. It has no opinion about whether your argument holds.

In every case the tool compares parts of your work to each other.

So ask one question of any check: what is it comparing to?

If the answer is the other parts of the same thing, you have consistency.

Consistency is not correctness, and only one of them is worth having.


The point of telling you

We could have fixed this quietly. Nobody had noticed and nobody was going to. Indeed, once fixed? No one would care.

The real value to others was in the recording.

The log exists because we were willing to document what went wrong based on our own work.

After our fix? The Nine Nazgûl were back in action... Nine items. Now on all seven pages, and in the printable document.


The nine, in full, with dates and both versions of the text: The Ghost in the Machine. There is a one-page printable.

Fast Help & More Resources