We showed machines a score we were deleting

SleekNova publishes a guide at llms-full.txt that tells AI models how our product data is shaped and how to quote us. While retiring an old field this week, we read that guide properly and found two things in it that were not true.

The first was a worked example showing how to cite a product. It ended with the line “Tier-1 verified with 85/100 authority score”. That number came from a scoring field we replaced eight months ago and have now removed entirely. We were handing models a sentence pattern built on a measurement we no longer keep.

The second was an example of our structured data. It showed a publisher block carrying the description “Tier 1 verification. Authority Score: 85/100”. We checked what our site actually sends. It sends a publisher name and an identifier, and no description at all. That example described output SleekNova has never published, and it had been sitting there long enough that nobody could say when it was last true.

Both are now gone from the guide. The retired field’s entry in the same document used to read “deprecated”, which sounds like a warning and is really an invitation, so it now says plainly that the field is retired and points at the one that replaced it. The migration table stays exactly where it was, because anything still holding the old name needs somewhere to be sent.

What makes this one worth publishing is how close it came to surviving. We map every place a field name appears before we touch it, and the map found neither of these. Both carried the idea without carrying the name, as a number inside an example rather than as the field it came from, and no search for the field could ever have reached them. A person reading the document found them.

Six more fields are queued for removal. Each one leaves its old values behind in examples the same way, so reading the published documents is now part of retiring a field, not a check we run afterwards.