Some of the writing here was spoken, not generated.

If you came from a piece with a mark under the date, that piece wasn't written by a machine on the writer's behalf. It started as a recording — a person talking through what they wanted to say — and what you read is that talk, assembled.

What the percentage counts

The number is the share of sentences in the finished piece that trace back to the recording word-for-word, or near enough that only grammar moved. Sentences are the unit. Nothing else is counted, and nothing is rounded in the writer's favor.

It can be 100%, and often is. Speech comes out with false starts and doubled-back sentences that no reader needs, and taking those out is subtraction — the piece gets shorter, not different, so every sentence left still traces. When the number lands lower, the gap is almost always the small connective words that hold two spoken thoughts together on a page. Those are capped, deliberately, so they can never quietly become writing the person didn't do.

When a sentence appears that the recording can't account for, it gets cut before publication rather than shipped and explained. That check runs in code, not in a model's judgment about its own output.

When a piece doesn't carry the mark

Then it says only Written in Verbaria, and that means what it says: the piece was written here, in the editor, with whatever help the writer chose to use. It is not a claim about whose words the sentences are, and we don't dress it up as one. A mark that appeared on everything would be a logo, not a credential.

Why bother

Because "a human wrote this" is about to be the least believable sentence on the internet, and saying it louder won't fix that. The only thing that helps is being able to show your work — so the tool that helped keeps a record of what it did and didn't do, and the writer publishes it.