The Rule We Enforced Where We Read And Not Where We Publish

We hold a rule that nothing we write for our own writers may contain an example sentence. No phrase that could be lifted onto a page. A rule here tells a writer what to check, never what to say, because every page has to be written once, from its own evidence, for its own object.

That rule was scoped to the documents our own system reads, since those are the ones we write from. Files we publish sat outside it.

That scoping was exactly backwards. What makes an example dangerous is that it gets executed word for word at scale, and the audience that executes at scale is precisely the audience our published files for machines are written for. So example copy was illegal in the documents one model reads and legal in the documents every model reads.

The rule now follows authorship rather than consumption. It binds every file we author, including the ones we publish, and including the copy our own templates assemble at the moment a file is requested. A description of a field may carry the shape of what that field holds. It may not carry an instance of it.

Our own published files are in breach today, and we would rather say so than wait until it is comfortable. Measured on 17 August 2026 against the theme currently in our mirror, 34 example markers sit across four of the seven files we publish, the largest group of them in llms-full.txt. That figure is a floor rather than a receipt, by this rule’s own terms, because searching a template cannot see what the template assembles when a machine asks for it. The cleanup is owned and has not run.

Which is the other half of what changed. The receipt this rule demands is a count taken over the file as it is emitted, not over the source that generates it, and where we cannot generate the file the audit says so and stays incomplete. An asserted count is not a count.