A claim that ships inside a picture escapes every check you have.

We write copy in Markdown and run it through a checker before anything goes out, a register of what is true today, what is conditionally sayable, and what is banned outright, with a pattern for each. It catches a banned phrase in a draft in about a second, and I trusted it more than I should have.

Then we shipped a PDF. The checker reads text files. A PDF, a social card, a screenshot with a caption baked into it, a video, every one of those is a claim that no text check will ever read again once it is rendered. We found a struck sentence sitting inside a rendered document twice in one week, and both times every check had passed.

What we do now is make the artefact tell on itself. Anything rendered writes a small file beside it transcribing exactly what it asserts, generated from the page rather than typed by hand, so it cannot go stale when someone recuts the words. The checker reads that. Our video renderer will refuse to write a single frame if the onscreen text has drifted from the cleared sentence by one character.

It generalises past compliance, any artefact your pipeline cannot read is an artefact your pipeline is not checking, and a green run on a subset reads exactly like a green run.