the document that notices it has been overtaken

Prose about a system has no way to notice that the system moved. The sentence written confidently one month and false the next looks identical, and the reader has no instrument except the age of the file, which measures when somebody last had a reason to open it rather than whether it is still true.

The hard part is that detecting staleness needs something normally discarded: what the reader believed at the moment it read. A claim and its consumer are two documents, and only their disagreement is interesting.

So each claim was given a version, and every consumer was made to name the version it was written against. When a claim advances, every consumer that named the older one stops building until somebody looks at the pair. It was demonstrated the same day by moving a single claim forward by one patch level and watching the build refuse to proceed. The whole design is in that gesture: the distance between a document and its reader became something that can fail loudly instead of quietly.

One further line matters. The claim files were declared human-maintained and read-only to the machine. A checker that may rewrite the material it checks is not a checker; it is a second author with no incentive to disagree with itself.

The cost is that every claim bump becomes somebody else's interruption, so bumps get batched, and batching buys back the delay the mechanism exists to remove. What it proves is that staleness here is a broken link in a dependency graph rather than a matter for memory.