the document acquired a reader

Turning a document into something a build can verify gives it a machine reader, and a machine reader is a new place to be wrong. The claims were prose — which components connect to which, what has to sit between them, the chains those connections compose — and to check them the build had to read sentences written by people in the habit of writing sentences.

Three defects surfaced on the first serious pass. An entry that wrapped onto a second line was read as two entries. A remark written beside an entry was read as an entry of its own. And a version belonging to one entry contaminated the reading of its neighbours, so a claim appeared to be at a version nobody had given it.

The direction of these failures matters more than the failures. A reader that mistakes a remark for a claim is noisy, and noise gets fixed. A reader that fails to see a claim is silent, and a claim it does not see is indistinguishable from a claim nobody wrote — the register reports coverage it does not have, with the same confidence as the coverage it does have.

The fixes went into the reader, not into a demand that people write differently. That is the more consequential decision: the document stayed prose, so the reader had to become forgiving of how prose is actually laid out, rather than the writers learning a notation.

The cost is a permanently more complicated reader, and every loosening is a place where two claims can be confused with one another. What it proves is that the team can tell a checker which is wrong from a document which is wrong.