The reasons

The chapters up to here describe what was measured. This one describes the rules that decide what counts as a measurement, and every one of them was written after something went wrong.

They are worth reading even if you never touch this repository, because the failures they are monuments to are ordinary. A prototype tuned against seven documents and then benchmarked against tools that got one attempt each — a model scored on its training set, and the comparison looked convincing. A number written in two places, stale in one, five times over. A detector that fired ten times on the word “NaN” inside legitimate prose.

None of those are exotic. They are what happens when a project measures itself without a rule about how.