Where this leaves you

You have a measurement of the whole pipeline against the live model, in a table a person reads, with an exit code a script can gate on, run by hand and on purpose. You know what it costs and what it cannot tell you.

What it cannot tell you is whether the pipeline would show an invented figure if the model wrote one, because the model rarely does, and a measurement that waits for the rare case is not a test of it. The last chapter writes the model’s wrong answers itself.