Evaluation

Prompt v19 of the claims assistant's coverage recommendation fixes the invented clause that Chapter 5 traced to v18, and on a golden set of 600 closed claims it passes 93.5% of cases against v18's 91.0%. Is that enough to ship it? The two numbers could be noise, they could hide a slice that got worse, and they say nothing about any behaviour the golden set does not test. An evaluation can only stop the failures it measures, and a green gate says nothing about the behaviours nobody wrote a test for.