#regression-testing

I deliberately broke my summariser's prompt, then failed to detect it two different ways: with an LLM judge over a golden dataset, and with a judge-free deterministic assertion. Removing the judge changed nothing. Here is the calculation that would have told me before I started.