Before we run the numbers on whether accurate shows stay accurate, here's the test, locked in public.
We rebuilt our scoring bar around startable weeks, regraded the whole corpus, and tested the results against two skill-less baselines.