A redesigned summary screen must use the same counting rule as the scoring screen, or the two contradict each other on stage
Situation: a UI redesign turned a detailed findings list into large 'problem' cards with a headline count ('5 problems found'). A later screen scored the report against ground truth ('4 of 5 caught, 1 missed').
Bug found by a visual QA pass on a recorded real-model run (not by tests, and not on the scripted fixture): one finding had been downgraded by a reviewing agent. The new cards filtered only on 'category is an allegation and not rejected', so the downgraded item was shown and counted as a problem. The scorer counts only accepted findings, so the reveal marked that same item as missed. On screen, the product listed a finding and then said it had missed it.
Fix: derive the card list from the same predicate the scorer uses (accepted/verified only), and show downgraded or rejected items in a separate, collapsed group labelled as not counted. The headline count then matches the score.
Lessons: 1. When simplifying a view, enumerate every status value the data can carry (grep the backend enum or dump statuses from saved real runs) before writing the filter. The scripted fixture only ever produced 'verified', which is why every test passed. 2. QA the redesign on a real recorded run as well as the fixture; real model output contains the awkward states. 3. Ask the QA agent to compare numbers across screens explicitly (count on screen A vs score on screen B).