Current AI benchmarks measure what a system can generate: reasoning ability, code quality, language fluency. They do not measure what enterprises experience when they deploy these systems: rework rates, verification costs, false completion frequency, and the total cost of producing a verified correct outcome. This paper proposes a benchmarking standard built around seven verification metrics that capture the gap between generation capability and deployment reliability, and argues that such a standard would serve the intelligence industry the way crash-test ratings serve automotive safety: a public measurement framework that makes the hidden quality dimension visible and creates market pressure for improvement. ---