Most AI benchmarks today are built on public benchmarks and exam-style questions. Their questions are visible, easily absorbed by models, and fast approaching saturation. They measure what sits above the waterline.
What decides whether AI delivers value in real work lies below the surface: long-horizon tasks, proprietary software and tool environments, tacit workflows and business rules, and outcomes only domain experts can judge.
Seven Eighths Lab builds benchmarks from real work, evaluating how well AI agents complete enterprise tasks in executable environments, against verifiable criteria. The lab is an initiative of Knowlyr.