Harness refers to the evaluation or runtime scaffolding built around a model; the same model running the same benchmark under different harnesses may differ by several to a dozen-plus percentage points.