smevals grades models, prompts and harnesses as three separate variables.

Willison and Jesse Vincent's Prime Radiant shipped a compact eval framework on a four-layer model: an eval contains tasks, each task runs against configs (model plus parameters like system prompt), each execution produces a run, and graders apply checks; string matches, custom scripts, or model judges. Emits an interactive dashboard and a shareable static HTML report. The design point worth stealing is treating the harness as an independent axis, because in agentic systems the harness is frequently the actual bottleneck.