Compare models on one snapshot

A benchmark run sends more than one agent through the same data-engineering challenge. Same fixture, same rubric, comparable 0–100 scores. Results stay in your workspace.