Reduce evaluation cost

Guide — reduce evaluation cost without losing signal.

Levers:

1. Smaller dataset — most regressions show up in the first 100-200 rows
2. Cheaper judge model — most rubrics work on a balanced model; reserve frontier judges for subtle dimensions
3. More scorers, fewer judges — code graders cost ~zero compared to judges
4. Cache + dedupe — don't re-judge identical (input, output) pairs

See also

• Cookbook: cost-aware evaluation