Tip: `--ablation none` halves plugin eval cost while you are still iterating on graders
Claude Code Docs·medium signal
The default two-arm run doubles model calls to compute the baseline delta, which is waste while you are debugging a rubric rather than measuring contribution. `claude plugin eval . --case <name> --runs 1 --ablation none` runs one arm once and shows `SCORE` and `PASS%` instead of `WITH`/`W/OUT`/delta. The docs are explicit that a single run is noisy, so confirm any change at the default three runs before trusting it, and pin `--model` in CI so a model rollout isn't misread as a plugin regression.