# Benchmark Operator

> Run, inspect, import, and export ControlKeel benchmark suites and multi-subject matrices. Use this when comparing governed and external agents or validating policy changes.

- Skill: `aryaminus/benchmark-operator` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add aryaminus/benchmark-operator`
- Raw SKILL.md: https://api.skillmd.com/api/skills/aryaminus/benchmark-operator/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: Apache-2.0
- Author: aryaminus (https://skillmd.com/u/aryaminus)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/aryaminus/benchmark-operator

---


# Benchmark Operator Skill

Use this skill when the task is benchmark orchestration instead of normal governed delivery work.

## Workflow

1. Select the suite and subjects.
2. Run the suite or import manual outputs.
3. Review catch rate, block rate, expected-rule hit rate, latency, and overhead.
4. Export the run if you need external analysis.

## Local observability feedback loop

For generated observability coverage, prefer the human-gated loop:

1. Inspect `ck_observability` reports for `saved_evals`, `benchmark_drafts`, `benchmark_scenarios`, and `benchmark_history`.
2. Use CLI-only commands for mutations: draft approval, materialization, and benchmark execution are not exposed through MCP in this skill.
3. Run generated observability benchmarks only after a dry-run review and explicit operator approval.
4. Treat `promotions` as advisory evidence; do not mutate policy, router, prompt, or autofix artifacts from benchmark results alone.

## Additional resources

- [Benchmark operator playbook](references/benchmark-playbook.md)