Autoresearch Experiment Loop (Andrej Karpathy)

Goal → Baseline → Hypothesize → Modify → Verify → Keep/Discard → Repeat. Andrej Karpathy의 autoresearch 원칙을 범용화한 자율 실험 루프. 하나의 메트릭, 제한된 범위, 빠른 검증, 자동 롤백, Git as Memory. ML 실험, 코드 최적화, 성능 튜닝 등 측정 가능한 모든 개선에 적용 가능.

Best For

• repeated experiments
• performance tuning
• hypothesis-driven iteration

Workflow Shape

• Step count: 5
• Opens with: 실험 목표 & 메트릭 정의
• Closes with: 학습 기록 & 실험 보고서
• Difficulty: advanced
• Estimated time: 35 minutes

What This Package Includes

• A publish-ready BlueKiwi workflow package in JSON form.
• Clear stage titles that make review and adaptation easier.
• A reusable structure you can fork for your own team process.

Notes

• Source inspiration: Andrej Karpathy
• Editorial note: Useful when one metric matters and you want fast keep-or-discard decisions.

Included Steps

1. 실험 목표 & 메트릭 정의 — gate
2. 베이스라인 측정 & Git 체크포인트 — action
3. 가설 → 수정 → 검증 → Keep/Discard 루프 — loop
4. 실험 결과 리뷰 — gate
5. 학습 기록 & 실험 보고서 — action