Rigor Reproduce skill in the RigorPilot skill set: First, read the target repository's README and select the smallest reproducible goal documented—prioritizing inference, then evaluation, then training launch, with full training requiring user confirmation. Use the built-in scripts to generate a plan, verify candidate commands, then run the selected commands and validate outputs. Assumptions, deviations, blockers, and points requiring human decision are all recorded in the standardized repro_outputs/ directory. Explicitly not intended for paper summarization, general environment configuration, or score gaming.
Why it matters During reproduction, settings are not secretly altered to boost scores; assumptions, deviations, and bottlenecks are all documented, allowing others to verify the results.
skills.sh install count 352560 (+1260 / 24h)
Security Audit: Risk warnings present
- References unscanned scripts: scripts/orchestrate_repro.py, scripts/doctor.py
- Audit scope: GitHub lllllllama/rigorpilot-skills/skills/ai-research-reproduction/SKILL.md (only the SKILL.md text was scanned)
Rationale
- skills.sh 24-hour install count 1260 (indicates attention level only)
Usage
- Trigger by telling the agent "reproduce this repository according to the README"; requires Python 3.11+ and Git
Risks
- The audit note references unscanned scripts orchestrate_repro.py and doctor.py; the skill executes commands from the target repository, which aligns with its reproduction purpose, but it should be run in a trusted environment
- No account required; the target repository may require additional dependencies, internet access, or a GPU