This is the exploration phase within the RigorPilot skill set, used only when explicitly authorized by researchers and when a traceable research baseline (branch, commit, checkpoint, or run record) already exists. The process consists of an outer loop and an inner loop: the outer loop involves understanding the repository, fixing evaluation metrics and budgets, and scoring ideas for quality control; the inner loop performs one bounded modification or trial run, conducts a smoke test, compares results against the baseline, and then decides whether to stop or continue. Results are written to explore_outputs/, with a note that novelty is merely a hypothesis and cannot be treated as a conclusion without literature comparison and ablation studies.
Why it matters Freeze the task, dataset, evaluation method, and budget first, then change only one variable at a time for comparison to avoid agents producing incomparable results after acting freely.
skills.sh installs 358690 (+1221 / 24h)
Security Audit: Risk warnings present
- References unscanned scripts: scripts/orchestrate_explore.py, scripts/write_outputs.py
- Audit scope: GitHub lllllllama/rigorpilot-skills/skills/ai-research-explore/SKILL.md (only the SKILL.md text was scanned)
Reasoning
- skills.sh installs 1221 in 24 hours, cumulative 358690 (indicating attention only); MIT license
- The unscanned scripts flagged by the audit are orchestrate_explore.py and write_outputs.py, responsible for orchestration and writing results, which aligns with their intended use
Usage
- In a research repository, first use the companion ai-research-reproduction skill to reproduce the baseline, then explicitly instruct the agent to "perform candidate exploration on the current baseline" while providing the dataset, evaluation metrics, and budget
Risks
- It will actually run training and experiments; computational resource consumption is determined by the budget settings; the two scripts are not within the scope of this scan
- The weekly install count for this skill set jumped from dozens to tens of thousands within a week, with an unclear source of popularity