Mechanism
When It Triggers · How It Works · What It Produces
Breaks down how the skill works and what it produces, so you can quickly judge whether it fits your scenario and value.
Trigger Timing
When a researcher has explicitly authorized candidate-only exploration on top of a durable current_research anchor (branch/commit/checkpoint) and chosen task family, dataset, benchmark, evaluation, and SOTA references.
Working Mechanism
- Two-loop rhythm: outer loop freezes context and gates ideas; inner loop makes one bounded candidate change/run and ranks evidence.
- Uses scripts (orchestrate_explore.py, write_outputs.py) and references for policy; delegates code/run to explore-code/explore-run.
- Writes auditable outputs to explore_outputs/ with changelog and comparability report.
Outcome
Candidate ideas with explicit gates, ranked evidence, fair-comparison caveats, and governed experiment records—no trusted reproduction or novelty proof.
Use Cases
Two Typical Use Cases
LoRA variant exploration on local model
A deep learning researcher has a trained local model and wants to explore LoRA variants on an isolated branch. They use ai-research-explore to freeze dataset/benchmark/SOTA, generate ranked candidate ideas, delegate code adaptation to explore-code and short trials to explore-run, then collect evidence in explore_outputs/ with comparability reports. This avoids uncontrolled changes and keeps exploration auditable.
Governed campaign candidate search
A team runs a research campaign with fixed task family and budget. Using ai-research-explore, they map sources, preserve human ideas, gate candidates by cost/gain, and run bounded sweeps via explore-run. The skill writes SCIENTIFIC_CHANGELOG and COMPARABILITY_REPORT, enabling fair comparison against frozen SOTA without overclaiming novelty. It structures collaborative candidate search with rigor.
Skill File