explore-run

Use this as the Rigor Improve / Rigor Explore run leaf skill. The installed slug remains explore-run for compatibility.

Use the shared operating principles in ../ai-research-reproduction/references/agent-operating-principles.md; this skill should guide candidate run planning while preserving model judgment about the active repo.

When to apply

• When the researcher explicitly authorizes exploratory runs.
• When the task is a small-subset validation, short-cycle training probe, batch sweep, idle-GPU search, or quick transfer-learning trial.
• When the output should rank candidate runs rather than certify trusted success.

When not to apply

• When the user wants trusted training execution or conservative verification.
• When there is no explicit exploratory authorization.
• When the task is repository setup, intake, or debugging.

Clear boundaries

• This skill owns exploratory execution planning and summary only.
• Use ai-research-explore instead when the task spans both current_research coordination and exploratory code changes.
• It may hand off actual command execution to minimal-run-and-audit or run-train.
• It should keep experiment state isolated from the trusted baseline.
• It should prefer small-subset and short-cycle checks before heavier exploratory runs.
• It should label run results as bounded evidence and explain when a comparison

is not directly fair.

Ranking Semantics

• Pre-execution candidate selection uses three factors: cost, successrate, and expectedgain.
• Default weights should stay conservative unless the researcher explicitly provides selection_weights.
• Budget pruning still applies after scoring through maxvariants and maxshortcycleruns.
• If runs are executed later, downstream ranking should switch to real execution evidence, not stay purely heuristic.

Variant Spec Hints

• Use variant_axes to define the candidate dimension grid.
• Use subsetsizes and shortrun_steps to express exploratory run scale.
• Use selectionweights to rebalance cost, successrate, and expected_gain.
• Use primarymetric and metricgoal so downstream ranking can order executed candidates consistently.

Output expectations

• explore_outputs/CHANGESET.md
• exploreoutputs/SCIENTIFICCHANGELOG.md
• exploreoutputs/COMPARABILITYREPORT.md
• exploreoutputs/TOPRUNS.md
• explore_outputs/status.json

Notes

Use references/execution-policy.md, ../ai-research-reproduction/references/explore-variant-spec.md, ../ai-research-reproduction/references/deep-learning-experiment-principles.md, scripts/planvariants.py, and scripts/writeoutputs.py.