ai-research-explore

Purpose

Use this as the Rigor Explore compatible skill slug after the researcher explicitly authorizes candidate-only work on top of a durable current_research anchor. The installed slug remains ai-research-explore for compatibility. Rigor Explore is for meaningful and potentially novel deep learning research candidates while preserving scientific rigor, comparability, reproducibility, and auditable collaboration. Novelty and significance remain hypotheses before literature contrast, ablation evidence, and fair comparison. The skill does not promise autonomous discovery, global benchmark completeness, novelty proof, or trusted reproduction success.

Start from the shared operating principles in ../ai-research-reproduction/references/agent-operating-principles.md, then load ../ai-research-reproduction/references/research-rigor-principles.md for research claims and ../ai-research-reproduction/references/deep-learning-experiment-principles.md when experiment details affect comparability or reproducibility.

Fit

Use this skill only when the request has both:

• Explicit exploration authorization such as candidate-only work, isolated

branch or worktree, sweep, several variants, or exploratory ranking.

• A durable current_research context such as a branch, commit, checkpoint,

run record, or already-trained local model state.

Keep narrow code-only requests on explore-code. Keep narrow run-only requests on explore-run. Keep passive repository analysis on analyze-project. Keep README-first reproduction on ai-research-reproduction.

Research Rhythm

Use a two-loop rhythm:

• Outer loop: understand the repository, freeze task/dataset/evaluation/budget,

preserve user ideas, map sources, gate ideas, and decide whether the next experiment is worth running.

• Inner loop: make one bounded candidate change or run, smoke-check it, collect

evidence, rank it against the current anchor, and either stop or return to the outer loop with the new evidence.

This rhythm is a guide, not a rigid autonomous loop. Stop at explicit blockers, unclear scientific meaning, exhausted budget, missing anchor/evaluation, or a human checkpoint.

Workflow

1. Confirm current_research and explicit explore-lane authorization.
2. Accept either legacy variantspec or higher-level researchcampaign.
3. In campaign mode, freeze the task, dataset, benchmark, evaluation source,

SOTA reference, and budget before candidate work.

1. Build only the repo-understanding artifacts needed for the current campaign,

usually through analyze-project.

1. Run bounded, cache-first source lookup when source support matters; prefer

local curated literature such as Zotero if available, then seed sources, repo-local locators, public locators, or optional web lookup. Treat lookup as source resolution, not an open-ended literature search.

1. Preserve researcher-provided ideas, optionally add a small bounded set of

single-variable seed ideas, and rank ideas with explicit gates and score breakdowns.

1. Prefer one clear candidate at a time. Use explore-code for bounded code

adaptation and explore-run for short-cycle trials or sweeps.

1. Use minimal-run-and-audit or run-train only when the exploratory plan

requires real execution evidence.

1. Write candidate-only outputs to analysis_outputs/, sources/, and

exploreoutputs/ as appropriate; never present exploratory gains as trusted reproduction success. Include SCIENTIFICCHANGELOG.md and COMPARABILITY_REPORT.md for candidate scientific meaning and comparison boundaries.

Ranking and Evidence

• Before execution, prioritize candidates by expected gain, cost, success

likelihood, patch surface, dependency drag, evaluation risk, and rollback ease.

• After execution, rank by real evidence first: command status, observed

metrics, artifacts, changed paths, smoke results, and reproducibility notes.

• Keep researcher-provided evaluationsource and sotareference frozen for

the campaign; do not claim they are globally complete.

• If the top ideas are too close or the implementation cannot be decomposed into

auditable units, stop for a checkpoint instead of silently choosing.

Campaign Inputs

research_campaign is preferred for Rigor Explore campaigns, but it should stay minimal. The durable core is:

• current_research
• task_family
• dataset
• benchmark
• evaluation_source
• sota_reference
• compute_budget

Use candidateideas, variantspec, researchlookup, ideapolicy, ideageneration, sourceconstraints, feasibilitypolicy, baselinegate, and execution_policy as optional guidance, not as fields the agent must fill for every campaign. See references/research-campaign-spec.md for the advanced schema and artifact expectations.

Reference Loading

• Load references/ai-research-explore-policy.md for lane safety and candidate

semantics.

• Load references/research-campaign-spec.md only when a campaign file is

present or the user asks for Rigor Explore campaign governance.

• Load ../ai-research-reproduction/references/explore-variant-spec.md for run-level variant matrix

details.

• Load ../ai-research-reproduction/references/research-thinking-loop.md before proposing or ranking candidate changes; it is the required greedy observe-ground-design-compare cycle.
• Load ../ai-research-reproduction/references/research-rigor-principles.md before making novelty, contribution, SOTA, or comparability statements.
• Consult ~/.rigorpilot/PERSONAL_RIGOR.md if present, under ../ai-research-reproduction/references/continuous-learning-policy.md (advisory only; core wins).
• Load ../ai-research-reproduction/references/deep-learning-experiment-principles.md when training,

evaluation, baseline, ablation, metric, checkpoint, or dataset details matter.

• Use scripts/orchestrateexplore.py and scripts/writeoutputs.py for the

existing deterministic artifact workflow.