{ "markdown": "# RigorPilot Skills\n\n[![Skillselion Top 100](https://skillselion.com/badge/skills/lllllllama/rigorpilot-skills/paper-context-resolver.svg?award=1)](https://skillselion.com/skills/lllllllama/rigorpilot-skills/paper-context-resolver)\n\nResearch-first Agent Skills for Deep Learning Experiments: a harness that turns README commands into bounded runs and\nauditable evidence for deep-learning repositories. Trusted reproduction is the\ndefault; exploration requires explicit authorization.\n\n

\n English |\n 简体中文\n

\n\n

\n \"CI\"\n \"Listed\n \"skills.sh\n \"GitHub\n \"MIT\n \"Agent\n \"platforms\"\n \"tests\"\n \"external\n

\n\n

\n Examples ·\n Real-repo Evidence ·\n Quick Start ·\n Skill Index\n

\n\n\n\n## 📄 At a Glance: How RigorPilot Annotates a README\n\nRigorPilot reads the target repository's original README directly. Every word,\nblank line, and line ending stays intact; RigorPilot only inserts a status card\nat the end of each section. You can see what ran, what happened, and why the\nagent stopped before opening the underlying evidence.\n\n| Original README | In-place RigorPilot verdict | Auditable evidence |\n|---|---|---|\n| Commands, prose, badges, images, GIFs, videos, and HTML remain unchanged | Success, partial, blocked, read-only, or authorization required | `SUMMARY.md`, `COMMANDS.md`, `LOG.md`, `status.json` |\n\n🟢 success · 🔵 not executed · ⚪ read only · 🟡 partial · 🔴 blocked · 🟣 decision required\n\n
\n \"Annotated\n
\n The same evaluation command moves from a missing checkpoint to ready assets; training remains untouched without authorization.\n
\n\n| Inspectable demo | What it shows |\n|---|---|\n| [First attempt](examples/annotated-readme-demo/first-run/ANNOTATED_README.md) | 🟡 missing checkpoint with the real error excerpt · 🟡 dataset not ready · 🟣 training awaits authorization |\n| [After assets are ready](examples/annotated-readme-demo/after-setup/ANNOTATED_README.md) | 🟢 the same evaluation command succeeds with observed `mIoU` / `aAcc` |\n\n### Verified on a Real Public Repository: micrograd\n\n| Actual execution | README fidelity | Inspect directly |\n|---|---|---|\n| 🟢 `2` tests passed in 7.62 seconds | `8` headings = `8` annotations; stripped SHA-256 exactly matches the original file | [original repository](https://github.com/karpathy/micrograd/tree/7bc720e951fe422b8f8814aa5aa1b64121d26b4c) · [full annotated README with retained repo files](benchmark_outputs/showcases/micrograd/repo/RIGORPILOT_README.md) · [benchmark report](benchmark_outputs/external_micrograd.json) |\n\n\n\n## 🧪 Reproduced on Real Public Repositories\n\n\n \n \n \n \n \n \n \n \n
\n \"RigorPilot
\n micrograd · correctness
\n 🟢 2 tests passed in 7.62 s · 8 headings = 8 annotations
repository files retained · SHA-256 exact
\n Open RigorPilot README inside the real repo →\n 		\n \"RigorPilot
\n minGPT · risk boundary
\n 🔵 Test selected · no implicit model download
11 headings = 11 annotations · SHA-256 exact
\n Open RigorPilot README inside the real repo →\n
\n \"RigorPilot
\n PyTorch MNIST · data and metric capture
\n 🟡 bounded startup · loss 0.038893
1 heading = 1 annotation · SHA-256 exact
\n Open RigorPilot README inside the real repo →\n 		\n \"RigorPilot
\n nanoGPT Shakespeare · bounded training
\n 🟡 train loss 4.1676 · validation loss 4.1649
11 headings = 11 annotations · SHA-256 exact
\n Open RigorPilot README inside the real repo →\n
\n\n

\n Browse all four reproductions in one evidence index →\n

\n\n**Recorded result:** `56/56` regression scripts and `4/4` external protocols\npassed; the external suite took `251.0 s`, used at most `98.67 MiB` per\nworkspace, made `0` API calls, removed every temporary workspace, and retained\nabout `17.9 MiB` of tracked repository showcase snapshots.\n\n[Latest suite JSON](benchmark_outputs/external_suite_latest.json) ·\n[history](benchmark_outputs/external_suite_history.jsonl) ·\n[case definitions](benchmarks/external_cases.json) ·\n[methodology and limits](benchmarks/README.md)\n\n> `partial` proves bounded startup, metric capture, timeout handling, source\n> integrity, and cleanup. It does not prove convergence or reproduce a paper score.\n\n\n\n## 🚀 Install\n\nAll skills:\n\n```bash\nnpx skills add lllllllama/rigorpilot-skills --all\n```\n\nOnly the trusted reproduction skill:\n\n```bash\nnpx skills add lllllllama/rigorpilot-skills --skill ai-research-reproduction\n```\n\n
\nOther install paths, agent commands, and runtime controls\n\nClaude Code commands: `/ai-research-reproduction`, `/ai-research-explore`,\n`/analyze-project`, `/safe-debug`.\n\nEach executed command receives a run ID and writes atomic state, append-only\nevents, resource samples, and complete stdout/stderr under\n`/_runtime//`. Cancellation, restart recovery, and explicit\nretry preserve process and attempt lineage. Model profiles record identity and\ncapabilities without credentials.\n\nThe preferred source is `lllllllama/rigorpilot-skills`; the legacy\n`lllllllama/ai-paper-reproduction-skills` slug remains a compatibility fallback.\n\n
\n\n## 📄 Output Bundle\n\nEvery run writes the original README plus section-level verdicts to\n`repro_outputs/ANNOTATED_README.md`. Each verdict links to `SUMMARY.md`,\n`COMMANDS.md`, `LOG.md`, and `status.json`; the header records a rubric-style\ncoverage score. 🟢 success · 🔵 not executed · ⚪ read only · 🟡 partial ·\n🔴 blocked · 🟣 decision required.\n\n## 🎯 Choose an Entry Point\n\n| What you want to do | RigorPilot display name | Current skill slug |\n|---|---|---|\n| Reproduce a deep learning repository from README commands | Rigor Reproduce | `ai-research-reproduction` |\n| Analyze repository structure, entrypoints, and risks without editing | Rigor Analyze / Audit | `analyze-project` |\n| Prepare environment, datasets, weights, and cache assumptions | Rigor Setup | `env-and-assets-bootstrap` |\n| Run documented inference or evaluation conservatively | Rigor Run | `minimal-run-and-audit` |\n| Start or verify training conservatively | Rigor Train | `run-train` |\n| Debug a failure safely, diagnose before patching | Rigor Debug / Audit | `safe-debug` |\n| Explore candidates on top of `current_research` | Rigor Explore | `ai-research-explore` |\n| Implement candidate changes on an isolated branch | Rigor Improve | `explore-code` |\n| Run small probes or short-cycle experiments | Rigor Explore / Improve | `explore-run` |\n\nBundled helper skills are usually called by orchestrators:\n\n- `repo-intake-and-plan`\n- `paper-context-resolver`\n\n## 🛣️ Lane Model\n\n### 🔒 Trusted Lane\n\nUse this lane for reproduction, setup, read-only analysis, conservative\nexecution, training verification, and safe debugging.\n\n- Primary entrypoint: `ai-research-reproduction`\n- Output directories: `repro_outputs/`, `train_outputs/`, `analysis_outputs/`, `debug_outputs/`\n- Core requirement: preserve scientific meaning, minimize semantic changes, and record assumptions, blockers, and evidence.\n\n### 🧪 Explore Lane\n\nUse this lane only when the researcher explicitly authorizes candidate-only\nexploration.\n\n- Primary entrypoint: `ai-research-explore`\n- Leaf skills: `explore-code`, `explore-run`\n- Output directory: `explore_outputs/`\n- Key anchor: `current_research`\n\n`current_research` should be a durable research state such as a branch, commit,\ncheckpoint, run record, or already-trained local model state. Explore outputs\nare always candidate results. They must not claim trusted reproduction success,\ncomplete benchmark results, or verified novelty.\n\n## 🔬 Core Research Principles\n\n1. Do not chase scores blindly: score gains must have explanatory value.\n2. Do not claim novelty lightly: novelty needs literature, code, or experimental evidence.\n3. Do not break comparability silently: if evaluation conditions change, say that results are not directly comparable.\n4. Do not disguise engineering fixes as research contributions.\n5. Do not leave collaborators out of control: important changes must be auditable, reversible, and explainable.\n\nSee [references/research-rigor-principles.md](references/research-rigor-principles.md)\nand [references/agent-operating-principles.md](references/agent-operating-principles.md).\n\n## 🔁 Lifecycle View\n\nThe repository follows a shallow lifecycle-oriented routing model:\n\n```mermaid\nflowchart LR\n A[Understand] --> B[Reproduce]\n B --> C[Set up]\n C --> D[Run or train]\n D --> E[Debug]\n E --> F[Report]\n B -. explicit only .-> G[Explore]\n G --> H[Rank candidates]\n H --> F\n```\n\nThe lifecycle helps the agent choose the right lane and evidence target. It\ndoes not force every repository into a fixed implementation sequence.\n\n## 🧠 Research Thinking Loop\n\nAgents implement well but often think in engineering steps. Once the\nresearcher freezes the evaluation contract and explicitly authorizes\nexploration, `ai-research-explore` runs a codified **greedy research cycle** —\nfrom observation to a fair keep-or-rollback decision\n([full contract](references/research-thinking-loop.md)):\n\n```mermaid\nflowchart LR\n A[Observe run evidence] --> B[Ground: paper · code · prior runs · intuition]\n B --> C[Falsifiable hypothesis]\n C --> D[Single-variable design]\n D --> E[Bounded run]\n E --> F{Fair comparison vs current best}\n F -- better --> G[Keep as new candidate best]\n F -- worse or unfair --> H[Roll back, record why]\n G --> I[Ledger entry]\n H --> I\n I --> A\n```\n\n- Every hypothesis carries a **labeled evidence anchor** — `paper`, `code`,\n `prior-run`, or `intuition`; unanchored ideas queue in the idea bank and\n never execute.\n- **Greedy applies to selection, not honesty**: a keep requires comparable\n evidence under the frozen contract; ties favor the simpler, cheaper change.\n- Underneath: hard-gated idea ranking, atomic idea decomposition, three-layer\n implementation fidelity (planned / heuristic / observed), and\n executor-emitted file-level evidence.\n- Lineage: adapts the greedy solution-space search of\n [AIDE](https://arxiv.org/abs/2502.13138) and the managed agentic tree search\n of [AI-Scientist-v2](https://arxiv.org/abs/2504.08066), constrained by\n RigorPilot's comparability-first gates.\n\n## 🌱 Continuous Learning\n\nThe shipped skills are an **immutable universal rigor core**; personalization\nlives in a user-owned overlay\n([policy](references/continuous-learning-policy.md)):\n\n- Failed runs — and their later fixes — are auto-recorded as one-line lessons\n in `~/.rigorpilot/lessons.jsonl` (opt out with `RIGORPILOT_LESSONS=0`).\n- `python shared/scripts/lessons_store.py summarize` distills them into\n `~/.rigorpilot/PERSONAL_RIGOR.md`, which skills consult at run start as the\n researcher's standing preferences and known pitfalls.\n- Hard rules: lessons are **advisory only** — they never relax rigor gates,\n never store secrets, never edit skill files. Delete the folder and the\n skills return to the universal base.\n\n## 🧾 Suggested Research Evidence\n\n| Artifact | Purpose |\n|---|---|\n| `SCIENTIFIC_CHANGELOG.md` | Records what changed, why it changed, whether it affects scientific meaning, and whether it remains comparable. |\n| `COMPARABILITY_REPORT.md` | Explains whether results can still be compared to the README, paper, baseline, or SOTA reference. |\n| `REPRODUCIBILITY_NOTES.md` | Records commands, configs, seeds, checkpoints, datasets, environment assumptions, and known gaps. |\n| `NOVELTY_CLAIM.md` | States possible novelty as a hypothesis, with supporting evidence, missing evidence, limitations, and required ablations. |\n| `ABLATION_PLAN.md` | Describes which variables must be isolated to validate a candidate change. |\n| `EXPERIMENT_LEDGER.md` | Records runs, metrics, commands, artifacts, changed files, and evidence status. |\n\n`SCIENTIFIC_CHANGELOG.md`, `COMPARABILITY_REPORT.md`, and `EXPERIMENT_LEDGER.md`\nare already generated by standard trusted / explore writers. The remaining names\n(`REPRODUCIBILITY_NOTES.md`, `NOVELTY_CLAIM.md`, `ABLATION_PLAN.md`) are\nfuture-compatible evidence concepts.\n\n## 📁 Output Directories\n\n| Directory | Contents |\n|---|---|\n| `repro_outputs/` | Trusted reproduction bundle, including `ANNOTATED_README.md` |\n| `train_outputs/` | Trusted training bundle |\n| `analysis_outputs/` | Read-only analysis, research map, change map, eval contract, idea seeds, atomic idea map, implementation fidelity, and related outputs |\n| `debug_outputs/` | Safe debug diagnosis and patch plan |\n| `sources/` | Free-first research lookup records, repo-local extraction, and auditable index |\n| `explore_outputs/` | Changeset, idea gate, experiment plan, manifest, ledger, candidate ranking, and related outputs |\n\n## 🧩 Campaign Inputs\n\n`ai-research-explore` still accepts `variant_spec.json`, but\n`research_campaign.json` or `research_campaign.yaml` is preferred for Rigor\nExplore campaigns.\n\nDurable core fields:\n\n- `current_research`\n- `task_family`\n- `dataset`\n- `benchmark`\n- `evaluation_source`\n- `sota_reference`\n- `compute_budget`\n\nOptional fields:\n\n- `candidate_ideas`\n- `variant_spec`\n- `research_lookup`\n- `idea_policy`\n- `idea_generation`\n- `source_constraints`\n- `feasibility_policy`\n\nSee [skills/ai-research-explore/references/research-campaign-spec.md](skills/ai-research-explore/references/research-campaign-spec.md).\n\n## 🌐 Multi-Agent, Multi-Model\n\nRigorPilot is model-agnostic by construction:\n\n- **Agent Skills standard** — every skill is a spec-compliant `SKILL.md`\n ([agentskills.io](https://agentskills.io)), the format adopted by Claude\n Code, OpenAI Codex, Cursor, VS Code, Gemini CLI, and 30+ other tools.\n `npx skills add lllllllama/rigorpilot-skills` works for any of them.\n- **`AGENTS.md` routing** — the root [`AGENTS.md`](AGENTS.md) gives\n AGENTS.md-aware agents (Codex, Cursor, Copilot, Gemini CLI, Aider, Zed, …)\n the lane model, entrypoint table, and hard rules without any install step.\n- **Same contract, any model** — SKILL.md instructions carry no\n model-specific tool syntax; the evidence bundles (`status.json`,\n `ANNOTATED_README.md`, …) are identical whichever model executes the run,\n so results stay comparable across GPT-, Claude-, and Gemini-based agents.\n- **Per-skill client mirrors** — `skills/*/agents/openai.yaml` and\n `.claude/commands/*` keep Codex- and Claude-specific entry points in sync\n with the canonical contract.\n\n## 🛠️ Local Install\n\nUse the Python installer only when developing locally, needing a project-scoped\ninstall, or manually targeting client directories.\n\n```bash\npython scripts/install_skills.py --client agents --target \"$HOME/.agents/skills\" --force\npython scripts/install_skills.py --client codex --target \"$HOME/.codex/skills\" --force\npython scripts/install_skills.py --client claude --target \"$HOME/.claude/skills\" --force\n```\n\nProject-scoped examples:\n\n```bash\npython scripts/install_skills.py --client agents --target ./.agents/skills --force\npython scripts/install_skills.py --client claude --target ./.claude/skills --force\n```\n\nThese commands are written to work in both Windows PowerShell and Linux shells.\n\n## 💬 Example Prompts\n\n**Trusted reproduction**\n\n```text\nUse ai-research-reproduction on this deep learning research repo. Stay README-first, prefer documented inference or evaluation, avoid unnecessary repo changes, and write outputs to repro_outputs/.\n```\n\n**Read-only analysis**\n\n```text\nUse analyze-project on this repo. Read the code, map the model and training entrypoints, and flag suspicious patterns without editing files.\n```\n\n**Safe debug**\n\n```text\nUse safe-debug on this traceback. Diagnose the failure first, propose the smallest safe fix, and do not patch until I approve.\n```\n\n**Candidate exploration**\n\n```text\nUse ai-research-explore with research_campaign.json. Treat the task family, dataset, evaluation source, and SOTA table as frozen inputs. Rank candidate ideas and write evidence outputs to analysis_outputs/ and explore_outputs/.\n```\n\n## ✅ Local Validation\n\nRun everything (CI entrypoint):\n\n```bash\npython scripts/run_all_tests.py\n```\n\nBasic checks:\n\n```bash\npython scripts/validate_repo.py\npython scripts/test_skill_registry.py\npython scripts/test_trigger_boundaries.py\npython scripts/test_operating_principles_structure.py\npython scripts/test_claude_command_wrappers.py\npython scripts/test_readme_selection.py\n```\n\nCore output and explore regressions:\n\n```bash\npython scripts/test_output_rendering.py\npython scripts/test_readme_annotation.py\npython scripts/test_train_output_rendering.py\npython scripts/test_analysis_output_rendering.py\npython scripts/test_safe_debug_output_rendering.py\npython scripts/test_research_explore_dry_run.py\npython scripts/test_research_explore_campaign_flow.py\npython scripts/test_research_explore_artifact_consistency.py\npython scripts/test_research_explore_variant_execution.py\npython scripts/test_research_explore_nontraining_execution.py\npython scripts/test_atomic_idea_decomposition.py\npython scripts/test_idea_seed_generation.py\npython scripts/test_implementation_fidelity.py\n```\n\nInstall-related regressions:\n\n```bash\npython scripts/test_bootstrap_env.py\npython scripts/test_install_targets.py\npython scripts/test_setup_planning.py\n```\n\n## 🧭 Current Repo Snapshot\n\n- `11` skills total: `9` public skills and `2` helper skills.\n- `6` trusted-lane public skills and `3` explore-lane public skills.\n- `4` project-scoped Claude Code wrappers under `.claude/commands/`.\n- `59` root Python scripts, including `55` test scripts.\n- Documentation and command examples are kept usable from both Windows PowerShell and Linux shells.\n\n## ⚠️ Current Limits\n\n- The persistent queue is a single-host, single-writer scheduler; resource\n requests provide admission control, not OS-level CPU, RAM, or GPU isolation.\n- The external suite covers four repositories, but minGPT is selection-only\n and the two training cases prove bounded startup rather than convergence or\n paper-result reproduction.\n- `run-train` remains a bounded training monitor; long runs must be submitted\n deliberately through the queue or an external scheduler.\n- Trusted reproduction avoids silent semantic changes.\n- Helper skills stay narrow and are not public catch-all entrypoints.\n- Exploratory work must stay isolated from trusted baselines.\n- `ai-research-explore` is the governed Rigor Explore compatible slug, not an open-ended autonomous research agent.\n\n## 📚 References\n\n- [Research rigor principles](references/research-rigor-principles.md)\n- [Deep learning experiment principles](references/deep-learning-experiment-principles.md)\n- [Shared operating principles](references/agent-operating-principles.md)\n- [Skill registry](references/skill-registry.json)\n- [Routing policy](references/routing-policy.md)\n- [Trigger boundary policy](references/trigger-boundary-policy.md)\n- [Client compatibility policy](references/client-compatibility-policy.md)\n- [Output contract](references/output-contract.md)\n- [Research pitfall checklist](references/research-pitfall-checklist.md)\n\n## 🧱 Scope\n\nRigorPilot Skills is a research-first skill repository for deep learning\nexperiments. It focuses on scientific meaning, comparability, reproducibility,\ncollaborator control, and auditable workflow boundaries. It helps agents move\nresearch forward more reliably, but it does not replace researcher judgment.\n", "bytes": 22726, "sha": "d39f6b5370a45b4dc4a9a5c76fcfb8fc6dd5615b64c021586772f93a8a4917f4", "repo_slug": "lllllllama/rigorpilot-skills", "fonte": "repo", "truncated": false, "api": "https://agentalog.com/api/listings/skl_lllllllama_rigorpilot_skills_minimal_run_5459fbe4/readme" }