🌐 Language: English | 简体中文 | 繁體中文 | 日本語 | 한국어

Stanford REAP × CoPaper.AI · An academic–industrial AI toolkit for empirical research
_{Built by Stanford's empirical-methodology team — the full pipeline from data cleaning to top-journal submission}

The empirical-research specialist's agent-skills distribution. Not a marketing list — 1,072 skills vendored and cataloged in this repo, wrapped in a numeric benchmark, an eval harness, a security audit, and CI, plus a curated map of 23,000+ skills across 119 repositories in the wider ecosystem.

AERS is two things at once: (1) a small set of first-party flagship skills that run the full empirical pipeline — data cleaning → identification → estimation → robustness → tables/figures → submission-ready draft — and (2) a curated, security-aware catalog of the empirical-research skill ecosystem, organized by research-workflow stage. The differentiator is not the count; it is that the flagship behavior is verified against known answers, not asserted.

git remote set-url origin https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills.git

• What you actually get (the numbers, precisely)
• Verify it yourself in 2 minutes
• Why trust this — three layers
• The flagship pipeline skills
• Start here — pick a skill in 30 seconds
• What makes this more than a 23K-skill dump
• Browse the landscape
  • By research stage
  • Comprehensive skill suites
  • Anti-AIGC detection & de-AI academic writing
  • Tools catalog (tools/)
  • Multi-agent systems · MCP servers · platforms · learning
• Security
• Changelog
• Contributing & citation

Numbers in this README are kept honest and disambiguated. "Vendored" means the files live in this repo and are tracked in a generated catalog; "cataloged ecosystem" means curated links to external repositories.

The security audit covered the original 52-collection / 2,940-file baseline (52/52 CLEAN). Skills vendored after that baseline are tracked in catalog/provenance.json, docs/LICENSE_AUDIT.md, and docs/SKILL_AUDIT.md; run make audit before relying on them in high-trust contexts.

The most persuasive thing here is not a number — it is that the flagship pipeline's behavior is checkable without an API key or paid model. Just Python 3:

git clone https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills.git
cd Auto-Empirical-Research-Skills
make check        # repo validation + unit tests + eval lint + numeric benchmark

The benchmark is the convincing part: it recomputes the gold answer from the raw dataset on every run, so a passing score cannot be faked by hard-coding a number. Out of the box it recovers:

• LaLonde (1986) / Dehejia–Wahba (1999) — the naive observational comparison gets the wrong sign (−$635); covariate adjustment flips it positive (≈ +$1,548) toward the experimental benchmark (≈ +$1,794).
• Card (1995) — IV return to schooling (0.131) exceeds OLS (0.075), with the first-stage F (13.3) reported rather than hidden.
• Plus staggered-DID (TWFE bias vs. group-time truth), sharp RDD, and a bad-control / post-treatment-bias trap.

A pipeline passes only if it surfaces the trap, refuses to headline the misleading number, and matches the recomputed truth. See benchmark/ and the full trust overview in docs/TRUST.md.

💡 Want it hosted and end-to-end? Skip the assembly — copaper.ai runs the empirical pipeline for you, built alongside this catalog by the same Stanford methodology team.

Four parallel implementations of the same 8-step empirical loop — data cleaning → variable construction → descriptives → diagnostics → estimation → robustness → mechanism/heterogeneity → publication-ready tables & figures — plus the submission and de-AIGC stacks. Each uses progressive disclosure: a thin canonical-call spine in SKILL.md, with deep per-step reference manuals loaded only on demand. They coexist; pick by stack and use case.

Why a DSL and explicit ports? Reach for StatsPAI when you trust the one-shot DSL; reach for 00.1/00.2/00.3 when you are teaching, auditing, or must swap every diagnostic by hand. AER-skills then takes a correct analysis to acceptance threshold — these solve different problems and compose.

More ways in:

• Not sure which skill? → docs/CHOOSING_A_SKILL.md · faceted search: docs/search.html
• First 10 minutes, end to end → docs/GETTING_STARTED.md
• Copy-paste a full workflow → docs/GOLDEN_WORKFLOWS.md
• Install into a runtime / use without installing → docs/INSTALL.md
• Machine-readable index → catalog/skills.json · taxonomy: docs/TAXONOMY.md · full catalog: docs/SKILL_CATALOG.md
• FAQ → docs/FAQ.md

Public-skill counts are easy to inflate, and recent studies show large skill indexes are often redundant and occasionally unsafe. AERS competes on verifiable quality, not raw count. Every layer below runs locally via make check and in CI.

make catalog     # regenerate catalog, provenance, audit, enrichment
make validate    # freshness + link / frontmatter checks
make check       # full gate: validate + Python compile + unit tests + eval lint + benchmark

The trust surface is necessary, not sufficient — regex rubrics don't certify prose and a small benchmark doesn't cover every design. It is built to fail fast on known high-cost mistakes. Read the honest scope in docs/TRUST.md and docs/QUALITY_GATE.md.

Topic Ideation → Lit Search → Deep Reading → Research Design → Data Collection
      │              │             │              │                │
      ▼              ▼             ▼              ▼                ▼
     01             02            03             01               04

Data Cleaning → Statistical Analysis → First Draft → Revision → Typesetting
      │              │                    │            │            │
      ▼              ▼                    ▼            ▼            ▼
     04             05                   06           07           08

Replication → Submission → Peer Review Response → Defense
      │           │              │                   │
      ▼           ▼              ▼                   ▼
     09          10             10                  10

Per-stage skill notes (bilingual): 01 Topic & design · 02 Lit review · 03 Paper reading · 04 Data & cleaning · 05 Causal inference · 06 Writing · 07 Revision · 08 Citation & typesetting · 09 Replication · 10 Review response

The pain point AERS exists to fix: ask an AI to "run a DID" and it gives the baseline regression and stops. "Parallel trends?" — it adds one. "Placebo?" — another. Every time, like squeezing toothpaste. A skill is a methodology playbook for the agent: it already knows a complete DID means parallel-trends → baseline → robustness battery → heterogeneity → mechanism, with a defined output at each step.

One of 2026's sharpest pain points: papers failing AIGC detection (Turnitin, GPTZero, CNKI) can be rejected outright. The skills below are the most complete open-source solutions — all MIT, all locally archived (skills/44-48).

Combos: 🇨🇳 Chinese (CNKI/Wanfang/VIP) → chinese-de-aigc · 🇬🇧 English → humanizer_academic · need an audit trail → avoid-ai-writing · general prose → stop-slop.

Unlike the skills above, tools/ catalogs the software and services an agent (or researcher) actually invokes — structured, license- and maintenance-aware, and wired into make validate. Source of truth: tools/tools.json; browsable list: tools/CATALOG.md.

335 tools across 6 categories (curated 2026-06):

• Causal-inference / treatment-effect libraries (32) — DoWhy · EconML · CausalML · DoubleML · CausalPy · causallib · grf · CATENets · TMLE family · Mendelian randomization …
• Econometrics / quasi-experimental libraries (170) — panel FE · DiD (incl. modern/staggered) · event study · RDD · IV · synthetic control/SDID · matching & weighting · sensitivity (fixest · did · HonestDiD · rdrobust · synthdid · reghdfe · csdid · sdid · pyfixest · linearmodels …); plus spatial econometrics (spdep · PySAL/spreg · GeoDa), local projections/IRF & (S)VAR (lpirfs · vars · svars), survey weighting/MRP/raking (survey · samplics · balance), and meta-analysis (metafor · meta · netmeta · metan) — across R/Python/Stata/Julia.
• Autonomous research / data-science agents (51) — end-to-end research & data analysis: AI-Scientist · data-to-paper · Agent Laboratory · RD-Agent · AI-Researcher · STORM · PaperQA2 · gpt-researcher · DeepAnalyze · MetaGPT (DI) · Biomni … (⚠️ includes non-OSI / no-LICENSE repos — confirm terms before use).
• MCP servers (48) — stats execution (StatsPAI · stata-mcp · R/Jupyter MCP) + data access (FRED · World Bank · IMF · OECD · Eurostat · Census · BEA · BLS · SEC EDGAR · OpenAlex · Semantic Scholar · PubMed · Zotero · arXiv …).
• Causal discovery / structure learning (25) — causal-learn · Tetrad/py-tetrad · gCastle · CDT · tigramite (PCMCI) · LiNGAM · NOTEARS/DAGMA · pcalg · bnlearn · pgmpy …
• Benchmarks & datasets (9) — causaldata · IHDP/Twins · ACIC competition data · RealCause · JustCause · Tübingen cause-effect pairs · bnlearn network repository …

Full write-up: tools/README.md.

The original 52 skill collections / 2,940+ files passed a systematic audit — 52/52 CLEAN, zero FLAGGED: no malicious prompts, viruses, reverse shells, or prompt injection. Every "sensitive" hit verified as one of three legitimate categories: defensive security rules, legitimate academic API calls (arXiv / CrossRef / PubMed / FRED / World Bank / OECD / BLS), or standard Claude Code workflow hooks (all local file ops, zero network IO).

Six-phase, defense-in-depth: automated grep across 13 risk categories → 100% manual review of all 6 hook-bearing skills and their 40+ hook scripts (no Bash(*) wildcards anywhere) → three parallel agent content audits → supplemental integrity checks (hidden Unicode, encoding anomalies, HTML injection, network imports).

Key insight: largest ≠ riskiest. The biggest skills all passed; 17-DAAF actually sets the bar for security-conscious design (14 defensive hooks + 32 deny rules + active credential scanning).

Newer vendored additions are tracked in catalog/provenance.json and docs/SKILL_AUDIT.md — run make audit. Full report: SECURITY-SCAN-REPORT.md.

The narrative changelog has moved to CHANGELOG.md. Recent highlights:

• 2026-05 — Vendored AER-skills (top-5 economics submission stack, 9 skills) with weekly upstream sync; expanded the numeric benchmark to 5 causal-recovery tasks and the eval harness to 17 scenarios / 95 rubric items.
• 2026-04 — Completed the 52/52 security baseline; shipped the four full-pipeline flagships (StatsPAI + explicit Python / Stata / R); launched the original chinese-de-aigc skill.
• Earlier — Grew from 43 collections to a curated map of 119 repos / 23,000+ skills; added bilingual README, academic data MCP servers, and multi-agent systems.

Contributions welcome — see CONTRIBUTING.md and the docs/SKILL_SUBMISSION_GUIDE.md. We especially welcome social-science skills (economics, political science, sociology, psychology, education, public health), new causal-inference implementations, MCP servers for academic/government data, Chinese-friendly skills, and multi-agent case studies. New submissions must declare source, license, and category for the provenance audit.

If AERS helps your work, please cite it (CITATION.cff) and star the repo so more researchers can find it.

AI is an amplifier, not a replacement. It handles the heavy lifting; you keep the core judgment.

_{Stanford REAP × CoPaper.AI · An academic–industrial AI toolkit for empirical research}

Visit copaper.ai

WeChat: CoPaper.AI

20 built-in methodology skills · 20-minute empirical paper · powered by StatsPAI (900+ functions, MIT)

Maintained by CoPaper.AI, incubated at Stanford REAP / SCCEI · AI Assistant for Empirical Research