Invisible Hands for Economists

imageLicense: MIT imageSkill: econ-research-lab imageVersion imageTheory Proof: Lean 4 imageReplication: AEA Standard imageArchitecture: PI DAG imageSelf-Improvement: Dream-RSI

An auditable, end-to-end Multi-Agent Research Lab for Empirical, Theoretical, and Structural Economics. One accountable PI coordinates bounded specialist agents, discovers and verifies data with cryptographic hashes, formalizes mathematical proofs with Lean 4, runs reproducible econometric pipelines, and stress-tests manuscripts via adversarial peer review.

Architecture • Trust Surface • Core Capabilities • Protocol Reference • Installation • CLI Tools

🏛 Trust Surface & Audit Rigor

Unlike generic prompt catalogs, Invisible Hands for Economists enforces deterministic, cryptographic, and mathematical verification gates. AI generation is halted whenever a gate fails:

🧭 Architecture

flowchart TD
    subgraph PI["1. Accountable PI & Research Contract"]
        Q["Research Question"] --> TS["Topic Survey Gate<br/>(Nearest literature · Contribution)"]
        TS --> C["Research Contract<br/>(Target outlet · Scope · Budget · Stop condition)"]
        C --> S["Shared Protocol<br/>(Definitions · Assumptions · Obligations)"]
    end

    subgraph Modes["2. Tri-Pillar Methodology Router"]
        S --> M{Research Mode}
        M -->|Empirical| E["Empirical / Causal<br/>(Estimand · Assignment mechanism)"]
        M -->|Pure Theory| T["Formal Theory<br/>(Lean 4 proof · Counterexample search)"]
        M -->|Structural| X["Structural Estimation<br/>(Moments · Solver tolerance · Holdout)"]
    end

    subgraph Harness["3. Execution & Falsification Battery"]
        E --> MR["Method Router<br/>(DiD · RDD · IV · SDiD · DML)"]
        MR --> DP["Data Provenance Gate<br/>(SHA-256 hashes · Source · License)"]
        X --> DP
        DP --> B["Reproducible Baseline<br/>(Deterministic harness)"]
        T --> B
        B --> FB["Falsification Battery<br/>(3+ Competing channels · Placebos · Oster bounds)"]
        FB --> K{Validate & Classify}
        K -->|Keep · Discard · Blocked| L[("Research Ledger & provenance graph<br/>(experiments.tsv & manifest.jsonl)")]
    end

    subgraph Delivery["4. Delivery & Review Gates"]
        L -->|Budget remains| B
        L -->|Stop reached| MC["Manuscript Coverage Matrix<br/>(Hash-bound section packets)"]
        MC --> AR["Adversarial Peer Review<br/>(Ref 1: Identification · Ref 2: Theory · Ref 3: Data)"]
        AR --> RA["AEA Replication Package Audit<br/>(Master script · Relative paths · Seed lock)"]
        RA --> EP["Verified Evidence Package<br/>(LaTeX manuscript · Figures · Tables · Code)"]
    end

    style PI fill:#1a237e,stroke:#3949ab,stroke-width:2px,color:#ffffff
    style Modes fill:#004d40,stroke:#00796b,stroke-width:2px,color:#ffffff
    style Harness fill:#263238,stroke:#455a64,stroke-width:2px,color:#ffffff
    style Delivery fill:#311b92,stroke:#512da8,stroke-width:2px,color:#ffffff

⚡ Capabilities

• Topic Survey Gate: Maps the closest literature and establishes novelty before choosing an empirical method or claiming contribution.
• Tri-Pillar Economics Coverage: Integrates reduced-form causal inference, formal microeconomic theory (with Lean 4 kernel verification), and dynamic structural estimation in a unified DAG.
• Database Cleaning Recipes: Eliminates institutional data traps in standard empirical databases:
  • WRDS (CRSP & Compustat): CCM linkage (linktype in ('LU', 'LC')), negative prices (bid-ask midpoint), Fama-French June fiscal lags, and Davis-Fama-French Book Equity.
  • Taiwan TEJ: Price dividend adjustments, quarterly flow variable cumulative-to-single-quarter decomposition, and industry exclusions.
  • China CSMAR: Board filtering (ChiNext, STAR), ST/*ST and financial firm screening, reinvested dividend returns (Dretwd), and quarterly flow adjustments.
  • Micro / Survey (CFPS & US Census/IPUMS): Immutable pid vs. dynamic fid, sampling weights (pweight/PERWT), 5-digit zero-padded County FIPS (09001), and boundary harmonization crosswalks.
  • Macro (FRED & PWT 10.x): Flow/stock aggregation rules and living standards (rgdpe) vs. productive capacity/TFP (rgdpo) selection.
• Modern Econometric Recipes: Pre-configured code patterns for Stata, R, and Python covering Staggered DiD (csdid, sunab, did_imputation), bias-corrected RDD (rdrobust), Montiel Olea & Pflueger effective weak-IV tests, Conley spatial HAC, and Double Machine Learning (DoubleML).
• Falsification & Competing Mechanisms Battery: Rejects confirmation bias by formalizing 3+ alternative confounding stories, temporal placebos, donor swaps, and Oster selection bounds.
• AEA-Grade Replication Audit: Enforces American Economic Association Data and Code Availability policies: zero absolute paths, pinned seeds, immutable raw data checksums, and single-click master execution.
• Pre-Submission Adversarial Peer Review: Simulates three classic referee archetypes (Identification Policeman, Theory/Mechanism Critic, and Data Hygiene Nitpicker) to stresstest manuscripts before submission.

📚 Protocol Index

🌌 Dream-RSI: Recursive Self-Improvement through Evolving Worlds

Based on Dream-RSI: Recursive Self-Improvement through Evolving Worlds (Google DeepMind / UMD / UVA 2026; dream-rsi.com), the lab converts accumulated research history into an exact replay simulator:

• History as an Exact Simulator: Every econometric specification trial, first-stage diagnostic, and proof branch is persisted into research/discovery_tree.jsonl.
• Zero-Compute Offline Dreaming: Candidate exploration and pruning policies are evaluated by replaying historical discovery trees at zero execution cost, screening thousands of specification curve configurations in milliseconds.
• Evolving World Pool: Each completed research run adds another world to the simulator pool, transferring evolved meta-policies and adversarial referee memories to accelerate future research projects.

Run the Dream-RSI offline simulator:

# Audit historical discovery tree integrity
python3 scripts/dream_replay_simulator.py --tree research/discovery_tree.jsonl --audit

# Dream offline specification curve across history at zero execution cost
python3 scripts/dream_replay_simulator.py --tree research/discovery_tree.jsonl --dream

📦 Install

The same release is packaged for Codex, Claude Code, and Google Antigravity. The plugin manifests are thin adapters around one shared econ-research-lab skill.

Codex

Codex plugins can be installed from this repository's marketplace:

codex plugin marketplace add Tohskcid/invisible-hands-for-economists --ref main
codex plugin add invisible-hands-for-economists@invisible-hands

# Refresh the marketplace and installed plugin later
codex plugin marketplace upgrade invisible-hands
codex plugin add invisible-hands-for-economists@invisible-hands

Claude Code

Claude Code plugins use the compatible marketplace in the same repository:

claude plugin marketplace add Tohskcid/invisible-hands-for-economists
claude plugin install invisible-hands-for-economists@invisible-hands

# Refresh the marketplace and installed plugin later
claude plugin marketplace update invisible-hands
claude plugin update invisible-hands-for-economists@invisible-hands

Google Antigravity

Antigravity CLI can import the shared Agent Skill directly from GitHub:

agy plugin install https://github.com/Tohskcid/invisible-hands-for-economists
agy plugin list

Direct git clone skill installs remain supported for existing users. Pull with git pull --ff-only to update those installations.

Direct skill installation paths

# Codex global skill
git clone https://github.com/Tohskcid/invisible-hands-for-economists.git \
  "$HOME/.agents/skills/econ-research-lab"

# Claude Code global skill
git clone https://github.com/Tohskcid/invisible-hands-for-economists.git \
  "$HOME/.claude/skills/econ-research-lab"

# Antigravity global skill
git clone https://github.com/Tohskcid/invisible-hands-for-economists.git \
  "$HOME/.gemini/config/skills/econ-research-lab"

🛠 CLI Tools

🛡 Verification & Standards

# Validate library cards and journal database
python3 scripts/validate_library.py --json

# Run package unit tests
python3 -m unittest discover -s tests -v

# Validate agent skills specification
uvx --from skills-ref agentskills validate "$(pwd)"

MIT Licensed. Maintained by Tohskcid.