Invisible Hands for Economists

An auditable economics-research skill that turns a research question into an evidence-backed empirical, theoretical, or structural manuscript. It uses tool-acquired literature and data, a provenance graph, design-specific validity checks, competing-mechanism tests, reproducible result bindings, and adversarial review. A failed hard gate returns the research process to the relevant stage instead of allowing unsupported prose.

Research flow

flowchart TD
    subgraph Scope["1. Research Question and Contribution"]
        Q["One-sentence research question"] --> BL["Baseline literature search"]
        Q --> FL["Frontier search<br/>new and low-citation work"]
        BL --> CP1{"Researcher checkpoint<br/>scope and contribution"}
        FL --> CP1
        CP1 --> RC["Research contract<br/>outlet, estimand, budget, stop rule"]
    end

    subgraph Design["2. Method and Evidence Obligations"]
        RC --> MODE{"Research mode"}
        MODE -->|Empirical| E["Identification design<br/>data and inference"]
        MODE -->|Theory| T["Definitions, assumptions<br/>proof obligations"]
        MODE -->|Structural| S["Moments, solver<br/>identification and holdout"]
        E --> CP2{"Researcher checkpoint<br/>design and data lock"}
        T --> CP2
        S --> CP2
    end

    subgraph Iteration["3. Bounded Research Iteration"]
        CP2 --> ACQ["Tool-acquired evidence<br/>sources, licenses, hashes"]
        ACQ --> RUN["Reproducible analysis<br/>tables, figures, proofs"]
        RUN --> MC["Mechanism challenge<br/>alternatives and falsification"]
        MC --> LEDGER[("Append-only research ledger<br/>provenance and result bindings")]
        LEDGER --> G{"Deterministic audit gates"}
        G -->|fail| FAIL["Classify failure stage"]
        FAIL --> DREAM["Dream-RSI policy<br/>recommend next bounded action"]
        DREAM --> ACQ
        FAIL -->|material scope change| CP2
    end

    subgraph Delivery["4. Delivery and Adversarial Review"]
        G -->|pass| COV["Manuscript coverage<br/>claim-evidence graph"]
        COV --> AR["Adversarial review<br/>identification, mechanism, data"]
        AR --> RA["Replication and layout audit<br/>master script and rendered PDF"]
        RA -->|fail| FAIL
        RA -->|pass| M["Verified manuscript<br/>and replication package"]
    end

Core guarantees:

• New and low-citation work remains eligible through separate frontier-search lanes.
• Model memory is not accepted as literature or data evidence.
• Quantitative claims require hash-bound data provenance and generated results.
• Research outputs must be separated into purpose-named folders for manuscripts, references, PDFs, figures, tables, code, data, models, and audit artifacts.
• Identification and economic mechanisms are audited separately.
• Formal claims distinguish paper proofs, proof sketches, and Lean-verified theorems.
• Manuscript delivery requires complete section coverage, central-claim review, and rendered LaTeX inspection.

Dream-RSI iteration

The optional Dream-RSI loop learns which action kinds most effectively resolve each failed research stage. It uses append-only discovery histories and deterministic package-gate progress as feedback. Coefficients, p-values, effect direction, and agreement with researcher intuition never enter the policy reward.

# Learn/update a stage-conditioned policy from prior trajectories
python3 scripts/dream_replay_simulator.py \
  --tree research/discovery_tree.jsonl \
  --policy-out research/dream-policy.json

# Run a bounded online loop through an explicit action adapter
python3 scripts/dream_replay_simulator.py --online \
  --tree research/discovery_tree.jsonl \
  --policy research/dream-policy.json \
  --actions research/dream-actions.json \
  --adapter scripts/research_action_adapter.py \
  --root . --config research/package.json \
  --max-steps 10 --max-cost 50

See Dream-RSI research iteration for the action and adapter contracts. Online execution stops on package completion, budget exhaustion, a block, or a researcher checkpoint.

Skill, MCP, and one-shot execution

The Agent Skill directs research from the initial question through sources, analysis, manuscript, review, and release. The execution guide defines the frozen contract and repeated next_tasks loop. run_research_workflow.py audits supplied artifacts; run_research_team.py launches separate Codex CLI research agents with a bounded budget and hard literature, evidence, and review gates. An original-publication goal cannot proceed to design on a signed replication/duplicate decision; the runner records a bounded contribution branch and retries the literature/prior-art stage.

python3 scripts/run_research_team.py --root /absolute/path/to/research-project \
  --contract research/contract-input.json --max-rounds 2 --json

New autonomous manuscripts use package protocol_version: "2", including a retained full-text prior-art review and a hash-bound exploration ledger. A passing file check does not certify scholarly originality or journal acceptance.

For clients that support MCP, install the optional mcp dependency (Python 3.11+) and launch a local stdio server bound to one project:

uv run --no-project --with 'mcp>=2,<3' python scripts/econ_mcp_server.py --root /absolute/path/to/research-project

The server exposes project_status, audit_package, run_workflow, and run_iteration, plus the econ://workflow resource and research_question prompt. run_iteration executes only declared research/dream-actions.json commands. A stopped action retains an in-flight receipt and is not blindly rerun. Configure the MCP host with the absolute path to scripts/econ_mcp_server.py and a project-specific --root.

Full manuscript profiles require a PI-accepted multi-agent handoff report, an independent substantive review, and a limitations section. Empirical, descriptive, structural, and replication research rejects proxy datasets whenever data is declared. Simulation is reserved for theory validation or isolated engineering tests. A failed workflow returns rework_plan with the earliest stage to revisit and the downstream stages to regenerate. The CLI runner actually dispatches agents and stores receipts; static package checks cannot establish the truth of self-reported research judgments.

Install

Codex

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

Claude Code

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

Google Antigravity

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

Direct skill installations remain supported:

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

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

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

Core tools

Verify

python3 -m unittest discover -s tests -q
python3 scripts/validate_library.py --json
uvx --from skills-ref agentskills validate skills/econ-research-lab
uv run --no-project --with 'mcp>=2,<3' python -m unittest tests.test_mcp_server

Research roadmap

See the harness optimization research plan (Traditional Chinese) for the proposed implementation work packages, controlled evaluation, milestones, and acceptance criteria grounded in the reconstruction pilot.

The workflow implementation status records engineering tests, the real-data pilot, failed research gates, and remaining work toward one-shot paper completion.

MIT Licensed.