Invisible Hands for Economists
Invisible Hands for Economists documents an economics agent skill and artifact-audit harness for empirical, theoretical and structural research. It prescribes source acquisition, frozen research contracts, provenance/result bindings and substantive review, but the deterministic runner does not itself create scholarly evidence.
Workflow level n/a Source: Repository-documented Outcome: Not empirically evaluated here
Research applicability
Directly addresses AI-assisted economics research design, literature acquisition, analysis and manuscript preparation. The documented process can inform a researcher-supervised pilot, but neither deterministic gates nor synthetic fixtures establish the scientific validity or real-world applicability of an economic conclusion.
Getting started
Editorial, untested suggestion: read SKILL.md, the execution guide, implementation status and relevant mode/data-boundary references before considering installation. Start with a bounded public-source feasibility question, explicit budget, declared data rights and independent disciplinary assessment. Treat blocked artifacts as blocked; do not turn engineering fixtures into empirical findings or run repository commands without separate review and authorization.
Maintenance
Dated primary-documentation snapshot only; continued maintenance and release availability were not independently established. Mutable HEAD and the live website were not commit-pinned.
License
README/SKILL declare MIT and the retrieved LICENSE contains the MIT grant and 2026 Tohskcid copyright. Dataset, paper, host, MCP, analysis-library and Lean/Mathlib terms remain separate; no right to redistribute restricted research material is inferred.
Workflow & human oversight
No fixed automation level assigned. The skill prescribes an accountable lead-agent process and researcher scope/design checkpoints; no open-ended autonomy without an explicit budget. Agents may conduct authorized tasks, while scientific interpretation and final release require accurately identified substantive review.
Workflow level is not an evidence score; higher is not better evidence.
Approach
The skill routes empirical, theory and structural modes around a scoped contract, source/data acquisition, claim-evidence records, design and mechanism audits, result bindings, replication and manuscript review. Optional MCP exposes local audit/workflow tools; the host must dispatch researchers and generate substantive artifacts. Empirical guidance separates predictive accuracy and diagnostic tests from causal identification. Current instructions bar simulated inputs as empirical manuscript evidence.
Limitations
Documentation-only review; no repository script, MCP server, research workflow, analysis, proof checker or experiment was executed. The implementation-status document reports synthetic/illustrative engineering cases, with all three manuscript packages blocked; these do not establish real empirical research, valid causal identification, completed publication-quality manuscripts or a one-shot success rate. Runner checks declared artifacts and hashes, not the correctness of economic interpretation or the quality/independence of agent review. Native agent dispatch and scientific work remain host responsibilities. Working-copy replication is explicitly not an OS sandbox; absolute paths can escape it. The status document also records no cross-process locking and wall-clock rather than API-money budget accounting. Token-saving or policy-learning benefits are not verified. Restricted-data approval, provider behavior and dependency terms must be assessed per deployment; the protocol is not proof of privacy enforcement.
Reproducibility
Exact HTTPS response bytes, UTF-8 bodies, requested/resolved URLs, UTC access timestamps, headers, raw-body SHA-256 and envelope SHA-256 saved. No upstream software or claimed outcome reproduced.
Adoption requirements
Needs a compatible agent host able to perform actual source acquisition and native task dispatch. SKILL.md lists Python 3.10+ with pandas/NumPy for profiling and Lean 4/Mathlib for formal-theory verification; README specifies Python 3.11+ and an optional MCP dependency for that server. End-to-end manuscript work additionally depends on permitted data, analysis/LaTeX tooling and substantive review. Approve command/filesystem/network scope first: the documented replication copy is not an OS sandbox. Follow DUA/IRB and approved model-endpoint rules; no confidential rows, logs, embeddings or small cells may leave the authorized boundary merely because a tool is local.
Comments
No comments yet.
Add a comment
Comments are held for moderation before appearing. Critical or corrective comments are welcome — especially evidence that an entry overstates what has been verified.