Skip to content

pAI-Econ-claude

external · status: active · focus: end-to-end · discipline: economics · started: 2026

Project page: https://github.com/maxwell2732/pAI-Econ-claude

Licence: unclassified (licence file present; see repository)

Source: projects/landscape/pai-econ-claude.yml

Positioning

A gated, human-in-the-loop multi-agent Claude Code Skill for AI-assisted economic THEORY development: a 10-stage sequential pipeline (research intake → puzzle refinement → literature → canonical-model matching → model primitives → assumption audit → propositions → proof sketches → counterexample search → economic interpretation → manuscript generation) with an optional numerical- simulation stage, 8 quality gates, and 6+ mandatory human decision points. Sits alongside theorist-toolbox (morankor) in the RISE formal-modeling layer, as an independent, differently-architected attempt at the same "trustable AI help on economic theory" problem.

Distinctive contribution

A 40+ template "model_library" of canonical economic-theory models (micro/labor/trade/IO) used for novelty-risk and canonical-fit gating — a structured prior that theorist-toolbox does not have — plus a companion arXiv paper (2607.21268) reporting a controlled before/after evaluation on 5 matched theory tasks against an ungated baseline: mean failure severity fell 1.58→1.16 and reported usefulness rose 2.60→3.10, the closest thing to a quantitative ablation in the catalog's economic-theory-agent niche.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Five declared stages (RQ/puzzle refinement, literature, formal modeling, optional code generation, manuscript drafting) with gaps — no dedicated literature-synthesis or referee-simulation stage as such (review lives inside internal gates, not as a pipeline stage).
Autonomy level 1 6+ mandatory human decision points across 8 gated stages — closer to copilot (human approves each significant step) than a supervised-and-review-only agent.
Architectural transparency 3 MIT license; 20+ stage-specific prompt files, SKILL.md routing logic, the model_library templates, and gate logic all published in the repo.
Inputs supported 1 Primary input is a free-text research idea/puzzle; an 'Empirical Companion' mode exists for applied-research inputs but no external literature-corpus or data-source connector is documented.
Outputs / reproducibility 2 Each stage persists a versioned intermediate record (propositions, proof sketches, optional simulation code/figures, manuscript files) gated by human checkpoints; no demonstrated end-to-end reproducibility across arbitrary runs.
Internal evaluation 2 Companion arXiv preprint (2607.21268) evaluates the gated pipeline against an ungated baseline on 5 matched economic-theory tasks with severity/usefulness scoring; not yet externally peer-reviewed or replicated by a third party.
Openness 2 MIT license, example inputs in examples/ for quickstart, but running the full pipeline requires a paid Claude Code/API subscription — not free-tier reproducible.
Maturity / traction 2 185 stars / 60 forks, tagged v1.4.0, 40 commits, last updated 2026-08-21 — beta-stage with growing external adoption but single-team origin.
Cross-family policy 0 Single-model by design — built specifically as a Claude Code Skill (name and commands are Claude-specific); no cross-family review mechanism.
Runtime assurance 3 8 in-pipeline gates (novelty risk, canonical-model fit, model coherence, proof integrity, economic meaning, mathematical review) that diagnose failure modes and trigger loopbacks before a stage can advance — a full runtime audit stack with gating on failure.
Cross-platform portability 0 Locked to Claude Code as a Claude Code Skill; no documented support for other IDEs, runtimes, or model providers.

Scored on 2026-10-01. See the evaluation rubric.

Tags

Pipeline stages: rq-formulation literature-discovery formal-modeling code-generation paper-drafting

Architectural features: multi-agent human-in-loop dag-orchestration iterative-loop

Inputs: research-idea user-dataset

Outputs: propositions proofs paper-draft figures

Data sources: user-provided

Knowledge sources: model-library-templates

Limitations

  • Locked to a single platform (Claude Code) and a single model family, unlike theorist-toolbox's Codex-adversarial-pair option.
  • Evaluation is a preprint self-reported on 5 matched tasks by the same team that built the system — no external replication yet.
  • Requires a paid Claude Code/API subscription to run; not reproducible on free infrastructure.

Papers describing this project

  • pAI-Econ-claude: A Gated Human-in-the-Loop Multi-Agent Architecture for AI-Assisted Economic Theory Development — Zhu, C., Wang, X., Zhang, W. (2026). arXiv. arXiv:2607.21268