An AI system for economic theorists, from idea discovery to theory paper development.

This is an English-first research workflow system with multilingual command understanding. You can talk to the assistant in Chinese or English, but research artifacts, referee reports, theorem notes, and manuscripts are English by default.

Originally developed with Econometrica-level theory standards in mind, this unofficial local workflow system helps researchers move from rough ideas to primitive hunting, example-to-theory model-base construction, model tournaments, theorem candidates, absorption tests, target journal calibration, human gates, simulated review, and controlled revision paths.

This project is not affiliated with Econometrica. The phrase Econometrica-level is used only as shorthand for high standards of theoretical clarity, novelty, and rigor.

The workflow is target-calibrated rather than Econometrica-only. A confirmed target_journal_profile.md can calibrate the reader path, referee mix, exposition style, and fit standard for RAND, JET, Theoretical Economics, GEB, ReStud, AER, or another venue. Target journal changes calibration, not quality.

If this workflow helps your research, please consider giving the repository a star so more researchers can find and improve their own research process.

active_context.md is only a dashboard, not a source of truth. Important claims must still be checked against the underlying workflow artifacts.

The router uses durable project artifacts such as human_decisions.md, field_profile.md, target_journal_profile.md, literature_evidence_ledger.md, and generality_ledger.md. Major research decisions stop at explicit human gates and are recorded before the workflow treats them as confirmed.

• Treats ordinary explanatory, translation, GitHub, software, or conceptual questions as ordinary Q&A.
• Runs research execution seriously by default once the user asks the system to act on an idea, model, theorem, literature, manuscript, review, revision, or project state.
• Uses simple user commands while keeping explicit human gates for major decisions.
• Searches for research directions without immediately forcing mainstream taste.
• Preserves 1-2 non-mainstream but internally coherent directions during discovery.
• Records closest-paper, anchor-paper, absorption-threat, and style-anchor evidence in literature_evidence_ledger.md.
• Runs Primitive Hunter / Theorem Generator panels when the primitive is unclear.
• Treats user-supplied model details as provisional modeling constraints until a model-base gate confirms them.
• Searches broadly over cheap model skeletons, then derives narrowly from the few model bases that survive.
• Requires a hand-solved micro-example before any general model, fixed point, or existence machinery enters.
• Builds models from small examples and economic tensions before formal theorem production.
• Shows economic progress, not only workflow progress, through economic_logic_map.md and chat summaries that explain what economic question became clearer.
• Compares model variants before manuscript writing and uses nonconvex branch generation inside the existing tree search.
• Requires absorption tests against closest literature and known theory families.
• Uses field_profile.md to assign field-sensitive simulated referees.
• Uses target_journal_profile.md to recommend and confirm a primary, stretch, and fallback target without lowering the quality floor.
• Records human gate decisions in human_decisions.md instead of relying on chat memory.
• Tracks assumption and generality drift in generality_ledger.md.
• Uses Scientific Judge / Nugget Test safeguards against defensive complexity.
• Runs target-calibrated simulated review with dynamic referee roles.
• Routes local-optimum traps back to discovery before manuscript polishing.
• Creates provisional Working Preview Notes when the human wants a short readable snapshot before gates are final.
• Runs an anchor-based Manuscript Architecture Compliance Gate before full manuscript drafting, major restructuring, or full style pass, so every section has a job and the section order matches comparable published papers.
• Calibrates exposition style through a Deep Style Anchor Pass after the contribution is locked.
• Uses full-text style anchors when legally available or user-provided, extracting exposition architecture rather than prose.
• Supports Python, Mathematica, Lean, LaTeX, and git-based verification workflows.

Only remember these commands:

Use the system: initialize this paper project.
Use the system: continue by the system.
Use the system: quickly screen this idea.
Use the system: run a full literature audit.
Use the system: what should I do today?

The system will ask explicit human-gate questions when important decisions are needed; you do not need to know the internal stage names.

For a fully automatic long run, use the guarded form:

Use the system: run guarded full-auto goal mode from idea discovery to a polished target-journal draft.

In this mode the system can keep working through the pipeline, but major decisions are saved as AI-delegated provisional rather than human-confirmed. The run ends with final_ratification_report.md, where you approve, edit, or reject the provisional choices before treating the paper as confirmed.

1. Download the repository as a ZIP file or clone it.
2. Copy these files into the root directory of your paper project:

AGENTS.md
ECONOMETRICA_ORCHESTRATOR.md
ECONOMETRICA_PANEL_PROTOCOL.md
ECONOMETRICA_DISCOVERY_WORKFLOW.md
ECONOMETRICA_VERIFICATION_WORKFLOW.md
ECONOMETRICA_AI_HUMAN_WORKFLOW.md
ECONOMETRICA_VERSION_CONTROL.md
FIRST_RUN.md
TOOLCHAIN_README.md
verify_toolchain.ps1
verification_templates/

3. Open the paper folder in Codex Desktop or another agent IDE. AGENTS.md should be read automatically, and ECONOMETRICA_ORCHESTRATOR.md acts as the router.

4. Run a first-run setup check:

Use the system: first-run setup check

Small Chinese command example:

按系统处理：初始化检测

5. Initialize the paper project:

Use the system: initialize this paper project

New users should skim the examples before running a real project:

examples/full_walkthrough_toy_project/
examples/model_base_mini_example/
examples/kill_or_pivot_mini_example/

The examples are toy walkthroughs, not publishable papers and not literature claims. They show how the workflow moves from a vague idea to artifacts, human gates, model-base choice, preview drafts, review, and kill-or-pivot decisions.

Use the system: first-run setup check.

Use the system: initialize this paper project.

Use the system: continue by the system.

Use the system: quickly screen this idea.

Use the system: run a full literature audit.

Use the system: I want to explore a new research topic.

Use the system: run a model tournament and absorption test before writing.

Use the system: find the minimal model base before formal derivation.

During D4/D4.5 model-base work, the system will automatically detect existing agent_runs/. If multiple model lanes exist, it should recommend a Judge Pass before the Minimal Model Base Gate. You can use the current model as judge, or ask the system to create agent_runs/_judge_prompt.md for another AI model or IDE opened in the same project folder.

Use the system: run Primitive Hunter and identify the deepest primitive.

Use the system: rigorously verify Proposition 1 with Python, Mathematica, or Lean if useful.

Use the system: recommend a target journal ladder for this project.

Use the system: run a full target-calibrated simulated review.

Use the system: what should I do today?

Use the system: where is this project stuck?

Use the system: if I only have two hours, what is the highest-value next action?

Use the system: revise with agentic tree search instead of a defensive patch.

Use the system: export a short working preview draft.

Small Chinese command examples:

按系统继续

快速看看这个想法

完整审查一遍

今天我该做什么

These files are created inside paper projects, not maintained as fixed files in this workflow repository:

Windows paths can be mangled by Markdown if backslashes are written as raw text. The workflow therefore treats path display as an output-safety rule:

• local paths and compiled PDF paths should be shown in backticks or fenced code blocks, not raw prose;
• file cards and Markdown file links should use the filename only as the visible title; the full absolute path belongs on a separate Full path: line;
• generated PDFs should be reported only after the exact path is verified with Test-Path -LiteralPath or Resolve-Path -LiteralPath;
• commands should quote paths with spaces, and path construction should use Join-Path, Resolve-Path -LiteralPath, pathlib, or an equivalent path API.

Example:

File card:
[main.pdf](<C:/Dropbox/Shufe/Research/Project/My Paper/output/main.pdf>)

Full path:
`C:/Dropbox/Shufe/Research/Project/My Paper/output/main.pdf`

Open command:
`Start-Process -FilePath "C:\Dropbox\Shufe\Research\Project\My Paper\output\main.pdf"`

The paper project is folder-portable. Copy or sync the whole paper project folder, open it on another computer, run the local toolchain check, and continue from the project artifacts.

Use this command after opening the copied folder:

Use the system: continue by the system.

Folder portability lets you continue elsewhere. Workflow artifacts tell the assistant what happened and what is confirmed. Computer-level setup still belongs to each computer: agent IDE login, Python, LaTeX, Lean, Mathematica, Git, and local tool paths should be checked locally.

Generated research files are saved locally in each paper project folder. They are not automatically uploaded to GitHub. The workflow works locally even if you never use Git.

Git is optional for running the workflow. For serious projects, Git checkpoints are recommended because they let you roll back file versions if an AI edit, manual edit, LaTeX rewrite, or model rewrite damages the project. Folder portability preserves project state; Git checkpoints preserve recoverable file versions.

Git commits are stored in the project folder's hidden .git/ directory. They stay local unless you choose to push them to GitHub or another remote.

If you use GitHub as a backup for a paper project, you may choose which generated research files to track. Durable research records that are often worth tracking include:

human_decisions.md
project_state.md
contribution_lock.md
field_profile.md
target_journal_profile.md
literature_evidence_ledger.md
model_tournament.md
model_base_design.md
micro_example_note.md
heuristic_derivation.md
manuscript_architecture_plan.md
absorption_tests.md
generality_ledger.md
risk_register.md
revision_tree.md
revision_log.md
version_log.md

Files that usually do not need to be tracked include:

active_context.md
economic_logic_map.md
toolchain_status.md
preview_drafts/
run_summary.md
auto_decisions.md
goal_run_plan.md
final_ratification_report.md
scratch_runs/
referee_reports/
verification/
formal/
lean/
large local caches

These runtime files are ignored in this workflow template so paper-specific process files do not accidentally enter the public workflow repository. In your own paper project, you may choose to track selected durable records if you want GitHub or another remote to back up the research process.

Keep Python, Lean/elan, Mathlib, and large package caches outside paper folders. On Windows the recommended shared tool root is:

C:\Tools\CodexVerification

You can choose another location by setting CODEX_VERIFICATION_HOME or by passing -ToolRoot to verify_toolchain.ps1.

Minimum setup check:

.\verify_toolchain.ps1 -WriteStatus

If the script cannot find Python, Lean, or Mathematica, the workflow still works as prompts and checklists, but mathematical verification is weaker until the tool root is configured. See TOOLCHAIN_README.md for details.

You can still use the research workflow. Missing tools only weaken mathematical verification, local compilation, or formal proof support. They do not block idea discovery, field profiling, target journal profiling, model tournaments, simulated review, style calibration, or revision planning.

The workflow files, project artifacts, referee reports, theorem notes, revision logs, and manuscripts are English by default.

Chinese commands are supported for convenience, but generated research files remain English unless the user explicitly requests a separate Chinese explanatory note outside the manuscript workflow.

Project artifacts are the source of truth. Cross-project memory, if used, is only an optional prior stored outside paper projects, typically under:

C:\Users\<user>\.econ-theorist-ai\

Possible files include researcher_profile.md, method_library.md, negative_knowledge.md, proof_technique_memory.md, project_postmortems/, and field_maps/. Literature evidence, proof verification, and current human gate decisions override researcher memory.

• Scientific taste is a filter, not the sole objective.
• Simple commands must not weaken human gates.
• Research execution is serious by default, while ordinary Q&A remains ordinary Q&A.
• Specificity is not stage advancement: formal model language from the user is provisional until artifacts and human gates confirm it.
• Exhaust broadly at the model-skeleton level; derive narrowly at the formal level.
• No general model before a hand-solved micro-example.
• Show economic progress, not only workflow progress.
• A model should have a visible economic lesson and scope before it becomes a formal baseline.
• Early review should develop the strongest simple version before it tries to kill an idea.
• A model is not ready because it is formal; it is ready when its smallest version explains the economic force.
• Nonconvex discovery expands branch generation inside tree search; it does not certify quality.
• Target journal changes calibration, not quality.
• Strong novelty, absorption, anchor, and style claims require evidence recorded in literature_evidence_ledger.md or must be marked provisional.
• Deep style calibration means deep style reading, not prose imitation.
• Style anchors are used to extract exposition architecture, not sentences, paragraph structures, or framing.
• Token economy must never override research quality.
• Main theorem discovery, proof verification, closest-literature checks, and simulated review require enough context and tools.
• Human gate decisions must be written to persistent artifacts.
• Simulated acceptance is a diagnostic benchmark, not a publication guarantee.
• If the paper is stuck in local polishing, return to primitive hunting and model discovery.
• The preferred path is theorem note first, manuscript second.

• Toy worked examples.
• Benchmark cases.
• Richer verification templates.
• Optional cross-platform toolchain notes.

This project currently uses an issue-only feedback model.

Please open a GitHub Issue if you find:

• unclear workflow routing
• internal contradictions
• missing safeguards
• confusing documentation
• toolchain setup problems
• examples where the system behaves poorly
• suggestions that may help researchers use the system better

Pull requests are not the preferred contribution path at this stage. The maintainer reviews issues, decides which suggestions to adopt, and integrates accepted changes directly to preserve consistency across the workflow system.

See CONTRIBUTING.md for the current contribution policy.

If this system helps your research, please cite the repository and consider starring it on GitHub so other researchers can find it.

GitHub should display a "Cite this repository" option from CITATION.cff.

Suggested plain-text citation:

viplee110. econ-theorist-ai: An AI workflow system for economic theory research.
GitHub repository, 2026. https://github.com/viplee110/econ-theorist-ai

Licensed under the Apache License 2.0. See LICENSE.