Econ Auto Research Pipeline

1. Inspect the project root.

   • Run pwd, rg --files, and find . -maxdepth 2 -type d.
   • Read existing README, PROJECT_LOG, RESEARCH_BRIEF, RESEARCH_DESIGN, DATA_DICTIONARY, REPLICATION_README, and AGENTS.md if present.
   • Report the current state before major edits.

2. Classify the entry point.

   • New idea: start with research question and literature scoping.
   • Existing data: audit provenance, structure, and sample construction.
   • Existing code: map pipeline entry points and logs.
   • Existing results: run empirical integrity and claim checks.
   • Existing draft: run manuscript integrity and peer review.
   • Reviewer comments: build a revision roadmap and response plan.

3. Load references as needed.

   • Research governance, topic tournament, gate decisions, or post-mortem: references/research-governance.md.
   • Project setup or messy files: references/reproducibility-standards.md.
   • Stata regressions or final tables: references/stata-first-pipeline.md.
   • Literature search, systematic review, citation chaining, source-density gates, or top-journal literature positioning: references/literature-review-protocol.md plus references/deep-research-method.md.
   • Paper writing or revision: references/manuscript-pipeline.md.
   • Chinese top-journal polishing: trigger chinese-econ-paper-polisher or load its style reference when drafting Chinese manuscripts.
   • Review, integrity, or claim validation: references/review-and-integrity.md.
   • External review services such as Stanford Agentic Reviewer / PaperReview.ai: references/external-review-systems.md.
   • AI-assisted research operating rules from Mihail Velikov's AI in Business & Economic Research wiki: references/ai-econ-wiki-integration.md.
   • Overnight automation: references/automation-runbook.md.
   • Artifact names and handoff: references/artifact-contract.md.

Integrated Workflow

Stage -1: Research Governance Contract

Create or update process artifacts before substantive execution:

• memos/research_tournament.md
• memos/decision_log.md
• memos/literature_review_protocol.md
• memos/ai_workflow_audit.md
• memos/review_report.md
• memos/revision_roadmap.md

Gate rule: if the project lacks these artifacts, initialize them before drafting a paper or running new formal regressions.

Apply the AI-econ-wiki operating rules:

• Treat the repo as context: important state must be in files, not chat.
• Use Prompt -> Plan -> Review -> Revise for nontrivial work.
• Use fresh-context review for design, method, data, and manuscript gates where possible.
• Record AI use, privacy/cost decisions, and skipped gates in memos/ai_workflow_audit.md.

Stage 0: Project Intake

Create missing project scaffolding conservatively. Do not duplicate an existing structure.

Deliver:

• Project map.
• Missing artifact list.
• Immediate risks.
• Next safe action.

Stage 1: Topic Tournament And Research Question

For a new or weakly specified project, compare at least three candidate research questions unless the user explicitly forbids alternatives.

Score each candidate on:

• substantive importance;
• theoretical contribution;
• literature gap credibility;
• data granularity and availability;
• identification strength;
• top-journal fit;
• reproducibility burden;
• risk of being a "regression-shaped essay."

Deliver memos/research_tournament.md and record the decision in memos/decision_log.md.

Gate verdict:

• GO: advance to design.
• REVISE: refine question before data work.
• PIVOT: choose another candidate.
• KILL: stop this direction.

Stage 2: Research Design Gate

Turn the topic into a precise economics or management question.

Require:

• Unit of analysis.
• Treatment/exposure/key explanatory variable.
• Outcome.
• Mechanism.
• Identification boundary: causal, quasi-causal, descriptive, predictive, or mechanism evidence.
• Feasible data path.

Deliver memos/research_design_brief.md.

Gate rule: do not collect data or run regressions until the design memo states what evidence would falsify or severely weaken the project.

Stage 3: Literature Review And Positioning Gate

Build a source matrix, not a pile of citations.

Deliver:

• memos/literature_review_protocol.md.
• memos/literature_matrix.md.
• Search strategy and inclusion/exclusion rules.
• Screening counts, rejected-source reasons, and saturation status.
• Competing explanations and gap map.
• Dangerous prior art list.

For Chinese top-journal writing, literature narration must mostly use natural inserted citations and problem-oriented synthesis, not a mechanical "Author (Year) found..." pile.

Gate rule: if literature density, journal quality, Chinese/English balance, or dangerous prior art coverage is insufficient, return to literature search before empirical execution.

Stage 4: Data And Design-Fit Gate

Make the project reproducible before chasing results.

Require:

• Raw data immutable.
• Derived data scripted.
• Metadata, codebook, sample rules, row counts, and checksums where useful.
• Every final dataset mapped to input files and generating scripts.
• Explicit judgment on whether data granularity matches the research question.

Deliver data/metadata/data_provenance.md or update equivalent.

Gate rule: if available data are too coarse for the claimed contribution, downgrade the claim, redesign the question, or pivot.

Stage 5: Empirical Plan Gate

Create a claim-driven empirical plan.

Deliver:

• memos/empirical_plan.md.
• Specification ladder.
• Inference plan.
• Robustness and placebo checklist.
• Mechanism and heterogeneity plan.
• Method-specific audit checklist: DiD / IV / RD / panel FE / text-as-data / prediction-ML as applicable.
• Pre-analysis lock: outcome hierarchy, timing, sample rules, controls, fixed effects, clustering, and falsification criteria.
• What would falsify the story.

Gate rule: do not write manuscript results before a methods-review pass has challenged identification, measurement, controls, fixed effects, clustering, and table format.

Stage 6: Implementation And Runs

Use existing code patterns.

Default data flow:

data/raw -> scripts/build or scripts/utilities -> data/final/*.dta -> do/*.do -> output/tables + output/figures + output/logs

Formal reported regressions should be Stata do-files unless the user says otherwise.

Deliver:

• Updated scripts/do-files.
• Logs under output/logs.
• Tables under output/tables.
• Figures under output/figures.
• Chronological update in PROJECT_LOG.md.

Stage 7: Empirical Integrity And Internal Review Gate

Do not write paper claims until this gate passes.

Check:

• Leakage.
• Sample filters.
• Metric or variable construction.
• Fixed effects and clustering.
• Baseline fairness.
• Robustness coverage.
• Whether causal language exceeds design.
• Whether top-journal table format is met: columns are regressions/specifications, rows are coefficients and standard errors, and controls/fixed effects/sample sizes/R-squared/clustering are reported.
• Whether AI-assisted data construction, code generation, or labeling has been audited and recorded in memos/ai_workflow_audit.md.
• Whether text-as-data or LLM-coded variables have human-audited labels, prompt/version logs, and leakage checks.

Deliver memos/empirical_status.md.

Also run a multi-perspective internal review before manuscript drafting. If verdict is PIVOT or KILL, stop or redesign.

Stage 8: Manuscript Pipeline

Write from evidence, not vibes.

Deliver:

• Paper outline and claim-evidence map.
• Draft sections in paper/.
• Citation audit.
• Figure/table integration notes.
• Limitations and scope conditions.
• For Chinese submissions, run chinese-econ-paper-polisher before review. Require terminology unification, de-AI rewriting, empirical narration correction, and top-journal risk notes.

Stage 9: Peer Review And Revision

Run a multi-perspective review:

• Editor-in-chief: contribution and journal fit.
• Methods reviewer: identification, statistics, reproducibility.
• Domain reviewer: literature and theory.
• Practical/policy reviewer: implications and external validity.
• Devil's advocate: strongest counterargument and overclaiming.

Deliver:

• memos/review_report.md.
• memos/revision_roadmap.md.
• memos/chinese_polish_report.md for Chinese manuscripts.
• Re-review after revisions.

Gate rule: a final manuscript requires at least one review report and one revision roadmap. If serious issues remain, label the output as a working draft, not submission-ready.

Stage 10: Optional External Review Triage

External AI review tools may be used only as supplemental signal after internal review. For PaperReview.ai / Stanford Agentic Reviewer, remember:

• it currently accepts PDF upload plus email and analyzes only the first 15 pages / 10MB;
• it warns AI reviews can contain errors;
• its own notes say arXiv grounding makes it more accurate in AI-like fields and less accurate elsewhere, and it supports English-language papers for now.

For Chinese economics/management manuscripts, do not treat it as a primary referee. Use it only after creating an English synopsis or translated draft and record limitations in memos/external_review_log.md.

Stage 11: Final Handoff

Update:

• README.md.
• PROJECT_LOG.md.
• MANIFEST.md.
• memos/handoff_brief.md.

End every substantial session with current question, data, scripts, latest results, limitations, next steps, changed files, and commands run.

Guardrails

• Never overwrite raw data.
• Never manually edit derived datasets.
• Never bury failed diagnostics.
• Never present exploratory correlations as causal effects.
• Never finalize claims without code, logs, evidence, and review status.
• Never skip topic/design/literature gates because a user asked for a polished paper quickly; label any rushed output honestly.
• Never write process/meta commentary inside the manuscript body (for example, "for Chinese top journals, this wording...").
• Stop before paid compute, remote credentials, destructive cleanup, or broad irreversible changes.