# Research Economic Frontiers

> Find, select, design, execute, and independently audit frontier economics research using verified literature gaps, claim-specific contracts, diverse theory and empirical routes, reproducible analysis, and conservative contribution gates. Use for open or underexplored economics questions, paper ideas, research-gap searches, theoretical model development, causal or predictive empirical designs, replication projects, pre-analysis planning, long-horizon multi-agent economics research, referee reports, and claims that a contribution is novel or publication-ready.

- Skill: `meleantonio/research-economic-frontiers` (Agent Skill, multi-file: 22 files)
- Install (CLI): `npx skillmds@latest add meleantonio/research-economic-frontiers`
- Raw SKILL.md: https://api.skillmd.com/api/skills/meleantonio/research-economic-frontiers/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: meleantonio (https://skillmd.com/u/meleantonio)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/meleantonio/research-economic-frontiers

---


# Research Economic Frontiers

Explore ambitiously and make claims conservatively. Economics contributions do
not share one proof standard, so classify the claim before selecting methods.

## Non-negotiable rules

- Verify central references from opened records. Never invent citations.
- Scope every novelty claim to sources, queries, and a search date.
- Freeze the question, claim, population, timing, estimand, and outcome before
  confirmatory analysis.
- Keep confirmatory, exploratory, and post-hoc work visibly separate.
- Never upgrade association to causation.
- Never fabricate data, results, access rights, consent, ethics approval, or
  disclosure.
- Do not access confidential, restricted, or personal data without explicit
  authorization and the required agreements.
- Record null results, deviations, failed routes, and unresolved threats.
- Require an independent audit for a candidate contribution and external human
  review for an externally validated contribution.

## 1. Define the run

Infer or obtain:

- field, topic, population, and decision context;
- whether to find a question or evaluate a supplied one;
- allowed claim lanes: `theory`, `causal`, `predictive`, `descriptive`,
  `policy`, or `mixed`;
- available data, model, compute, time, and agent capacity;
- research mode: `literature`, `replication`, or `blind-benchmark`;
- registration, ethics, privacy, disclosure, and data-access constraints; and
- desired output: ranked questions, design, analysis, replication package,
  paper, or referee report.

Use `literature` mode by default. Blind work cannot pass the novelty gate until
a post-hoc literature audit is complete.

## 2. Initialize persistent state

From this skill directory, run:

```bash
python3 scripts/init_workspace.py <project-directory> \
  --title "<short title>" \
  --claim-type <claim-type> \
  --mode literature \
  --budget "<resource budget>"
```

Do not overwrite an existing project unless the user explicitly requests it.
Maintain:

- `QUESTION.md`: frozen question, scope, outcomes, and decision relevance;
- `NOVELTY.md`: reproducible search and closest-work ledger;
- `CONTRIBUTION_CONTRACT.md`: claim lane, success test, and non-counting output;
- `ROUTES.md`: question, theory, data, and design route registry;
- `THEORY.md`: mechanism, assumptions, propositions, and predictions;
- `DATA.md`: provenance, access, legality, measurement, and sample construction;
- `IDENTIFICATION.md`: estimand, assumptions, threats, and falsification;
- `ANALYSIS_PLAN.md`: locked and exploratory analyses;
- `RESULTS.md`: complete results, deviations, nulls, and diagnostics;
- `REPRODUCIBILITY.md`: environment, master command, inputs, and outputs;
- `REFEREE_REPORT.md`: independent objections and verdict;
- `PAPER.md`: claim-aligned draft; and
- `STATE.json`: machine-readable gate status.

Before setting `reproducibility_status` to `passed`, create and verify a
content-addressed receipt:

```bash
python3 skills/research-economic-frontiers/scripts/research_receipt.py create <project-directory>
python3 skills/research-economic-frontiers/scripts/research_receipt.py verify <project-directory>
```

The receipt covers every workspace file except its own output and local cache
artifacts. It rejects symlinks and records only relative paths, sizes, SHA-256
hashes, and a deterministic tree hash. Regenerate it after any auditable file
changes; do not treat a stale receipt as reproducibility evidence.

Validate periodically:

```bash
python3 skills/research-economic-frontiers/scripts/validate_workspace.py <project-directory>
```

Use `--strict` before any candidate-contribution claim.
Validation requires a current `RESEARCH_RECEIPT.json` whenever
`reproducibility_status` is `passed`.

## 3. Search and select a question

Read [discovery-and-novelty.md](references/discovery-and-novelty.md).

Build a candidate pool before choosing. Search exact questions, mechanisms,
outcomes, data settings, methods, and adjacent-field terminology across:

- recent working papers and registered studies;
- publisher and journal records;
- economics indexes and series;
- conference programs and research networks; and
- citation chains around the closest verified papers.

Verify every central paper with two independent records when available,
including one durable identifier or authoritative record.

Score each candidate 1–5 on:

1. economic or policy importance;
2. verified novelty distance;
3. theory or design tractability;
4. data access and measurement quality;
5. identification or model discipline;
6. falsifiability and time-to-signal;
7. reproducibility; and
8. ethical, legal, and disclosure feasibility.

State the strongest rejection case for every candidate. Prefer a smaller clean
question over a broad question that the available evidence cannot answer.

## 4. Freeze the contribution contract

Classify the primary claim and read
[claim-gates.md](references/claim-gates.md). Write:

1. the exact question and primary claim;
2. the closest prior work and the incremental contribution;
3. what would count as support, contradiction, or an informative null;
4. outcomes that do not count, including weaker or different claims;
5. the model, data, estimand, design, or prediction target;
6. load-bearing assumptions and rival explanations;
7. confirmatory analyses and the lock date;
8. allowed exploratory work and labeling rules;
9. registration, ethics, privacy, and disclosure obligations;
10. reproducibility outputs; and
11. the independent-audit standard.

If the question changes materially after locking, create a versioned contract
and label the change. Do not silently rewrite success after seeing results.

## 5. Build diverse routes

Read [orchestration.md](references/orchestration.md) when multiple agents or
isolated contexts are available.

Create genuinely different route families across:

- mechanisms and theoretical primitives;
- datasets, populations, and measurement strategies;
- assignment or quasi-experimental sources;
- structural, reduced-form, experimental, descriptive, and predictive designs;
- alternative estimands and falsification tests; and
- replication or reanalysis routes.

Require a concrete artifact from each route: a verified gap, proposition,
counterexample, data audit, estimand, design diagram, power calculation,
simulation, falsification result, or executable specification.

Use route states:

- `active`
- `promising`
- `blocked`
- `refuted`
- `merged`
- `design-ready`
- `evidence-ready`

Block a route when it depends on inaccessible data, an assumption as strong as
the conclusion, an underpowered design, or a measurement strategy that cannot
support the claim. Reopen only for a materially new mechanism or source.

## 6. Lock, register, and execute

Before confirmatory analysis:

- complete the data and identification audits;
- declare exclusions, transformations, outcomes, specifications, inference,
  clustering, multiple-testing adjustments, and stopping rules;
- separate the untouched evaluation sample for predictive work;
- document power or minimum detectable effects when applicable;
- register field experiments as required and use an appropriate registration
  for other designs when useful; and
- record the lock or registration identifier in `STATE.json`.

Execute from raw inputs through a master command when legally possible. Preserve
immutable raw data, transformation logs, environment versions, seeds, and
machine-readable outputs. For restricted data, document the access path and
provide a synthetic or test fixture where allowed.

Run confirmatory analyses first. Label every deviation and all exploratory
work. Do not drop null or adverse results.

## 7. Apply the claim gate

Read [claim-gates.md](references/claim-gates.md) and downgrade the claim when a
gate fails:

- theory: prove the proposition and audit assumptions and boundary cases;
- causal: defend identification, diagnostics, falsification, uncertainty, and
  external-validity limits;
- predictive: pass locked out-of-sample evaluation against a declared
  benchmark with no leakage;
- descriptive: verify provenance and uncertainty and use no causal language;
- policy: add incidence, equilibrium, distributional, and normative
  sensitivity;
- mixed: pass every lane used in the headline claim.

An interesting estimate is not enough. Match the words in the conclusion to
what the gate supports.

## 8. Audit independently

Read [referee-checklist.md](references/referee-checklist.md). Give a fresh
reviewer only the frozen contract, source ledger, paper, results, code and
reproducibility record—not the authors' advocacy.

Require line-specific objections and one verdict:

- `REJECT`
- `MAJOR-REVISION`
- `MINOR-REVISION`
- `CANDIDATE`

Repair and re-audit every downstream claim affected by a change. A different
agent can support `CANDIDATE`; only external human review can support
`externally-validated`.

## 9. Package the outcome

For a candidate contribution, deliver:

- exact contribution and claim lane;
- verified novelty ledger;
- theory or identification dependencies;
- complete results including nulls and deviations;
- reproducible code, environment, and data instructions;
- independent referee report;
- limitations, external validity, ethics, and disclosures; and
- an explicit request for human expert review.

For an incomplete run, deliver:

- strongest supported result;
- exact unresolved gap;
- failed and blocked routes;
- reusable data, code, propositions, or diagnostics;
- what changed from the original contract; and
- the most informative next test.

Never force a publishable story from an uncooperative result.

## Research integrity reference

Read [sources-and-ethics.md](references/sources-and-ethics.md) before using
registries, third-party data, restricted data, human-subjects material, or
policy recommendations.