A library of structured quality checks for academic research papers — analogous to unit tests in software engineering.
A research unit test specifies a single quality criterion, how to check it, and what constitutes a pass. Tests range from deterministic (does the replication package run?) to judgment-based (is the contribution interesting?). They are organized by methodology and severity.
Tests are declarative: each is a Markdown file an agent can read and follow. No code execution required.
| Dimension | Values |
|---|---|
| Methodology | universal, did, rdd, iv, ols, synth, experiment_lab, experiment_field, theory, ml_prediction, survey |
| Scope | paper, proposal, replication |
| Severity | blocker (paper cannot pass), warning (flag for author), info (suggestion) |
| Clarity | deterministic (clear yes/no), heuristic (objective anchors), judgment (requires experience) |
| ID | Name | Method | Severity | Clarity |
|---|---|---|---|---|
universal-tables-have-n-obs |
Regression tables report N | universal | blocker | deterministic |
universal-replication-reproduces-results |
Replication package reproduces results | universal | blocker | deterministic |
universal-contribution-is-new |
Contribution is new | universal | blocker | judgment |
universal-contribution-is-interesting |
Contribution is interesting | universal | blocker | judgment |
did-parallel-trends-plot |
Pre-trends visualization shown | did | blocker | heuristic |
did-staggered-heterogeneous-effects |
Staggered adoption uses robust estimator | did | blocker | heuristic |
iv-first-stage-f-stat |
First-stage F-statistic sufficient | iv | blocker | deterministic |
rdd-bandwidth-sensitivity |
Estimates robust to bandwidth choice | rdd | blocker | heuristic |
Add this repository's .claude/ directory to your project, or reference it in your CLAUDE.md. Then:
Review this paper using research unit tests: [paper.pdf or description]
Methodology: did
Scope: paper
The agent will load applicable tests from the registry, run each check, and produce a structured pass/fail report.
See CONTRIBUTING.md. New tests go in community/{your-username}/. Propose to core via PR.
MIT. Tests in community/ are owned by their respective authors under the same license unless otherwise stated.