Replication Package (rt-replication-package)
The step that now blocks acceptance at the top (AEA / JF / Management Science Data Editors
re-run your code and check every number). Full manifest + validation checklist:
shared-resources/empirical-methods/replication-package.md.
When to trigger
- Approaching final submission, or a conditional accept with a data-and-code requirement.
- A Data Editor / verification service needs the package.
What it does
-
Read the venue policy from the target pack's
resources/official-source-map.md
(code at submission vs. acceptance? Data Editor? AsCollected page? restricted-data rule?
hosting?). Match it exactly.
-
Assemble the manifest: master script (raw → all exhibits), ordered steps
(clean → construct → estimate → robustness → tables), pinned environment, README/roadmap,
data (or pseudo/synthetic data for restricted access + access instructions), the
script→exhibit output map, and the disclosure statement.
-
Validate with the checklist: runs clean end-to-end on a fresh machine; every reported
number reproduced (body + appendix); seeded/deterministic; restricted-data plan; output
map complete; venue policy satisfied.
Hard rules
-
Every reported number traces to a script step (body and appendix).
-
Deterministic — pin versions, seed RNGs.
-
Restricted data → pseudo/synthetic data + access instructions, never confidential data committed.
-
Match the venue's exact policy from the source-map; re-verify before submission.
-
Keep it in sync with the execution bridge — when a revision re-runs an estimate,
regenerate the affected outputs.
Output format
【Venue policy】(code timing, Data Editor, restricted-data rule, hosting)
【Manifest】master · steps · env · README · data · output-map · disclosure — status each
【Validation】end-to-end ✓/✗ · all numbers reproduced ✓/✗ · seeded ✓/✗ · restricted-data plan ✓/✗
【Blocking gaps】what would fail a Data-Editor check, ranked
【Ready?】GO / NOT-YET — deciding items
Anti-patterns
- "Works on my machine" (absolute paths, unpinned versions, manual steps); reported numbers
the script doesn't reproduce; committing confidential data; leaving assembly to post-accept crunch.
