replicate-paper

name	replicate-paper
description	Run or debug the main paper and Internet Appendix replication pipeline. Use when Claude needs to choose the right script boundary, resume after a failure, or map a run request to the implemented main or IA output surface.
argument-hint	[--ndraws=50000] [--quick] [--skip-estimation] [--step=N]

Replicate Paper

Read these sources in order:

1. AGENTS.md
2. docs/agent-context/replication-pipeline.md
3. docs/manifests/data-files.csv
4. docs/manifests/exhibits.csv
5. docs/manifests/manuscript_exhibits.csv
6. docs/validation/validated_runs.csv
7. docs/agent-context/prompt-recipes.md
8. docs/agent-context/paper-results-ia.md

Use When

• the task is to run the main replication pipeline
• the task is to resume after a failed run
• the user wants the smallest script boundary for a table, figure, or IA subset
• the user wants to know what output a given runner script should produce

Do Not Use When

• the task is environment setup or package installation only
• the task is explanation of the paper without running code
• the task is a code change unrelated to pipeline execution

Inputs

• repo root access
• readiness information from tools/doctor.R when available
• user scope such as full run, quick run, specific step, or IA subset

Outputs

• exact script boundary to run
• status of completed versus failed stages
• output folders or files produced so far
• precise rerun boundary after a failure

Workflow

1. If readiness is uncertain, run tools/doctor.R --check-only first.
2. If packages are missing, use tools/bootstrap_packages.*; if bundle-managed required data are missing, use tools/bootstrap_data.* before launching replication.
3. Use docs/manifests/exhibits.csv to map user-facing exhibits to the smallest runnable boundary.
4. Use docs/manifests/manuscript_exhibits.csv when the user asks about the full paper inventory or about paper-only IA exhibits.
5. Prefer reduced-draw or help/list boundaries first when validating changes.
6. Stop at the first failing stage and report the exact rerun boundary instead of restarting everything.
7. Distinguish clearly between full main-paper coverage and the implemented IA subset.
8. For an IA replication prompt, use tools/run_ia_smoke.* or ia/_run_ia_full.R --ndraws=500 as the first acceptance boundary before scaling up.
9. Use docs/validation/validated_runs.csv when you need to distinguish a maintainer-validated boundary from a merely documented command.

Example Prompts

• "Run the paper in quick mode and stop at the first failing step."
• "Replicate the main text. If packages or data are missing, bootstrap them automatically first."
• "Replicate the Internet Appendix. Bootstrap what is needed, run the 500-draw IA smoke boundary first, and stop at the first failing model."
• "Which script generates Table 5?"
• "Resume the IA pipeline from the first missing output."

Failure Boundaries

• do not claim a replication completed unless the relevant scripts actually finished
• do not describe every manuscript IA exhibit as implemented when the repo only generates a subset
• do not jump straight to the full pipeline when a narrower boundary will answer the question
• do not treat the weighted treasury IA model or ia/data/w_all.rds as optional