Script Registry

Script Registry

Single source of truth: the registry: YAML block inside each Code/*.Rmd script.

Derived view: Code/REGISTRY.md, rebuilt by Code/_Claude Scripts/build_registry.R.

Scope

The four paper-producing sections cover only Code/*.Rmd (non-recursive). The builder additionally indexes the working .R scripts:

• Code/_Claude Scripts/*.R — working helpers (analysis drafts, diagnostics, utilities). Indexed in REGISTRY.md Section 5 (name, last-modified date, purpose from a # purpose: header) so past analysis is findable; NOT part of the numbered pipeline. Give each a # purpose: header line.

Out of scope (never scanned):

• Code/_Archive/ and Code/_Claude Scripts/_Archive/ — retired scripts.
• Code/_Claude Logs/ — session progress logs.
• Anything in Data/, Output/, Paper/.

YAML convention

Extend each Code/*.Rmd's front-matter with a top-level registry: block:

---
title: "Descriptive Title"
author: "Author Name"
output: html_document
registry:
  purpose: "Brief description of what this script does"
  inputs:
    - Data/Processed/input_file.parquet
  outputs:
    - Output/Figures/figure_name.pdf
    - Output/Tables/table_name.tex
  paper_target: "Main paper — Table 1, Figure 3"
  notes: "Any caveats, memory requirements, or known issues"
---

• script_id is optional — inferred from filename prefix (3. Claude_Foo.Rmd -> 3; 12b. Claude_Bar.Rmd -> 12b). Only set it explicitly if the filename has no numeric prefix.
• inputs and outputs are lists, one path per line. List every figure, table, CSV, and parquet the script writes or reads.
• paper_target is free text — include the table/figure number(s) in the paper.
• notes is for caveats (memory quirks, partial-status, known issues).

REGISTRY.md structure

The builder produces five sections:

• Per-script cards — one section per Code/*.Rmd, with Purpose, Paper target, In-paper flag, Last modified, Inputs, and Outputs split into three buckets: Figures, Tables, Processed Data.
• Processed Data — Dependency Graph — for each intermediate file: which script produces it, and which scripts consume it.
• Drift Warnings — mismatches between the YAML and the actual I/O calls detected in each script's code. Four categories: undeclared inputs, undeclared outputs, unused inputs, unused outputs. The builder NEVER auto-edits YAML — it flags for human review. (Detection classifies each quoted path by its nearest read/write call: reads include read_parquet/read.csv/readRDS/read_dta; writes include write_parquet/write.csv/saveRDS/ggsave/etable/writeLines/sink. Glob declarations like dir/*.csv, and any path referenced elsewhere in the code, are exempt from "unused" warnings.)
• Scripts Without Registry Metadata — any Code/*.Rmd that hasn't been given a registry: block yet.
• Claude Working Scripts — an index of every Code/_Claude Scripts/*.R (name, last-modified date, purpose) so a future session can find prior analysis.

Workflow

Before writing a new Code/*.Rmd script

• Read Code/REGISTRY.md.
• Scan the per-script cards and the Processed Data dependency graph for matches on: output filename, paper_target keywords, purpose keywords.
• If a match exists, surface it with the script name, purpose, and outputs, and ask: "edit this one, or create a new script?" Never silently start a new file.
• If no match, confirm a new script is appropriate, then proceed.

After creating or editing a Code/*.Rmd

• Make sure the registry: block is filled in (purpose, inputs, outputs at minimum). If an edit changed the I/O calls, update the block in the same turn.
• Run the builder:

   Rscript "Code/_Claude Scripts/build_registry.R"

• Confirm the "Registry updated: N scripts" line.
• Check the Drift Warnings section at the bottom of Code/REGISTRY.md. If the script you just touched appears there, the YAML and the code disagree — fix the YAML (or the code) and rerun.

When moving a Code/*.Rmd to _Archive/

• Move the file into Code/_Archive/ (never delete).
• Rerun the builder. The script drops out of REGISTRY.md automatically because _Archive/ is not scanned.
• If the script produced paper outputs, its replacement should inherit the paper_target text.

On demand

• "Update registry" — run the builder, report the summary line.
• "Show registry" — read Code/REGISTRY.md and show the relevant section.
• "Which script makes X?" — read Code/REGISTRY.md, scan the dependency graph or per-script Outputs sections.

Paper verification

The builder auto-checks each output against .tex files in Paper/ and writes In paper: yes/no per script. Matches on the filename stem (without extension) via fixed-string search. Stems shorter than 4 characters are skipped to avoid false positives.

Where outputs go

Three destination folders, chosen by output purpose:

• Output/Figures/ — any .pdf / .png / .jpg figure.
• Output/Tables/ — terminal paper tables: LaTeX .tex files and CSVs that feed the paper narrative directly and are not consumed by any other script.
• Output/Processed Data/ — intermediate artifacts: outputs of one script that are then inputs to another .Rmd (or to a helper .R). Use proc_dir <- file.path(project_dir, "Output/Processed Data") as the path variable.

Rule of thumb when writing a new output: if any other script reads this file, put it in Output/Processed Data/. Otherwise Output/Tables/ (for tabular) or Output/Figures/ (for graphical).

Good habits

• One registry: block per Code/*.Rmd. The block lives with the code — never edit REGISTRY.md by hand.
• After changing YAML, rerun the builder in the same turn. Don't defer.
• If a rename or move changes a path, rerun the builder so the dependency graph stays accurate.
• If two scripts claim the same output, that's a duplication problem — flag it rather than hiding it.
• Never register scripts that live in _Archive/. Archival = deregistration.
• Give every working Code/_Claude Scripts/*.R a # purpose: header line so it indexes cleanly (Section 5).
• Canonical .Rmd are decade-numbered (00s data, 10s descriptives, 20s results, 90s non-paper); Code/00_run_pipeline.R runs them in order. Promote an approved working .R into a numbered .Rmd, fill its registry: block, then archive the .R.