Backed by Y CombinatorThree documents and two sheets that trace every exhibit to its script, then flag one edited after its output was last written.
Free download · No account needed
What you get
Pack of 5 files: 3 documents, 2 spreadsheets
Research Code Documentation TemplateOutput Map
| Script | Script modified | Output file | Output modified | Status |
|---|---|---|---|---|
| 03_reg_compliance.R | 19 Nov, 15:47 | table2.csv | 24 Oct, 14:33 | Stale |
The script gained a referee-requested control on 19 November. Its output file has carried the 24 October timestamp since, twenty-six days earlier, so the number in the draft is what the script produced before the control existed.
Every reproducibility guide gives the same fix: rewrite the pipeline into a dependency tool like Make or Snakemake, or start from a Cookiecutter skeleton. Both organize a project. Neither checks whether a record already written is still true. The AEA's data and code policy requires a submitting README to map every program to its output and exhibit. The Social Science Data Editors' template it endorses turns that requirement into a table, naming the program and line behind each one. This pack builds that table, then checks it against two timestamps every file already carries.
The pack ships filled in for the Ashgrove Wage Compliance panel. Five scripts turn a raw payroll extract into a summary table, two regressions and an event-study figure across 9,920 establishment-quarter observations. A referee asked for an establishment-size control on the compliance regression. The script was edited 19 November and never re-run, so its output file carries a 24 October timestamp, twenty-six days earlier. Output Map catches the gap from those two dates alone. A sixth script, checking a threshold the paper never reports, produces a file no exhibit claims and the master never calls.
Open the pack in River and the agent reads your scripts folder and draft manuscript. It builds the File Register from what it finds, then checks every claimed exhibit against its script's and output's real modified times. Or download the documents and sheets and fill them in by hand. The panel this pack verifies usually starts at the dataset cleaning and codebook pack. The same numbers feed the methods section and get reformatted by the results table and figure pack for a target journal. More research packs sit alongside this one.
File RegisterFour of the pack's seven rows, including the master script and the two the next check turns into flags.
| File | Purpose | Reads | Writes | Called by master |
|---|---|---|---|---|
| 00_run_all.R | Sources every script below in sequence and nothing else | — | — | n/a |
| 01_clean_payroll.R | Cleans the raw payroll extract into the analytic panel | payroll_audit_raw.csv | payroll_panel.csv | Yes |
| 03_reg_compliance.R | Estimates the main compliance regression, Table 2 | payroll_panel.csv | table2.csv | Yes |
| 06_appendix_placebo.R | Runs a placebo test at a threshold the paper does not report | payroll_panel.csv | placebo_test.csv | No |
03_reg_compliance.R was edited 19 November to add a referee's control and has not been re-run since. 06_appendix_placebo.R writes a file but 00_run_all.R never calls it, which is what makes it a candidate orphan before Output Map checks it against the paper.
Output MapEvery exhibit's script and both files' real modified times, with a status computed from the two rather than typed in.
| Exhibit | Script | Script modified | Output file | Output modified | Status |
|---|---|---|---|---|---|
| Table 1 | 02_summary_stats.R | 22 Oct, 09:14 | table1.csv | 22 Oct, 09:16 | Current |
| Table 2 | 03_reg_compliance.R | 19 Nov, 15:47 | table2.csv | 24 Oct, 14:33 | Stale |
| Table 3 | 04_reg_robustness.R | 25 Oct, 10:00 | table3.csv | 25 Oct, 10:04 | Current |
| Figure 1 | 05_event_study_figure.R | 26 Oct, 16:00 | figure1.png | 26 Oct, 16:05 | Current |
| — | 06_appendix_placebo.R | 2 Nov, 11:20 | placebo_test.csv | 2 Nov, 11:22 | Orphaned script |
| — | n.a. | — | table3_v2.csv | 20 Oct, 09:50 | Orphaned output |
Table 2's script is newer than its output by twenty-six days, the exact gap the check computes rather than estimates. The last two rows have no exhibit to check against until File Register lists what actually produces a file.
Every exhibit's script, both files' real modified times, and a status computed by comparing them: current, stale or orphaned, never typed in by hand.
Every script's purpose, what it reads and writes, and whether the master script actually calls it, inventoried before anything gets checked against it.
The pipeline's run order and contents, copied from the master script's own source lines. The panel it describes is what the results section writer later turns into prose.
The exact software and package versions used, any random seed and where it is set, and a runtime estimate, so a replicator knows what to install.
Where every data source came from, what access it requires, and how a future researcher could obtain it again, including a source behind a fee or an agreement.
The standing space rule every prompt reads first. A journal's own README template leaves the exhibit-to-script table to be filled in by hand. This rule is what gets it checked instead.
Open the pack in River and the agent reads your scripts folder and draft manuscript with you. Or download the blank documents and CSV sheets instantly, filled in with the worked example.
Every script in the folder, not just the ones you remember, plus the draft naming its tables and figures by number. A folder listing beats a description of the main scripts.
File Register inventories every script's purpose and dependencies first. Output Map then traces each exhibit in the paper to the exact script and line that produces it.
The script's own modified time against its output's. A script edited after its output was last written is flagged stale, and a file or script nothing claims is flagged orphaned.
Yes. Download the pack's three documents and two sheets with no signup and no credit card, filled in for the worked example. Edit with AI installs the same pack as a private River workspace. There, the agent reads your own scripts folder and draft manuscript, and builds the register and the map from your files.
Word documents for the README, Environment Note and Data Provenance Statement, and CSV for File Register and Output Map, all in one zip. Nothing needs converting: they open in Word, Pages, Google Docs, Excel, Numbers and Sheets. Inside River the same content opens as live Docs and Sheets that update as you send more files.
A build tool re-runs a pipeline it already understands as a formal dependency graph. That is a real fix, once every script has been rewritten into that tool's own rule syntax. This pack does not require that rewrite. It reads the modified time already sitting on every script and output file in an ordinary folder, so it works on the pipeline you already have.
Stale means the script producing a claimed exhibit was edited after its output file was last written. The number on the page may not match what the code would produce today. Orphaned means a script or output file exists that neither the paper's own exhibit list nor the master script's run order accounts for, in either direction.
Yes, with one honest gap. File Register records the run order you can describe rather than inventing one, and Output Map still traces each exhibit to its script. What it cannot check without a master script is whether a script nobody calls was meant to run at all, and the register says so rather than guessing.
No. It checks whether the record of what produced what still matches the files you have, not whether the specification behind any exhibit is defensible. Testing that specification against reasonable alternatives is a separate question a dedicated robustness checks pack answers. Re-running a reviewer's requested control is what would make Output Map current again.
Written by
Co-founder & CEO, River
Chandler has spent the last six years working in AI, building tools for millions of users and Fortune 100 companies. In 2025, he co-founded River to build a new kind of operating system for AI agents.
Send your scripts folder and your draft. The register and the map come back with every exhibit checked against the files' own timestamps.
Edit with AI