Nothing.
This is the environment file as it stands today, not a record of this run. The run predates environment capture, so what it actually had on hand was never recorded and cannot be recovered.
# Find brief — papers worth reproducing
Paste this whole thing. Nothing to fill in.
**Environment:** `night-shift-map`. The exact settings are on this mission's
page — the env file is the only place they live.
---
## The job
Find candidate papers worth reproducing. A good candidate has a replication
package that is public, complete, and actually runnable, and a result worth
checking.
For each candidate: the paper, where its replication package lives, what
language the code is in, roughly how large it is, and why it is worth checking.
Say plainly which ones you could not verify are complete — an unverified
package is a finding, not a gap to paper over.
Do not pad the list. Six real candidates beat twelve with six invented.
## Output — two files, and both are required
**1. `output.md`** — the prose. Write it for a reader. Structure it however
reads best, and end with a section called **Notes on this run**: what you
searched, what failed to load, what you are unsure about. Blunt.
**2. `candidates.yaml`** — the same candidates as structured data:
```yaml
candidates_file:
mission: FIND-002
source: output.md
count: <number of entries below>
candidates:
- id: C01
title: <short title>
domain: <field>
package_url: <where the replication package lives>
package_read: yes | no # did you open it?
language: <R, Python, Stata, …>
verified_complete: yes | no | could-not-check
```
**`package_read` and `verified_complete` are different questions.** One is a
network fact, the other a judgment about what you found:
- `package_read: no` + `verified_complete: could-not-check` — you never
reached the package. Nothing here is a statement about its contents.
- `package_read: yes` + `verified_complete: could-not-check` — you opened it
and still could not tell whether it is complete. That is a finding about
the package.
Without the first field those two collapse into the same answer, and a
listing URL you never opened reads exactly like one you did. `package_read`
is checkable: an opened page leaves a fetch in this session's transcript,
which you do not write.
It also marks which of the other fields are observed and which are inferred.
With `package_read: no`, `language` and `verified_complete` are guesses from
the paper or a search result, not from the package — say so in the prose
rather than letting the row imply otherwise.
**These must agree.** Every candidate in the prose appears in the YAML and vice
versa. The board counts entries in `candidates.yaml` — it never parses your
prose, because a count derived from headings is a count that can silently be
wrong. If you can only produce one of the two files, produce neither and
deposit a finding explaining why.
## Where to put them
Push a branch to **`night-shift-network/inbox`** with both files inside a
directory named for this mission:
```
FIND-002/output.md
FIND-002/candidates.yaml
```
The directory is what matters. A file at the root of the branch is not a
deposit and will not be collected — it will sit there looking delivered while
nothing has been.
The branch name does not matter and is deliberately not specified: the relay
reads every branch, so whichever branch your session is already on is fine.
Never delete or force-push a branch, yours or anyone else's; cleanup is
maintainer-side only.
**Paper.** Baker, Callaway, Cunningham, Goodman-Bacon & Sant'Anna, "Difference-in-Differences Designs: A Practitioner's Guide," *Journal of Economic Literature*. **Package.** https://github.com/pedrohcgs/JEL-DiD — **opened.** R and Stata, both provided. The README lists all raw data as coming from public sources (CDC, BLS, Census Bureau, Kaiser Family Foundation), ships analysis-ready datasets in CSV and Stata formats, and drives the whole pipeline from a master script. Stated runtime ~10–15 minutes on standard hardware, ~4 GB RAM, 3,143 US counties over 2010–2019 (Medicaid expansion as the running example). **Why it is worth checking.** It is the cleanest first reproduce run available here: fully public data, self-contained, fast, and two independent language paths (R and Stata) that should agree — a disagreement between them would itself be the finding. The figures *are* the paper's product, so reproducing them checks the thing the paper is actually for. **Completeness.** `verified_complete: yes` — from the README and file tree, not from executing. Only caveat: Stata is proprietary, but the R path is free, so an outsider is not blocked. ---https://github.com/pedrohcgs/JEL-DiD
**Paper.** Herbert, Kingi, Stanchi & Vilhuber, "Reproduce to Validate: a Comprehensive Study on the Reproducibility of Economics Research," *Canadian Journal of Economics*. **Package.** https://github.com/larsvilhuber/aej-applied-replications-public — **opened.** R with shell automation, Docker-based setup recommended. The README says all data are public; the underlying Kingi et al. (2019) dataset is not bundled but is downloaded automatically from Zenodo, with Crossref / OpenAlex / Web of Science data provided or reached via cached APIs. Stated runtime under 10 minutes on a desktop, 25–250 MB storage, ~16 GB RAM for some steps. Repository archived read-only on 2024-09-25. **Why it is worth checking.** A paper whose subject *is* reproducibility, that is itself reproducible — reproducing it is a recursive check of the field's own claims about itself, which is squarely what this board exists for. The archived status is a plus: the target cannot move under you. **Completeness.** `verified_complete: yes` — public data auto-fetched, Docker path documented. Same "read, not run" caveat as every row here. ---https://github.com/larsvilhuber/aej-applied-replications-public
**Paper.** snpArcher, *Molecular Biology and Evolution* (2023) — a Snakemake workflow for highly parallel variant calling in non-model organisms. Python / Snakemake. Ships an `example/` directory with sample data and a `tests/` directory; documentation at snparcher.github.io; the associated Comparative Population Genomics datasets are public (distributed via Globus). ~1,300 commits, a mature project. **Why it is worth checking.** Breadth outside economics, and a different kind of reproducibility claim: not "these numbers" but "this pipeline runs and gives the same variants." The bundled small test data makes a bounded check feasible without pulling terabytes. **Completeness.** `verified_complete: yes`, with a real gotcha the repo itself flags: a March-2026 note pins Snakemake to v9.12 because a later version has a resource-parsing regression. That is exactly the kind of silent version dependency a reproduce run should record — worth reproducing partly *to* pin it. ---https://github.com/harvardinformatics/snpArcher
**Paper.** Dingel, Miscio & Davis, "Cities, Lights, and Skills in Developing Economies," *Journal of Urban Economics*. **Package.** https://github.com/jdingel/DingelMiscioDavis — **opened.** R, Stata, and shell, orchestrated with `make`; originally involved ArcGIS. The README lists explicit software versions (R 3.5.1, Stata 15, GNU bash 4.2.46) and system dependencies (GEOS, GDAL, PROJ, udunits), and supports both cluster (Slurm) and local runs. Four country directories (Brazil, China, India, USA). **Why it is worth checking.** A production-grade, pinned, `make`-based package — a good stress test of whether a multi-language, multi-country pipeline still runs years later. The precise version pins make "does it still build" a real, answerable question. **Completeness.** `verified_complete: yes` with two caveats stated plainly: Stata 15 is proprietary, and the Brazil branch needs gigabytes of IBGE microdata that are not bundled — but the README says that step can be skipped, so the rest is runnable without it. ---https://github.com/jdingel/DingelMiscioDavis
**Paper.** Barberá et al., "Tweeting from Left to Right: Is Online Political Communication More Than an Echo Chamber?", *Psychological Science*. **Package.** https://github.com/pablobarbera/echo_chambers — **opened.** R and Python. The code is present; the raw data is not in the repo — it lives on Dataverse (doi:10.7910/DVN/F9ICHH), and because of Twitter's terms the package ships tweet-ID collections and processing scripts rather than tweet text, so the full dataset (~150M tweets, ~3.8M users) has to be re-collected through the Twitter API. **Why it is worth checking.** A widely cited, genuinely contested empirical claim — the "echo chamber" result is exactly the loud-claim-thin-independent-check pattern this board hunts. High relevance. **Completeness.** `verified_complete: no`, and this is the important honesty: the reproduction pathway is largely broken today. Re-collecting tweets from IDs now requires paid/closed Twitter API access, so an outsider likely cannot rebuild the dataset. The *code* is checkable; the *result* is not, from the public materials alone. Map it knowing that up front. ---https://github.com/pablobarbera/echo_chambers
**Paper.** Bonhomme, Lamadon & Manresa, "A Distributional Framework for Matched Employer-Employee Data," *Econometrica*. **Package.** https://github.com/tlamadon/blm-replicate — **opened.** R, built as an R package with `packrat` pinning exact library versions, plus a Docker container — three documented run paths (Docker, local + packrat, direct install). Well-organised: core functions in `R/`, estimation scripts, figure/table generation, a `data-simulation` path, and `inst/main.R` as the entry point. **Why it is worth checking.** Methodologically important, and unusually careful about environment reproducibility (packrat + Docker) — the estimator can be exercised on simulated data, so the *method* is checkable end to end. **Completeness.** `verified_complete: no`. The headline empirical results use Swedish administrative data that requires separate access through IFAU and is not in the package. So an outsider can reproduce the machinery on simulated data but cannot reproduce the paper's actual numbers — a clean "code public, data gated" split, and the notes should say so rather than imply otherwise. ---https://github.com/tlamadon/blm-replicate
# Run 0 — papers worth reproducing **Prepared:** 2026-08-17. **Method in one line:** keyword search to find candidate papers with public replication packages, then — for every candidate — the package's own repository page opened and read before it was listed. > **Read this before you trust a row.** This run was executed from a session > whose network reached **GitHub** (`github.com`, `raw.githubusercontent.com`) > but was blocked by the egress proxy for **arxiv.org, openICPSR, Zenodo, and > Dataverse** — every one tested returned a policy denial. Two consequences, > and both are load-bearing: > > 1. **Every package below lives on GitHub, and every one was actually opened.** > That is why all six carry `package_read: yes`. It is not a coincidence of > taste — it is the shape of what this session could reach. Strong packages > that live only on openICPSR/Zenodo/Dataverse could be *found* by search but > not *opened*, so they are not listed: an unopened package is exactly the > row this brief tells me not to pretend I checked. > 2. **`package_read: yes` here means "I opened the repository page and read its > README and file tree," not "I cloned it and ran it."** No package was > executed. So `verified_complete` is a judgment about whether the package > *contains* what it needs — data, code, instructions — not a claim that its > results reproduced. Where the data is public and present, that judgment is > `yes`; where the data is gated or must be reconstructed, it is `no`, and > the row says which. One notch of doubt on top of all of it: the pages were > read through a fetch-and-summarise tool, not by my own eyes on the raw > tree, so file-level details are as-surfaced. > > More in the notes. The one thing this run can stand behind fully is that all > six repositories exist, resolve, and contain broadly what each row says. --- ## 1. Difference-in-Differences Designs: A Practitioner's Guide — *econometrics / methods* **Paper.** Baker, Callaway, Cunningham, Goodman-Bacon & Sant'Anna, "Difference-in-Differences Designs: A Practitioner's Guide," *Journal of Economic Literature*. **Package.** https://github.com/pedrohcgs/JEL-DiD — **opened.** R and Stata, both provided. The README lists all raw data as coming from public sources (CDC, BLS, Census Bureau, Kaiser Family Foundation), ships analysis-ready datasets in CSV and Stata formats, and drives the whole pipeline from a master script. Stated runtime ~10–15 minutes on standard hardware, ~4 GB RAM, 3,143 US counties over 2010–2019 (Medicaid expansion as the running example). **Why it is worth checking.** It is the cleanest first reproduce run available here: fully public data, self-contained, fast, and two independent language paths (R and Stata) that should agree — a disagreement between them would itself be the finding. The figures *are* the paper's product, so reproducing them checks the thing the paper is actually for. **Completeness.** `verified_complete: yes` — from the README and file tree, not from executing. Only caveat: Stata is proprietary, but the R path is free, so an outsider is not blocked. --- ## 2. Reproduce to Validate: a Comprehensive Study on the Reproducibility of Economics Research — *meta-reproducibility* **Paper.** Herbert, Kingi, Stanchi & Vilhuber, "Reproduce to Validate: a Comprehensive Study on the Reproducibility of Economics Research," *Canadian Journal of Economics*. **Package.** https://github.com/larsvilhuber/aej-applied-replications-public — **opened.** R with shell automation, Docker-based setup recommended. The README says all data are public; the underlying Kingi et al. (2019) dataset is not bundled but is downloaded automatically from Zenodo, with Crossref / OpenAlex / Web of Science data provided or reached via cached APIs. Stated runtime under 10 minutes on a desktop, 25–250 MB storage, ~16 GB RAM for some steps. Repository archived read-only on 2024-09-25. **Why it is worth checking.** A paper whose subject *is* reproducibility, that is itself reproducible — reproducing it is a recursive check of the field's own claims about itself, which is squarely what this board exists for. The archived status is a plus: the target cannot move under you. **Completeness.** `verified_complete: yes` — public data auto-fetched, Docker path documented. Same "read, not run" caveat as every row here. --- ## 3. snpArcher: a reproducible variant-calling workflow — *genomics / bioinformatics* **Paper.** snpArcher, *Molecular Biology and Evolution* (2023) — a Snakemake workflow for highly parallel variant calling in non-model organisms. **Package.** https://github.com/harvardinformatics/snpArcher — **opened.** Python / Snakemake. Ships an `example/` directory with sample data and a `tests/` directory; documentation at snparcher.github.io; the associated Comparative Population Genomics datasets are public (distributed via Globus). ~1,300 commits, a mature project. **Why it is worth checking.** Breadth outside economics, and a different kind of reproducibility claim: not "these numbers" but "this pipeline runs and gives the same variants." The bundled small test data makes a bounded check feasible without pulling terabytes. **Completeness.** `verified_complete: yes`, with a real gotcha the repo itself flags: a March-2026 note pins Snakemake to v9.12 because a later version has a resource-parsing regression. That is exactly the kind of silent version dependency a reproduce run should record — worth reproducing partly *to* pin it. --- ## 4. Cities, Lights, and Skills in Developing Economies — *urban / development economics* **Paper.** Dingel, Miscio & Davis, "Cities, Lights, and Skills in Developing Economies," *Journal of Urban Economics*. **Package.** https://github.com/jdingel/DingelMiscioDavis — **opened.** R, Stata, and shell, orchestrated with `make`; originally involved ArcGIS. The README lists explicit software versions (R 3.5.1, Stata 15, GNU bash 4.2.46) and system dependencies (GEOS, GDAL, PROJ, udunits), and supports both cluster (Slurm) and local runs. Four country directories (Brazil, China, India, USA). **Why it is worth checking.** A production-grade, pinned, `make`-based package — a good stress test of whether a multi-language, multi-country pipeline still runs years later. The precise version pins make "does it still build" a real, answerable question. **Completeness.** `verified_complete: yes` with two caveats stated plainly: Stata 15 is proprietary, and the Brazil branch needs gigabytes of IBGE microdata that are not bundled — but the README says that step can be skipped, so the rest is runnable without it. --- ## 5. Tweeting from Left to Right: Is Online Political Communication More Than an Echo Chamber? — *political communication* **Paper.** Barberá et al., "Tweeting from Left to Right: Is Online Political Communication More Than an Echo Chamber?", *Psychological Science*. **Package.** https://github.com/pablobarbera/echo_chambers — **opened.** R and Python. The code is present; the raw data is not in the repo — it lives on Dataverse (doi:10.7910/DVN/F9ICHH), and because of Twitter's terms the package ships tweet-ID collections and processing scripts rather than tweet text, so the full dataset (~150M tweets, ~3.8M users) has to be re-collected through the Twitter API. **Why it is worth checking.** A widely cited, genuinely contested empirical claim — the "echo chamber" result is exactly the loud-claim-thin-independent-check pattern this board hunts. High relevance. **Completeness.** `verified_complete: no`, and this is the important honesty: the reproduction pathway is largely broken today. Re-collecting tweets from IDs now requires paid/closed Twitter API access, so an outsider likely cannot rebuild the dataset. The *code* is checkable; the *result* is not, from the public materials alone. Map it knowing that up front. --- ## 6. A Distributional Framework for Matched Employer-Employee Data — *labor economics / econometrics* **Paper.** Bonhomme, Lamadon & Manresa, "A Distributional Framework for Matched Employer-Employee Data," *Econometrica*. **Package.** https://github.com/tlamadon/blm-replicate — **opened.** R, built as an R package with `packrat` pinning exact library versions, plus a Docker container — three documented run paths (Docker, local + packrat, direct install). Well-organised: core functions in `R/`, estimation scripts, figure/table generation, a `data-simulation` path, and `inst/main.R` as the entry point. **Why it is worth checking.** Methodologically important, and unusually careful about environment reproducibility (packrat + Docker) — the estimator can be exercised on simulated data, so the *method* is checkable end to end. **Completeness.** `verified_complete: no`. The headline empirical results use Swedish administrative data that requires separate access through IFAU and is not in the package. So an outsider can reproduce the machinery on simulated data but cannot reproduce the paper's actual numbers — a clean "code public, data gated" split, and the notes should say so rather than imply otherwise. --- ## Ranking — which to reproduce first **Tier 1 — fully runnable on public data, do one of these first.** 1. **#1 Difference-in-Differences guide.** The best first run: public data, R and Stata, ~15 minutes, two language paths that should agree. Nearly nothing can go wrong that would not itself be a finding. 2. **#2 Reproduce to Validate.** Public data auto-fetched, Dockerised, under ten minutes, and archived so it cannot drift. Reproducing a reproducibility study is a satisfying recursive check. 3. **#3 snpArcher.** The breadth pick, and the one with a live version-pinning gotcha already documented — reproduce it partly to confirm that pin. **Tier 2 — runnable with a proprietary or heavy dependency.** 4. **#4 Cities, Lights, and Skills.** Runnable but wants Stata 15 and a geospatial stack; the version pins make "does it still build" the real question. **Tier 3 — code checkable, result not, from public materials.** 5. **#5 Echo chambers.** High relevance, but the Twitter-API data pathway is effectively closed now. Reproduce the code; do not expect to rebuild the data. 6. **#6 Matched employer-employee.** Elegant, well-pinned code runnable on simulated data; the actual results rest on IFAU-gated Swedish records an outsider cannot get. --- ## Notes on this run **Blunt version, for me, not a reader.** **The single biggest constraint: I could open GitHub and nothing else that matters.** `github.com` and `raw.githubusercontent.com` resolved; `arxiv.org`, `openicpsr.org`, `zenodo.org`, and `docs.github.com` were all blocked by the egress proxy, each tested and each returning a policy denial. So this list is GitHub-shaped by force, not by preference. Every strong replication package that lives *only* on openICPSR (most AEA economics packages), Zenodo, or Dataverse is absent — not because it is not worth checking, but because I could not open it, and this brief is explicit that an unopened package is a finding, not a row to fill in anyway. **`package_read: yes` is honest but narrow.** I opened each repository's page and read its README and file tree. I did not clone or run any of them. So `verified_complete` is a judgment about the package's *contents*, never a claim that the results reproduced. And the reading itself went through a fetch-and-summarise tool, not my own eyes on the raw files — so the exact language mix and completeness of each are as-surfaced, one notch below hand-inspected. Before committing a night to any of these, clone it and read the tree yourself. **Provenance mismatch worth flagging for the maintainer.** This run's mission env is `night-shift-map`, whose allowlist is journal and preprint hosts plus a few party domains — it does **not** include GitHub or any package host. I did not run under that env; I ran from this session, which happened to reach GitHub and not the journals. So neither `night-shift-map` nor any single existing env describes what this run could actually touch. FIND-002 is, concretely, the mission that wants its own environment: one that can reach the places replication packages live — GitHub *and* Zenodo *and* openICPSR *and* Dataverse — which is a different allowlist from the map env. That gap is the real infrastructure finding of this run. **Economics is over-represented (4 of 6).** That is not a thumb on the scale — it is where public, mandated, runnable replication packages genuinely cluster, because the AEA and other journals require and pre-verify them. The two non-economics picks (snpArcher, echo chambers) are real, but if you want more spread, the honest way to get it is to fix the environment so Zenodo and Dataverse open, not to pad this list with fields where I could not confirm a package. **What I did not do.** I did not verify author lists or exact titles against the published articles — those are as they appear on the repository pages and in search results, and a couple (the full author list on #5, the exact venue on #3) I am carrying partly from prior knowledge. Confirm them when you write a repro brief. I also did not attempt any package on a non-GitHub host, so I cannot say what the openICPSR/Zenodo landscape holds — only that it is large and unreached. **Bottom line.** Six real packages, all on GitHub, all opened, honestly split three-and-three on whether their data is public enough to reproduce the actual result versus only the code. The three Tier-1 picks (DiD guide, Reproduce to Validate, snpArcher) are the ones I would start a reproduce night on: public data, bounded runtime, and nothing gated standing between you and the numbers.