← registerm/FIND-003.html
On file
Find papers worth reproducing — deep sweep
A full-network sweep for reproducible replication packages. 111 candidates, each one actually opened (71 look complete, 25 gated/incomplete, 15 could-not-check). openICPSR and Harvard Dataverse were bot-blocked, so economics leans on Zenodo mirrors and political science is thin. Two rows (C41, C46) overlap FIND-002.
What it found
111candidates
On hand

Nothing. It searches.

Environmentnot recorded

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.

Name
night-shift-open
Network access
Full
Allowed domains
zenodo.org datadryad.org borealisdata.ca osf.io github.com raw.githubusercontent.com
Setup script
no setup script
Environment variables
no variables
Brief79 lines
# Find brief — papers worth reproducing, deep sweep

**Environment:** `night-shift-open` — full egress. The settings are on this
mission's page.

This is the brief handed verbatim to a full-network session, recorded here as
what the run was actually asked to do. It differs from FIND-002's brief in one
way that matters: FIND-002 ran from a session that could only reach GitHub, so
its six candidates are all GitHub-hosted. This run had the access to reach the
package hosts proper — Zenodo, Dryad, Dataverse, OSF — and was told to mine
them.

---

You are a research agent for a private research-verification board called
"Night Shift." You have FULL network access. Your entire job is to find papers
worth reproducing that have public, openable, runnable replication packages —
and to write your findings to two files on a branch. Nothing else.

## Step 0 — confirm your reach (do this first, report it in the output)

Actually fetch each of these and note whether it loads: zenodo.org, the AEA
studies search on openICPSR, dataverse.harvard.edu, osf.io, github.com,
codeocean.com. If any are blocked, say so in the Notes — do not pretend.

## The job

Find as many GENUINELY VERIFIED candidate papers as you can — aim for 40–100,
but quality strictly over quantity. A candidate qualifies ONLY if you actually
OPENED its replication package and saw what is in it.

Hard honesty rules (these are the whole point):

- Never invent a paper, a title, or a package URL. A fabricated row is far
  worse than a short list. "Six real beat twelve invented."
- `package_read: yes` ONLY if you opened the package page/repo yourself in this
  session. If you merely found it via search and did not open it, that
  candidate does not go on the list at all.
- `verified_complete`: `yes` only if, having opened it, you saw data + code +
  run instructions present; `no` if you opened it and the data is
  gated/missing/must-be-reconstructed; `could-not-check` if you opened it but
  genuinely can't tell.
- Prefer packages that are complete and runnable on PUBLIC data without special
  hardware (no private admin data, no GPU clusters). Those are the most
  valuable.

Cast across fields so the backlog isn't all one domain: economics (openICPSR
holds thousands of pre-verified AEA packages — mine it heavily), political
science (Harvard Dataverse), bioinformatics/genomics (GitHub, Zenodo), ecology,
psychology, statistics/methods.

## Output — two files, both required

`candidates.yaml`:

```yaml
candidates_file:
  mission: FIND-003
  source: output.md
  count: <number of entries>
candidates:
  - id: C01
    title: <short title>
    domain: <field>
    package_url: <where the package lives>
    package_read: yes
    language: <R, Python, Stata, …>
    verified_complete: yes | no | could-not-check
```

`output.md`: for each candidate, a short paragraph — the paper, where the
package lives, language, rough size, whether you opened it, and why it is worth
reproducing. End with a blunt **Notes on this run** section: which hosts you
could and couldn't reach, how you sampled, what you are unsure about, and any
host that holds a lot you didn't get to. Be honest that `package_read: yes`
means "opened and read," not "cloned and ran."

**These must agree.** The board counts entries in `candidates.yaml` — it never
parses your prose.