Cross-repo recurrence only guesses whether code is public. A repo's own .git/config holds its git remote URL, which is proof: a live public GitHub remote means the code is not novel; a remote that is gone/private, a non-GitHub host, or no remote at all (never pushed) marks genuinely novel code. Where .git/logs/HEAD survived we also show the last commit (author, message, count). Below, every repo in the download that preserved .git/.
| Repo | Git remote | Last commit | Status |
|---|
This is the whole recoverable set, not a sample. The code-only extraction stripped .git/ from the catalog (0 configs survive there), and the raw ingest/ archives are now purged (a re-harvest of 1,200 returned zero, all NoSuchKey). So these 18 are every repo whose .git/ survived in the full-extraction download. Scaling row-count corpus-wide needs a cold re-extraction of the raw archives (a paid restore). Even at n=18 the signal is stark: 13 of 18 are not currently public (gone, private, non-GitHub, or never pushed), so a recurrence-based "public" label would have been wrong for most of them.