Deleted Is Rarer Than Abandoned: Keep a DOI Next to Every AI Repo URL
I run a radar of open-source AI tools on my site. Between 11 June and 31 October 2026 it selected 757 entries.
This article covers what I measured on those entries and what you can do with your own list of repositories this week.
What I measured
Of the 757 entries, 733 were projects on GitHub or Hugging Face. I followed them in 50 daily snapshots from 12 September to 31 October 2026.
I fixed the protocol on 5 October 2026, before the analysis. That means the rules were set before the results were known.
The rule for "unavailable" is simple. GitHub has to answer 404 or 451, or Hugging Face has to answer 401 or 404, in two snapshots in a row. One failed check is not enough.
For the statistics I used Kaplan-Meier curves and a log-rank test. I counted every horizon from the day each project was selected.
What happened to the projects
23 projects became unavailable. 5 of those 23 came back later. So a "gone" project is not always gone.
A returned project also shows why one failed check is not counted as unavailable.
This is the share still available, counted from the selection day:
| Weeks after selection | Share still available |
|---|---|
| 4 | 96.7% |
| 8 | 94.6% |
| 12 | 93.0% |
Roughly 1 in 14 projects disappeared within 12 weeks.
Deleted is rarer than abandoned
I also counted abandonment, which I defined as no push for 90 days. 67 projects were abandoned, against 23 that became unavailable.
A repository with no push for 90 days still opens. A link check will pass, and the project can still be inactive.
So "does the URL work" and "is the project maintained" are two separate questions, and they need two separate columns in your notes.
Stars and platform
I looked at the top quarter of projects by stars at selection, which means 4,988 stars and more. In that group there was 1 event out of 144 projects.
The log-rank p value was 0.004, and the result held when I stratified by month of selection. But stars and the radar's changing selection cannot be fully separated. So I do not say that stars protect a project. I only report what I saw in this data.
Hugging Face versus GitHub showed no difference (p = 0.21).
Limits
This is one radar with one selection rule. It does not describe all of GitHub.
After week 12 the combined curve drops. That drop is an artefact of the 90-day inactivity rule, not projects vanishing. The curve has a step because of the threshold.
Week 20 has only 5 projects under observation, so I do not interpret it.
What you can do this week
If you keep a list of open-source AI tools, depend on repositories from GitHub or Hugging Face, or cite repositories in papers and docs, you can run a small version of this on your own list.
- Keep a dated list of the repositories you depend on or cite.
- Check each one on a schedule and record the response code.
- Mark a repository as missing only after two misses in a row.
- Keep a separate column for "no push in 90 days".
- When a missing repository comes back, record the return.
- Add a DOI or an archive link next to every URL you cite.
Step 3 follows my own rule. Step 5 matters because 5 of my 23 unavailable projects returned, and a list that deletes them would hide that.
Step 6 is the practical conclusion of the report. Cite a DOI or an archived copy, not only the repository URL.
The related Internet-Draft
I also published an IETF Internet-Draft, draft-arsentev-repository-availability-reporting-00. It is an individual Informational draft. It is not an IETF standard, and it has not been adopted by a working group.
The draft proposes these points for anyone who reports repository availability:
- Classify every check as available, missing or error. An error is not evidence that a repository is gone.
- List per platform which response codes count as "no repository".
- Call a repository unavailable only after N misses in a row, and state N.
- Count returns instead of deleting them.
- Report inactivity and archiving separately from unavailability.
- Count horizons from the selection date, and state the estimator and the number under observation.
- Put a DOI or an archived copy (Memento, RFC 7089) next to the repository URL.
My checklist above is a small version of the same ideas. In my radar I used two misses in a row. If you pick another number, state it, so the reader knows what "unavailable" means in your list.
Summary
In my data, 23 projects became unavailable and 67 were abandoned. Deleted was rarer than abandoned, and some of the deleted ones came back. Keep the two cases apart in your notes, and keep a DOI or an archive copy next to the URL.
The report with the curves and the data: How Long Do Trending AI Projects Stay Available and Maintained?, DOI 10.5281/zenodo.24101932.
The draft: draft-arsentev-repository-availability-reporting.
Evgenii Arsentev, MD, PhD - creator of Arsminal; founder of ArsentevAI and Dad4Kids. ORCID 0000-0002-9120-7298 ยท Wikidata Q141398931
Comments
No comments yet. Start the discussion.