Adopters
This page is the sign-up sheet: a self-reported list teams add themselves to by pull request. It is not a measurement of anything, and it is one of two pages named "adopters" — see the box below before you cite either.
It exists so a new user can see the tool in real use, and so adopters can find each other. An entry means a team said they use Provael to red-team a VLA policy — in CI, in research, or as part of a compliance-evidence workflow.
Two pages are called 'Adopters' and they mean different things
This page (docs/adopters.md) |
provael.com/adopters | |
|---|---|---|
| What it is | a sign-up sheet | a measurement |
| Where the data comes from | teams adding themselves by PR | PyPI downloads, GitHub stars, forks |
| Current state | empty — nobody has signed up | populated, and deliberately unflattering |
| What it can tell you | who is willing to say so publicly | how much the package is actually pulled |
They are not versions of each other and neither supersedes the other. An empty list here is not evidence of no users — it is evidence that nobody has opened a PR, which is a different and much weaker claim. The download figures on the site are the number to look at if you want to know whether anyone runs this; the ratio between them is the point of that page.
Status: PLANNED — no results claimed. This list is community-maintained and self-reported. An entry is a statement by the adopter, not an endorsement by Provael, and carries no measured Attack Success Rate or transfer claim.
Reproduce a result — the statement this project is short of
Ran one of the published results yourself? That is a different and stronger statement than using the tool, and it is the one this project needs: the reproduction register on the errata page holds zero third-party reproductions, and the sign-up sheet below is empty. The reproduction request points at one result — the delivery pack generated from the published body — with everything needed to attempt it at three levels. Level 1 needs no GPU:
git clone https://github.com/provael/provael && cd provael && git checkout v0.43.0
sha256sum -c examples/delivery-pack/smolvla-libero-object-2026-09-14/shards.txt # level 1: verify
Level 2 regenerates the derived artifacts from the committed shards (make check-delivery-pack);
level 3 re-executes the policy with the pinned inputs (REPRODUCE.md in the pack; Linux, GPU,
[lerobot]). Whichever level you reach, and whichever way it comes out, report it with the
reproduction result form
— a failed reproduction is published the same way as a successful one.
Add yourself
Open a pull request that adds one row to the table below. Keep it plain text — no logos, no marketing copy:
- Organization — your team, company, lab, or handle, exactly as you want it cited. A link on the name is welcome but optional.
- Use case — one honest line on how you use Provael (e.g. "CI gate on our SmolVLA fork", "coursework on embodied-AI security", "pre-deployment evidence for an internal review").
- Since — the month you started, as
YYYY-MM.
Add the row in alphabetical order by organization, and sign the commit off (git commit -s) like
any other contribution.
Adopters
| Organization | Use case | Since |
|---|---|---|
| No public adopters have signed up yet. |
What this empty table does and does not say. It says zero teams have opened a PR to add themselves. It does not say zero teams use Provael — the measured distribution figures on provael.com/adopters are the number for that question, and they are not zero. Reading an empty sign-up sheet as an empty user base is the mistake this note exists to prevent, in both directions.