Red-team your robot policy before a pilot.
Before a pilot customer relies on it, find out how often an unusual instruction pushes your policy out of its safe zone. Provael tests it in simulation, beside a harmless control. The tool is free; the fixed-price Checkpoint Report adds a reproducible report and an engineering review.
Who this is for
You are a small team shipping a learned robot policy. The next thing between you and a deployment is often a document: a customer’s security questionnaire, a CE technical file, an insurer’s application or a diligence list. None of them asks a settled question about learned robot policies yet. A measurement anyone can re-run gives you something to hand over.
What applies to you, and when
- EU Machinery Regulation 2023/1230 - Annex III 1.1.9 and 1.2.1
- If you sell into the EU, the machine your policy drives must meet the Annex III requirements that apply to it, 1.1.9 and 1.2.1 among them, from this date: protection against corruption and resistance to malicious attempts. The technical file carries the evidence. The delegated acts adding the AI Act robustness requirements apply by 2 August 2028; until then the evidence is your own.
- Product Liability Directive (EU) 2024/2853
- Software and AI systems are products. After this date defectiveness may be presumed where relevant evidence is not disclosed, which is a reason to have a dated measurement before you are asked.
- The customer’s security questionnaire
- It often arrives before any regulation does. Provael publishes its own answers, the questions it answers "No" included, and the same shape works for your answers about the policy. Provael’s answers →
What you get
- The report and its SARIF file: the attack-success rate (how often the policy left its safe zone) with 95% confidence intervals, the benign and clean-task-success controls, and the release decision under your acceptance protocol.
- The GitHub Action that fails your CI build when that decision is a fail.
- The signed attestation over the whole pack, which you verify offline.
- The technical-file annex: the test report in the ISO/IEC 17025 clause-7.8 shape (not accredited, no statement of conformity) and the crosswalk.
- Or run the same measurement yourself, free:
pip install provael, run the full sweep on the CPU stub in about twenty seconds, then your own checkpoint on a GPU with the committed recipe. Run the quickstart →
The sample evidence pack is a real pack from the committed run, published in full. The one measured result, and what it does not mean, is on the results page.
Start with one checkpoint
The Checkpoint Report covers one eligible checkpoint at a fixed price; for more, start with a scoping call. Eligibility is confirmed by a short form and a human reply before anything is sent or paid.
Feature and status claims on this page were read against the product repository on . This page was rewritten on , after that check; the re-check is pending. Unlike the numbers on /results and /leaderboard, these claims are not enforced by a build check: this is a human review with a date on it, not a guarantee. The maintained source is the repository; where it disagrees with this page, it wins.