WHO IT IS FOR
Find the page written for you.
Provael tests one thing: how often an unusual instruction pushes a robot policy out of its safe zone, in simulation, beside a harmless control. What differs by reader is the deadline you face and what you do with the result.
Pages by reader
- Robotics startupsYou ship a learned robot policy to pilot customers. Test it before the pilot, free with the tool or as a fixed-price report.
- Humanoid buildersYour humanoid runs on a whole-body learned policy. What the EU Machinery Regulation asks of it, and what can be tested in simulation today.
- Automotive & logistics deployersYour fleet’s navigation or manipulation is a learned policy. Where a measured rate fits in the cybersecurity work you already do.
- Certifiers & insurersYou review or underwrite machines run by learned policies. What the evidence contains, and where it stops.
The same scope applies to every reader. There are zero real-hardware results: the physical-transfer study is pre-registered and has not been run. What is published is one measured result, with its limits: simulation only · 10 tasks · 5 seeds per task · uncalibrated keep-out predicate · in the control run, the frame with no target named does the same: 27 of 30 against the attack’s 26 of 30. Evidence toward an assurance case, not a certificate.