BY SEGMENT · PV-015
Same measurement. Four different reasons you need it.
Provael measures one thing: how often an adversarial instruction drives a robot policy out of its safe envelope, with a benign control beside it. What changes by segment is the deadline you are answering to and the document the answer has to go into.
Segments
- PV-016Robotics startupsYou ship a policy and a customer asks what happens when someone rewords the instruction. Evidence you can produce this week, in CI.
- PV-017Certifiers & insurersThe Machinery Regulation binds 20 January 2027 and the standards are not ready. Independent, reproducible evidence you can file.
- PV-018Humanoid buildersWhole-body and balance failure modes that a manipulation benchmark does not have a row for, against a standard still in working draft.
- PV-019Automotive & logistics deployersAMR and AGV fleets under UN R155/R156 type approval, already mandatory - the segment with the clock that has already run.
The same honest scope applies to every segment. Provael measures in simulation. There are zero real-hardware results - the physical-transfer study is pre-registered and has not been run. The keep-out predicate is uncalibrated. What we publish is one measured result and the nulls beside it: sim-only · ten tasks · 5 seeds/task · uncalibrated keep-out predicate · benign-reword arm measured: 1/50 vs the attack’s 44/50. Evidence toward an assurance case, not a certificate.