Has anyone run an adversarial evaluation on the released checkpoints?

[sattyamj-attri]

Question rather than a proposal.

Has anyone here run a structured adversarial evaluation against the released pi checkpoints, with a benign control arm? I have run a matched-pair protocol against SmolVLA and got an 88% roleplay ASR against a 4.0% benign control, and I do not know whether that number is characteristic of the policy family or of that one model.

If nobody has, I would rather ask what a useful result would look like to you before I run anything, because a number nobody asked for is not evidence of anything.

Tooling I use is Apache-2.0 if that matters, but I am not pitching it here.