PiPER probe evidence v0.1 preview

An experimental local CPU test lift for PiPER, with continuous grip checks and a reproducible native-evidence archive.

• 24/24 valid native attempts; 12 fixed scenes per controller.
• Evidence gate: 6 admissions, 0 false initial qualifications; four later losses, 2 final qualified holds.
• Timer: 12 admissions, 8 false initial qualifications; 2 final qualified holds.
• All 480,000 native rows, commands, source identities and failed outcomes retained; independent readback included.

Both controllers stop at a short probe/hold. Neither executes full transport or learns from video. These finite simulator results do not establish hardware safety, material properties or general accuracy. Closed-loop trajectories differ. No improvement in final hold success is demonstrated.

Clone this tag or the experimental branch and follow benchmarks/piper-probe-evidence-v01/README.md. Python 3.12+, NumPy and MuJoCo are required; no GPU or API key. Provision pinned public PiPER assets once, then run offline. No new wheel or stable-main promotion is included.

Research · Hugging Face evidence

Core runtime frozen at a6a1d15. The release commit adds documentation, guarded archive tools and publishing code; runtime bytes remain identical. Original code and synthetic records are MIT; upstream notices are retained.