Real-world eval runs for a manipulation policy

What we buy: a setup that runs a policy checkpoint we supply on a fixed task suite and reports results. Not a dataset.
Embodiment: a common arm (UR, Franka, SO-100 or similar) with a parallel gripper.
Run: 5 pick and place task variants, 20 trials each, fixed object set and reset protocol, documented once.
Deliverables per run: success rate per variant, video of every trial, the logged trajectories in LeRobot v3.0.
Environment: lab or real cell; held constant across runs.
Format: LeRobot v3.0 (we can convert v2.1, MCAP or raw video with Prep). Acceptance: passes Soul Verified validation (decodes, streams in sync, every episode labelled) on a 2 hour sample before the full order. Rate is final after the sample is validated.
Rate: $300 to $600 per eval run, final on validation.