Moonlake is building evaluation and training infrastructure for robot deployment, so teams can turn a captured station into a physics twin and test policies before the next round on hardware. This launch is an early research preview for labs and operators who already have a robot and a task, and who lose days to teleoperation, resets, and manual diagnosis. A small set of real trajectories calibrates the robot, cameras, and object physics. The same actions can then be replayed in simulation, and the twin is used to find where generalization fails so the next real collection, or a round of reinforcement learning, targets those failures. The preview covers three bimanual YAM tasks: sorting small industrial parts from a tote, placing bottles in a bin, and flipping mugs upright. Across fifteen policies, simulation and real-robot progress agreed closely. A policy that did well in simulation also did well on the robot, and a policy that did poorly in simulation did poorly on the robot. Starting from 50 teleoperated episodes on the contact-rich sort, rejection fine-tuning on successful simulated grasps raised task progress by 11 points. Reinforcement learning in 384 parallel environments, with a reward of plus one per correctly placed part and minus one per part knocked out, then added 38 points in simulation and 39 on the robot. On one layout of 45 parts, the policy went from sorting 6 parts to 23. On the bottle and mug tasks, policies trained only on simulated demonstrations from a privileged code policy transferred zero-shot to the real robot. The only real data used was a single teleop trajectory to calibrate the scene. Cofounders Fan-Yun Sun and Sharon Lee, both Stanford AI PhDs, are putting these calibrated scenes into an Environment Hub so evaluation, data generation, and reinforcement learning run against the same physics. Moonlake reports that, on these tasks, iterating in simulation took a fraction of the real-only process's resources to reach similar reliability.