Which Policy is Better? Scaling Policy Decisions with Partial Credit From Real and Sim Rollouts

You have a new policy and you need to know whether it beats the one you had – last week's checkpoint, another experiment's, the baseline you ship today. That question arrives every time a training run finishes, so the evaluation has to be cheap enough to ask often and sensitive enough to answer. On DROID-style manipulation, binary success has none of these properties – most rollouts score zero, and twenty episodes produce ±20-point interval, which covers any improvement worth arguing about. So the call falls back to watching video. We show the path from "here is the new one and the old one" to an answer with numbers behind it – two inference endpoints in, rollouts on a real Franka FR3 and in simulation out, and the rule for what counts as better fixed before the rollouts run. It is the loop Positronic Robotics already runs for Runway: checkpoints in, rollouts on real hardware and in simulation back, at whatever cadence the training sets.