Run one evaluation, end to end, and get the report.

You give us one artifact — a robot policy, a set of world-model clips, or both. We agree the criterion and the thresholds with you before anything runs, and we freeze them. Then the pipeline runs the same way it ran on the published results, and the report comes back with a verdict, the per-clip evidence, the calibration state, the limitations and the hashes.

The point of a pilot is not the green or red in the report. It is that you end up with a criterion you agreed to, a number you can check, and a limitation list you did not have to ask for.

A pilot runs two weeks. That covers agreeing the criterion, the run itself, and the report.

The run is compute-bound rather than engineering-bound, and it is the same pipeline the published results came from, so there is no bespoke work to schedule around.

The report, in the shape the published runs are published in. It carries:

• the verdict, and the numbers behind it, each with its interval
• the per-clip evidence — what was tracked, what was measured, what was compared against which threshold
• the overlay videos: the footage with the measurement drawn on the pixels that were judged
• the calibration state — what the thresholds were fitted on, and whether the calibration and the test set were disjoint
• the limitations, including the ones that make the result weaker than it looks
• the hashes: the digest of every evidence file the report rests on

The harness needs the artifact and the thing to judge it against. In practice that means:

• the artifact — the policy, the model, or the clips to score
• its family and the task it is for, so the right geometry and the right detectors apply
• the success criteria, or a starting point for them: what counts as a violation, in your terms
• the footage, and whether it may be published — the published runs use public datasets, and a private run is reported the same way but with the evidence withheld
• the timeline you need the answer by

The standard scopes are priced on a published ladder, which the deck sets out: a policy-stress run at four tasks and four tiers, a focused run, a continuous re-run, a world-model physics pass, and a fixed-scope audit. A pilot outside those scopes is scoped on a call.

The price of a particular pilot depends on the artifact, the scope and how long the evaluation runs, so we agree that on a conversation rather than quoting a number that would not apply to your case.

A pilot is not a consulting engagement, and it is not a rewrite of your system. It is one frozen criterion, one run and one report.

• it does not fix the artifact — it tells you what the artifact does, and where it fails
• it does not certify anything: a verdict against an agreed criterion is not a certification, and the report says so
• it does not cover real-robot execution — the policy stress runs in simulation, and the world-model gate runs on video
• it does not promise a passing result. Two of the three published runs failed, and they are published with the same prominence as the rest

This is a real report, generated by the harness from the real public ALOHA run: the verdict, the per-track evidence, the overlay frames, the calibration state, the limitations and the hashes. The footage is LeRobot's public dataset, so nothing in it is sensitive and nothing is redacted.

Open the sample report (PDF) →