Your data stays in your bucket.
Refine runs locally or in an AWS workspace. Checkpoint intake stores a reference to the model. Customer self-serve onboarding and publication of contract-approved anonymised aggregates are planned.
One test, the same for everyone.
- We sell no footage and no models. Cupel has nothing to gain from any vendor's data or any lab's policy scoring well.
- No controlling stake is held by a robot lab, a data vendor or a model company.
- Vendors can't buy a better score. Buyers pay for audits. Vendors pay only a flat, published fee to be listed.
- No data moves between customers. No code path sells, exports or shares one customer's footage, checkpoints, world models or fleet data with another.
What stays, and what leaves.
Your cloud account
- Raw and blurred footage
- Refined sets, labels and exports
- Checkpoints
- World models fitted to your robots
- Fleet recordings and episodes
Cupel's shared ledger
- Aggregated scores with score versions
- Pooled A/B success counts and paired trial counts
- Aggregated prediction accuracy
- Fleet summaries
Aggregates reach the ledger only under the outcome-sharing clause in your contract. It gives Cupel the right to anonymised scores and outcomes, never the underlying data. Contracts with the clause cost less. Each customer's footage sits in a bucket of its own. Upload and download happen in the dashboard, so customers need no AWS account. A customer can connect a bucket in their own account instead.
Workspace admins can download pooled Refine Gate A outcomes for private review when the contract permits it. Publication is a separate release. Export workflows for the other categories are planned.
Consent records and mandatory blurring
- A consent record on every clip. No clip enters the pipeline without a consent record ID, and the ID stays with the clip in every manifest.
- Face and screen detection runs on every frame. Refine blurs the detected regions before later stages, previews and exports read the clip. Blurring is mandatory.
- Captions stay private. An open model writes captions on the machine running Refine or in its AWS workspace.
- Known gap: screen detection misses some monitors and tablets. We will measure how many it misses on real footage before any pilot.
Every model and dataset, on the record.
We record the licence of every model and dataset before it enters the pipeline. We don't use EgoDex (non-commercial) or Ego4D (commercial use unclear), or any hand model that depends on MANO.
| Model or dataset | Used for | Licence |
|---|---|---|
| V-JEPA 2 ViT-L | Embedding and scoring | MIT (model card, checked Nov 2026) |
| Qwen3-VL-2B-Instruct | Captions, run on our own GPUs | Apache 2.0 |
| MediaPipe | Hands, hand pose, screen detection | Apache 2.0 |
| YuNet | Face detection for blurring | MIT |
| Build AI Egocentric-100K | Open footage for the first proof | Apache 2.0 |
| CaptainCook4D | Open kitchen footage for testing the quality filter | Apache 2.0 |
On-device safety and rollout rules
- On-device watchdog. The development runtime holds or stops on late actions or a lost link. This behaviour has been exercised with a simulated driver. Physical verification is planned.
- Test records and overrides. The development control plane requires a Test record by default and records overrides in the audit log.