Capture, validation, delivery

How an hour of real work becomes licensed training data.

The homepage “How it works” tabs are the same loop that sits under Collect missions, expert review, and enterprise delivery. An hour is not “in the dataset” because someone pressed record. It is in the dataset when it has passed capture rules, validation, and a licence the buyer can use.

Capture

Egocentric multimodal intelligence from real operational environments. First-person video, spatial audio, and contributor context captured in the field.

Capture is egocentric: the camera looks at the work the way the person doing it looks at the work. Collect missions are typically one continuous clip, 5–30 minutes, hands and arms in frame for about 90% of the take. Faces, selfies, third-person room shots, and other people are out of spec unless a programme explicitly allows them. Children are never in spec.

Commercial field capture (a mill, a bay, a line) uses the same idea with kit and consent written for that site. Wearables, spatial audio, and extra sensors appear when a managed programme specifies them—not as a default Collect toggle.

Validation

Reviewers check quality before any hour is released. Automated checks catch duration, upload integrity, and obvious framing failures. Human review applies the rubric: is this the task, is the take continuous, is consent valid, is the environment allowed. Footage that misses the bar is rejected or sent back for rework. It does not silently enter a training dump.

Expert Network validation is rubric-based judgment on model outputs or labels, not phone video. The same rule holds: unpublished, unreviewed work does not ship.

Delivery

Licensed hours arrive digitally to teams building models that act in the world. Harbor Passport keeps identity, consent, and provenance on the asset so a lab can answer “who filmed this, and were they allowed to?” without guessing.

Your data stays yours. Datasets stay in the buyer’s infrastructure unless a programme says otherwise. Optional API access is for hours Harbor owns and can licence—not a scrape of Collect. Delivery timing follows enterprise validation, which can take weeks after capture. Contributor rewards follow that acceptance, not the upload click.

What you see on the site

The three drawings on the homepage (intersection + phone, modalities into a car check, globe + package into a model) illustrate this loop. They are not a live operations dashboard.