Annotation

Annotation is labels, spans, boxes, preferences, and comments on captures or model outputs so a training run is not guessing. Harbor’s workspace is built so labeling, ranking, and QA sit together. The buyer does not get a spreadsheet from a random crowd and a second vendor for QA.

Work against the pinned rubric version. Updated rubrics apply to new batches, not retroactively. Partial submissions are rejected. Do not train personal models on client data. Delete local copies when instructed.

When annotation is the wrong tool

If the lab needs new first-person hours, that is Data Network / Collect / managed capture—not more labels on internet stills. If the lab needs specialists recruited and run as a cohort, that is Managed programs.