AI Training
How Consented Capture Is Delivered
What a buyer receives when Harbor records real work: the file, the task, the licence, and the consent record.
Most training data starts as a pile that already exists. Someone labels it. The vendor’s job is the label, the workforce, and the throughput. That is a real business, and it is the right business when the corpus is already on a disk.
Harbor starts one step earlier. The file does not exist until a person records it. They are doing a task they accepted. Both hands and the work stay in frame when the brief is first-person video. Speech rows name the language and the accent, and a speech-to-text row includes the transcript. Photos are tied to the task, not pulled off the public web.
Payment follows acceptance. A file that fails review is not a payout and it is not delivered. The buyer sees the hours, the clips, or the patient count on the spec before the work starts, and they receive the file with the task label, the licence, and the consent record. The person in the clip is the person who agreed to record it.
That is a worse fit when the job is a labeling floor. A large managed operation, the kind Scale AI or Appen already run, will put more people on documents you already hold. Expert writing and language feedback, the kind of bench Surge is known for, is a different craft again. Harbor does not hire a crowd to click through an existing corpus, and it does not keep a roster of prompt writers.
The work Harbor does is the scene that still has to be filmed. A buyer who needs labels on data they already own should hire the labeling operation. A buyer who needs the recording, with a person who consented and a licence that says so, is in a different purchase.
Harbor Enterprise is where that programme is scoped.