Inspect complete sequences
A polished highlight can hide missing starts, interrupted tasks, or long gaps. Review complete sequences that reflect the expected delivery. Note whether the beginning, transitions, and final outcome provide enough context for your intended model task.
Check what the camera actually sees
Look for hands leaving the frame, objects hidden at critical moments, and motion that obscures an action. These are not universally disqualifying; their importance depends on whether the task is recognition, prediction, or manipulation learning.
Read the metadata alongside the video
Confirm that timestamps, activity names, participant identifiers, and file references mean the same thing across the sample. If labels describe time spans, establish whether endpoints are inclusive and which clock is used. Record unanswered questions before scaling a purchase.
Make the project boundaries explicit
Discuss commercial training, evaluation, redistribution, retention, and any exclusions with the data provider. Ask what evidence supports the proposed permissions. A file being accessible is a technical fact, not a statement about what a buyer is allowed to do with it.