Clip acceptance criteria
Version 2026-08-10
When you fund a bounty you pay in full for clips that do not exist yet. What follows is exactly what your money buys: the rules a clip must meet to be accepted and charged to you. Every clip is graded by the same code against these numbers, automatically, at the moment it is submitted — and these are the same values that code reads, not a description of them.
The criteria
Clip length
At least the minimum length set on the bounty. Clips are scored higher as they approach the target length; anything short of the minimum is rejected.
A clip that ends before the task does shows an incomplete action, which is the one thing a demonstration cannot be missing.
Resolution
Shortest side at least 720 pixels. Full marks at 1080p or above.
Below 720p, small objects and hand contact points stop being legible at the scales a manipulation policy needs.
Camera and interaction movement
Measured from the device accelerometer. Observation tasks must show sustained but controlled movement (0.15–4 m/s² of dynamics away from gravity) — neither a static shot nor a shaken one. Tasks involving physical interaction must show at least 0.1 m/s².
Separates a real recording of someone doing something from a phone propped up pointing at a room, without needing to inspect the pixels.
Framing steadiness
Scored from how much the number of detected objects varies across the clip; must reach 40%.
A scene whose contents keep changing is usually a camera swinging around rather than a task being performed.
How a bounty settles
Overall score. Each criterion is scored from 0 to 1 and combined by the weights above. A clip must reach 70% overall, or the higher bar you set on the bounty, whichever is greater. Failing any criterion marked blocking rejects the clip regardless of its overall score.
Distinct homes. When a bounty requires clips from more than one home, a single capturer can fill only one slot. Node Data treats one account as one home; we do not verify addresses, and you should read the number as a floor on diversity rather than a certified count.
You are charged only for accepted clips. The full bounty is charged up front and held by Node Data. Accepted clips draw the amount down one at a time. Rejected clips cost you nothing and are never delivered.
Unfulfilled clips are refunded. When a bounty closes — because it filled, you closed it, or it hit its deadline — everything not spent on an accepted clip is refunded to the card that funded it.
Grading is automatic and happens once. Every clip is graded by the same code against the numbers above, at the moment it is submitted. No human reviews a clip to decide whether it is accepted, and we do not re-grade a clip after the fact.
What we do not check
Publishing only the checks that pass would let you infer guarantees we have not made. These are the limits of what an accepted clip tells you.
No people in frame is enforced on the device, not by the grader. Node GO rejects a recording on-device when it detects a person in frame, and clips are captured without audio — no microphone permission is requested on either platform. Both are properties of the capture app rather than checks the server repeats on the uploaded file. Treat them as strong defaults, not as a warranty that no person ever appears.
No semantic check that the task was done correctly. The grader verifies that the named object is present and that something physical happened. It does not judge whether the chore was performed well, completely, or the way you would want a robot to perform it.
No location data. Location is never collected — it is discarded server-side even if a client sends it. Clips cannot be filtered or grouped by geography, and no clip carries a home's location.
Fixed third-person viewpoint. Clips are recorded from a phone the capturer positions themselves. There is no egocentric, wrist-mounted, multi-view or depth capture, and no camera calibration or pose data accompanies a clip.
No action labels or robot proprioception. A clip is human video with quality metadata, motion samples and a detection trace. It is not a teleoperated demonstration and carries no joint states, gripper signals or per-frame action annotations.
If you disagree with an acceptance
You have 14 days from first downloading a clip to dispute it against one of the criteria above. Where our own quality report records that check as failed, the dispute is upheld and refunded immediately. Otherwise a person reviews it. See the buyer terms for the full mechanics.