Home · How trust works
Two results on every listing, never mixed
Kinetic Blocks answers two separate questions about every dataset: is the seller telling the truth, and is the data any good. Keeping them apart is the whole point.
01 · Verification
Is the card true?
Every listing starts with a dataset card: the seller's own description of what the data contains. It is mandatory, and it is the canonical claim. Every factual statement on it, frame rate, resolution, duration, task, environment, modality, and license, is checked against the actual files.
Three stamps, always with evidence
verified
The observed value matches what the card claims.
doesn't match
The files contradict the card. The evidence shows expected against observed.
can't confirm
The files cannot prove the claim either way. Shown to buyers as declared by the seller, not yet verified.
From stamps to a listing verdict
The stamps roll up into one verdict for the listing: certified, needs review, or rejected. Only certified listings carry the verification seal.
- Hard facts are checked exactly. A mismatch on frame rate or duration rejects.
- Judgment calls about task and environment route to review, never to automatic rejection. Model judgment never gets the last word.
- Provenance and licensing signals that disagree route to review.
Flags go to the seller first
When a check flags something, the evidence goes back to the seller before anything else happens. They can fix the card and resubmit for free, withdraw the listing, or contest the finding, which is the only path that escalates to a human reviewer at Kinetic Blocks. Seller files stay read-only throughout. We measure, we never modify.
02 · KB Quality grade
How good is the footage?
Separately from verification, every dataset gets an independent KB Quality grade from A+ to F. It runs on the same versioned weights for every seller, so a grade means the same thing across the whole marketplace, and it is what sets the quality tier and therefore the price band.
Four pillars, four plain questions
- Technical quality. Is the footage physically well captured? Sharpness, exposure, frame-rate stability, resolution, and compression.
- Temporal integrity. Is the recording complete and honest over time? Dropped frames, timestamp sanity, and alignment between video and any accompanying data.
- Content usefulness. Is there actually useful robot-learning content in the frames? Action density is the largest single signal.
- Completeness. Is everything the card declares actually there?
Dead time is measured, not cut
Robotics data is priced by the hour, so idle stretches are how buyers get overcharged. We map dead spans with timestamps and price them into the grade. We never cut the video: transitions and resets are training signal too, and seller files are read-only on principle. The buyer sees exactly where the dead time is and judges for themselves.
How it works
Card, check, grade
The seller writes the card
Task, environment, robot, modalities, frame rate, resolution, license. The card is the claim, and it is mandatory.
We check it against the files
Claims are verified against the actual bytes and stamped with evidence. Flags return to the seller first, with a free resubmit after a fix.
The grade sets the price band
Four quality pillars produce one grade from A+ to F. Buyers compare like for like before they pay.
See it on a real listing
Browse the marketplace and open any dataset card.