Provena Gate
The trust layer for AI training video

Every frame verified before a dollar moves

AI runs on video, and most of it is bought by the hour and never checked. Provena Gate is the checkpoint. It fingerprints every clip, searches 163 million frames in seconds, catches the copies a human would miss, even cropped or mirrored ones, and lets an AI agent settle the invoice inside rules it can’t break.

Delivered clip, shrunk and compresseddelivered
Original frame in the Vript datasetpublic dataset
Match found in Vriptshrunk and compressed · seconds 0.0–3.2
155.6 hof duplicate footage hiding in one famous AI dataset, OpenVid-1M (~8% of it)
163Mvideo frames searched for every clip you receive
~2 sto find a copy, even after it’s been shrunk, compressed or trimmed
0false matches on unrelated video in our tests (13 of 13 edited copies caught)
The index

We fingerprinted the datasets AI video is built on

We mapped the 300 most-downloaded video datasets on Hugging Face, then fingerprinted five of the biggest, frame by frame, on our own C++ engine. Every delivery is checked against all of it.

300video datasets on Hugging Face surveyed (~870 TB)
1,630,518clips fingerprinted and searchable
~11,500 hof footage indexed, 163M frames
155.6 hof duplicates found inside OpenVid-1M alone

OpenVid-1M

Text-to-video generation
1,019,956 clips2,051 hourslicence: CC BY 4.0

LLaVA-Video-178K

Video understanding
165,969 clips2,510 hourslicence: none stated

Vript

Video captioning
370,854 clips2,218 hourslicence: none stated

ShareGPT4Video

Video captioning
40,303 clips305 hourslicence: CC BY-NC 4.0

VideoChat3-LV116k

Long-video chat
33,436 clips4,404 hourslicence: Apache 2.0

Fingerprints only: no video is stored. Coverage per dataset is 84–100% of its files. The OpenVid-1M audit compared all of its 1M clips against each other on rented cloud machines for about $10.

The blind spot

Bought by the hour. Checked by no one.

Every duplicate hour is money burned and a model trained on the same footage twice. Today the “check” is a person scrolling through a folder.

The same clip twice

One recording delivered again, renamed or re-encoded.

Copies of footage you own

Cropped, mirrored, recoloured or trimmed versions of clips you already bought.

Sold to someone else

“Exclusive” footage that another buyer already paid for.

From a public dataset

Clips lifted from free AI training sets and sold as new.

Why now

Where training data comes from is about to matter a lot

The lawsuits are here

Creators are suing AI companies over scraped video. Buyers need to know exactly what they bought and where it came from.

Regulators want receipts

The EU AI Act’s training-data rules apply from August 2026. “We didn’t check” stops being an answer.

The buying is exploding

Robotics and world-model labs are starting to buy real-world video at a scale no human team can review by hand.

Edits don’t hide copies

Crop it. Flip it. Stamp a logo on it. It still gets caught.

The oldest trick in reselling footage is to change it just enough to slip past a simple check. Provena Gate reads what’s actually in the frame, so the trick stops working.

Original frame
Original
Cropped copy
Cropped matched
Mirrored copy
Mirrored matched
Copy with a logo and captionEXCLUSIVE: aerial rice terraces 4K
Logo + caption matched

Exclusive finally means exclusive

Buyers pay a premium for footage nobody else has. Now a recut, reframed copy sold to someone else shows up too.

Gaming it stops paying

When a crop or a flip no longer hides a resold clip, selling the same footage twice stops being worth the risk.

Two levels of proof

An exact match is a confirmed copy. A match that survives the edits is a likely copy, laid out frame by frame for a person to confirm.

How it works

From delivery to settled invoice in under a minute

1

Fingerprint

Two fingerprints per clip: one exact, one that survives edits. The video itself is never kept.

2

Match

Compared with your library, other buyers who opted in, and 1.6 million public clips.

3

Fixed rules

Plain rules decide what may be held and what a person must look at.

4

Claude reviews

Decides inside those rules, explains why, and asks a person when unsure.

A real run

Watch it catch what a human would miss

A replay of a real run of the app. The matches, frames, money and Claude’s words are exactly what it produced. The vendor and names are made up, and most clips are test footage; one is a real public-dataset video we shrank and compressed.

Loading the run…
Guardrails

Autonomous where it’s safe. Human where it matters.

It makes real decisions, explains every one, and stays inside rules written in code that it can’t touch.

Unclear cases go to a person

Weak matches, still footage and likely edited copies are never held automatically, so honest contributors don’t lose pay.

Only a person releases payment

Claude can hold or escalate. Paying out stays a human decision.

Other buyers stay private

Buyers compare fingerprints only if they opt in, and never see each other’s footage.

Every step is logged

What the rules decided, what Claude looked at and chose, what a person changed.

Works without the AI

If Claude is unavailable, the rules’ decision stands and disputes go to a person.

Money is done in code

Claude never adds up amounts. Each second counts once, and overbilling is held.

The trust layer for AI training video

Every clip fingerprinted. Every match proven. Every dollar accounted for. Built at Cal Hacks 13.0.

View the code