Sealed plans and a manifest you can check

What makes AI video verifiable?

AI video is verifiable when anyone can check which plan made it, from which files and with which models. Everyframe makes that check possible with two records, a fingerprint of the plan and a job manifest.

Every plan gets an identifier made from its own contents. The identifier is a SHA-256 hash, a 64-character fingerprint that changes completely when a single byte of the plan changes. We call a plan with this fingerprint sealed.

The machine that runs the plan works out the fingerprint again before it starts, and a plan whose fingerprint does not match does not run. So a plan either runs exactly as it was sealed, or it does not run at all.

The same fingerprints cover the rest of the job. The plan records a fingerprint of every AI model's weights, the numbers the model learned in training, and the machine checks them when it loads the model. Inputs are fingerprinted as they arrive and outputs as they are uploaded.

When the job finishes, all of it goes into one record, the job manifest. It lists the plan, every input and output with its fingerprint, every model with the fingerprint of its weights, and the machine the job ran on. A command-line tool checks a finished video against its manifest with no network connection and no account.

The check never renders the video a second time. That matters for the AI models that run on a graphics card, where two runs of the same model can differ in the last few bits.

What the manifest does not prove

The manifest records how a video was made. It says nothing about whether the content is correct or lawful. Its description of the machine comes from the machine itself, so a reader has to take that part on trust.

The manifest is also a separate file, so on its own it does not meet the AI Act's duty to mark the video itself. Manifests are not signed yet. Every plan has been sealed since July 2026, and the fingerprints and the manifest arrived the same month.