This page is a demonstration chain. Each section is hashed with SHA-256 over the previous section's hash, the signer label and its text, the same idea the app uses for custody. It is not the app's statement format, and the signers are labels, not keys.
The same picture from several sources. One name spelled four ways. A file nobody can place.
Sealbook is evidence and documentation management where AI reads everything first and proposes what it finds, each claim pinned to the exact words, page or second it came from. People make every factual decision. Every step can be checked later.
A file arrives
Pick one of the synthetic case files. Sealbook records its fingerprints and how it was acquired, checks its content credentials and reads its text. Then the AI proposes claims, each one pinned to a span, a region of a page or a stretch of audio. Nothing it proposes counts as fact until a person accepts it.
Accept a claim below. It turns solid, a seal is stamped, and this page gains a new link right after this one.
Search by meaning
Ask in Arabic, in dialect or in English. Exact identifiers rank first. Matches by meaning are labelled as leads, not answers, and each result opens at the exact place it came from: a character span, a page, or a time in the audio.
One local model places text, pictures, scanned pages and audio in the same space, so a written question can find an image or a recording. It has been measured on the synthetic retrieval set only, where scanned pages rank weakest.
Same footage, different sources
Re-encoded and resized copies of a photo are found by picture fingerprint. A short audio excerpt is found by sound landmarks, with the time it was cut from recovered. Cropping is different: a cropped copy moves past the fingerprint threshold and is not caught, and the demo shows that too. A distance is not a probability, and nothing is linked until a person confirms it.
One street, two years apart
Two real, openly licensed clips of the Jdeideh quarter in Aleppo: an explosion over its rooftops in April 2015, and the same quarter in ruins in September 2017. Sealbook samples a frame every second, fingerprints each one, and matches re-uploaded copies frame by frame and by sound, recovering where in the original each copy starts.
Re-encoded, trimmed and logo-stamped copies match. Cropped and mirrored copies do not match by picture; their unchanged soundtrack still does. A written question, in English or Arabic, finds the moment in the footage. Every match here is a proposal for a person to check.
One name, two people
The case holds two men with the same name, the same father's name and different mothers and birth years. Sealbook proposes identity candidates across Arabic and Latin spellings, and shows what supports each one. Similar names are never merged on their own.
Timeline
A person's detention and transfers, in order. Where two sources place him in two places on one day, the conflict is shown, not resolved for you. Where no evidence covers a stretch of time, the gap is shown as a gap. Every point opens the span that supports it.
Review together
Ask a colleague to check a claim. The discussion stays with the people in it and does not appear in general views, history or exports. The decision is signed, and it can be undone with a signed undo that restores what came before.
Proof
Statements are grouped into Merkle batches. One root stands for the whole batch, and any change to any statement changes the root. The app can request timestamp receipts for a root and copy it to partner archives, and an exported package can be verified offline with the bundled verifier.
Those receipt and partner paths are implemented, but this is not a configured or certified deployment. Erasure removes a person's values from views and search while the chain still verifies. It excludes access; it does not yet physically delete stored bytes.
What we learned
- One source is not two. Several uploads of one recording are one origin. Corroboration counts only after a person confirms they are the same recording, and undoing that link changes the tiers back.
- Similar names are not one person. Embeddings and fuzzy matches propose. They never merge identities.
- Erasure must reach every copy. Search entries, prepared media and anything that shares bytes with the erased source all have to be blocked, or the value leaks back through a side door.
- A model's HTTP 200 does not prove it read the image. One endpoint accepted media and silently ignored it. Distinct test images catch that; a success code does not.
- Search must cite the original text. Arabic is normalized to find matches, but the citation points back to the exact characters in the stored evidence.
- Private means private everywhere. Hiding a discussion in the inbox was not enough. It also had to be kept out of history, reveals and exports.
- A loose threshold can make a broken matcher look right. The audio test asserts the known time offset of an excerpt, not just that something matched.
Where it stands
Everything on this page runs on a synthetic case built for testing. The numbers come from real runs over those files, not from estimates. What works today, what was measured, and what comes next are listed below.
All people and organisations in the case are fictional. Place names are public geography.