Check a real AI record yourself.
Below is a real, sealed record from Raptor's production log. Its content stays private — but your browser can prove it's genuine and untampered, using only Raptor's published key. Nothing is sent anywhere.
The record
Loading…
—Show the raw record (fingerprint, proof, signature)
—
Verify your own receipt
It's checked entirely in your browser against Raptor's published key. Nothing is uploaded.
Checked in your browser against Raptor's published production key — verify the key yourself. Your pasted content never leaves this page.
Now try to break it
Change the record and re-verify. A real proof catches any tampering — watch which check turns red.
Honest note: this is a real record from Raptor's production log, sealed by Raptor's
real production key (id f31d3b60…, published at
/.well-known/raptor/trust),
and the checks run Raptor's real Go verification code compiled to WebAssembly. The record's content is private —
only its fingerprint and proof are public, which is exactly enough to prove it's sealed and signed without revealing what it was.
That's why you see two checks here, not three; paste your own receipt (which includes your content)
to check all three.
How this works
Every record Raptor governs is hashed into a Merkle tree; the tree's root is signed with an
Ed25519 key whose public half is published. For a record whose content is private you verify the two
things public data proves: (1) fold the published fingerprint up the tree's inclusion proof to the root, and (2) check
Raptor's published production key signed that root. Both run in your browser — inside
raptor-verify.wasm, which is Raptor's own verification packages
(internal/verified, internal/canon, internal/crypto) compiled to WebAssembly. The
glue is verify.js and the verifier compiles from cmd/verify-wasm in the
Raptor repo — both open for you to read. Verify our verifier.
See the model mark its own grounding.
The check above proves a sealed record is genuine. This shows a different thing: give Raptor a prompt with facts in it, and as the model writes it marks each claim — green where it's drawing on what you gave it, grey where it's reasoning on its own. Raptor records this labeling on every governed answer, so grounded fact and inference are separated on the record, where anyone can audit them.
These are the model's own labels — Raptor's job is to make them explicit, consistent, and tamper-evident, not to grade the model. Read each one and judge it yourself; a label the model got wrong is one you can catch.