These layers answer different questions. A practical export pipeline can use all three; none makes the content itself true or lawful.
| Layer | What a recipient gets | Important boundary |
|---|---|---|
| Visible disclosure | A readable indication that AI created or altered the media | Placement, wording and accessibility depend on your use and applicable duties |
| Invisible watermark | A machine-readable recoverable identifier in supported pixels/frames | Recovery is bounded by content, transforms, scan budget and deliberate removal |
| Signed C2PA manifest | A signed statement bound to the published asset, with declared source/actions | Signature validity, asset binding and trust in the signer are distinct checks |
| Recovery registry | A route from a supported recovered ID to a stored signed manifest | A copied ID alone does not prove the new asset matches the original |
| Signed verifier report | A checkable record of a verifier's observations about exact bytes | Pin the report public key and asset hash; the report is not a legal certificate |
The CAI durable-credentials documentation describes combining provenance manifests with watermarks or fingerprints to recover information lost when metadata is removed. Ariadne's current kit uses its supported watermark/registry route. It does not search every vendor's marks.
Publish the final export and a reachable verifier; add appropriate visible disclosure in the app/content where required. Preserve the structured result when verification is incomplete or a recovered claim is untrusted.
Disclosure component · Verification results · EU requirements · California requirements