Thanks for writing this up. Getting provenance to survive federation is worth doing. Some C2PA-specific notes on the discussion so far, then a review of the draft itself.

On the thread

• @mczachurski on verifiability. Agreed. It’s worth adding that a server-signed derivative doesn’t fully close the gap either. A parentOf ingredient binds to the parent’s manifest (hashes of its activeManifest and claimSignature), not to the parent’s pixels. A signed derivative is therefore the server attesting “this file came from that one” under its own certificate. What it adds over a plain summary is accountability: the claim is signed and attributable to a named signer, and the parent’s manifest and signature can still be checked independently. Receivers should present it as “signed by server name”, not as the camera’s credential.
• A middle ground for servers that won’t sign. A server could publish the original’s manifest store as a sidecar .c2pa file, advertised with Link: <…>; rel="c2pa-manifest" (§15.5.3.2). Receivers could then check the signature, signer and editing history themselves. They couldn’t confirm the manifest binds to the displayed pixels, so the asset won’t be a Valid asset, but they would no longer depend on the server’s validationState. The catch: the sidecar carries the same personal metadata, so the redaction rules still apply, and redacting means re-signing.
• @dansup on “no different from stripping”. I’d push back a little. Stripping produces an absence; a server-reported Trusted produces a positive claim. It may be worth requiring receivers to show “reported by origin” differently from “verified locally”.
• FenTiger on geotags. The exif:GPS* fields are among those allowed in c2pa.metadata, so offering this as an author opt-in seems reasonable. Once the location is copied into an ActivityStreams location, though, it is no longer signed, so it should be presented as author-provided.
• lukasmueller on withdrawn consent. C2PA makes this stricter, not looser. Manifests are signed and immutable. Redaction produces a new manifest on a new copy; it doesn’t reach copies that have already federated, or any manifest repository the creator registered with for soft-binding lookup. Consent has to be settled before the first federation. An Update that swaps in a redacted derivative is best-effort only, and I agree with @mczachurski that the FEP should say so.
• @mczachurski on author intent. Agreed. Since redaction in the FEP is a SHOULD, it could add an explicit exception: “unless the author has chosen to publish the credentials as exported”, with a warning about any personal metadata that would be retained.

On the draft

I read it against the C2PA 2.4 specification it cites, and against the current 2.5 working draft. The section references (§6.8, §14.3) are correct, but a few things don’t line up with C2PA, and one rule produces the wrong label. Section numbers below refer to C2PA 2.4.

digitalSourceType

1. Rule 1 labels composites as fully AI-generated. Rule 1 looks for c2pa.created + trainedAlgorithmicMedia anywhere in the ingredient chain. Take a camera photo with an AI-generated element placed into it (c2pa.placed, componentOf ingredient). That element’s own manifest has c2pa.created + trainedAlgorithmicMedia, so Rule 1 fires and the photo is labelled AI-generated. Rule 2 would have given the correct compositeWithTrainedAlgorithmicMedia. A manifest has at most one c2pa.created/c2pa.opened (§18.15). So Rules 1 and 3 should follow only the parentOf chain from the active manifest down to its root c2pa.created. componentOf and inputTo ingredients should feed only Rule 2. Ingredients removed with c2pa.deleted should probably be skipped.

2. Rule 3 can produce values that aren’t IPTC terms. C2PA allows its own values alongside IPTC ones: http://c2pa.org/digitalsourcetype/empty and http://c2pa.org/digitalsourcetype/trainedAlgorithmicData (§18.15.4.5). empty is required for anything started from a blank canvas or template, so Rule 3 will often return it. The FEP should either allow these values or say to leave the property out.

3. Ingredients without a manifest can carry their own digitalSourceType. When an ingredient has no active manifest, the ingredient assertion may have a digitalSourceType field (§18.16.11). The rules only look at actions, so they miss these.

4. Please define “AI source type”. At minimum, list trainedAlgorithmicMedia, compositeWithTrainedAlgorithmicMedia, compositeSynthetic and C2PA’s trainedAlgorithmicData. Also say whether algorithmicMedia and algorithmicallyEnhanced count.

5. Only the AI label is protected. The FEP keeps receivers from dropping an AI label, but says nothing about how much to trust digitalCapture. “This is a real photo” is the label people have a reason to fake. I’d suggest receivers SHOULD NOT present a non-AI source type as verified unless the publisher reports Valid or Trusted.

contentCredentials

6. The validationState values don’t match C2PA. §14.3 defines three manifest states: Well-Formed, Valid and Trusted. There is no “Invalid”. Well-Formed (structurally fine, but the signature isn’t valid, for example because the certificate expired or was revoked) is missing, and it’s a common result in practice. C2PA also separates manifest state from asset state (§14.3.3): a manifest can be Valid while the file’s content hash no longer matches. I’d suggest the three C2PA terms plus an explicit “no valid manifest” value, or carrying the C2PA status codes.

7. signer should come from the certificate subject, not the issuer. The issuer is the certificate authority, so as written most attachments would show the CA’s name rather than “Example Labs” (§14.2).

8. claimGenerator source field. v2 claims use claim_generator_info, a map with name and version (§10.2.3). v1 claims use a claim_generator string. Worth saying which field to read for each.

9. The summary describes a different file from url. (This is the same gap @mczachurski describes.) contentCredentials describes the original upload, but url serves the re-encoded copy. Unless that copy is signed, a receiver can’t check the claim. Two possible fixes: include the original active manifest’s label (its URN) so receivers can match things up, and mention the sidecar/Link option described above.

Signed derivatives

10. Use c2pa.transcoded for server resizing, not c2pa.resized. c2pa.transcoded explicitly covers resolution scaling and is a non-editorial transformation. c2pa.resized isn’t classed as non-editorial in 2.4, and the 2.5 draft classes it as perceptible. An actions list containing only c2pa.published, c2pa.transcoded and c2pa.repackaged tells consumers that nothing editorial changed (§18.15), which is exactly what a federating server wants to say. c2pa.resized.proportional (non-editorial) is another option.

11. Spell out the manifest structure. The first action needs to be c2pa.opened, referencing a c2pa.ingredient.v3 assertion with relationship: parentOf (§18.16.3). It would also help to say:

• Personal data has to be removed before signing, not after.
• The JUMBF/APP11 data has to survive any later pipeline step, such as a CDN or image optimizer.
• A server certificate that isn’t on the C2PA Trust List gets Valid, not Trusted.

12. The redaction list.

• stds.exif has been deprecated since 2.0. It still shows up in older manifests, so keep it, but it isn’t the main risk.
• The parent manifest’s thumbnail (c2pa.thumbnail.claim.* / c2pa.thumbnail.ingredient.*) is the biggest leak. In the “then cropped” example, it would show the uncropped original.
• Also worth listing: custom *.metadata assertions (any namespace is allowed), identity assertions, and stds.schema-org.CreativeWork (author).
• It may help to describe how redaction works (§6.8): the assertion is removed and listed in the new claim’s redacted_assertions, and adding a c2pa.redacted action is strongly recommended. c2pa.actions can never be redacted, so the provenance history always survives.
• For the “can’t redact” fallback: partial redaction, for example keeping camera info and dropping location, is possible through an update manifest (§18.17.4).

Minor

• C2PA 2.5 is in progress, so it’s worth citing sections by anchor as well as number.
• “A publisher MUST only include…” may read more clearly as “A publisher MUST NOT include contentCredentials unless it validated the manifest itself.”

I’m happy to help with wording on any of these.