The c2pa.ai-disclosure assertion: what fields it asks you to declare

The C2PA implementation guide, updated 11 August 2026 on the C2PA site, describes a c2pa.ai-disclosure assertion. By the guide's own summary it covers model provenance, scientific domain and the degree of human oversight. If you plan to sign Content Credentials for AI video, those are the three things to have written down for each batch.

This post is a preparation list, not a how-to for any signer. The Sume docs do not say that Sume output carries a manifest, so the facts below are what you can gather on your side from jobs you ran.

The three declarations

Model provenance is which model made the media. Scientific domain is the field the model operates in. Human oversight is how much a person directed or reviewed the output. C2PA's page names these; it does not give a scoring scale on the page I read.

Where Sume data fits

A Sume job gives you a job id, the model id you asked for, and usage.cost. For provenance, avoid sume/auto: the docs say it never discloses which model ran, so a record would say only sume/auto. Use a catalog id from GET /v1/videos/models when the record has to name the model.

The human-oversight field is your own process. Nothing in the Sume docs says what level of oversight a job had, so log it where you approve the clip.

Cheap habit

Add three columns to your job log: model id, approver, approval time. That is enough to answer the three fields later. See Sume pricing for the billing rule on the same jobs.