What draft gives you
Most tools bolt provenance on afterwards, if at all. draft writes it with every article, as part of the run that produced the text:
- A C2PA manifest. Every article set ships a manifest built on the C2PA Content Credentials standard — the interoperable, cross-industry format for machine-readable content provenance.
- Per-sentence attribution. A separate file maps each sentence in the article to the claim, and the verbatim source quote, that backs it.
- A verification command.
draft --verify <final.md>recomputes the manifest digests and reports whether the article is still exactly what the ledger was verified against. Anyone with the files can run it.
Nothing here is a paid add-on or a separate step. It is what a run outputs.
Why this matters now
The regulatory direction of travel for AI-assisted content is clear and converging: disclosure and machine-readable provenance. The EU AI Act introduces transparency obligations for AI-generated and AI-manipulated content, and several US states have passed or proposed AI-disclosure laws. The common thread across them is that AI involvement should be detectable and declared, and C2PA Content Credentials are the standard the industry has lined up behind to do exactly that.
A tool that produces AI-assisted text with no provenance leaves that entire burden on you. draft produces it with the provenance layer already attached and checkable.
How draft maps to the requirements
| The direction rules are moving | What draft ships |
|---|---|
| AI involvement must be disclosed | A C2PA manifest declaring it, per article |
| Provenance should be machine-readable | C2PA Content Credentials, not prose |
| Claims should be traceable to sources | Per-sentence attribution to verbatim quotes |
| Provenance should be verifiable, not asserted | draft --verify recomputes and checks the digests |
Download a real, verifiable sample
Do not take our word for it. Here is a complete article set draft produced
from a rights-cleared test paper, published exactly as written:
- Download the full set (zip) — source PDF, article, attribution, C2PA manifest and claim ledger.
- View in place: the generated article, its C2PA manifest, the per-sentence attribution, and the source paper.
Install draft and, from inside the unzipped folder, run:
It recomputes the digests and confirms the article is unchanged since it was written, the claim ledger matches, and the source is unchanged since it was read. This set is honest about its limits: of 36 sentences, 13 rest on a verified claim and the rest are connective prose, and the gate dropped 9 of the 15 candidate claims. That is what grounded-by-construction looks like on a real run, not a polished demo.
Honest scope
draft provides the provenance layer: it makes AI-assisted articles carry checkable, standards-based credentials by default. It does not, by itself, make you compliant with any specific law — compliance depends on your full publishing workflow, your role, and your jurisdiction, and this page is not legal advice. Confirm your own obligations. What draft removes is the hardest part to retrofit: producing the provenance in the first place, bound to the text, at the moment the text is written.
See how grounding works for the verification gate, and security for the local-first, keyless posture.