Reviewing a whole draft with Attest

A draft review is the main way to use Attest. You upload the documents you want to be judged against, you upload the draft, and Attest finds the factual assertions in it and checks each one for you. Coverage then does not depend on which sentences you recognised as claims, which matters because the claims a writer omits are usually the ones they did not perceive as claims at all.

Running one

1. Upload every source document as one batch. Absence can only be judged against what you loaded, so a document left out makes every "no", "never" and "only" claim unanswerable.
2. Write the finished draft to a file and upload it. Retyping it into a request is a generation step that can drift, and the review would then describe text nobody will read.
3. Start the review with the draft and the batch. Over MCP that is attest_draft; over HTTP it is POST /v1/attest/draft.
4. Read the cited sentences before changing anything, and clear the batch when you are done.

While it runs

A short draft finishes inside the request. A long one is hundreds of model calls, so Attest hands back a job id instead of holding the connection open, and the work continues whether or not you wait.

Progress moves while the review runs. The count of claims found grows as each passage is read, with the passages read and their total beside it, and the answer carries the time the counts last changed, so a long passage looks different from a job that has died. You can poll for that, or read the event stream the review gives you and let each change arrive as it happens. On the API dashboard, the Usage page shows every review your workspace has in flight with those counts moving.

The claim limit, and continuing past it

A review checks at most 200 claims by default, and at most 500 if you ask for more. Attest reads passages in order and stops reading once the ones it has read already fill the limit, so nothing is extracted only to be thrown away.

When a draft holds more claims than that, the result says so, lists the passages it left unchecked separately from any it could not read, and gives you a cursor. Send the same draft and batch again with that cursor and the review carries on from where it stopped, checking nothing twice. Repeat until the cursor comes back empty. Each pass is charged for the claims it checks, so a long draft costs the same in several reviews as it would have in one.

What comes back

One result for every claim, in the order the claims appear in your draft, each carrying the sentence it came from so you can find it in your text. Contradictions come first, then partial support, then what the documents could not settle, so a reader who stops after the first section has still read the findings.

Each result also says whether your documents could speak to the claim at all. A claim marked out of scope usually means the right document was not loaded rather than that anything is wrong with the draft.

Two things save you money without changing an answer. The same claim written twice in one draft is checked once and reported at both places. And a claim Attest has already measured, in the same wording against the same uploaded batch, is answered from that earlier measurement at no cost. Those measurements are held in memory for a few hours, they belong to the batch they were made against, and nothing of yours is ever written to disk.

Reading a verdict

• Supported means every part of the claim is stated in your documents or follows from them.
• Partially supported means some part holds and another part is unverified.
• Insufficient evidence means your documents do not address the claim.
• Contradicted means your documents make the claim false. A contradiction always names the sentence it rests on, so you can read it and judge for yourself.
• Disputed means your documents conflict about the claim and do not establish which of them governs.

Supported means supported by what you sent, never true in the world. Attest never searches for evidence, so it cannot tell the difference between evidence that does not exist and evidence you did not upload. There is no confidence score anywhere in a response, and a result marked as not measured is not a finding: change nothing on it.

Passages that could not be read

Occasionally a passage defeats the extractor and is reported rather than skipped silently, so a stretch nothing looked at is never presented as a stretch with nothing wrong. The one text shape this happens to reliably is a very long line dense with links and punctuation, such as a reference list or a row of URLs. Breaking it into one entry per line before you upload avoids it.

Cost and privacy

What a review costs is in the article "Enterprise API credits and pricing". Attest never keeps the claims or documents you send, and the dashboard records counts and routes, never your content.