Verdicts

Every claim check ends in one of five verdicts. Each is a statement about your evidence, never about the world.

| Verdict | Meaning | | --- | --- | | SUPPORTED | Every part of the claim is established by the evidence. | | PARTIALLY_SUPPORTED | Some part is established and another part is not. | | INSUFFICIENT_EVIDENCE | The evidence does not address the claim. It may share the topic or the entity and still not answer the question. | | CONTRADICTED | The evidence makes the claim false, whether stated outright or derived from what it says. | | DISPUTED | Evidence items conflict about the claim, and nothing in the evidence establishes which one governs. |

A sixth value, INDETERMINATE, means Attest could not measure. It is returned with HTTP 503, never with 200, so a failure to measure can never be read as a finding about the claim.

PARTIALLY_SUPPORTED is the one to watch

A compound sentence usually gets this verdict, and it is the most common honest answer. "Revenue rose to CHF 94m and operating margin improved", checked against evidence that covers only revenue, is PARTIALLY_SUPPORTED, with supportingSpans showing the half that held and explanation naming the half that did not.

If you generate text from retrieved documents, this is usually the verdict that identifies the sentence to fix: drop the unverified part, qualify it, or retrieve evidence for it.

Split compound sentences

Send one claim per assertion. A compound claim gets a single verdict for the whole sentence, so checking it whole reports that something is wrong without saying which part. Checked as two claims:

The Meridian cloud migration completed in November 2024.        CONTRADICTED
The Meridian migration produced no client-impacting incidents.  SUPPORTED

Checked as one sentence, that is a single CONTRADICTED, and a reader acting on it may delete the half that is true. The draft review endpoint splits sentences for you.

CONTRADICTED and DISPUTED are different claims about the world

CONTRADICTED says your evidence refutes the claim. DISPUTED says your evidence disagrees with itself and Attest will not pick a side.

Attest resolves a conflict where the evidence resolves it. Final minutes supersede a draft, an amendment with an effective date supersedes the clause it replaces, and a primary decision that explains another figure settles it. Only when nothing establishes precedence do you get DISPUTED. Two executed counterparts of a contract naming different governing laws, neither marked as superseding, is DISPUTED.

INSUFFICIENT_EVIDENCE cannot tell you why

It means the documents you sent do not speak to the claim. It cannot distinguish "no such evidence exists" from "your retrieval missed it". Treat it as a prompt to look for evidence, not as a judgement that the claim is false.

There is no confidence score to act on

An earlier version returned a probability. Measured over real checks nearly all of them landed between 0.95 and 0.99, so the number separated nothing while reading to a caller as "95% true", a claim Attest cannot make. The verdict plus the named unverified assertions carries everything the score was meant to. A groundingConfidence field may still appear in responses for compatibility. Do not branch on it.

How verdicts are decided

A single grounding call reads the claim and the evidence under reading rules chosen for that request (a unit rule for a claim with figures, a precedence rule for several sources, and so on). Deterministic code turns the measurements into the verdict. The explanation is produced separately and does not decide the verdict, which is why the audit can report explanationDisagrees. See Explanation and audit.

Measured on 190 hand-labelled examples, verdict accuracy is 95.9% and the overclaim rate, how often Attest asserts more grounding than the evidence supports, is 1.6%. Run-to-run variation is about 3.7 points, because hosted inference is not bit-reproducible, so treat small differences as noise.