"Verification" is easy to misread. Here is exactly what it is, and what it is not.
What it is
For multi-model questions, a reviewer model reads two things together: the draft synthesis and the individual responses. The reviewer is the plan's verifier, or the judge when the panel disagreed and a judge was planned. It returns a short list of checks. Each check has:
- a kind:
reasoning(a step that does not hold),contradiction(two statements that cannot both be true, within a response or between the draft and the responses) orunsupported(a claim some models made that the others did not support); - a finding, in a sentence;
- the models concerned;
- an outcome: held, revised, removed, or kept as dissent;
- how the synthesis resolved it.
What it is not
The verifier does not fetch external sources, does not open the links models cite, and does not look anything up. It weighs the evidence and reasoning that already exist inside the responses. That makes it a consistency and reasoning check, not a fact check. Keplar's pages say "weighs the evidence and reasoning behind each response" for exactly this reason.
What you see
The Verification section lists the checks with their outcomes. For single-model answers it says no cross-model verification was needed. If the review failed or its output could not be read, the section says that. It is never presented as "all clear" by default.
How findings change the answer
If the reviewer marks a claim as removed, the claim is taken out of the evidence the synthesis uses and the claim's support is recomputed. A claim kept as dissent stays in the answer, labeled as a minority view. A revised claim is rewritten before the final answer.
Disputed claims
The reviewer also returns a short list of disputed claims: claims from minority positions or claims left unsupported. In a design that includes optional web checking, these are the only claims that would be looked up. Web checking is off today, so nothing is looked up and the product does not claim to check claims against sources.
Cost and speed
The review is one model call. It is metered like every other call, and it shares the run's time budget. When time is short, the stage says so instead of being silently skipped.
What this is good for
- Catching a step in a chain of reasoning that does not follow.
- Catching a draft that quietly contradicts one of the models.
- Catching a number or name that only one model asserted.
What it is not good for
Facts that every model gets wrong, recent events, and anything that needs a live source. For those, follow the checklist in Sources and citations.
Related
- How the final answer is written: The draft, review and final synthesis steps: weighing evidence over headcount, keeping minority views honest, and what happens if synthesis fails.
- Sources, citations and web lookups: What the Sources section really contains, why Keplar does not open or check the links, the status of web lookups, and a checklist for verifying what matters.
- Read a Keplar answer: A tour of the six sections under every multi-model answer, what each one means, and what it does not mean.