DEV Community

Verity Research
Verity Research

Posted on AI-assisted

Checking AI answers claim by claim, without another LLM

If you ship anything built on an LLM, you have probably written a second prompt that asks a model whether the first answer was right. That check is made of the same stuff as the thing it checks.

Verity takes a different route. A separate, non-generative and fully auditable model reads each answer, splits it into claims, and marks every claim supported, contradicted or unsupported against your own documents, with the passage that decides it.

That changes the budget too. The checking no longer runs on a premium model, and the model that writes your answers doesn't have to be the priciest one, because a claim that fails goes back with its source and is rewritten before anyone reads it.

You can call it from an API or connect it over MCP to the agents you already run.

Free to start: https://veritylabs.io/signup?r=kk

Top comments (0)