About
Every generated claim is bound to a located quote, exposed to Claude as an MCP service.
Most tools that summarise a paper ask you to trust the summary. The model produces confident prose, the prose sounds like the paper, and the only way to know whether it is right is to go and read the paper yourself, which is the work you were trying to avoid.
Pepiros takes the opposite position. Every claim it surfaces is matched against the exact sentence it cites, with a score you can see. Claims that clear the threshold show their quote, their page, and their citation id. Claims that do not are labelled inference and get no citation at all.
A matched quote tells you the sentence is real. It doesn't tell you the conclusion built on it is right. A model can quote the Methods section correctly and still draw the wrong conclusion from it.
That's why the badge stops at quote located and never says verified. Claim and quote sit side by side so you can judge whether the claim actually follows, not so Pepiros can tell you it already checked that part for you.
An entailment overlap floor catches the sharper failure, a real quote attached to a reversed or overstated conclusion. It's a floor, not a guarantee; the mechanics are on how it works.
The same grounding is callable over MCP, so an agent can check its own claims against a source before asserting them to you, and say so when one comes back unsupported. Documented on the agents page.
Pepiros is built by Anay Dhawan and Yash Kewlani, under the StudentSuite org, under active development. The roadmap says what is real today and what is not, without rounding up.