Questions people actually ask

No, and the distinction matters. Pepiros checks whether a generated claim actually quotes the source it cites. That is quotation provenance. It says nothing about whether the paper is right, current, or well designed.

Because a fuzzy match proves quotation, not entailment. A model can attach a genuine sentence from the Methods section to a conclusion the paper never draws and still score a perfect match. “Quote located” is the strongest thing the check can honestly support.

A claim scoring 0.92 or above against the source text is badged quote located. Between 0.75 and 0.92 it is paraphrase. Below that the anchor is dropped entirely. A claim with no checked excerpt at all is badged inference, meaning the model wrote it and nothing has been checked yet.

An extra check on top of the fuzzy match. Every number, unit, and comparator in a claim has to appear in the anchored span too. It catches the failure a text match alone misses: a real quote attached to a reversed or inflated conclusion.

Yes, over MCP. Eight tools are live today, including verify_claim, which lets an agent check its own output mid-conversation. See the MCP page.

No. An uploaded paper stays private to your workspace. The public library only ever contains open-access and CC-licensed work.

Yes. Sign in as guest / guest for a fully populated demo account. Nothing you do there is saved.

Partly. The verification spine is real and tested; several surfaces around it still run on a fixed demo workspace. The status page lists exactly what works.