Guide

How to use Pepiros

Seven steps, in the order you’d actually do them. If you want the argument for why the checking works the way it does, how it works covers that instead.

Open the demo workspace Add your own paper

Open a workspace
A workspace holds a few papers and everything derived from them. The demo workspace is already populated with three papers on circadian rhythm and cognition, including a deliberate contradiction between two of them.
You don’t need an account to look around. Signing in is what makes your own work persist — without it, anything you add is gone when you close the tab.
Read in the doc view first
The default view is a reader, not the graph: summary at the top, the source PDF beside it, related papers on the right. Claims and their quotes sit next to each other so you can judge whether the quote actually supports the claim.
Click any citation chip — the small monospace C7-style ids — to jump to the exact sentence it came from.
Open the graph when you want structure
Explore graph switches to the canvas. Pillars start collapsed, so you see the shape of the argument before its details; click a pillar to expand its claims.
If the colours and line styles don’t mean anything to you yet, open What am I looking at? in the bottom-left. It only lists what that particular graph uses.
Zoom out and cards deliberately shed detail: at a distance you are reading structure, not sentences.
Ask a question
The chat dock answers from the papers in the workspace and cites what it used. Every citation it returns is re-checked against the source after the answer is written, so a fabricated reference is reported as unsupported rather than passed off as evidence.
If the papers genuinely don’t cover your question, it says so instead of guessing. You can still ask it to answer without sources, and that answer is marked so it can never be mistaken for a grounded one.
5
Check the grounding
Every claim carries one of three badges, and the difference matters:
- quote locatedThe quote was found in the source. This is not a statement that the claim follows from it — that judgement is yours, which is why the two render side by side.
- paraphraseClose to the source wording, but not verbatim.
- unsupportedThe quote didn’t match, so the anchor was dropped. Nothing in the app will show you this as evidence.
The Audit view lists every claim in the workspace with its badge and a drop rate, which is the fastest way to see how well-grounded a workspace is overall.
Add your own paper
Add a paper takes a PDF or an arXiv / PMC / DOI link. It checks the file properly: that it really is a PDF, that it’s within the size and page limits, that it has a text layer rather than being a scan, and that the workspace doesn’t already have it.
If something is wrong it tells you which thing, in those words. Note that parsing an accepted paper into a graph is the next piece of work — today the file is validated and queued, not yet analysed.
Call it from Claude
Pepiros also runs as an MCP server, so Claude can search the papers and check its own claims against them mid-conversation. Point Claude Code or Desktop at it, then ask Claude to summarise a paper and follow up with “now verify every claim you just made.”
It calls verify_claim on its own sentences and reports which ones the source actually supports. Setup is in the agents guide.

One thing worth repeating

Nothing here is ever labelled verified. A matched quote proves the sentence exists in the paper; it doesn’t prove the claim built on it is true. Pepiros does the part a machine can do reliably, and puts the claim and its quote next to each other so you can do the part it can’t.

Open a workspace

A workspace holds a few papers and everything derived from them. The demo workspace is already populated with three papers on circadian rhythm and cognition, including a deliberate contradiction between two of them.

You don’t need an account to look around. Signing in is what makes your own work persist — without it, anything you add is gone when you close the tab.

Read in the doc view first

The default view is a reader, not the graph: summary at the top, the source PDF beside it, related papers on the right. Claims and their quotes sit next to each other so you can judge whether the quote actually supports the claim.

Click any citation chip — the small monospace C7-style ids — to jump to the exact sentence it came from.

Open the graph when you want structure

Explore graph switches to the canvas. Pillars start collapsed, so you see the shape of the argument before its details; click a pillar to expand its claims.

If the colours and line styles don’t mean anything to you yet, open What am I looking at? in the bottom-left. It only lists what that particular graph uses.

Zoom out and cards deliberately shed detail: at a distance you are reading structure, not sentences.

Ask a question

The chat dock answers from the papers in the workspace and cites what it used. Every citation it returns is re-checked against the source after the answer is written, so a fabricated reference is reported as unsupported rather than passed off as evidence.

If the papers genuinely don’t cover your question, it says so instead of guessing. You can still ask it to answer without sources, and that answer is marked so it can never be mistaken for a grounded one.

Check the grounding

Every claim carries one of three badges, and the difference matters:

quote locatedThe quote was found in the source. This is not a statement that the claim follows from it — that judgement is yours, which is why the two render side by side.
paraphraseClose to the source wording, but not verbatim.
unsupportedThe quote didn’t match, so the anchor was dropped. Nothing in the app will show you this as evidence.

The Audit view lists every claim in the workspace with its badge and a drop rate, which is the fastest way to see how well-grounded a workspace is overall.

Add your own paper

Add a paper takes a PDF or an arXiv / PMC / DOI link. It checks the file properly: that it really is a PDF, that it’s within the size and page limits, that it has a text layer rather than being a scan, and that the workspace doesn’t already have it.

If something is wrong it tells you which thing, in those words. Note that parsing an accepted paper into a graph is the next piece of work — today the file is validated and queued, not yet analysed.

Call it from Claude

Pepiros also runs as an MCP server, so Claude can search the papers and check its own claims against them mid-conversation. Point Claude Code or Desktop at it, then ask Claude to summarise a paper and follow up with “now verify every claim you just made.”

It calls verify_claim on its own sentences and reports which ones the source actually supports. Setup is in the agents guide.