Pepiros
A publishing platform for researchers, with a summariser you can actually check. Publish a paper you have read, and every claim in the write-up stays bound to the sentence it came from.
What it is
Two things, in one place.
A publication. Researchers post the papers they have read, with their own write-up. Readers follow people whose judgement they trust, and the library fills with papers someone has actually worked through rather than a search index.
A summariser you can check. Each write-up is generated from the paper and then verified against it. Every claim shows the exact sentence it came from, at the page it came from, or says plainly that it has none.
The second part is what makes the first part worth reading. A feed of AI summaries nobody can check is just a faster way to spread a misreading.
The problem
A confident summary is not evidence.
Ask a model to summarise a paper and you get fluent prose that sounds exactly like the paper. Whether it is faithful is invisible from the output. The only way to check is to read the source, which is the work the summary was supposed to save.
Pepiros does not ask you to trust its output. It shows you the sentence each claim came from, and tells you when there is no such sentence.
The mechanism
Checked against the source, not asked of a model twice.
Each claim is matched against the exact sentence it cites, and scored. Above 0.92 it is quote located and shows its quote, page and citation id. Between 0.75 and 0.92 it is paraphrase. Below that the anchor is dropped and the citation is stripped rather than left dangling.
None of this asks a second model whether the first one was honest. It is a deterministic string match against the real text, which is why the same claim always gets the same score.
What you read
The claim and its source, side by side.
A grounded claim sits next to its quote, not on top of a tooltip containing it. That is a deliberate choice: putting them adjacent is what lets you decide whether the claim actually follows.
Claims with nothing behind them are labelled inference and carry no citation at all, rather than a hedge that looks like one.
Ingest
Structure lands before any model runs.
The page is never a spinner. The skeleton graph appears first, then related work, then metadata, then the summary and pillars, then the rest of the generators.
- Skeleton graph<300ms
Paper node and ghost pillars pulse in, before any AI has run.
- Related papers<1s
The Related Papers rail populates from Semantic Scholar, no LLM involved.
- Metadata badge<2s
Metadata and an archetype badge appear, a fast pass.
- Summary + pillars~5-10s
Archetype classification and pillar planning resolve -- the upload view's live progress reflects this in real time, stage by stage.
- Generators + graph expansion~15-45s
Remaining generators fill in (6 of 21 types implemented so far), and citation-graph expansion becomes available.
For agents
An agent that can check itself, mid-answer.
Connect over MCP and your agent calls verify_claim on its own sentences before asserting them. When one comes back unsupported, it says so in the same breath.
Limits
What this does not prove.
A fuzzy-matched quote proves quotation provenance, not entailment. A model can attach a real Methods sentence to a wrong conclusion and still score 1.0 on the match.
So the badge always reads quote located, and never verified. Saying that here, rather than in a footnote, is the point.
Start with someone else’s reading.
The library is open. Every paper in it opens into a write-up where each claim sits next to its source.