How it works

How each claim gets checked against its source.

Every claim is matched against the exact sentence it cites and scored. No second model grades the first one's work.

Ingest

A paper becomes a graph, in stages you can watch.

Structure appears before any model has run, so the page is never a spinner. The skeleton graph lands first, then related work, then metadata, then the summary and pillars, then the remaining 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, shown live as they happen.
Generators + graph expansion
~15-45s
Remaining generators fill in, and citation-graph expansion becomes available.

Structure

Each paper's sections are generated from its own content.

A clinical trial and a machine-learning paper don't have the same shape, so they don't get the same headings. Pepiros classifies the paper first and plans its sections from what's actually in it.

Every leaf under a pillar carries its own evidence, which is what makes the graph navigable rather than decorative.

Grounding

The claim and its source sit side by side.

A claim that cleared the match shows its badge, its citation id, and its score, and the sentence it came from is highlighted in the source pane next to it.

A claim with nothing checked behind it is labelled inference and gets no citation at all, rather than a hedge.

Worked example

One claim, scored

Morning bright light advances circadian phase by about 1.4 hours in shift workers.

Source excerpt

Participants receiving 30 minutes of 10,000 lux morning light advanced dim-light melatonin onset by 1.4 hours (95% CI 0.9-1.9) after five days.

0.0 dropped0.75 paraphrase0.92 quote located

quote locatedC7score 0.97

For agents

An agent can check its own claims, mid-answer.

Connect over MCP and the agent calls verify_claim on its own sentences before it asserts them to you.

When one comes back unsupported, it says so, in the same answer. That is the whole point: see the MCP tools.

Said on stage, not just in the docs

Where the guarantee stops

A matched quote proves the words are really there. It doesn't prove the conclusion follows from them.
A model can attach a real sentence from the Methods section to the wrong conclusion, and the match still scores 1.0.
The badge reads quote located. It never reads verified.
Claim and quote sit side by side on purpose, so you're the one who decides whether the claim follows, not the matcher.

An entailment overlap floor helps: every number, unit, and comparator in a claim also has to appear in the anchored span, checked against the numeric ledger. That catches the failure a fuzzy match alone misses, a genuine quote attached to a reversed or overstated conclusion.

Try the demo workspace See what to click Read the docs