Agentic Science

Lemma

An auditor of scientific claims that refuses to trust itself.

A paper becomes a self-judged evidence trail: supported, falsified, or inconclusive, with the receipts attached. Success criteria may only be derived from a claim's own wording — no added thresholds. Every LLM call and tool result lands in an append-only trace, so the reasoning is auditable end to end. Built for the re:AGENT End-to-End Agentic Science track at Founders Inc.

Claims, not vibes. Up to six testable claims are extracted per paper, ordered by importance, each with a success criterion drawn only from the claim's own statement.
Honest audits. Numerical audits are written and run per claim. Where reproduction can't reach the paper's number, the gap is recorded — not the threshold.
Judge before human. A judge grades the assembled logbook's trustworthiness first; every call and result sits in an append-only trace.

4 stages

Pipeline

≤ 6

Claims / paper

Append-only

Trace

Open Lemma Ask

In product

Paper under audit.

Judged before a human looks.