Check an AI agent's research, claim by claim

How it works

Paste a memo from a research agent. With the literature criterion on, Litmuz runs the memo through four stages and returns a verdict for every claim, with the evidence attached. Two of the stages are rule-based and run with no model in the loop. The stage that decides whether evidence supports a claim uses Claude, and it has to quote the exact sentence it relied on. Genomic evidence takes a different path: no citation registries, no retrieval, no model, and the same safety gate and traffic lights at the end.

• 1

  Decompose the memo into atomic claims

  Claude splits the memo into single checkable propositions and attaches each citation to the claim it is meant to support. Nothing is verified at the paragraph level, because a paragraph can be half right.

• 2

  Resolve every citation deterministically

  Each identifier is resolved by rule against PubMed, PMC, and Crossref, with no language model involved. A fabricated identifier, a retracted source, or metadata that does not match the claim is caught here, and cannot be argued away downstream.

• 3

  Retrieve the evidence and judge whether it supports the claim

  Litmuz pulls the passages from the resolved source, and Claude assesses whether that evidence actually entails the claim. A supported verdict must quote the exact sentence it relied on; when the sentence is not there, the claim does not pass.

• 4

  Apply the safety gate and return a per-claim verdict

  A separate deterministic lexical check re-derives whether a claim is safety-critical (a dose, a dosing regimen, a clinical indication, a molecular target), and those claims are always routed to a human, however confident the model is. Every claim ends on green, yellow, or red, with a diagnostic code and a link to the citation status and evidence behind it.

What the verdicts mean

Every claim gets one of three traffic lights, plus a diagnostic code (D1 to D5) and a link to its citation status and evidence. Green is not the only good outcome. Yellow and red are first-class results: they tell you where the memo is not standing on evidence.

• Grounded (green)

  The evidence supports the claim and the citation resolves cleanly against PubMed, PMC or Crossref, or (in genomic mode) the region matches the reference. In literature mode Claude has to quote the exact supporting sentence, so you can check the call yourself.

• Needs review (yellow)

  The claim is unverifiable, under-supported, or safety-critical, so it goes to a human in the review queue. A safety-critical claim (a dose, a dosing regimen, a clinical indication, a molecular target) always lands here, even when the evidence looks supportive and the model is confident. A retracted source or an expression of concern can never come back green.

• Flagged (red)

  The evidence contradicts the claim, or the citation is fabricated. Citation resolution is deterministic, so a made-up identifier is caught by rule, not by judgment.

Two ways to verify

Pick the criterion that fits the memo you are checking. Each run is scored against one of them. Every verdict still carries its diagnostic code and links back to the evidence that produced it.

• Literature

  Citations are resolved deterministically against PubMed, PMC and Crossref, with no model in the loop. A fabricated identifier, a retracted source, or metadata that does not match the claim is caught by rule, not by judgment. Only then are evidence passages retrieved for Claude to judge entailment, and it must quote the exact supporting sentence.

• Genomic evidence

  Genomic claims are checked against Gladstone Institutes reference data: the Pollard lab's Human Accelerated Regions and Zoonomia mammalian conservation. The check is deterministic, with no LLM involved at any step, so a real HAR asserted truly comes back green and a real HAR denied comes back red. The reference is a curated subset of well-characterized HARs rather than the full catalogue, so a region it does not cover returns an honest needs-review and is routed to a person, never a silent pass.

What it checks against

The sources are public. The prompts, thresholds and rubric that combine them are not.

Who it is for

Litmuz is built for the people who have to stand behind what an AI agent wrote, and for the agents that would rather not hand over weak work in the first place.

• The computational biologist reviewing an agent's output

  An agent returns a fluent memo where every sentence sounds equally confident and every citation looks equally real. Litmuz splits it into atomic claims and marks each one green, yellow or red, so you can start with the claims that are contradicted or unsupported instead of reading every sentence at the same level of suspicion. Green claims still carry the quoted sentence they rest on, so you can check any call yourself.

• Teams running AI agents in drug-discovery and genomics workflows

  When a target or a dosing claim gets questioned three months later, "the model said so" is not an answer. Litmuz stores every verdict with the evidence behind it, and every human decision is written to an append-only audit trail with its rationale, reviewer and timestamp. Safety-critical claims never auto-pass into that record without a person seeing them.

• The AI agent itself, over MCP

  Any Claude-powered agent can call Litmuz as a tool and get its own claims back with per-claim verdicts and diagnostic codes. It can then drop the fabricated citation or soften the overreaching sentence before a human ever reads the draft.

Frequently asked questions

How the verification works, what the verdicts mean, and where the limits are.