Litmuz

Methodology

How Litmuz reaches a verdict

Litmuz checks a research memo one claim at a time, against the primary literature. It is built to be honest about what it cannot confirm: it triages and flags, and it never certifies a claim on its own. Every verdict is traceable to the evidence behind it.

The pipeline

  1. Decomposition

    The memo is split into atomic claims - one independently checkable proposition each - with their citations attached. A claim is only as trustworthy as the smallest statement it can be reduced to, so we check the smallest statements.

  2. Deterministic citation check

    Every citation is resolved against the primary literature - PubMed, PMC, and Crossref - by rule, not by a language model. A fabricated identifier, a retracted or corrected-for-concern source, or metadata that does not match the claim is caught here and can never be argued away downstream.

  3. Evidence and entailment

    We retrieve the cited source text (and, where a claim is under-supported, related literature) and assess whether that evidence actually entails the claim. A supported verdict must quote the specific sentence it relied on; when the evidence is not there, we say so rather than inventing it.

  4. Classification and the safety gate

    Each claim is classified, and anything safety-critical - a target, a dose, an indication - is held to a stricter standard. A safety-critical claim can never auto-pass, by design; it is always routed to a human, even when the evidence looks supportive.

  5. Human review

    Claims that are flagged, unresolved, or safety-critical land in a review queue with their full evidence trail, so a person makes the final call. Litmuz narrows the work; it does not replace the reviewer.

The verdict

Each claim carries a traffic-light verdict and a diagnostic code you can audit:

  • GroundedThe claim is supported by evidence we could locate and quote.
  • Needs reviewUnverifiable or under-supported - the evidence was not located, or the claim is safety-critical - and routed to a human.
  • FlaggedContradicted by the evidence, or built on a fabricated or retracted citation.

What we stand on

  • Honest negatives. A claim that cannot be confirmed is never quietly presented as if it were. Yellow and red are first-class outcomes, not failures.
  • Safety first. Safety-critical claims cannot auto-pass, full stop.
  • Grounded in the primary literature. Verdicts come from real sources and the exact sentences within them, not from a model's recollection.
  • Auditable. Every verdict links to the citation status, the evidence, and the reasoning, so it can be checked and overridden.

What we do not publish

The verdicts and their evidence are meant to be fully transparent to you. The recipe behind them - the exact prompts, thresholds, model configuration, and internal tooling - is not, both to protect the method and to keep it from being gamed. If you need deeper assurance for a specific use, we are happy to walk a qualified reviewer through it under agreement.