Preprints.ai

A machine-screening index for bioRxiv and medRxiv preprints

One preprint, three separate signals.

Every screening run records Evidence, Trust and Novelty as three independent readings. Each one carries the provenance that decides whether it can be used at all.

A first-pass reading aid. Not peer review, not scientific verification, and not a recommendation to publish.

The index today

Loading live figures…

Evidence

letter, A to E
Waiting for the live catalogue.

Machine-detected strength of methods, results and claim support. A is the strongest recorded support, E the weakest.

Trust markers

letter, A to E
Waiting for the live catalogue.

A meets nearly all applicable trust checks, C some, E minimal. Transparency markers, reproducibility, statistical rigour and citation health; fraud checks cap the grade.

Novelty

number, 1 to 10
Waiting for the live catalogue.

1 landmark, 5 important, 10 minimal. Machine-generated, and lower means more novel.

Recent screening signals

See the whole pipeline
Signal Paper Source Run Generated
Loading recent signals…

How a signal is made

Automated checks run first, model-reviewer roles read second, and explicit gates decide what can be published. Every step leaves a record on the report.

Read the full methodology
  1. Deterministic checks

    Paper-mill detection, statcheck p-value verification, GRIM tests, data-availability and retraction cross-referencing all run before any language model reads a word.

  2. Eleven model-reviewer roles

    Methods, statistics, ethics, validity, domain, adversarial and positive-evidence roles inspect routed manuscript context. Their outputs are correlated machine signals, not independent peer reviews.

  3. Rubric-guided synthesis

    Structured model outputs become separate Evidence, Trust and Novelty axes. Deterministic safety gates can cap or withhold a signal when a critical check fails.

  4. Provenance and abstention

    Full-text input, role participation, automated-check health, judge execution and literature grounding decide whether each axis is available, limited or withheld.