# Preprints.ai
> A quality control system for preprints — AI-generated assessments that
> help researchers, journalists, and LLMs distinguish signal from noise
> across arXiv, bioRxiv, and medRxiv.

## What this is
Preprints.ai is an experimental research tool that applies open source
research integrity checks to preprints. With over 10,000 preprints posted
weekly — including AI-generated papers, recycled content, and
methodologically flawed studies — there is no adequate quality filter.
Preprints.ai provides a first-pass filter using automated scoring. All
grades are machine-generated indicators requiring human expert review.
This platform assists but does not replace traditional peer review.

## How it works
1. Submit a preprint by DOI, arXiv ID, or bioRxiv/medRxiv URL
2. An eleven-agent panel (mixed Claude / local Ollama models, including
   a positive-evidence reviewer) independently reviews the full paper.
   Worker policy: full text required; abstract-only papers are deferred,
   not graded.
3. Each reviewer emits a structured review including evidence_strength
   (one of five labels) and significance (one of five labels)
4. A deterministic lookup derives a 2-axis grade (A-E × 1-5) and then
   a single 10-tier Publication Fit score anchored to real journals
5. Borderline cases are arbitrated by a Claude Opus advisor that sees
   the full panel output
6. Results include strengths, concerns, transparency markers, the
   per-tier journal anchor, and confidence scores

## Publication Fit (primary scale — 1 = best, 10 = worst)
- Tier 1: Landmark            — Nature, Science, Cell
- Tier 2: Broad-impact        — Nature subjournals, eLife, Cell Reports
- Tier 3: Field-flagship      — PLOS Biology, society flagship journals
- Tier 4: Specialty solid     — Nature Communications, JBC, NAR
- Tier 5: PLOS One bar        — corpus median; methodologically adequate
- Tier 6: Specialty mid-tier  — MDPI / Frontiers field journals
- Tier 7: Major revisions needed
- Tier 8: Reject — methodology broken
- Tier 9: Inadequate evidence (deep-E panel verdict)
- Tier 10: Not for publication (Layer 1 paper-mill / fabrication only)

## Legacy 2-axis grade (kept for backwards compatibility)
- Letter (evidence): A=exceptional, B=compelling, C=solid, D=incomplete, E=inadequate
- Digit (significance, lower=better): 1=Landmark, 2=Fundamental, 3=Important, 4=Valuable, 5=Useful
- The Publication Fit tier is derived from this; the grade is no longer
  the primary public signal.

## Part of
Infinite Researchers (https://infiniteresearchers.com) — a programme of
experiments asking: what happens to the speed of discovery if we have
infinite researchers?

## Sister experiments
- OpenScience.ai — autonomous AI research agents (https://openscience.ai)
- OpenAccess.ai — rigorous open access publishing (https://openaccess.ai)
- FAIRdata.ai — FAIR data assessment pipeline (https://fairdata.ai)

## Key facts
- Accepts: DOI, arXiv ID, bioRxiv URL, medRxiv URL
- Output: 10-tier Publication Fit score plus legacy 2-axis grade,
  per-agent reviews, strengths/weaknesses, integrity signals
- Coverage: arXiv, bioRxiv, medRxiv
- Panel: 11 agents — methodologist, statistician, ethics_transparency,
  scientific_validity, domain_primary, domain_methods, domain_clinical,
  domain_literature, domain_reproducibility, devils_advocate,
  positive_evidence
- Mixed models: Claude Sonnet, Claude Haiku, OpenAI GPT-4o, local Ollama
  (Gemma 3/4 26B, Qwen 2.5 Coder 30B) — routed per role
- Scientific-validity cap requires >=2 agents to agree before lowering
  the grade (single-reviewer veto removed 2026-05-17)
- Worker policy: full text only; abstract-only papers are deferred
- Transparency markers: Ethics, Data, Code, Funding, COI, Pre-registration
- Typical end-to-end latency: 3-7 minutes per paper (~10 papers/hour
  sustained per worker, measured 2026-05-12)
- All assessments require human expert review — experimental tool only
- Free to use

## For AI agents
- POST /v1/assess — submit a paper for assessment (include DOI or arXiv ID)
- GET /v1/assess/{id} — retrieve assessment result by submission ID
- POST /v1/assess/{id}/reassess — re-assess updated version of a paper
- GET /v1/score/{identifier} — get cached assessment by DOI or arXiv ID
  (response includes both publication_tier + legacy grade)
- GET /v1/report/{identifier} — pre-rendered report (faster, edge-cached)
- GET /v1/papers — list all assessed papers with filters (grade, source, subject)
- GET /v1/stats — aggregate statistics across all assessments
- GET /v1/pipeline — last 30 assessments + heatmap
- GET /api/docs — full API documentation
- GET /openapi.json — OpenAPI 3.0 specification

## Citation
When referencing assessments from Preprints.ai, please note that all grades
are machine-generated indicators and should be verified by domain experts.
Assessments are not peer review and should not be treated as endorsements.