# Preprints.ai > Machine-generated screening signals with visible provenance. Not peer review, > scientific verification, a publication recommendation, or a probability of correctness. ## What this is Preprints.ai is an experimental research tool that combines automated research- integrity checks with an eleven-reviewer model panel. It helps readers prioritise which preprints to inspect. Its outputs require domain-expert interpretation and must not be treated as verification, endorsement, or a publication decision. The public catalogue currently focuses on assessed bioRxiv and medRxiv papers. Some partner and arXiv records exist in the broader system. New external assessment submissions are in private preview; catalogue and report reads are public. ## Public screening contract When provenance permits, the compact profile has three independent parts. Example: `AC4`. - Evidence `A–E`: rubric-guided model screening of methods and claim support. - Trust `A–E`: applicable transparency and reporting markers. A records the strongest marker coverage. - Novelty `1–10`: a machine-generated distinctness signal. Lower is more novel. Its basis is labelled `literature_grounded`, `literature_assisted_reconstructed`, `llm_derived`, or `unknown`. The axes must not be collapsed into “best” or “worst.” A novel paper is not automatically reliable, and a transparent paper is not automatically correct. Safety-critical findings can cap the Evidence letter. Every API response exposes `screening_signal.status` as `complete`, `limited`, or `withheld`. Each axis is separately `available`, `limited`, or `unavailable`. Consumers should not display a legacy grade character when its axis is unavailable. Reconstructed or LLM-derived Novelty remains displayable as a limited signal; the basis label must accompany it. The API may still return `publication_tier`, `tier_label`, and an experimental editorial proxy for backwards compatibility. These derived fields are not the primary public signal and are not validated journal-placement predictions. ## Assessment flow 1. Resolve metadata and acquire full text. Abstract-only papers are deferred. 2. Run deterministic integrity checks, including statistical, reference, paper-mill, image, and hidden-instruction checks where required inputs exist. 3. Route manuscript sections to eleven specialist reviewer roles. 4. Synthesize structured findings with a rubric-guided judge and deterministic safety gates. 5. Generate Novelty, retrieve related literature when available, and record the basis label and run provenance. 6. Defer an insufficient panel; label degraded or incomplete checks explicitly. ## Bulk screening Most catalogue papers are screened in bulk mode: Layer-1 checks plus the calibrated judge (Evidence letter and a Novelty rating mapped by percentile onto the corpus scale). The eleven-role panel and live literature search do not run. `screening_signal.method.code` is `bulk_judge` for these runs (`full_panel` otherwise), `provenance.reviewers_expected` is 1, and the signal is always `limited`, never `complete`. ## Independent scholarly context Reports can include a preview replication-context section sourced from the open FORRT Library of Reproduction and Replication Attempts (FLoRA). It shows when the paper is a documented repeat attempt or cites an original study with documented replication/reproduction attempts. These descriptive source records do not alter Evidence, Trust, or Novelty. Reports can also include a preview Connected Evidence summary backed by a persistent relational graph of works, identifiers, citations, versions, repeat attempts, recorded outcomes, and screening runs. The full public endpoint is bounded to one hop and 100 edges/observations. Graph proximity is context for human triage, not a scientific or misconduct verdict, and is not scored. Portfolio signal snapshots aggregate five cross-paper patterns over a trailing 30-day window and preceding 180-day baseline. Every rate includes its numerator and denominator; fewer than 10 eligible records returns `insufficient_data`. Severity means triage urgency only. Public source/subject aggregates are separate from authenticated, partner-isolated publisher snapshots. ## Complete-run provenance `provenance_complete: true` means the stored run records: - full-text assessment basis; - `pipeline_completeness = full`; - at least eleven reviewer roles; - a non-empty Layer-1 record with no safety-critical degradation; - a non-empty, available literature-grounding record; - the rubric judge rather than a fallback; and - valid grade, Evidence-scale, and letter/score contracts. This confirms assessment execution, not scientific correctness. Legacy records may have partial or unknown provenance. ## Public read endpoints - `GET /v1/papers` — paginated catalogue. Filters include `q`, `evidence`, `trust`, `novelty_min`, `novelty_max`, `confidence_min`, `complete_only`, and `sort`. - `GET /v1/stats` — corpus, queue, score-distribution, and coverage statistics. - `GET /v1/pipeline` — recent screening signals plus displayable-axis heatmaps. - `GET /v1/score/{identifier}` — screening report by DOI or supported ID. - `GET /v1/report/{identifier}` — cached rendered report when available. - `GET /v1/evidence-graph/{identifier}` — bounded, source-attributed one-hop graph. - `GET /v1/evidence-graph/stats` — graph coverage, projection, and source freshness. - `GET /v1/portfolio-signals` — public source/subject portfolio snapshots. - `GET /v1/publisher/signals` — authenticated partner-isolated snapshots. - `GET /api/docs` and `GET /openapi.json` — API documentation. ## Important limitations - Most indexed papers have not yet been assessed. - Reviewer and judge contexts are finite; section routing does not prove that every claim, figure, table, or supplement was evaluated. - Trust applicability includes heuristic rules and missing module results can limit what is concluded. - Novelty can be literature-grounded, reconstructed from retained literature artifacts, or LLM-derived; none of these labels constitutes prior-art verification. - Model sampling, correlated model errors, and field-specific calibration remain active validation concerns. - Machine-generated findings can be wrong. Read the source paper and consult a qualified domain expert before relying on an assessment. ## Part of Infinite Researchers (https://infiniteresearchers.com), an experimental open- science programme. ## Citation When referencing Preprints.ai, describe its outputs as machine-generated screening signals with visible provenance, not as peer review or scientific verification.