Built for responsible evidence synthesis

Citation integrity benchmark

We re-verify every DOI cited in real AutoSearch deliverables against Crossref / doi.org with an independent, re-runnable harness (scripts/benchmark_citations.py). Results are published as measured — including failures.

Benchmark generated: 2026-06-10 · Methodology v1.0.0, mode: live. The harness is deterministic and re-runnable on demand against the same artifacts. /methodology

Context: how generalist tools compare

Independent evaluations report that generalist deep-research systems hallucinate 3–18% of their URL citations (arXiv:2604.03173). AutoSearch targets zero fabricated DOIs by construction: a citation either resolves to a real record before it enters the deliverable, or it is flagged and excluded.

RAISE recommendations — how AutoSearch responds

RAISE (Responsible AI in Evidence Synthesis, 2025) urges tool developers to provide clear documentation, transparency about limitations and biases, and evidence that the tool does what it claims. How each expectation maps onto a shipped AutoSearch feature: (cochrane.org)

• Clear documentation of what the tool does → Public methodology page plus a Methods section in every report documenting sources queried, query strings, and record counts
• Transparency about limitations and potential biases → Every manuscript ships with a Limitations section; preprints and weak bibliographic metadata are flagged, never silently upgraded
• Evidence that the tool does what it claims → Published citation-integrity benchmark, re-run on real deliverables with an open, deterministic harness — failures included
• Human oversight of AI-assisted synthesis → Deliverables are drafts for expert review: inclusion decisions, interpretation, and sign-off remain with the human author
• Transparent reporting of AI use → A ready-to-paste AI use statement is generated with every manuscript: tool, version, models used, and the exact scope of AI work
• Reproducible, auditable methods → Reproducibility statement (exact query, keywords, sources, dates, engine version, run ID), PRISMA 2020 flow with real counts, and a record-level screening log export
• Verifiable outputs → Live DOI verification against Crossref before a citation enters the report, plus semantic relevance verification of citation-claim links

The 2025 joint position statement (Cochrane, Campbell, JBI, CEE)

The joint position statement on responsible AI use in evidence synthesis expects human oversight, transparent reporting, and justification of AI use. How AutoSearch aligns:

• Human oversight — AutoSearch output is decision support for a human author, not a finished verdict. The author reviews the deliverable, owns inclusion decisions, and remains accountable for the synthesis.
• Transparent reporting — every manuscript includes an AI use statement declaring the system, version, models, and which steps were AI-assisted (search, screening, extraction, drafting) versus human (review, inclusion decisions, sign-off).
• Justification and methodological soundness — the published benchmark, the PRISMA 2020 flow with real counts, and the reproducibility statement give authors the evidence they need to justify using the tool in their own methods section.

Data residency and GDPR

AutoSearch runs on UK and Swiss infrastructure with GDPR- and Swiss nLPD-aligned processing. Your uploads are not used to train models. Some AI processing can involve external model providers under appropriate safeguards such as standard contractual clauses — the privacy policy documents processors and international transfers in full. /privacy

Frequently asked questions

Is AutoSearch endorsed by Cochrane or RAISE?

No. There is no Cochrane or RAISE certification for AI tools, and we claim none. AutoSearch aligns with the publicly available RAISE recommendations and the 2025 joint position statement, and this page documents that mapping feature by feature.

How do you prevent fabricated citations?

Every DOI is verified live against the Crossref API before it enters a report; a citation that does not resolve is flagged and excluded. We additionally publish an independent benchmark that re-verifies the DOIs in real deliverables — including any failures.

What does the AI do, and what stays human?

The AI runs the search, screens and filters records, extracts evidence, and drafts the manuscript. The human author reviews the deliverable, makes the final inclusion and interpretation decisions, and signs off. Every manuscript includes an AI use statement spelling this out.

Independence disclaimer

AutoSearch is not certified, endorsed by, or affiliated with Cochrane, Campbell, JBI, CEE, or the RAISE authors. No such certification exists for AI tools today. We align our product with their published recommendations and link to the originals so you can verify the mapping yourself.