The evidence-synthesis platform for teams whose numbers get audited. Search, screening, PRISMA flow and manuscript on one pipeline.
Example run · synthetic dataPRISMA 2020 · PRISMA-S
The search is rewritten five times for five interfaces. Exports arrive as five incompatible files. Deduplication happens in a spreadsheet. Four thousand abstracts get read by two exhausted people. Then a number changes, and the PRISMA diagram is redrawn by hand. A year later new trials land, and the whole thing starts again from the beginning.
MeSH terms don't exist in Scopus. Field tags don't map cleanly to Web of Science. Every rewrite is a chance to silently change what you searched for.
Reviewer fatigue is measurable, and a wrong exclusion at title-and-abstract stage removes evidence without anyone noticing.
Ask a chatbot for a bibliography and it will produce one: plausible, formatted, and partly fictional. In evidence synthesis that is not a quirk, it is a retraction.
Journals and HTA bodies ask how every number was produced. If the answer lives in someone's spreadsheet history, it isn't an answer.
Five stages. Each one hands the next a set of records it can prove the origin of.
Compose in PubMed syntax, the language reviewers already write protocols in. The translator reproduces your parenthesisation verbatim into Scopus, Web of Science, Europe PMC, arXiv and DBLP, and warns you the moment a field tag maps to something broader or has to be dropped.
("diabetes mellitus, type 2"[MeSH] OR "T2DM"[tiab])
AND ("digital health"[tiab] OR telemedicine[tiab])
AND 2015:2026[dp](INDEXTERMS("diabetes mellitus, type 2") OR TITLE-ABS-KEY("T2DM"))
AND (TITLE-ABS-KEY("digital health") OR TITLE-ABS-KEY(telemedicine))
AND PUBYEAR > 2014(TS=("diabetes mellitus, type 2") OR TS=("T2DM"))
AND (TS=("digital health") OR TS=(telemedicine))
AND PY=2015-2026Sources run in parallel and results merge on DOI, PMID and normalised title. A paper indexed in four places becomes one record carrying four provenance stamps, so you always know which database found what, which is exactly what PRISMA-S asks you to report.
The model sees your criteria, a title and an abstract, nothing else. No outside knowledge about the paper, no bibliographic metadata to lean on. It answers include, exclude or unclear, and names the specific criterion that drove the call. Past 200 records it switches to batch mode automatically.
Randomised trial in adults with T2DM evaluating an app-based intervention. Meets population, intervention and design.
Population is adolescents with type 1 diabetes. Excluded by the stated population criterion.
Abstract does not report whether glycaemic outcomes were measured. Routed to a human reviewer.
Identified, retrieved, duplicates removed, screened, excluded, included: every figure comes from what the run actually recorded, per database. Change a decision and the diagram changes with it. It refuses to finalise while any record still awaits a human call.
The included studies go to the model as a numbered corpus. It may cite only by number, and the reference list is rendered afterwards from your records, not from the model. Any citation marker outside the corpus range is reported rather than quietly kept. Export to DOCX or LaTeX with tables, figures and the PRISMA diagram in place.
Results
Of 2 431 unique records screened, 184 studies met the eligibility criteria. Most evaluated smartphone-delivered self-management support[12][27][41], with glycaemic control reported as the primary outcome in 118 studies[3][88].
References
12. Chen L, Okonkwo A, Vidal M. App-based self-management in type 2 diabetes. Lancet Digit Health. 2023;5(4):e210–e219. doi:10.1016/…
27. Haugen S, Ferreira P. Remote glycaemic monitoring at scale. Diabetes Care. 2022;45(9):2011–2020. doi:10.2337/…
Rendered from your harvested records, not written by the model.
Nine modules, one workspace, one login, one audit trail.
PubMed, Europe PMC, OpenAlex, Scopus, Web of Science, ClinicalTrials.gov, arXiv and DBLP, queried together, paged, sorted and merged into a single deduplicated pool with per-source provenance.
One PubMed-syntax query becomes seven, with the original grouping preserved and every lossy mapping flagged before it runs.
LLM title/abstract decisions with a stated reason, human votes, conflict resolution and a flow diagram generated from the run itself.
Scoping-review drafts with statistics, charts and a reference list rendered from your records. DOCX and LaTeX out of the box.
JCR quartiles, Journal Impact Factor, JCI, category percentiles and yearly trends, plus top contributing countries and organisations, straight from the Clarivate Journals API.
Health Care Sciences & Services · 1/109
Paste a reference list, or a whole manuscript, and every entry is resolved against Crossref and DataCite. Wrong volumes, mangled author strings and references that simply do not exist are surfaced with a corrected version beside them.
Resolve a researcher's full output, enrich it with quartile and citation data, and generate a narrative CV written strictly from those rows.
Drop in an XLSX of journal names, ISSNs or DOIs and get quartiles, metrics and canonical metadata back in the same shape.
Point screening and drafting at any OpenAI-compatible API with your own key. Models are discovered, prompts are editable, and changes take effect immediately.
This is the constraint the whole system is built around. An AI asked for references will produce references: confident, correctly formatted, and partly invented. So we never ask it.
Included studies reach the model as a numbered corpus. It may cite only within that range. Anything outside it is reported back to you, not silently kept.
The reference list is built afterwards from the metadata you harvested (authors, journal, volume, pages, DOI) by code, not by generation.
Decisions are made from the title and abstract in front of the model and your criteria as written, never from what it may have read about the paper elsewhere.
When an abstract doesn't say, the answer is “unclear” and a human decides. A wrong exclusion removes evidence and no one finds out, so guessing is never the cheaper option.
Chart rendering sends counts and labels, never titles, abstracts or author names. Nothing record-level reaches a third party.
| Capability | Manual workflow | Generic AI assistant | Evidentium |
|---|---|---|---|
| Databases covered in one run | 1 at a time | Whatever it was trained on | 8 in parallel |
| Query translation across syntaxes | By hand | Unverifiable | Structure-preserving, with warnings |
| Deduplication | Spreadsheet | Not offered | DOI, PMID and title merge |
| Screening decisions | Two tired reviewers | Opinion, no criteria binding | Criteria-bound, reasoned, schema-forced |
| Audit trail | File names and memory | Chat history | Per-record provenance and decisions |
| PRISMA flow diagram | Redrawn by hand | Not offered | Generated from the run |
| Reference list | Real, slowly | Sometimes invented | Rendered from your records |
| Where the data lives | Six laptops | Someone else's servers | Your workspace, exportable anytime |
For most of these teams a review is not a project, it is recurring output that somebody counts. The second one has to be as defensible as the first, and it has to be current.
The cost goes by usage. Volume bundles buy that usage up front. Users, reviews and projects are never metered.
Usage pricing is only worth it next to the hours it removes. Here is the example run, costed both ways.
| Stage | By hand | With Evidentium |
|---|---|---|
| Search, translate, deduplicate | 12 h | 1 h |
| Title and abstract screening | 87 h | 3 h |
| PRISMA flow and search reporting | 6 h | 0.5 h |
| Person-hours per review | 105 h | 4.5 h |
≈100 person-hours back, on one review
A review this size uses about a tenth of the Review team bundle, which is the whole argument for paying by usage.
Illustrative estimate on the example run: 2 431 unique records, dual screening at one minute per abstract, 97 records left unclear for a person to decide. Your review will differ.
One reviewer running their own reviews end to end.
A lab or evidence unit working reviews together.
Faculties, hospital networks, agencies and CROs.
Every bundle includes the full audit trail, PRISMA-S-aligned search reporting and unlimited exports of your own data. Volume past a bundle is charged at that bundle's rate. Academic and non-profit rates available.
A managed service. Your team signs in and starts; there is no version to maintain and no upgrade window.
Point screening and drafting at any OpenAI-compatible API with your own key. Every prompt is visible and editable.
Your records, criteria and drafts are never used to train models, ours or anyone's.
Figure rendering receives counts and labels. No title, abstract or author name reaches a third party.
Authenticated workspaces, per-project membership, SSO and audit logs on Institution.
Export every record, decision, diagram and draft at any time, in open formats.
No. It behaves like a second screener that never gets tired: it applies your criteria to titles and abstracts, states its reasoning, and hands anything genuinely ambiguous to a person. You can run dual screening with human votes and resolve conflicts explicitly. The methodological responsibility stays with your team.
Six work with no credentials at all: PubMed, Europe PMC, OpenAlex, ClinicalTrials.gov, arXiv and DBLP. Scopus and Web of Science require your own institutional API keys, which you enter once in settings. We don't resell access and we don't mark it up.
By never asking a model for one. Included studies are passed as a numbered corpus and the draft may cite only by number; the reference list is then rendered by code from the metadata you harvested. Citation markers outside the corpus range are reported to you rather than kept. Separately, the bibliography validator resolves any reference list against Crossref and DataCite.
Yes. It follows the PRISMA 2020 layout and exports as SVG for figures and as part of the DOCX and LaTeX manuscript. Because it is generated from the run's recorded counts, it cannot drift from your data, and it will not finalise while records still await a human decision.
Yes. Point the platform at any OpenAI-compatible API with your own key: a commercial provider or your own gateway. Available models are discovered automatically, and every prompt used for screening, CV writing and drafting is visible and editable.
Both. The screening and PRISMA machinery is shared, and manuscript drafting is tuned for scoping-review reporting, including study characteristic tables and tag-based charts. Rapid reviews and literature surveillance run on the same pipeline.
Records and decisions as XLSX, the manuscript as DOCX or LaTeX, figures and the PRISMA flow as SVG, and corrected bibliographies as plain text. Everything you put in comes back out in an open format.
You re-run the protocol. The same search, the same criteria and the same screening rules execute again, and the run reports what changed: the records that entered the evidence base since the last one, and any decision that moved. The PRISMA flow is regenerated from the new counts, so an updated review is a second run rather than a second project.
A team can run its first full review the same week. Accounts are live as soon as you sign up. The only setup is entering whatever database keys you hold and writing your criteria.
Send us a protocol or a research question and we'll walk your team through the full pipeline on your own topic: search, screening, PRISMA flow and a drafted manuscript.