A self-contained systematic review search & appraisal workbench. Search across major biomedical databases, screen and de-duplicate results, run AI-assisted critical appraisal, and assemble PRISMA-ready records — all from a single HTML file that runs in your browser. Installable as an app, offline-capable, light or dark theme, and with no project-operated backend — your projects stay in your browser (see Privacy & data flow for what does travel to external services).
Status: Beta. Features and interfaces are still changing between builds.
🔗 Live: https://evidentum.app/
Evidentum walks a review through the full early workflow — search → screen → appraise — without a backend, an account, or a build step. Everything is one index.html.
A two-stream search builder lets you compose a query once and fan it out across sources:
- Primary stream — structured search against PubMed, Europe PMC, and Cochrane, with MeSH and Boolean syntax.
- Keyword stream — an auto-derived natural-language query for the broader engines: Scopus, OpenAlex, Semantic Scholar, Crossref, and ClinicalTrials.gov.
Each database is queried in its own native advanced syntax (PubMed field tags and MeSH, Europe PMC TITLE_ABS/MESH, Scopus TITLE-ABS-KEY, ClinicalTrials.gov AREA[] expert queries, and per-concept search for OpenAlex and Semantic Scholar), translated deterministically from a single set of concept blocks so the same PICO produces the same, reproducible queries every time.
Includes a concept-block builder, a live count preview before you commit, set history with Boolean nesting (combine prior sets into a pool), per-source result counts, and deep pagination (Europe PMC cursorMark, Semantic Scholar token paging) so large searches aren't silently capped at the first page. Optional, clearly-labelled recall levers — Europe PMC synonym expansion and cross-engine phrase proximity — are off by default for precise, reproducible searching.
Study-design limits (e.g. RCTs only) are applied with each engine's native filter where one exists, and with a Cochrane-style sensitivity hedge on the keyword engines that lack one — so the restriction actually holds across every source, not just PubMed. For the paywalled databases the app can't call, the builder emits copy-and-paste strategies for Ovid Embase and Cochrane CENTRAL in their native line-numbered syntax, translated from the same concept blocks.
- Review protocol — a structured, pre-registration-style protocol per project (background, PICO, outcomes, eligible designs, information sources, risk-of-bias plan, synthesis plan, registration), with a completeness counter and a standalone protocol export formatted to PRISMA-P. Define the method before you search, then let the search trail and reports document what you actually did.
- De-duplication across sources.
- Keyboard-driven screening — Rayyan-style shortcuts in a project:
J/Kmove through records,I/M/Xrecord include/maybe/exclude (auto-advancing to the next undecided), number keys1–0set the PRISMA exclusion reason,Njumps to the next unscreened record. A live tally shows progress at a glance. - Dual-reviewer screening — two reviewers screen independently under blinded A/B lenses; agreements auto-fill the final decision, disagreements surface in a Conflicts filter for adjudication, and inter-rater agreement is reported live as Cohen's κ (with the Landis–Koch label). Reviewer votes travel in the CSV export.
- Two-computer dual review — no shared account needed: the host exports a reviewer packet file (blinded — the host's votes and final decisions are stripped from the file itself), Reviewer B imports it into Evidentum on their own machine, screens under a locked Reviewer B lens, and sends back a votes file that merges with conflict detection and a live κ update. Merges are newest-wins per record, so a stale file can never roll back fresher work.
- A decision filter (To Screen / Included / Maybe / Excluded / Conflicts) that turns the record list into a shrinking screening queue.
- A screening burndown — a progress bar with coverage (screened / total / %) and, from your decision timestamps, an honest pace and time-to-finish estimate taken from active screening bursts so breaks between sittings don't skew it. Lens-aware, so each reviewer sees their own progress.
- Send all to project — push a whole result set into a project for screening in one step (no more adding records one at a time).
- Update searches (Cochrane MECIR C37) — re-running a saved strategy later and sending again adds only new records to the screening queue (existing ones are skipped); update arrivals are badged ✨ NEW. The project records each search date and yield, and shows an amber/red reminder when the search is ageing past ~6–12 months.
- A PRISMA-style identification panel with per-source counts and a "Copy for methods" export (structured and prose) that you can paste straight into a manuscript — now including any active search options (synonym expansion, phrase proximity, field limits) so the documented strategy is reproducible.
- A PRISMA-S search report export — a submission-ready supplement with the full per-database strategies as executed, limits, search dates, update-search yields, and the deduplication method, keyed to the PRISMA-S reporting items.
- A complete PRISMA 2020 flow diagram with screening stages, plus structured exclusion reasons.
- Projects — group searches and saved records into named projects.
- Retraction sweep — scan your reference set against retraction data.
- Citation tooling: cited-by and related articles lookups, copy citation, reference-manager export (RIS / BibTeX), and CSV export.
- A structured extraction table per project — define your own columns once (seeded with a standard set: design, population, N, intervention, comparator, outcomes, results, funding), fill them per study, and export the whole thing as a wide evidence matrix (one row per study, one column per field) for your appendix.
- AI first-draft — when the appraisal agent runs, it pre-fills the fields it can already answer (design, population, sample size, key result) and flags them with a 🤖 VERIFY badge until a human confirms them; reviewer edits are never overwritten.
An optional AI agent reads full text (PMC XML, PDF, or fetched HTML) and drafts structured critical appraisals against standard instruments:
- Risk of bias — RoB 2 (randomized trials) and ROBINS-I V2 (non-randomized studies of interventions), with traffic-light plots you can download as a robvis-style export; QUADAS-2 for diagnostic-test-accuracy studies (four bias domains plus the three applicability judgements); and the Newcastle-Ottawa Scale and Jadad scale.
- Appraising included reviews — AMSTAR 2, for overviews/umbrella reviews that include other systematic reviews, with its proper overall confidence rating (High / Moderate / Low / Critically low) derived from the seven critical domains rather than a raw item score.
- Reporting quality — CONSORT 2025, PRISMA 2020, STROBE, CARE 2013, and AGREE II.
- Certainty of evidence — GRADE, with a Summary of Findings export.
- Validation harness — compare the AI's RoB 2 / ROBINS-I V2 judgements for a project against a reference standard (e.g. a published review's consensus judgements): export a reference-template CSV, fill it in, import it back, and read per-domain agreement (raw %, Cohen's κ, linearly weighted κ) with a disagreement list, results CSV, and a copy-ready methods paragraph — the measurement instrument for a formal validation study.
The agent is bring-your-own-key with a provider toggle for OpenAI, Anthropic / Claude, Google Gemini, and xAI Grok:
- OpenAI (Responses API) and Anthropic (Messages API) — strong, well-calibrated appraisal judgements and reliable structured output.
- Google Gemini (generativelanguage API) — large context windows and low cost, a good fit for high-volume abstract screening.
- xAI Grok (OpenAI-compatible API) — an additional low-cost option.
Pick a model per provider in Settings (or enter a custom model ID). The active model is recorded alongside your search provenance.
- PDF Scanner — drop in a local PDF (plus supplementary files), run any of the appraisal checklists against it, and add the result straight to a project — checklist metadata and any detected DOI / full-text links carry across exactly as for a database result.
- Reference Scanner — extract a paper's bibliography, locate each reference online, and bulk-add the located set to a project.
- LibKey integration for institutional library access.
- Optional Scopus, NCBI/PubMed, and Semantic Scholar API keys for expanded access and higher rate limits.
No install needed — just open the live link. To run it yourself:
git clone https://github.com/AnkiAnaesthesia/evidentum.git
cd evidentum
# open index.html in a browser, or serve locally:
python3 -m http.server 8000
# then visit http://localhost:8000A local server (rather than opening the file directly) is recommended so the service worker and PWA install work, and to avoid
file://CORS limits on some API calls.
In Chrome/Edge, open the live site and use the install icon in the address bar (or browser menu → Install Evidentum). On iOS Safari, use Share → Add to Home Screen. Once installed it launches in its own window and works offline for the parts that don't need network access.
Evidentum has no project-operated backend — no server run by this project ever receives your data. Your projects, screening decisions, appraisals and API keys live only in your browser's local storage. Because the app talks directly to external services, though, some data does leave your device while you use it:
- Search queries go directly from your browser to whichever scholarly databases you tick — PubMed, Europe PMC, Cochrane, Scopus, OpenAlex, Semantic Scholar, Crossref and ClinicalTrials.gov.
- AI appraisal sends the selected abstract, full text or supplementary text to the AI provider you configure (OpenAI / Anthropic / Google / xAI), authenticated with your key. Nothing is sent to any AI provider until you enter a key and start an analysis.
- Full-text and metadata lookups (DOI resolution, open-access PDF discovery) contact services such as Crossref and Unpaywall.
- Semantic Scholar can fall back through a public CORS proxy (AllOrigins / CodeTabs / ThingProxy) when a direct request is blocked — which would route that one query, including its search terms, through a third-party service. This fallback is opt-in and off by default: enable it under ⚙ Settings → Semantic Scholar if you want it. With it off, a blocked Semantic Scholar request simply fails and is reported, with no third party in the path. No other database ever uses a proxy.
Keys and saved data are never sent to or stored by this project — but note that browser extensions and other same-origin scripts running in your browser can read local storage, including any saved keys. Enter keys under ⚙ Settings → AI Agent — API Keys (OpenAI, Anthropic, Google Gemini, xAI), along with Scopus, NCBI and Semantic Scholar credentials. By default keys persist in this browser; untick “Remember keys on this device” to keep them for the current tab only (cleared when you close it). Because there's no backend, clearing site data clears your projects, so export anything you want to keep.
Evidentum has been reviewed for common web-security issues — cross-site scripting, credential handling, Content-Security-Policy, and injection at its import and render boundaries — through several rounds of automated review using large language models, alongside the maintainer's own review. Automated review is not a substitute for a professional security audit, and it is not a guarantee that the app is free of vulnerabilities. If you find a security problem, please report it — see SECURITY.md.
- Single file. The shipped app is one
index.html— markup, styles, and logic. No bundler, no dependencies to install at runtime. - Generated, not hand-edited. The single
index.htmlis generated from a modular source (HTML/CSS/JS partials) and validated by an automated build and test suite maintained in a separate development workspace. This repository ships the built, self-contained file rather than the build tooling — so there is nothing to install or compile to run it. - PWA assets in the repo root:
manifest.webmanifest— app metadata.icon-192.png,icon-512.png,icon-512-maskable.png— app icons.sw.js— service worker. Network-first for same-origin app assets (page and icons alike) so a new deploy is always served fresh, with the cache used only as an offline fallback. Cross-origin requests are never intercepted or cached — every scholarly-database call, AI-provider request, CORS-proxy fetch and credential-bearing URL (e.g.?api_key=…) goes straight to the network and never lands in the Cache API. The service worker only touches the Cache API; it never reads or clears your projects, appraisals or keys (those live in IndexedDB / local storage). Bump theCACHEversion string insw.json release to evict old caches.
- Fonts are embedded (WOFF2 data URIs) — no font-CDN requests, so typography survives offline and on
file://.pdf.jsis vendored (pdf.min.js/pdf.worker.min.jsin the repo root, loaded from the same origin) — the only third-party runtime library. There is no CDN fallback: if the vendored files can't load, PDF text extraction fails closed (appraisal falls back to the title/abstract) rather than fetching an unpinned third-party script. The page's Content-Security-Policyscript-srcallows only same-origin scripts plus the two hash-pinned inline blocks.
Evidentum is a research aid, not a substitute for methodological judgement. AI-generated appraisals and extracted data must be verified by a human reviewer before use in any review or publication. Database coverage and API behaviour change over time; always confirm counts and records against the source.
See Licensing below and the in-app Settings → References & licences panel for the appraisal-instrument citations and their licence terms, and always consult the original source for full guidance.
If you use Evidentum in a review, please cite it (or use GitHub's Cite this repository button):
McNally A, Duckworth A. Evidentum: an AI-powered, self-contained systematic review search and appraisal workbench (v1.0.0-beta). Zenodo. https://doi.org/10.5281/zenodo.21159986
Built and maintained by AnkiAnaesthesia. Successor to the PubMed Lab Assistant.
The application code is released under the MIT License — see LICENSE. © 2026 Angus McNally.
The embedded appraisal instruments and reporting guidelines — including RoB 2, ROBINS-I V2, QUADAS-2, AMSTAR 2, the Newcastle–Ottawa Scale, the Jadad scale, PRISMA 2020 / PRISMA-S / PRISMA-P, CONSORT 2025, STROBE, CARE, AGREE II and GRADE — are not covered by that MIT licence. Each remains the copyright of its respective authors and is reproduced here, with attribution, solely for non-commercial research use under the terms each instrument's licence permits; those under CC BY-NC-ND (e.g. RoB 2, ROBINS-I V2) are reproduced verbatim, without modification. Evidentum is a non-commercial academic instrument — it is not sold and has no paid tier. Full per-instrument citations and licence terms are listed in the app under Settings → References & licences. If you reuse or host Evidentum, keep that non-commercial scope and the instrument attributions intact.