User guide

From "New project" enter the research data (working title, topic, research type, manuscript language), then the author name, then the journal and template. Everything except title, journal and template then runs automatically.

The platform drafts the research question and scope in an understanding memo, then a search protocol (queries, inclusion and exclusion criteria). Any deviation during execution is recorded as a dated, locked protocol amendment.

OpenAlex, Crossref, PubMed and Europe PMC are searched, then records are screened automatically with append-only decisions (corrections are added as linked overrides). After screening the platform pauses at a checkpoint showing sources, inclusions and cost, and continues after you acknowledge it.

Open full texts are fetched via Unpaywall and Europe PMC only when an open licence is available; closed sources are used as bibliographic references without claiming they were read.

Each piece of evidence is extracted with the span that supports it, and support is checked per claim–evidence pair. A link to a source alone does not make a claim supported.

The solution is built from explicit relations between evidence. If a core relation has a gap the solution is not accepted and is automatically reworked up to three times; a rejected solution never passes.

Figures and tables use separate counters in document order. Conceptual papers require an activity-system drawing and an operational contribution drawing with a traceability table; a render failure blocks completion.

You can upload a Word template (DOCX). Only validated styles and page settings are applied, never its body content. Broken, hidden or unsafe templates are rejected; DOTX files are download-only.

A fully automated scientific and methodological review runs, then an editable Word file is exported with a DRAFT watermark. The file hash is computed server-side and its text matches the stored snapshot.

• "Draft": automated output not finally approved; review it before submission.
• "Untested": a function or item not yet proven by an actual test, so it is never shown as passed.
• "Needs re-check": something changed after the last check (e.g. evidence or template), so it must be re-checked before approval.

• The post-screening checkpoint awaits your acknowledgement.
• The project's AI cost budget is exhausted.
• An external source or step failed; a message shows the cause and next step.
• A decision needs the project owner.
• Runs execute on the server under a lock that prevents duplicates and continue with the browser closed; a background worker resumes them about every two minutes, and you can resume from the project page once the cause is handled.

The cost of each AI call is reserved before it runs; settled spend afterwards is authoritative and shown on the project page. While an automated run is active, manual writes to project data are refused until it ends.

Implemented:

• All stages above, from creation to Word export.
• Policy pages and contact form.
• Arabic and English interface.

In progress or not yet tested:

• Full result-page coverage for every search query.
• Visual check of figures in all three templates.
• Background worker load test on many projects.
• Self-service account deletion (currently on request via Contact).