Method
Why it holds up
A systematic review has to be repeatable by someone who was not there.
The published method tells you what a review must do and what it must report. It does not tie that report back to the work: the methods section is written afterwards, by the same people, from notes and spreadsheets. Julie's record is produced by the run itself, while the work is done: the standards applied as published, the additions declared as additions, and your decisions recorded under the name you gave.
EXAMPLE · A GUIDELINE UPDATEThe F2F consensus meeting is not the last reading this work gets. After the panel there is a journal, and after the journal a methods editor who was not in the room.
-
Checked against PRISMA 2020, item by item
The chain has been compared with the official checklist and flow diagram, item by item. What you get from a run is a methods report built from that run's own record, and where it cannot produce something the checklist asks for, it names the item and hands it back to you rather than reading as complete.
-
CEBM and GRADE as published, not a simplified reading of them
Ratings follow Oxford CEBM (March 2009) or classic GRADE, chosen per project. Every rating includes its source support and methodological rationale. Any project-specific additions are identified separately from the published standard.
-
Decisions are recorded under a name
Julie asks who is deciding before she writes anything down, and that is the name the record carries and the methods report reads from. Her own name never goes into those fields, because a machine name there would misstate how the review was made.
-
It stops rather than guesses
Missing or inconsistent information is brought to your attention before the work continues. You can see what needs resolving and which decisions remain yours.
-
Keeps the review current and traceable
Your results remain connected to the material and decisions behind them. When that material changes, Julie shows you which results need review.
-
AI use, in a form you can publish
Every paid stage records, while the work is done, which model ran, at which stage and under whose approval. Your AI-use statement is built from that record, not reconstructed afterwards.
What it costs, and who decides
Setting the review up, checking what is missing, seeing where it stands, and estimating what a stage would cost. None of it spends anything.
Drafting the criteria, screening and its validation sample, the full-text eligibility check, extraction and risk-of-bias appraisal, a second look at ratings marked uncertain, linking the evidence to the recommendations, the brief and the drafting. Rating itself spends nothing.
Every paid stage is estimated first, runs only under a ceiling you set and an approval in your name, and every paid run is recorded under your approval.
Julie is in pilot use at the moment, so there is no licence fee: you pay the processing costs directly, and nothing on top. What a stage comes to depends on how many records you screen and how many full texts you process, which is why none of them starts without your approval.
What Julie does not do
- You approve and run the searches. Julie helps draft strategies for PubMed, Embase and CENTRAL from your questions. You settle the final strategies and run them in the databases. The full strategies and search details still belong in the manuscript.
- She does not make the decisions that are yours. No inclusion call, no proposed meeting outcome, no rating in your place.
- She does not quietly substitute one instrument for another. The instrument follows the study design. Where a question could not be answered the way the instrument intends, the answer is named and flagged rather than quietly recoded into a milder one.
- She does not approve spending. The ceiling and the name are yours.
- She does not write clinical sense. Whether a sentence is clinically right, or accurate but misleading, is something you read.
- She does not replace expert reading. The checks establish that the material is internally consistent and traceable. The scientific judgement stays with you.
What Julie brings to your workflow
| Your question | What Julie brings |
|---|---|
| I already find papers with Consensus. Why Julie? | Develop and approve search strategies for PubMed, Embase and CENTRAL, then work through screening and evidence assessment towards a manuscript draft. |
| Elicit and SciSpace already offer screening and extraction. What should I compare? | How the workflow handles unresolved decisions, distinguishes evidence levels from certainty, and carries your team’s agreed recommendations into the Discussion. |
| What is Julie’s particular focus? | Clinical evidence synthesis and guideline work, with supporting and contradicting evidence linked to the recommendations your team accepts. |
What does the benchmark actually measure?
A headline accuracy score needs context. Elicit’s May 2026 extraction benchmark reported 95.6% correct on methods, participants and interventions, with outcome extraction excluded. SciSpace’s 200-query benchmark measured search relevance using AI judges; it did not establish complete study retrieval or GRADE accuracy.
These are useful measurements of specific tasks. For a systematic review, the questions go further: were eligible studies retained, were critical results extracted correctly, and were certainty judgements justified?
If you are running a systematic review, sitting on a guideline panel, or working in an evidence synthesis unit, tell us what you are working on.