AI fidelity and parallel sessions
How many records the AI judges in one prompt, how many screening sessions run at once, and who may choose what.
Records per prompt
At title and abstract screening, the AI can judge several records in one prompt. Fewer records per prompt is more careful: at Max every abstract gets a fresh context window of its own, so nothing from one abstract colours the judgement of the next. More careful means more AI calls, so a run takes longer and costs more.
| Mode | Records per prompt | AI calls compared with Default |
|---|---|---|
| Minimum | 200 | Fastest and cheapest; each record gets the least attention |
| Low | 100 | Fast and cheap |
| Default | 20 | The usual balance |
| Medium | 10 | About twice as many |
| High | 5 | About four times as many |
| Max | 1 | Twenty times as many: a fresh context window for every abstract |
Full texts are always judged one at a time.
Parallel sessions
A review can run several structured screening sessions at the same time. Each session takes records that no other session has taken (under a database lock), so every record is judged exactly once and the run ends sooner. Sessions share no context window, so they do not change the judgements. Cancel stops them all. Agent mode and full-text screening use one session.
Where to set them
On the review's AI tab: Fidelity and Parallel sessions. A job keeps the mode it started with.
Who may choose what
Each run is limited to what the person who starts it may use:
| Who | Modes | Sessions at once |
|---|---|---|
| Administrators | Every mode | 8 |
| Student plan | Default | 1 |
| Researcher plan | Low to High | 2 |
| Lab plan | Every mode | 4 |
| Institution plan | Every mode | 8 |
| Anyone else | The site's default mode | 1 |
An administrator can also change it for a person, and time-limited grants (for example from a promotion or an event) can add a mode or sessions until they expire. If you choose a mode you may not use, the run uses the closest one you may.
In the paper
When a review screened with a mode other than Default, or changed mode during the review, the methods and the AI-use statement say how many records the AI judged per prompt.