This is the first stage that cannot start until you verify your address. If screening asks you to do so, use the button in the welcome email or request another email from the app. Your existing work stays in place.
How decisions are calculated
The AI evaluates each required criterion as “yes”, “no” or “uncertain”. Code then calculates the study decision from those verdicts:
- Any “no” for a required criterion: the study is excluded.
- Every criterion is “yes”: the study is included.
- There is uncertainty and no “no”: the study goes to Review. It is never silently discarded.
Each study receives s = (number of “yes” + 0.5 · number of “uncertain”) / number of required abstract criteria. A score of 1.0 means everything is confirmed; 0.5 represents complete uncertainty.
Adjust the “Review” band
“Full screening” shows the real score distribution and two thresholds between 0.5 and 1.0.
- Check the histogram. You need search or import records and at least one required “Abstract” criterion.
- Set the lower threshold. Below it, studies with low scores inside the uncertainty zone are excluded.
- Set the upper threshold. Studies at or above it are included. Those between the thresholds go to “Review”.
- Click “Apply band and recalculate”. Decisions are reordered without calling the AI again, and the history is preserved.
A “no” for a required criterion always excludes. The band only decides how to handle uncertainty.
Resolve the “Review” queue
The queue shows one study at a time with verdicts and quotations for each criterion. Review the evidence and choose:
- Include or Exclude: records an override with an editable reason.
- Next: leaves the case undecided for later.
- Shortcuts: I to include, E to exclude and S to move on.
Resolve the queue before moving to full text.
Double screening and overrides
- Double screening. You can decide blind as an independent second reviewer. The AI verdict is then revealed and agreement is measured with Cohen's kappa.
- Conflicts. Disagreements remain in that filter and your decision prevails.
- Override a decision. Change the outcome with a required reason that remains in the audit log.
- Exclusions table. Excluded studies are grouped by criterion in the next stage. View the exclusions table.
Troubleshooting
- Too many studies are in “Review”. Raise the lower threshold to exclude more low-scoring cases. At 0.5, uncertainty never causes exclusion, which is the most conservative setting.
- A study changes between “Include” and “Review” on repeat runs. The model is not perfectly deterministic. Any variation ends in human review, never silent exclusion.
- Screening ended with an error. Run it again. Processed records do not consume allowance twice.
- The AI budget is exhausted. Manual screening and the “Review” queue remain available.