Screening source
| Imported | Database | Records | Duplicates | Delete |
|---|
Pilot screening
Learning from pilot corrections
Confirm the criterion and correction for each disagreement. All corrections are saved; relevant rules and examples are selected for each subsequent request. The screening model uses this feedback; model weights are not changed. Approval freezes the feedback for formal screening.
Confirmed correction rules
Main screening
Latest AI readings
Method and research references
Complete at least one pilot round and approve it before formal screening. Additional rounds are optional. One unmet criterion supported by original text is sufficient for exclusion; missing information alone is not. When readings directly disagree on every supported exclusion criterion, the reviewer resolves the conflict. Clear exclusions are not audited by default; the user can opt into an exclusion sample or a full audit.
Human check defaults to conflicting and uncertain records. Agreement statistics in the pilot inform the reviewer's approval; they do not mandate additional rounds or guarantee recall. Explicit title evidence can support exclusion without an abstract. Reasons are retained in all result export formats.
- Cochrane Handbook v6.5.1, Chapter 4 §4.6.4 — pilot the eligibility criteria on a sample of reports before screening.
- Flemyng et al. (2025), Cochrane, Campbell, JBI and CEE position statement on AI in evidence synthesis — pilot or calibrate an AI tool within the review, keep human oversight and report its use.
- Tran et al. (2024), Ann Intern Med — high-sensitivity LLM screening rules.
- Cao et al. (2025), Ann Intern Med — optimised criteria-based prompts reach about 98% sensitivity; zero-shot prompts much lower.
- Xie et al. (2026), J Evid Based Med — meta-analysis: few-shot or reasoning prompts improve sensitivity.
- Affengruber, Gartlehner et al. (2026), BMC Med Res Methodol — false exclusions rise when abstracts lack information; uncertain records go forward.
- Garritty et al. (2024), BMJ: updated Cochrane rapid review methods guidance — when one person screens, a second checks all exclusions.
- Gartlehner et al. (2025), Cochrane Rapid Reviews Methods Group position statement on AI — do not fully automate a step; use AI with human checking of exclusions.
- Callaghan and Müller-Hansen (2020), Syst Rev — statistical checks of missed records in a random sample of unscreened or excluded records.
- PRISMA 2020 explanation and elaboration, item 8 — report the automation tool, its role and how its risk of missed studies was assessed.
- Covidence: Maybe votes — uncertain citations advance to full text.