Screen · local citation screening

Screen titles and abstracts, then assess retrieved full texts. Review independently, reconcile disagreements, and keep a dated history with your project backup. Local by default; AI assistance is optional.

What to do on this page — decide which records to include

Read each record and decide whether it belongs in your review. This is a human judgement the tool cannot make for you.

  1. Import records (from Search or a file), or click Load example.
  2. For each card, read the abstract. Where a journal (PubMed) abstract exists it is shown in preference; records that only have a ClinicalTrials.gov brief summary are labelled as such — use the ClinicalTrials.gov / PubMed / DOI buttons to open the source and check it.
  3. Press i (include), e (exclude) or m (maybe); add a reason for excludes.
  4. In dual-reviewer mode, resolve any conflicts so each record has one agreed decision.
  5. When done, send the included records on to Extraction.

The tool does

Flags likely duplicates, an offline RCT-likelihood score and your search-term matches — these are hints.

You must check

Every include / exclude decision, and that the final included set is correct, before moving on.
This browser only — export regular backups

0 of 0
Decision history

Dated local history; editable backup files are not signed or authenticated. Older imported decisions have no reconstructed history.

    Local by default — with AI assist off, records never leave your device. The optional bring-your-own-key path is the only exception: it sends the title and abstract of each screened record to the AI provider you choose (which is why connect-src allowlists those three hosts), and your key is stored only in this browser. Dedup: DOI exact + trigram-Jaccard title similarity (≥0.85), with a multi-field assist (relaxed title threshold when first-author surname AND year both match) that catches reformatted titles and same-work/different-DOI duplicates. Candidate pairs come from blocking + a sorted-neighbourhood window, to reduce pair comparisons; this upgrade does not claim measured 100k-record throughput; earliest-imported record wins. Agreement: Cohen's κ over records both reviewers decided. Effective decision in dual mode: adjudication, agreement, conflict, or pending until both reviewers decide. PRISMA push writes the prisma-flow-v1 key. Relevance is a transparent term-frequency score over your own include/exclude terms — no cloud model, no black box.