Six stages. Every claim traceable; every draft reproducible.
BibTeX, RIS, a folder of PDFs, or a pasted list of DOIs. Each paper is read to its abstract; full-text is parsed when available. Duplicates are collapsed by DOI.
Author, year, venue, citations in and out, section structure. Nothing inferred; unreadable fields surface as gaps for human review.
Hierarchical clustering on abstract-level embeddings. Themes emerge bottom-up. Four themes per twelve papers is typical; the silhouette score is shown alongside.
A draft paragraph per theme. Every sentence carries an inline {{citeKey}} marker. No sentence ships without at least one marker; hallucinated citations do not appear.
Under-cited intersections between themes surface as gap candidates. These are potential research directions, not conclusions — they are flagged, not asserted.
The review is signed with a pipeline version and a corpus hash. Rerun the same corpus against a new pipeline; see what changed.
No sentence in the synthesis exists without a paper behind it. That is the single rule.
Read a review