Community · E2 · artifact verified
Extract clinical review data as verbatim quotes
A browser app for systematic-review extraction: Jev never writes the answer - it points at line ids in trial reports and supplements, and code copies the quote out with its file, page, row, or slide, highlighted where it sits.
01 · Role in the system
What Jev does here
Files are read and segmented entirely in the browser - PDF, Word, Excel, PowerPoint, OpenDocument, saved pages, CSV, text - into sentences and table rows with stable line ids. Pass one asks, per chunk of two or three pages, a Choice of which line answers each extraction question (none is an option) plus a Noul of whether the passage answers it at all. Pass two re-verifies the best lines and their neighbors, one Noul each; quotes are the lines that clear 0.5, adjacent lines merged, table labels attached. Because Jev only points and code copies, nothing can be paraphrased and so nothing can be invented. Templates ship for trial characteristics, RoB 2, ROBINS-I, QUADAS-2, TIDieR, and outcome data, and studies import from EndNote, Zotero, Mendeley, PubMed, Scopus, Web of Science, Covidence, or Rayyan.
Point, never write: the design that forbids invention
Systematic-review extraction punishes paraphrase, so this app removes the possibility entirely: Jev never produces the answer, it selects line ids. Documents are segmented in the browser into sentences and table rows with stable ids, a first pass asks which line answers each extraction question per few pages, and a second pass re-verifies the best candidates and their neighbors. Code copies the winning line out verbatim, with file, page, row, or slide attached.
Because quote text is always copied and never generated, the failure modes shrink to wrong selection — visible, checkable against the highlight — rather than confident fabrication. A question the documents do not answer comes back as unanswered instead of inferred, which for evidence synthesis is the only acceptable default.
Methodology templates make it a tool, not a demo
The extraction templates ship for the instruments working reviewers actually use — RoB 2, ROBINS-I, QUADAS-2, TIDieR, and trial characteristics — and studies import from the reference managers and screening platforms the field already runs on. That is the difference between a technology demonstration and software a review team could pilot.
The boundaries are practical rather than scientific: everything lives in browser storage with no server, so clearing site data loses projects unless exported, and the hosted demo depends on the author's own hosting. The extraction quality itself is bounded by an honest rule — the tool can only report what the files state.
02 · Control boundary
Where Jev sits
Two passes over line-id-indexed segments - a screening Choice per question per chunk, then a verification Noul per candidate line; quote text is always copied by code, never generated.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Answers are bounded by what the files state; a question the documents do not answer comes back as such rather than inferred.
- Projects live in the browser's storage with no server, so clearing site data loses them unless exported.
- The hosted demo's availability depends on the author's own hosting.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.