Jev Reviewer

Open a trial report in Chrome together with its supplements, protocol, analysis plan or data tables, then ask for what your systematic review extraction form needs: inclusion criteria for age, baseline age, how many were randomized, who funded it. Ask by voice, by typing, or with a questions file (CSV, TXT or a spreadsheet) or a ready-made template (trial characteristics, RoB 2, ROBINS-I, QUADAS-2, TIDieR, outcome data, transparency). Every answer is a verbatim quote with its file and page, paragraph, row or slide, highlighted where it sits. You check each one, write the value for your form beside it and tick it; the extraction table shows every study against every question, and exports one row per study or one row per quote.

Files can be PDF, Word (.docx, .doc), Excel (.xlsx, .xls), PowerPoint (.pptx), OpenDocument (.odt, .ods, .odp), RTF, saved web pages (.html), CSV, TSV, plain text or Markdown. Work is kept as projects that hold studies, and studies hold their files and answers, all stored in your browser: nothing is uploaded or kept on a server. A project's studies can be imported from a reference manager (EndNote, Zotero, Mendeley) or a database export (PubMed, Scopus, Web of Science, Covidence, Rayyan), with their PDFs, and its questions answered in every study at once.

Use it at jevreviewer.xera.ac or choxos.github.io/jev-reviewer. No key, no install. The guide walks through a whole review.

_{The first question of the tour. Watch the full tour at 1080p, silent with captions: a three-file study, a quote from a Word supplement, a Table 1 answer with its row label, an 18-question extraction form in five seconds, a question the files do not answer, the CSV export, the projects sheet and the dark theme.}

The model is Jev (TypeSafe's System One model, jev-1.13.0). Jev never writes text: it answers typed questions with probabilities. Here it only points at line ids, and code copies the quote out of the file. Nothing is paraphrased, so nothing can be invented.

How it works

 files, read in the browser          segment.js                         jev.js
 ─────────────────────────────       ─────────────────────────────      ──────────────────────────────────────
 PDF (pdf.js): text with places ──▶  sentences and table rows           pass 1, screen: one request per chunk of
 Word, OpenDocument, RTF, web        running headers dropped            2 or 3 pages of one file, all questions
   pages, text: paragraphs, tables   hyphens rejoined when safe           Choice "which line answers q?" (+ none)
 Excel, CSV: rows by sheet           reference lists flagged              Noul   "does this passage answer q?"
 PowerPoint: slide by slide          ids by file: A001, B001, C001   ─▶ pass 2, verify: per question, the best
 (textfile.js, office.js)                                                lines and their neighbors, one Noul each:
                                                                         "does line B129 itself answer q?"
                                                                         code: quotes = lines with Noul ≥ 0.5,
                                                                         adjacent lines merged, table label added

• A study is several files. The article, its appendix, the protocol, the statistical analysis plan, the data tables, a conference slide deck. Each file gets a letter that starts its line ids, every question is asked of every file, and each quote says which file it came from. The sample study is a PLoS Medicine trial with its analysis plan and CONSORT checklist, both Word files.
• Every format gives the same thing: blocks. Paragraphs, headings and table rows in reading order, whatever the file. A spreadsheet row keeps its row number and shows its cells as the spreadsheet does (percentages, decimals, dates); a slide's text keeps its slide number. Files are read by their content, so a .doc that is really RTF or a saved web page still opens. There are no libraries for this: .docx, .xlsx, .pptx and OpenDocument files are zip files read with the browser's own decompression, .doc and .xls are read from their binary formats in docs/office.js, and web pages with the browser's HTML parser.
• Select, don't generate. Candidate lines come from the files; Jev picks ids; the quote is copied verbatim with file, page or paragraph, section, and its place to highlight.
• Speculative fan-out. Each screening request carries every question against the same text, so the 18-question template costs about a cent and five seconds over three files.
• Two kinds of judgment. The screening Choice is relative (which line, if any). The verifying Nouls are absolute (does this line answer), which is what multi-row answers such as "Mean (SD)" and "Median (IQR)" under "Age" need.
• Not reported is an answer. When no line passes, the card says Not found (or Unclear, with the closest lines) instead of guessing. The strip under each question shows, file by file, how likely each stretch was to hold the answer; click it to go there.
• Projects, studies, files and answers stay in your browser. A project holds studies; a study holds its files, its answers and their highlights. The column on the left lists every project with its studies: pick one to open it, add a study or a project there, delete one with its × (pressed twice), or fold the column to a rail for more room. Come back later and the study you left opens again, at the file you were reading, with its answers. If an improved reader splits a file differently, saved quotes are found again by their exact text; one that no longer matches word for word stays, grayed, instead of being lost. Any saved answer can be deleted with its own × (pressed twice).
• A project is a review. Its questions belong to the project. Ask: Every study (in the panel; Ask in every study in the projects sheet; Ask the missing answers in the extraction table) asks each study only what it still lacks: questions it never answered, questions reworded since, and every question when a file was added after the answer (about a cent a study for 18 questions). Ask: This study does the same for the open study, and Ask again redoes one answer. An answer you have checked or annotated is never overwritten by a reworded question: it stays, under a new id, beside the new one.
• You check every answer. Every quote has three small buttons (with a mouse, the pencil and the squares wait until the quote is pointed at). The green tick checks it as the question's answer: it turns green, it becomes your answer (shown under the quotes; words you wrote yourself stay), and the other quotes fold away behind Show N other answers. The pencil opens a quote's words in the editor for your answer, where Done or Escape closes it and Cancel puts back what was there; your answer has its own pencil, and Write an answer starts one from scratch. The squares copy a quote with its file and place. The quotes themselves stay verbatim, since they are the evidence; your answer is what you edit. The Check button under a question checks it as it stands, for example when the files do not report it. The panel counts answers and checked ones, and the column shows checked/answers per study, with a green tick once a study is done. Find looks up exact words in the files at once, without Jev, to check an answer or a Not found. Keys speed this up when no field is being typed in: j and k move between answers, c checks, e edits the answer, n goes to the next one not yet checked, / goes to the question box; once every answer is checked, the next study to check is one press away.
• The extraction table shows a project's studies down and its questions across, each cell the answer's verdict, green with a tick once checked and hatched when it is due to be asked again; a cell opens the study at that answer. It exports the table (one row per study: its reference, how many answers are checked, and each question's value and quotes) and the quotes (one row per quote). Under each question it counts the studies that report it (hover for unclear, not found, not asked and checked), to spot what the studies rarely report. The rest sits in its tabs: Risk of bias (once the project has a tool's questions), Retractions and PMC, Second reviewer, and Report, where Write the methods paragraph drafts the text PRISMA 2020 items 9 and 10 ask for, from the project's own numbers (the model version, the dates, the answers checked, the cost, the agreement with a second reviewer when compared, and the reports excluded with their reasons) and Back up the project saves it.
• Reporting. The extraction table's Report tab draws the PRISMA 2020 flow diagram from the project's own numbers: the records each search found and the duplicates removed (from the deduplicator), the records Jev marked ineligible in bulk (PRISMA's "automation tools" box) and those screened and excluded (from the screening), and the reports sought, not retrieved (studies still without a full text), assessed and excluded with their reasons, and the studies included. It downloads as SVG for the manuscript, or as the PRISMA2020 R package's own template with the numbers filled in, for its Shiny app or PRISMA_flowdiagram() to redraw and edit (docs/prisma.js). The table of included studies puts your checked answers to the questions you pick in a row per study, copied as a table Word pastes, or downloaded as HTML.
• Eligibility and notes. The icons under a study's file tabs open its abstract (when it came with its reference), a note for things to remember (a companion report, a question sent to the authors), its risk of bias, and Exclude, which asks for a reason (the usual ones are offered, and the project's own). An excluded study stays, struck through: runs skip it, the extraction table lists it apart and counts the PRISMA flow (full reports assessed, excluded with reasons, included), and the exports carry the reason and the note.
• Citations. The open study's full citation (Vancouver style, with volume, issue and pages, and DOI and PubMed links) stands at the top of the right column. A study added from its files gets its reference from the DOI the article prints (Crossref, with OpenAlex for the PubMed id and the abstract), and a study named after its file takes the name reviews cite it by; a title match without a DOI is only offered, to confirm or dismiss, and a DOI can be typed in.
• Retractions and open access. Each study with a DOI or PubMed id is checked, after an import or with Check every study in the extraction table's Retractions and PMC tab, the way the retraction package for R does it: Retraction Watch through XeraRetractionTracker (with the reasons), Crossref, OpenAlex and PubMed, on exact identifiers only. A retracted study, or one with an expression of concern, is marked in the column, the table and its citation (date, reasons and notice on hover); a reinstated one or a notice is named as such. The same check finds an open access copy in PubMed Central (through the study's own PubMed record): the citation gets a PMC link beside DOI and PubMed, and Get N files beside it brings in the article and its supplementary files, only when you press it (or for every study at once from the table). The tracker and PubMed Central's copies send no CORS headers, so the relay asks them (GET /v1/retractions, GET /v1/pmc/... in server.js).
• Risk of bias. The shield under a study's file tabs, shown once the project has a tool's template questions, judges each domain of the project's tool (RoB 2, ROBINS-I or QUADAS-2), with your answers to the domain's template questions beside it (a press opens the study at one) and a line for the support for each judgment. The overall judgment is suggested from the most serious domain until you set it. The extraction table shows the judgments as a traffic-light grid and exports them in the table robvis draws its figures from (Study, D1..., Overall, Weight); backups keep them.
• A second reviewer. For independent double extraction, Send a copy to extract independently (in the extraction table's Second reviewer tab) saves the project without your answers, ticks, exclusions and notes. The second reviewer restores it in their browser, extracts, and sends back a backup; restored here, Compare with matches the studies (by DOI, PubMed id or name) and lists every answer that differs, every answer only one of you gave, and every study one of you excluded, with the agreement to report (such as "212 of 240 answers agree, 88%"). Keep mine or Use theirs settles a disagreement (for outcome data, their numbers come onto your arms of the same names), and Open goes to the study; the answers each of you gave first are kept, so the agreement stays the one before consensus, and the methods paragraph says how many disagreements were settled and how many are still open. Screening is compared the same way: Compare with in the screening sheet matches the records (by DOI, PubMed id or title, never two whose DOIs or PubMed ids differ) and gives the agreement and Cohen's kappa on include against exclude, on the decisions made independently, with a Conflicts tab where a decision settles each one (the same decision again takes the settlement back).
• Question lists grow in the app. A question typed into one study can join the project's list with Add to the project's questions. In the panel's list, pressing a question's wording changes it in place (studies that answered the old wording are asked again on the next run), ↑ and ↓ reorder it (the order is the extraction table's), × removes it, and the list downloads as CSV, to share with a second reviewer.
• Not applicable. A question that does not apply to a study (blinding in an open-label trial, say) is marked Not applicable under its answer: it counts as checked, reads "n/a" in the extraction table and "Not applicable" in the exports, and is never asked again.
• Numbers in the quotes. The answer editor lists the numbers found in the checked quote (or in every quote), citation marks left out; one press puts a number in the answer at the cursor, for outcome data such as 55.6 (12.7).
• Outcome data, arm by arm. A question can be answered with numbers for each arm instead of words: its small toggle in the questions list cycles 123 (words), M, SD (continuous: N, mean and SD) and n/N (dichotomous: events and N), or a data column in the questions file says continuous or dichotomous. Its answer then has a grid of the study's arms, named once for all its outcomes, and the numbers in the quotes fill the cell in use with one press each, moving on to the next, while the table row stays highlighted in the file. The extraction table shows the values as the answer (iCBT: 8.1 (4.2), n = 120; Waiting list: 11.3 (5.0), n = 118), and Export the outcome data writes one row per study, outcome and arm (study, authors, year, doi, outcome, outcome_question, kind, arm, n, mean, sd, events, checked): the long layout meta::pairwise() and netmeta::pairwise() read, with treat = arm and studlab = study.
• Importing references. Import references (the column's import icon, the projects sheet, or the empty desk) takes a reference list with its files: RIS, BibTeX, EndNote XML or tagged .enw, PubMed (.nbib), Web of Science, CSL JSON, or CSV and Excel with a title column. Pick the list together with its PDFs, a zip of them, or the folder the reference manager exported. Each reference becomes a study named as reviews cite it (Smith 2024, then Smith 2024b); files find their reference by the attachment names the list records, then by the DOI, the title's first words, or the first author and year in their own names. A preview says what was found and what matched before anything is created; references already in the project are left alone, except that one imported earlier without its files gets the files matched to it now (import the list first, add the PDFs once you have them). Abstracts come in from every format; a reference that arrives without its full text gets its abstract as a small text file (Park 2022 abstract.txt), so the study can be asked about until the full text comes, and every quote from it says so. The bar under the file tabs shows the abstract on request, and once a project has eight studies or more, the column offers Find a study by name, title or author. The code is docs/references.js. In EndNote, export the library as XML or RIS and add its .Data/PDF folder; in Zotero, export the collection as BibTeX or RIS with its files.
• Deduplicating search results. Before screening, Deduplicate search results (in the column) takes the exports of every database searched (RIS, BibTeX, EndNote, PubMed, Web of Science, CSL JSON, CSV or Excel) and finds duplicates two ways. The rules: the same DOI or PubMed id (certain); the same title in the same year, or titles 90% alike with the same first author at most a year apart (high). Records whose DOIs differ are never duplicates, and a protocol, an erratum, a reply or a notice is not the duplicate of the article it shares a title with. Jev: every pair the rules consider, down to titles 75% alike, is asked whether the two records are the same publication (a Noul, 20 pairs a request, well under a cent for a hundred pairs). A pair both call duplicates, the rules with high certainty and Jev at 0.9 or more, is removed; a pair only one of them calls a duplicate is listed for you to mark Same or Different; the rest stay apart. When Jev cannot be asked, identifier matches are removed and the other pairs listed. Each group of duplicates keeps its most complete record, with what it lacks taken from the others; Download the list (RIS) writes the records that remain, for any screening tool, and Download the log says what was removed, as what, and why (docs/dedupe.js).
• Screening titles and abstracts. Screen titles and abstracts (in the column, the funnel among the open project's tools, or Screen these records after deduplicating) takes a project's search results and its eligibility criteria, one per line. Jev reads each record's title and abstract and answers, for each criterion, whether the record meets it, does not, or does not say (a Choice; twenty records to a request, about two cents for a thousand records with three criteria; for thousands of records, paste your own TypeSafe key in Settings). The likeliest records come first, and a reviewer decides each one with a press or a key (j and k move, i includes, m marks maybe, x excludes, the same key again takes it back); Jev disagrees lists the decisions that go against Jev's. Records Jev finds clearly ineligible (0.95 or more that a criterion is not met, with an abstract to judge from) can be excluded in bulk, recorded as excluded on Jev's judgment and said so in the methods paragraph. The included records become the project's studies, each with its abstract to ask until the full text comes (import the list again with the PDFs, or take an open access copy from PubMed Central); the decisions download as CSV and the included records as RIS (docs/screen.js). The counts (records found by each search, duplicates removed, records screened and excluded) are kept with the project and its backups.
• Backups. Everything lives in the browser's IndexedDB (docs/library.js), on this device, for this site. Back up the project (in the column, the extraction table and the projects sheet) writes it to one zip file with its studies, files, answers, values and checks, plus its two extraction sheets as CSV to read without the app (docs/backup.js); Back up all projects does every project, and Restore a backup adds them back as new projects, in this browser or another one. The sheet also says how much storage the site uses, and the column says when the open project was last backed up (louder when it never was, or changed a week or more since). When the browser's storage for the site is full, the app says so and keeps the work on screen instead of losing it quietly; a file it cannot keep is not added. A study open in two tabs stays in step: a change in one shows in the other, which never saves an older copy over it.
• What it cost. The band at the bottom counts requests and dollars this session and in all (in this browser), and each project keeps its own total, shown in its extraction table and kept in its backups: the number a methods section or a budget asks for.
• Voice uses the browser's speech recognition. Each finished phrase becomes a question; a small Jev check (is_request, 0.59 to 0.98 for questions, about 0.01 for side talk) drops chatter such as "hmm let me see". Say next or previous to step through quotes.

All question texts and thresholds live in docs/jev.js, in one place, like the constants.js of jev-voice-browser. Both keep code in control and ask Jev narrow, typed questions; the voice browser acts on partial speech and drives Playwright, while here the hard part is reading papers well enough that every quote is a clean sentence or table row.

Run it on your computer

Requirements: Node 20 or newer. Chrome or Edge for voice.

git clone https://github.com/choxos/jev-reviewer.git
cd jev-reviewer
cp .env.example .env              # paste your key from https://console.typesafe.ai/keys
npm start                         # http://localhost:8787
npm start -- path/to/paper.pdf    # opens that paper on start

The server has no dependencies. It serves the app from docs/ and relays requests to TypeSafe with the key from .env, so the key never reaches the browser. It listens on 127.0.0.1 only and refuses requests from origins and hosts it does not know. Open several files at once with Choose files, add more with Add file, or drop them on the page. The file tabs wrap to a second line; a study with more files than the tabs show (a trial with its PubMed Central supplements, say) also gets a list of every file beside Add file, naming the open one in full; on a phone that list stands for the tabs once there are two files, with a × for the open one. Links in a PDF work as in a PDF reader: a web address opens in a new tab (an address only printed, such as a page footer's DOI, is found in the text and linked too). A link within the file shows what it points to without leaving the page: hovering or focusing a citation shows the reference in a popover (the entry from its number to the next, its DOI clickable; a supporting file or a caption the same way), a press keeps it open, and Go to it goes to its place.

The hosted copies

• https://jevreviewer.xera.ac: the full app with its relay.
• https://choxos.github.io/jev-reviewer/: the same page from GitHub Pages (published from docs/), sending its questions to the relay on jevreviewer.xera.ac.

Files are read in the browser and never uploaded; only their text and your questions go to TypeSafe. Projects are saved in the browser you use, per site: the two copies keep separate projects, and clearing the site's data deletes them. The hosted copies count visits with Google Analytics (docs/analytics.js): page views only, with a fixed page title and address, so no study name, file address, file or question reaches it. It is not loaded on localhost or in automated browsers. The TypeSafe API does not accept requests straight from web pages, so both pages go through server.js, which adds a shared TypeSafe key on the server. To keep a public key affordable, each address can send only so many requests a second (enough for a batch), and the server stops spending the shared key after DAILY_TOKEN_BUDGET input tokens per day. A visitor who pastes their own key in Settings uses their own quota and is not capped. Screening thousands of records is a job for your own key: 10,000 records with three criteria take about 5 million input tokens (about $0.20), which is likely much or all of the shared key's day.

Hosting your own copy

server.js is the whole back end. Run it with a TypeSafe key in .env (see .env.example) behind any HTTPS reverse proxy, and list the sites allowed to use it as their relay in ALLOWED_ORIGINS.

Questions files

Upload a list under the question box (or Upload questions in a project's row in the projects sheet) takes any of these, and Replace the list swaps it for another; once there is a list, it and Templates shrink to icons. The panel lists the project's questions, Ask: This study asks the open study what it lacks, and Ask: Every study asks every study of the project.

Templates adds a ready-made list to the project (questions already on it stay once), or downloads it to edit. Each asks for the quotes a reviewer needs, in plain words; the judgments stay yours.

template	questions	file
Trial characteristics	18: design, setting, age criteria, inclusion and exclusion criteria, number randomized, baseline age and sex, intervention, comparator, primary outcome, follow-up, sequence generation, allocation concealment, blinding, attrition, funding, registration	questions-template.csv
Risk of bias in randomized trials (RoB 2)	14, by domain: randomization, deviations from the intended interventions, missing outcome data, measurement of the outcome, selection of the reported result	questions-rob2.csv
Risk of bias in non-randomized studies (ROBINS-I)	10: confounding, selection, classification of interventions, deviations, missing data, measurement, reporting	questions-robins-i.csv
Diagnostic accuracy (QUADAS-2)	12: patient selection, index test, reference standard, flow and timing	questions-quadas2.csv
Intervention description (TIDieR)	12: what, why, materials, procedures, who, how, where, when and how much, tailoring, modifications, fidelity, comparator	questions-tidier.csv
Transparency and reproducibility	7: competing interests, funding and the funder's role, registration and its timing, the protocol, data sharing, code sharing, the reporting guideline	questions-transparency.csv
Outcome data for meta-analysis	10: time points, the measure and its direction, numbers analyzed, means and standard deviations, medians, change or final values, events, the effect with its confidence interval, adjustment, clustering; the means and the events are answered arm by arm	questions-outcomes.csv

Your other projects' lists are offered there too, to start a new review from an old form.

• CSV with a header: a question (or query) column, optionally an id column, a data column (continuous or dichotomous for numbers arm by arm) and a guidance (or notes) column: the coding manual's rule for the question, shown under it in every study (the trial characteristics template has them). An id given twice becomes age, age_2, so two questions never share their answers.
• CSV without a header: id,question rows.
• A spreadsheet (.xlsx, .xls, .ods, .tsv): its first sheet, read like a CSV, so an extraction form kept in Excel loads as it is.
• TXT: one question per line; lines starting with # are comments.

A questions file is kept with the project, for all its studies.

The download icon in the panel writes the open study's answers, one row per quote, best first: study, id, question, verdict, best_score, file, location, section, excerpt, excerpt_score, line_ids, checked_quote, checked, note, asked_on, model, then the study's reference (authors, year, title, journal, doi, pmid), empty for a study not imported from a reference list. location reads p. 4 in a PDF, para. 129 in a Word or text file, row 12 in a spreadsheet and slide 3 in a slide deck; checked_quote marks the quote you checked as the answer; checked and note (your answer) are yours; asked_on and model say when and with which Jev version the answer was found. A question with nothing found gets one row with an empty excerpt, so the sheet always has every item. The extraction table's Export the quotes writes the same sheet for every study of the project, and Export the table writes one row per study: study, its reference, checked (such as 12 of 18), then for each question your answer and its quotes (only the checked one, once you checked it). The files carry a byte order mark, so Excel reads them as UTF-8 (quotes are full of ≥, ± and µ).

Tests, measurements and the tour

npm install          # dev only: pdfjs-dist for the tests, playwright-core for the tour
npm test             # segmenter, every file format, reference lists, requests, policy, CSV, backups, server and relay,
                     # retraction and PubMed Central lookups, deduplication, screening, the PRISMA diagram
npm run live         # real API: 9 questions on the sample study (about half a cent)
npm run live -- paper.pdf supplement.docx --questions my-form.csv --debug
npm run tour -- https://jevreviewer.xera.ac   # writes documentation/tour.mp4 and tour.gif

Measured on the sample study (a 17-page article PDF, its 12-page analysis plan and its CONSORT checklist, 712 lines to search) in September 2026:

run	requests	time	cost
1 question	10	1.2 to 2 s	$0.0016
9 questions	17	2.3 s	$0.0052
18 questions (template)	27	4.6 s	$0.0101

Screening chunks of 12,000 characters gave the same answers as chunks of 7,000 with a third fewer requests. Across the sample study, the age criterion (the article and the analysis plan's "18 years and older"), baseline age (text and Table 1 rows), sample size and missing data (the analysis plan and the article), randomization, primary outcome and funding came back as the top quotes; the CONSORT checklist item that points to the missing-data paragraph came back too; "dose of metformin" came back Not found. Treat these as spot checks, not a validation study: check quotes against the files before they enter your review.

The file readers are tested on test/fixtures: one report, one workbook and one slide deck, each written by LibreOffice as .docx, .doc, .odt, .rtf, .xlsx, .xls, .ods, .pptx and .odp (and by macOS as .doc and .rtf), must all give the same blocks, with footnotes, tracked deletions and field codes left out and numbers shown as their cells show them.

The tour is recorded by record-tour.mjs against the live site, so every answer in it is one the app gives. Playwright drives Chrome at a device scale of 1.5, which draws the 1280 by 720 layout with 1920 by 1080 real pixels, and a Chrome screencast saves each frame as it is painted; ffmpeg joins the frames with their own timing. Headless Chrome has no pointer and no microphone, so the recorder draws a pointer and captions; voice is mentioned, not shown. It warns and exits with status 1 when a step does not happen: a question that never comes back, a best quote from the wrong file, a Table 1 answer without its rows, a template run that is not 18 of 18, a CSV with too few rows, or a projects sheet without the study.

Design

The page follows the design of game-of-life: warm oat paper, one vermilion accent, Instrument Serif for headings, Geist for everything else, pill controls and a thin data band at the bottom. It follows the system's light or dark theme until you pick one with the switch in the header. All colors and fonts are tokens in docs/tokens.css; the fonts are served from docs/fonts under the SIL Open Font License.

Limits

• Scanned PDFs have no text layer: run OCR first (the app warns when it finds almost no text).
• Figures are images, so their contents are not searched; captions are. Images, charts and equations in Office files are skipped too, and so are speaker notes.
• Not read: PowerPoint 97-2003 (.ppt), Word and Excel 95 or older, password-protected files (save an unprotected copy, or a PDF). A spreadsheet is read up to 5,000 rows. Number formats are applied without their literal text, so 54.2 kg in a cell formatted 0.0 "kg" reads 54.2.
• A web page is read from its main content and its first heading on; the Node scripts (npm run live) read every format but web pages, which need the browser's parser.
• An EndNote library file (.enl) is not read directly: export it as XML or RIS. A reference list's attachment paths only help to match files; the files themselves have to be picked, since a web page cannot open paths on the computer.
• Projects live in one browser on one device. Clearing the site's data, or a private window, loses them: download a backup to keep them, or to move them to another browser or site.
• PDF text order follows the file's content stream, which is reading order in publisher PDFs (checked on single and two-column layouts). Unusual layouts can merge or split sentences.
• English works best. Thresholds were tuned on jev-1.13.0; re-check them if you move the model version.
• In Chrome, speech recognition sends audio to Google.

Layout

docs/index.html      the page (GitHub Pages serves docs/)
docs/guide/          the guide: a whole review in Jev Reviewer, for readers and search engines
docs/robots.txt      with sitemap.xml, manifest.json and the icons, what search engines and
                     phones read about the site
docs/app.js          projects and studies, viewer, highlights, questions by voice, text or file, export
docs/library.js      projects, studies, files, answers and screened records in IndexedDB
docs/backup.js       backups: projects with their files in one zip, and restoring them
docs/references.js   reference lists (RIS, BibTeX, EndNote, PubMed, Web of Science, CSL JSON,
                     tables) and matching the files that come with them
docs/lookups.js      retractions (Retraction Watch, Crossref, OpenAlex, PubMed), PubMed Central
                     copies, and a study's reference from the DOI its PDF prints
docs/dedupe.js       duplicates across search exports, by rules and by Jev, and RIS out
docs/screen.js       title and abstract screening: Jev's judgment of each criterion, suggestions
docs/prisma.js       the PRISMA 2020 flow diagram from a project's numbers, as SVG and as the
                     PRISMA2020 package's template
docs/segment.js      PDF text and other files' blocks to sentences and table rows, with places
docs/textfile.js     every format but PDF as blocks: zip-based Office and OpenDocument files,
                     RTF, web pages, CSV and TSV, text and Markdown
docs/office.js       Word 97-2003 (.doc) and Excel 97-2003 (.xls), and spreadsheet number formats
docs/jev.js          questions, thresholds, two-pass requests, result policy, which answers a
                     study still lacks, CSV in and out
docs/tokens.css      colors, fonts, spacing, motion; docs/styles.css uses only these
docs/theme.js        the light and dark switch, and the projects column's first state
docs/analytics.js    Google Analytics page views on the hosted copies
docs/samples/        the sample study (CC BY 4.0) and the question templates
server.js            app server and TypeSafe relay, local or on the server (no dependencies)
record-tour.mjs      the tour recorder; documentation/ holds its video, gif and the screenshot
test/                node --test suites, their fixtures, and the live check

The sample study is Johnson E, Hyde A, Corrick S, et al. (2026) Effect of a digital intervention on mental health symptoms in adults with chronic conditions: A three-arm randomized controlled trial. PLoS Med 23(8): e1005198, doi:10.1371/journal.pmed.1005198, with its S1 File (statistical analysis plan) and S1 Checklist (CONSORT 2025, Hopewell and colleagues), published under CC BY 4.0.

MIT license.