Paper Radar
A weekday arXiv radar: Jev scores every new paper against your interests, and you read the short list.
Overview
Paper Radar (github.com/Eliot5566/JEV-Paper-Radar) asks Jev to score every new paper against interests you write. It then publishes the few that matter. You fork the repo and edit radar.toml. Add a TypeSafe key, or an OpenRouter key for typesafe/jev-1.13, and turn on Pages. GitHub Actions runs on weekdays at 02:00 UTC. You get a page, an RSS feed, and optional Slack, Discord, Telegram, or email digests. Every run keeps an audit trail in data/. A public set of radars is already up at eliot5566.github.io/JEV-Paper-Radar/public/. 1kpapers is a public atlas of about a thousand papers. Paper Radar is the fork-and-run daily triage.
Problem: A weekday arXiv listing is too long to read, and keyword alerts miss papers that use different words.
Built for: Researchers who will fork a repo, write their interests in plain English, and read a short list each morning.
First indexed on Jev Directory: 2026-09-23
Creator and team
- Name
- Eliot5566
- Handle
- @Eliot5566
How Jev is used
- Role in the product flow
- One call per paper: a Noul per interest and per exclusion, a Choice for paper type, and a Noul for whether code was released. Your code applies the thresholds.
- Primitives
- ChoiceNoul
- State in
- Title, abstract, and categories only. The README says author names and affiliations are left out.
- Decision out
- Per-interest probabilities. Your code combines them with max of weight times probability, or noisy-OR, into must-read, maybe, and near-miss bands.
- Sources are deduped against the last 14 days. News near-duplicates are folded before Jev runs.
- One system call per paper sends every interest, exclusion, paper type, and code-released question together.
- Thresholds you set, or paper-radar calibrate fitted to thumbs-up and thumbs-down GitHub issues, place each paper in a band.
- An optional LLM writes a one-sentence TL;DR for the top 10 only. Jev does not write the sentence.
Paper Radar follows the shape the README attributes to TypeSafe: one small question per interest, all of them in one call, then your code decides. Relevance defaults to the max of weight times the yes probability, so one strong interest is enough. noisy-OR is there when several weak matches should add up. Exclusions are their own Nouls, written as positive statements. paper-radar check warns about interests that ask for counts or dates, because the README says Jev does not count or compare dates. Calibration reads labels you leave as GitHub issues (radar-label) and reports Brier score, expected calibration error, and thresholds for a precision and recall target. Screening mode is a different command: every inclusion criterion has to hold, one exclusion vetoes the record, and the output is a PRISMA 2020 count block. 4 reviews the README says were never used while building the tool reproduced 96.9% of included studies and removed 78% of the reading, for $0.60. The README calls that the optimistic case, because the thresholds were fitted on the same judgments they are scored against. The development set did far worse, and about a third of new PubMed records have no abstract and are never auto-excluded. 1kpapers classifies a fixed atlas with DeepSeek summaries and one Jev topic Choice. Paper Radar is a weekday filter you fork, with feeds, calibration, and that screening benchmark. Public radars on the GitHub Pages site cover AI, agents, efficiency, robot learning, neuroscience, clinical AI, and a signal feed. The AI radar's feed.xml was live on 2026-10-05.
Sourced performance claims
- README footnote, run on 2026-09-23: 50 papers in 5 seconds, 46,584 input tokens, $0.0020, 932 tokens per paper, model jev-1.13.0. The README scales that to about $0.06 a day for about 1,500 weekday papers, and says output tokens are free.Source: github.com/Eliot5566/JEV-Paper-Radar README
- README held-out Cochrane table: 4 reviews, 19,447 records, 127 eligible studies, 96.9% recall, 78.0% pooled work saved, $0.60. The README calls this the optimistic case, because those thresholds were fitted on the same judgments they are scored against. The 8 development reviews pooled 38% to 48%.Source: github.com/Eliot5566/JEV-Paper-Radar README benchmark
- README: Jev sees title, abstract, and categories. Interests combine with max of weight times probability, or noisy-OR. Screening mode uses a Noul per criterion. The tool is described as a second screener, not a replacement for a person.Source: github.com/Eliot5566/JEV-Paper-Radar README
- Public GitHub repo Eliot5566/JEV-Paper-Radar had 28 stars and 7 forks on 2026-10-05. MIT license. Created 2026-09-23. Python.Source: GitHub repo metadata, 2026-10-05
Features and stack
Features
- radar.toml interests in plain English, weekdays at 02:00 UTC via GitHub Actions
- Page plus RSS. The README says Zotero can subscribe. Optional Slack, Discord, Telegram, or email
- Sources: arXiv, bioRxiv, medRxiv, PubMed, any RSS, Reddit, and Bluesky
- Thumbs on the page become GitHub issues. paper-radar calibrate fits your thresholds
- Screening mode with PRISMA 2020 counts and a Cochrane replay in the README
Stack
- Python
- GitHub Actions
- TypeSafe Jev or OpenRouter typesafe/jev-1.13
- GitHub Pages
- RSS
Pricing: MIT code. You pay the TypeSafe or OpenRouter key. The README's 50-paper run was $0.0020, and it estimates about $0.06 for a full weekday arXiv pass. paper-radar check estimates a profile before you spend.
Links
FAQ
- How do you filter new arXiv papers with Jev each day?
- Fork the repo, write one interest per idea in radar.toml, store TYPESAFE_API_KEY or OPENROUTER_API_KEY, and enable Pages. The README says Actions runs on weekdays at 02:00 UTC. Each paper gets one call: a Noul per interest and exclusion, a Choice for paper type, and a Noul for code released.
- How much does it cost to read all of arXiv?
- The README's measured run on 2026-09-23 judged 50 papers in 5 seconds for $0.0020 (46,584 input tokens, 932 per paper, jev-1.13.0). It scales that to about $0.06 a day for about 1,500 weekday papers. Output tokens are described as free. Your own runs log cost in data/runs.jsonl.
- How is Paper Radar different from 1kpapers?
- 1kpapers is Hassan El Mghari's public atlas. A launch clip describes DeepSeek summaries plus one Jev Choice over about 1,018 papers. Paper Radar is a repo you fork. It triages new papers every weekday, publishes RSS, and includes a Cochrane screening benchmark.
- Can this screen a systematic review?
- The README's screening command treats each criterion as a Noul and prints PRISMA 2020 counts. On 4 held-out Cochrane reviews it reports 96.9% recall and 78.0% pooled work saved across 19,447 records, for $0.60. The same section says it is a second screener, not a replacement, and that the thresholds were fitted on the judgments they are scored against. That is the optimistic case.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.