Replace Jev Deep Research with JevGPT in Built with Jev

[sunyasheng]

Passes Built with Jev: JevGPT calls hosted Jev in its released evidence-finding path and publishes code, an Apache-2.0 license, the authors' own measurements, and reproduction commands. Replaces one entry in place, so the Built with Jev count stays at 41.

Scope note: the Choice / Noul locator described here is the BrowseComp-Plus question-answering path. For NeuCLIRBench ranking, Jev rates documents with Score and keeps pages with Noul (jev_document_ranking_policy.ts).

Disclosure: I am Yasheng Sun, first author of the JevGPT paper and maintainer of github.com/sunyasheng/JevGPT. This replaces the existing Jev Deep Research entry: that repository is no longer maintained and its README now points to JevGPT. Development and release preparation used coding-agent assistance, and this pull request was prepared with Claude Code.

JevGPT is the code release for the paper "JevGPT: Thinking Fast and Slow in Autonomous Research" (arXiv 2610.18272). The entry follows the paper entry format: the arXiv link, the quoted title without the JevGPT: prefix, then arXiv, Code, Daily Papers and Website badges. The Code badge carries github.com/sunyasheng/JevGPT, which separates this entry from the unrelated idlivada/JevGPT and Bewinxed/jevgpt. Because the entry now carries an arXiv badge and a Daily Papers badge, npm run counts moves the papers badge from 26 to 27 and the daily papers badge from 20 to 21. Entries (170), with code (127) and every section count stay the same.

GPT-5.5 plans the research and integrates findings. In BrowseComp-Plus question answering, for each retrieved document region, Jev answers typed questions in parallel: Choice picks the numbered source line or range that answers each query, and Noul estimates whether any line in the region addresses it. Further Choice and Noul gates decide whether to read further, expand the ranking, or stop scanning. Code returns the original passages with source coordinates to GPT; Jev never writes text. It runs on the Pi-Serini agent runtime.

• Typed Choice and Noul questions
• Parallel Jev API calls
• Read, expand and stop gates
• Results table
• Reproduction guide
• Run outputs on Hugging Face

The numbers in the entry are the authors' own measurements, from the README results table and Tables 1 and 2 of the paper: 90.48% accuracy on 830 BrowseComp-Plus questions, at 91.7 s and $0.6681 per question, and nDCG@20 0.5390 on 105 NeuCLIRBench topics, at 79.3 s and $0.5838 per topic. Cost covers both the GPT and the Jev API calls.

Validation: npm run lint (65 rules active, no problems), npm run counts (15 counts checked, all match), npm run links:new (the 4 added URLs answer 200, no dead links), and node scripts/check_authors.mjs origin/main..HEAD (exit 0; one warning for the Co-Authored-By trailer, which CONTRIBUTING says never reaches the contributors list).

• The entry passes one of the four tests, and the PR says which.
• The paper or primary page is linked, and the arXiv id in the badge matches it.
• Official code or model links are included when they exist.
• The description is one clause, concrete, and at the level the source supports.
• npm run counts and npm run links:new pass.
• Section placement is right and the entry is not duplicated elsewhere.

🤖 Generated with Claude Code