Add LitJev (Jev-schema decision layer for the Qwen family)
Adds LitJev to the decode section: shared prefill + forked KV/recurrent state, one cached branch per question, option logits read from lm_head on any Qwen checkpoint (default Qwen3.8-27B). Uses the exact Jev /v1/systemone schema (choice/score/noul), with a playground, MMLU-Pro eval, optional temperature calibration, and Doom/chess from screenshots.
GitHub: https://github.com/zhengxuyu/litjev
No announcement post on X yet. Star count is a 2026-09-18 snapshot.
Added the announcement post on X (https://x.com/yuzxfred/status/2100652136878981337), the author handle and a quote from it. Likes are a 2026-09-18 snapshot.
Added a blog link to the write-up: an X article (in Chinese) on how Jev reads decisions from one forward pass, LitJev's design, Jev's limitations, and what the reproduction does and does not cover: https://x.com/yuzxfred/status/2100896621173047367
Merged — thanks! LitJev is in the tracker under decode.
Verified github.com/zhengxuyu/litjev and both X posts (the announcement and the linked write-up).
Two corrections while merging:
- Stars 1 → 4, and I added the missing view count (125) from the announcement post.
- Description length trimmed to 256 chars to match the card ceiling used across the tracker.
There was a merge conflict with a concurrent PR (both appended after the same entry); resolved by keeping both additions and the current star count on the shared entry. Your commit is preserved in history as the second parent of the merge.
The Chinese quote renders fine on the card alongside the rest.