WHAT NEW IN X11.03 TUE

X 上的 AI 合集档案 · 22 个合集 · 65,263 条

jevVIDEO

把 Jev 换成本地 1.5B 模型守 Claude 写盘

作者用本地小模型加 Jev 阈值替代云端裁判,拦截 Write/Edit 与密钥泄露,月 1.8 万次决策零成本。

Asteri
@Asteri_eth2026.10.03收录原推发布于 2026.10.03
Asteri @Asteri_eth原推发布于 2026.10.03
在 X 上打开原推

I STOPPED SENDING MY REPO TO AN API THE DAY I MOVED JEV ONTO MY LAPTOP I used to let every Write, Edit and bash command go out to a hosted judge, 600 times a night -> now Claude only builds. a 1.5 GB model on 127.0.0.1 makes the calls, code enforces them, and nothing leaves the disk here's what's inside the folder: ◦ the input > hooks/pre_tool_use.sh - catches every Write and Edit before it lands on disk > state.json - the command, the folder, the diff. about 4,000 tokens, never uploaded ◦ code goes first (rules/) > http://regex.py - known key formats: AWS, Stripe, private keys. costs nothing > thresholds.yaml - deny at 0.8, ask at 0.3. a missed secret costs more than one extra prompt ◦ the local judge decides (server/) > von/ - 395M ModernBERT, Apache 2.0, about 18 ms a call on a Mac > bound to 127.0.0.1 - these servers ship without auth > pinned to a commit hash, never main ◦ the router (route/) > http://choice.py - picks the cheapest Claude before claude -p starts > unsure means a tier up, never down ◦ Claude writes (workers/) > gets the deny reason back and knows exactly what to fix on the next try > the only place in the folder where text gets generated ◦ the guards > fail closed - server down means ask, never allow > shadow.log - a week of Von next to Jev before it decides anything. Jev thresholds don't carry over ◦ the catch > Von 93.5% vs Jev 97.2% on its own 78 cases. narrow questions stay local, fuzzy ones go out 18,000 decisions a month for $0. 18 ms a call Claude builds, a laptop decides, code acts

这条是引用着另一条写的:x.com/asteri_eth/status/2104901781725598103

x.com/asteri_eth/status/2106089034942746762采集于 2026.10.03 10:03