Local 1.5GB model replaces Jev for Claude Code judgments
An open-source setup makes Claude Code's typed decisions locally in about 18ms, covering secret detection, model choice and read-only checks without uploading data, and falls back to Jev or a human for fuzzy judgments.
this is a f*cking gift two days after Jev launched, open source started shipping its own versions. now your Claude Code loop can make every judgment call on your own laptop in ~18 ms and upload nothing Claude only does what needs code. everything else is a typed decision that stays on your machine: > secret in the edit? every Write and Edit checked before it touches disk > known key? AKIA, ghp_, sk- caught by regex, the model never even sees them > which Claude? haiku, sonnet or opus, one Choice before the run starts > read-only? short bash commands sorted in milliseconds, hard rules in code first > server down? the hook asks you, it never quietly allows Claude writes the code, a 1.5 GB model judges it, nothing leaves 127.0.0.1 > unsure? the router goes up a tier, never down > fuzzy? "is the goal really met" still goes to Jev or to you the split comes straight from Von's own README: 93.5% against Jev's 97.2% on 78 cases, great at narrow questions, weaker on fuzzy ones and the privacy is the whole point > Jev: every command, path and diff goes to TypeSafe's API > Von: goes nowhere, Apache 2.0, runs on a Mac, GPU or CPU a night of decisions on a hosted API costs 10 cents and ships your repo out. the same night locally costs $0 in tokens and ships nothing you stop sending your diffs out for a yes or no. the judge stays home
This post quotes another post: x.com/asteri_eth/status/2104901781725598103