A local 1.5GB model handles safety hook decisions instead of Jev
Narrow questions in the state-to-decision-to-action loop go to a local model while fuzzy judgment stays with Jev, so the safety hook no longer uploads bash commands and .env lines.
Jev Engineering doesn't need the API anymore. state → decision → action → verification → next state 1.5 GB model. ~18 ms per call on a Mac. $0 in tokens. everyone is asking what Jev decides. the better question is where the decision happens. your safety hook ships every bash command, the folder it runs in, sometimes a line from .env. 600 times a night. to one more company's API. two weeks after launch, there are 38 open alternatives. some speak Jev's exact wire format, so old hooks keep working after one base_url change. the split is simple: narrow question → local model fuzzy judgment → Jev or you so the loop becomes: Claude → builds code rules → catch the obvious local model → decides the rest hook → allow, ask or deny state → never leaves your disk the honest part: Von scores 93.5% against Jev's 97.2% on its own benchmark. not a replacement for every question. just for the ones you never wanted to upload. secret gate code, model router and the full bill in article ↓
This post quotes another post: x.com/asteri_eth/status/2104901781725598103