ENPT

"Jev engineering" for Claude Code: a safety gate, a stop hook and a triage cron

Last updated:

Source · Model card: Jev

An X article by @polydao, “Loop Engineering Meets Jev Engineering: How to Cut Your Claude Bill From $765 to $3 a Month” (26 September 2026, 00:33 BRT), drew about 739,000 views and 700 bookmarks by 3 October. The author pointed to it again on 3 October. That newer post is a general pre-ship checklist for Claude Code and has nothing about decision models.

The idea is simple: Claude writes the code, and Jev answers the small typed questions inside the loop. The article calls moving every judgment call out of the coding model “jev engineering”. It shows three builds on Claude Code’s existing hooks:

  1. Safety gate (PreToolUse hook). Every bash command is classified before it runs. Hard rules run in code first, because Jev trusts the text in its state and a command can be written to mislead it. Anything external always goes to the human, and other is never auto-approved. For LangChain users, the same gate already exists as AutoModeMiddleware in langchain-typesafe.
  2. Stop hook. When Claude tries to finish, a script prints evidence (tests, TODOs in changed files, the README diff) and Jev judges whether the goal is really met. This addresses the usual problem of an agent grading its own work.
  3. Triage cron. Jev labels new issues in a cron job, and Claude only starts a session for bugs it can reproduce, instead of spending a full turn on every /loop tick.

It also lists seven rules for writing questions: meaning goes in the instructions (the question ID is not sent to the model), one judgment per question, describe situations rather than moods, give every Choice an other exit, ask everything in one call, keep math and dates in code, and pin the version and run in shadow mode before switching over.

The numbers are the author’s (UnverifiedClaim). The bill assumes 200 turns a night, three decisions per turn, about 4,000 tokens of state and list prices over 30 nights. That comes to about $3 a month on Jev with three calls per turn, or about $1.08 with one call per turn, against $765 on a frontier chat model. With cache reads the chat model drops to about $225. The math is internally consistent but describes a hypothetical loop, not a measured bill. The speed and cost multiples it quotes (193.6× faster, 444.6× cheaper) are TypeSafe’s own figures. The community builds it lists (session compaction, Browser Use’s jev-ultrafast, jev-trader, and others) are already in the Jev page ecosystem table.

What changes on the site. Nothing in the scores. It is a usage guide, not a new model. It is a sign of reach for Jev that we will count in the next adoption recalculation.