Question 1

What is the Bouncer judgment layer?

Accepted Answer

A judgment layer for agents that starts with Claude Code. A PreToolUse hook hands each proposed tool call to TypeSafe Jev — a fast System One model rather than a frontier LLM — and the answers become allow, deny or silence through thresholds in a YAML file you own.

Question 2

How does one tool call get judged?

Accepted Answer

Bouncer hands Jev a state — the tool, the command, what the paths in it mean — plus a map of plain-English questions. Each noul question comes back as a single number between 0 and 1, so the thing that acts on the answer is a threshold rather than a parser.

Question 3

What does it cost to run on every call?

Accepted Answer

The README prices it at about four cents a day for seven questions in one call, against reasoning prices for if-statements: sending those small, repeated judgments to a frontier model charges a few hundred milliseconds and a few thousand tokens each time.

Question 4

What happens with --dangerously-skip-permissions?

Accepted Answer

It never prompts you — that is the point. The baseline is zero interruptions, and seatbelt mode keeps it there: silent until one of nineteen deterministic rules matches something never okay, like a private key or a live credential on the command line. Judgments still run and get logged.

Question 5

What is bouncer judge for?

Accepted Answer

Batch work, where the README says the bill actually goes down: a policy of questions runs over a file or a directory of items and produces a judgments log plus an escalation manifest naming which items failed which criteria, at what p. Only the escalations reach the reasoning model.

Question 6

Can I tune the policy without touching code?

Accepted Answer

Yes. Questions are plain English and thresholds are numbers in a YAML file you own, so re-tuning never means editing a source file, and the policy can sit in your repo next to your CLAUDE.md. Dry-run is the default, with the calibration table published.

Question 7

What is the project's status?

Accepted Answer

The README says the gate (v0.2) and the batch judge (v0.3 in the PRD roadmap) have both shipped, with the plugin at 0.2.9. The hook runs end to end and ships observing: judgments run and get logged before anything is allowed to interrupt you, unless you enable a deny threshold yourself.

Bouncer

Highlights

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community