pi-jev uses Jev's typed judgments for two separate decisions: which previous tool outputs still matter during context compaction, and whether a new request is trivial, moderate or complex enough to justify a cheaper or stronger model. Both extensions fail open to Pi's normal behavior when Jev is unavailable.

jev-mcp exposes a jev_classify tool that sends a class catalog once and turns each document or message into an independent Choice question. Its documented live example routes support messages into billing, technical and sales classes while separating automatic decisions from review cases.

The public Jev project index documents Notra's production GEO path: a NOTRA_JEV_CLASSIFIERS flag moves brand-visibility classifiers from an LLM to Jev Boolean decisions using a 0.5 threshold, with a stated 300 ms p50 target. This is a concrete production integration rather than a prompt-only demo.

Jev-Mail runs on user-owned Google Apps Script and asks Jev to score incoming mail for urgency, importance and category. The workflow keeps uncertain or suspicious messages in a Review path instead of treating every classifier output as safe to automate.

Jev Wrapped reads up to 1,500 posts from a public Telegram channel, chooses among ten post types and separately scores paid-ad, clickbait and emotional-pressure questions. Thresholded results feed a shareable monthly card while links to the highest-scoring posts preserve a manual-check path.

AI-decision-maker uses Choice questions to map CSV columns into a 13-code type vocabulary and datasets into six scenes, then performs the actual writes locally. Its author also reports that Jev cost more tokens than an LLM on this particular task because repeated criteria outweighed the tiny one-character output, making it a useful counterexample rather than a blanket cost claim.

jev-skill-router evaluates the installed skill roster with a Choice plus gating decisions, but defaults to shadow mode: it logs the recommendation without injecting it. Only sufficiently confident candidates cross the configured threshold, providing a practical staged rollout pattern for decision models.

hyspacex's jev-router is a self-hosted OpenAI-compatible router. Jev assesses a fresh task, then inspectable YAML policy chooses a model, provider and reasoning effort; the selected model remains bound for the session rather than changing silently during a tool loop.

This skill-router asks Jev to choose among installed agent skills and uses confidence thresholds to avoid unnecessary loads. Its published exploratory evaluation reports 68 of 72 synthetic requests routed correctly versus 51 of 72 for a lexical baseline, while explicitly warning that reused-data results do not establish production accuracy.