Why use this instead of prompting an LLM to classify tickets?

Jevalyn wraps typed Jev judgments with calibrated probabilities inside Rails control flow, while an LLM prompt returns prose you must parse. Jev answers in about a request round-trip and bills on input tokens only with free output tokens.

Will adding speculative questions multiply my latency and bill?

All questions in one Decision go out in one request and are evaluated in parallel against the state once, so extra speculative questions cost a few tokens each. Results expose values, distributions, thresholds and margins for logging.

How do I stop a low-confidence refund from auto-approving?

Each question can name its own floor with the class threshold as default, and call-site thresholds: or confidence_threshold: override for one call. Noul has no confidence so its floor is measured against certainty distance from 0.5.

What input can I pass and when will bad rubrics fail?

Anything JSON-shaped works including String, Hash, Array or ActiveRecord, and every question validates when the class body runs so an 11-level score or choiceless choice fails on boot. Install writes the initializer and app/decisions, then jevalyn:ping checks the key.

What does it need, and how do I install it?

Ruby 3.1+ and Rails 7.0+. Add the gem, bundle install, run bin/rails g jevalyn:install, export TYPESAFE_API_KEY, then bin/rails jevalyn:ping checks the key.

Why is confidence nil on a noul?

Jev returns confidence on Choice and Score only; for a noul the probability is the answer and 0.5 means the model does not know. Jevalyn adds certainty from the distance to a coin flip so certain? works uniformly, while confidence stays nil so you can tell which is which.

How is a question's confidence floor resolved?

Most specific first: thresholds: at the call site, then confidence_threshold: at the call site, then the question's, then the decision's, then Jevalyn.config.default_confidence_threshold. With none set, certain? is true.

CommunityJev Integrations & Adapters

Jevalyn

Rails decision layer that evaluates noul, choice and score questions about app state in one Jev request, with per-question confidence floors that gate auto-action versus human review.

See highlights

Open github.com

Stars: 19
Language: Ruby
License: MIT
Published by: Community
Author: Ray-Hughes
Last verified: Oct 2026

Ray-Hughes/jevalyn

User repository on GitHub

View on GitHub

The decision layer for your Rails app. A Rails-native wrapper around TypeSafe's Jev System One API: typed, calibrated decisions in your control flow.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Fast, cheap, structured decisions baked into your Rails app's control flow. Install · Decisions · Guardrails · Router · Testing. Jevalyn is the decision layer for your Rails app, built on Jev. That is a routing decision made by a model, inside a Rails request, in about as long as a database query. Not a prompt, not a parsed JSON blob, not a retry loop around something that might return prose this time.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Ruby

Highlights

Three typed questions with probabilities, distributions and weighted scores in one request.
Per-question confidence floors with call-site overrides and certainty margins to log.
Class-body validation so malformed rubrics fail on boot rather than at request time.
Generator, initializer and ping command for Rails setup with environment-key config.

Watch out

Jev never generates prose or calls tools, so free-text summaries and open-ended reasoning still need an LLM routed separately, and pricing and rate limits are still moving. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-feels

Ruby gem that models Jev as field predicates: feels and decide bind to model attributes with shared definitions, thresholds, confidence gates, batching, and test tapes.