Explainer library · independent
Learn Jev and decision models
Short, sourced explainers for the questions people actually search: what the Jev model is, what a decision model returns, which open-weight alternatives exist, and how the pieces fit into an agent workflow.
This site is an independent third-party directory. Figures quoted here are attributed to the vendor or project that published them; this directory does not run its own benchmarks.
Try it in the browser
Paste state, ask a typed question, read the structured answer.
Compare the models
Jev vs Clef, Jev vs Laya, and decision models vs LLMs.
See the patterns
Routers, guards, triage, and verification built from the primitives.
Core concepts
What the model is, what the category is, and what the alternatives look like.
What is the Jev model?
What the Jev model is: TypeSafe System One, a decision model that returns typed Choice, Score and Noul answers. Version names, model IDs, and what it does not generate.
Open source alternatives to Jev
Jev is a closed hosted API. These are the open-weight decision models with the same typed-question shape: Cloudflare Clef, Laya, AWS Strands Decider 2B, and where rules or LLMs are the better answer.
What is a decision model?
A decision model scores a fixed answer space and returns a typed value with a probability. What the 2026 family looks like, where it belongs in an application, and where it does not.
How a decision API works
The decision API contract: one state, several typed questions, constrained answers with probabilities. Question types, thresholds in code, and the failure modes that structured output does not have.
What is Clef?
Clef is Cloudflare's open-weight decision model family. What Clef, Clef-flash and Clef-omni are, how they run on Workers AI, published pricing and latency, and what the launch benchmarks do not prove.
Strands Decider 2B, explained
Strands Decider 2B is AWS Strands Labs’ open 2B-parameter decision model. Architecture, published latency and JevBench results, local install, and where it sits in an agent loop.
Ecosystem and integrations
Gateways, agent frameworks, and what running these models yourself actually involves.
Calling Jev through OpenRouter
OpenRouter serves Jev under typesafe/jev-1.13, typesafe/jev-1.14 and the ~typesafe/jev-latest alias. Model IDs, the Decisions route, why chat completions do not work, and published pricing.
Where a decision model goes in a LangGraph graph
How a typed decision model fits into a LangGraph conditional edge: router nodes, guard nodes, and scoring nodes that branch on a value instead of parsing generated JSON.
Decision models in agent workflows
Why agent loops put small decisions on a hot path, where a typed decision model sits relative to an agent framework, and how to build an approve / deny / escalate gate with a confidence threshold.
Self-hosting a decision model
What running Clef, Laya or Strands Decider 2B yourself actually involves: published hardware notes, local server defaults, private network access, and the real cost ledger versus a hosted API.
Reference
The primitives, the training objective, and the vocabulary behind typed decisions.
RLCD vs RLHF
RLHF aligns generated text with human preferences; RLCD trains a decision model so its probabilities are useful as thresholds. Why calibration is the load-bearing claim in this category.
Choice, Score and Noul
The three question shapes behind Jev, Clef and Laya: what you supply, what comes back, and how to combine several questions against one state in a single request.