The Unofficial Jev AI Handbook
Jev doesn't write. It decides.
Jev is a judgment-only AI model from TypeSafe AI — no paragraphs, no chat, just fast yes/no answers, choices, and scores. This handbook explains the paradigm and shows how to pair Jev with LLMs to make AI pipelines dramatically cheaper. Early Access launched September 15, 2026.
What is judgment-only AI?
Generative models produce text. Jev produces decisions. You send structured input; Jev returns a verdict — and nothing else.
Yes / No
Binary judgments: is this review spam? Does this lead fit our ICP? Does this email need a human reply today?
Choice
Pick one of N labels: route a ticket to billing, support, or sales; bucket content by topic or risk level.
Score
Numeric ratings: score a YouTube video idea, rank candidate leads, grade retrieval quality from an AI memory store.
Built to pair with LLMs
Jev is the cheap front filter; GPT-class models handle only what passes. Reported savings: up to hundreds of times cheaper.
Start here
Your entry point: the core guides
API key from console.typesafe.ai, your first curl call, and the official Python/TypeScript SDKs.
System One, the three primitives, state design, calibration, models & pricing, and the API reference.
Ten field-tested classification patterns: moderation, lead scoring, email triage, sentiment, data filtering, ticket routing, RAG reranking, and more.
When a judgment model beats a generative one, when it does not, and why the two together cost less than either alone.
Pipeline blueprints: speculative fan-out, confidence gating, cascade routing, and the Jev→LLM front filter.
Noul, RLCD, jaggedness, calibration — every term in the paradigm, defined in one line with links deeper.
Try the paradigm
Feed a state, pick a primitive, get a typed verdict — simulated locally in your browser. No API key, no network, no cost.
Simulated in your browser — not a live Jev call
Request
yes/no → probability 0–1
Response
Fill in a state and instructions, then run the judgment.
Noul returns a probability · Choice returns probabilities + confidence · Score returns legend + probabilities + confidence
This is the *shape* of a real Jev exchange. The real thing needs an API key from console.typesafe.ai — walkthrough in Getting Started.
What would it cost?
Estimate a monthly judgment workload on Jev vs doing the same judgments with a GPT-class LLM. Runs locally — nothing is sent anywhere.
Estimate based on published pricing — verify at jev.com
Your workload
That is 300,000 judgments and 150M input tokens per month.
Estimated monthly cost
- Jev
- $0.042 / M input · output free (reported early-access pricing)
- $6.30
- GPT-class LLM (same judgments)
- assumes $2 / M input, $8 / M output, ~150 output tokens per verdict
- $660
~105× cheaper for pure judgment work
Assumptions: 30-day months; the LLM figures are placeholder rates for a frontier model, not any vendor’s current price list; real workloads add retries, cache misses, and your judgment error costs. Full numbers and caveats in Models & Pricing and Jev vs alternatives.
Latest from the handbook
Every page is version-stamped and sourced — see the Sources section at the bottom of each guide.
Jev Alternatives: Structured Outputs, Fine-Tuned Classifiers, and Rules
An honest comparison of Jev against LLM structured outputs (JSON mode), fine-tuned BERT-class classifiers, and rules engines — including the cases where you should not use Jev at all.
Jev News & Timeline
A living timeline of the Jev early-access launch: dates, model versions, official demos, and known caveats — each entry sourced. Updated as new information surfaces.
Document Triage: Type and Urgency on Two Axes
Sort inbound documents — contracts, invoices, resumes, tickets — by type with Choice and by urgency with Score, so each document class gets its own downstream pipeline and deadline.
Intent Classification: The Front Door of a Chatbot
Classify user messages into intents (and urgency) before any LLM touches them — so simple requests get cheap handlers and only genuinely open-ended messages reach the expensive model.
RAG Reranking: Filtering Retrieval Results by Relevance
Use Jev-style Noul judgments to batch-score retrieved chunks for relevance and keep only the top-k — so the LLM answers from signal, not from whatever the vector search coughed up.
Explore the handbook
Guides, use cases, and recipes — organized by what you are trying to do.
Guides
Core concepts: what Jev is, how to get access, and how it compares to LLMs.
Concepts
The mental models: System One, the three primitives, state design, calibration, models & pricing, and the API reference.
Use Cases
Where a judgment-only model fits: moderation, scoring, triage, and filtering patterns.
Recipes
Reusable pipeline patterns for combining Jev with generative LLMs.
News
Launch timeline and tracked updates: model versions, pricing, demos, and caveats — each entry sourced.
Frequently asked
A few common questions. Full list on the FAQ page.