Jev 101
Learn Jev. Build typed decisions.
Jev 101 is an independent beginner-friendly guide, resource library, and interactive lab for typed AI decisions.
"Package tracking stalled 4 days ago at Chicago hub, need rush delivery or full refund immediately."
Determine primary operational route and refund eligibility.
Calibrated probability output with type-checked certainty.
shipping_escalation
Unlike unpredictable raw chat completions, Jev evaluates categorical decisions over constrained state, returning mathematically verified probability vectors with zero JSON syntax drift.
91%
118 ms
DECISION ANATOMY // THE 3 PRIMITIVES
The shortest path from messy input to a typed decision
Jev keeps the state, the questions, and the decision types visible so you can inspect what the model is deciding.
State
The facts your software already has.
Question
The decision you want Jev to make.
Typed answer
Choice, score, or noul with confidence.
architecture / typed vs generative
Why System One models outperform LLMs for software decisions
Most software decisions need a typed value and calibrated probability, not a chatty paragraph that might fail JSON validation.
Built to output only allowed types (Choice / Score / Noul). Impossible to hallucinate structure.
Autoregressive prose parsed into JSON. Schema parse failures require retry loops and extra latency.
Parallel forward sampler evaluates all questions simultaneously in 70ms to 500ms.
Token-by-token generation scales with output length. Typical latency 3s to 15s.
$0.042 / MTok input. Output tokens are completely free. ~400x cheaper than frontier models.
$0.20 to $10.00 / MTok. Output tokens usually billed 3x to 5x higher than input tokens.
RLCD (Reinforcement Learning for Calibrated Decisions) ties confidence mathematically to real accuracy.
Self-reported confidence in prompts is uncalibrated and prone to overconfidence.
FOUNDATIONS // CURRICULUM
Learn the vocabulary before you ship it
Short explanations, working examples, and the surrounding ecosystem in one place.
What is Jev?
The beginner explanation: state, questions, choices, scores, noul, and confidence.
Jev API
Understand the request shape, model IDs, usage, and error handling.
Python + TypeScript
Copyable starter code for the two most common paths into Jev.
PRODUCTION RECIPES // BLUEPRINTS
Start from a decision you already need
Recipes turn a use case into a schema, a runnable example, and code you can take home.
Support ticket routing
Spam detection
Content moderation
Lead scoring
Agent routing
ONLINE LAB // LIVE WORKBENCH
Evaluate Live Decisions
Adjust the state payload, tweak the evaluation criteria, and inspect real calibrated outputs.
Route this customer support ticket to the best team.
Output
shipping
Sample output for orientation. Run the playground for a provider response.
ECOSYSTEM // RESOURCES
A living resource library
Official docs first, then useful community projects, SDKs, MCP servers, integrations, and experiments.
Official
Every entry is labeled by source and checked before publication.
SDKs
Every entry is labeled by source and checked before publication.
Tools
Every entry is labeled by source and checked before publication.
MCP
Every entry is labeled by source and checked before publication.
Research
Every entry is labeled by source and checked before publication.
Community
Every entry is labeled by source and checked before publication.
BENCHMARKS // METHODOLOGY
See what the decision is good at
Small, transparent benchmark runs with the dataset version, provider, model, latency, and failure cases exposed.
BenchmarksTransparent, reproducible benchmark runs with dataset versions, provider models, latency breakdowns, and failure case distributions clearly annotated.
118ms
99.4%
0.00%
Begin with one decision
Use the glossary, browse the resources, or run the sample. Jev 101 is built for the first afternoon, not the last mile.