What is Jev? An independent take on the System One model

Jev is the first model in a category its maker calls System One: it does not generate a single token of text. You send it a state (structured facts) and a batch of typed questions, and it returns structured answers with calibrated probabilities. Your code reads the answers and branches. That is the whole product.

Where the name comes from — and why it matters

The name is a deliberate nod to Kahneman's Thinking, Fast and Slow. System 1 thinking is fast, automatic, intuitive; System 2 is slow, effortful, deliberate. Generative LLMs are System 2 machines: they are brilliant at producing text, but when you press them into service as a classifier or a gate, you pay for slowness, cost, and a text output you must parse back into structure.

TypeSafe's thesis is that most production AI work is not generation — it is decision. Routing a ticket, scoring a risk, checking whether a step is safe, re-ranking a retrieval: thousands of small, fast, yes/no-ish judgments per minute, machine to machine, no human reading the output. Jev is built for exactly that layer, and nothing else.

How it is trained: RLCD, a third path

Post-training has had two famous recipes: RLHF (which produced chat assistants) and RLVR (verifiable-reward RL, which produced step-by-step reasoners). TypeSafe's CEO, Diogo Almeida, co-invented RLHF (InstructGPT). Their new recipe is RLCD — RL for Calibrated Decisions — which trains the model to emit a decision plus a calibrated probability rather than text.

Calibration is the load-bearing feature. A model that says "80%" and is right about 80% of the time can be trusted as a control signal. That is what lets you use its output as a threshold in production logic instead of eyeballing a paragraph.

How it runs: one forward pass, no autoregression

Every question in a request is evaluated in parallel and in isolation in a single forward pass. There is no token-by-token decoding. The consequences are practical, not academic:

• Output tokens are free — you only pay for the input (state + questions).
• Latency is milliseconds, and it barely moves as you add questions — adding a question is nearly free, which changes how you design calls (see Speculative fan-out).
• Questions cannot leak into each other — isolation means one tricky question doesn't contaminate another's answer.

What this is not

It is not a chatbot, not a replacement for your LLM, and not a general-purpose "AI brain." It cannot generate text, it is weakest at anything numeric or multi-hop, and it has no built-in defense against prompt injection in the state. If you arrive expecting a smarter GPT, you will be disappointed; if you arrive looking for a cheap, fast, trustworthy judgment primitive to wire into software, this is a genuinely new tool. The rest of this site is a field guide to using it that way.