Lesson 03 · 8 min

How it works

Two core concepts, three question types, and one training idea. That’s the whole model.

1 · State in, decisions out

Every call has two parts: a state (the data — a string or nested JSON) and a set of questions about it. Jev answers every question in a single parallel query.

State

ticket + order + policy

Typed decisions

answers + probabilities

2 · The three question types

NoulYes / No

Answers a true-or-false question. Returns the probability that a statement is true.

Does this ticket need immediate escalation?probability: 0.95
ChoicePick one

Single-label classification from a defined set of options. Returns a probability per option plus an overall confidence.

Which department should handle this?billing · confidence 0.80
ScoreRate it

Rates the input against an ordered scale (e.g. 0–2). Returns a continuous score, the distribution, and a confidence value.

Customer frustration, 0–2?1.04 · confidence 0.94

You can ask many questions about the same state in one request. Because Jev evaluates them in parallel, extra questions barely change the response time — they only cost their few tokens.

3 · System One vs System Two

Jev (System One) is not a replacement for LLMs (System Two). They divide the labour:

Jev · System OneLLM · System Two
OutputTyped decisions (data)Free-form text (strings)
Latency~70–500 msSeconds to minutes
Cost per decision~$0.00008~$0.0139
SamplingParallel — all answers at onceSequential — one token at a time
ConfidenceCalibrated probability on every answerOften overconfident, inconsistent
Best atClassify, route, score, gatekeepReasoning, writing, planning

4 · How it’s trained: RLCD

Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD). Where RLHF tunes a model to match human preference, RLCD tunes it so the probabilities it reports genuinely reflect its accuracy.

That calibration is the whole point: if a model can do a task 95% of the time but can’t tell you when it’s in the failing 5%, you can’t safely automate it. Calibrated confidence is what makes “auto-act above X, escalate below X” a viable strategy.

Putting it together: the decision flow

1A business event arrives (ticket, order, log, agent state).
2Your code builds a state and defines Noul / Choice / Score questions.
3Jev evaluates them all in parallel and returns typed answers + confidence.
4If confidence ≥ your threshold → act automatically (route, block, escalate).
5If confidence < threshold → send to a human.
6Log the decision and its probability — every call is auditable.