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
Answers a true-or-false question. Returns the probability that a statement is true.
Single-label classification from a defined set of options. Returns a probability per option plus an overall confidence.
Rates the input against an ordered scale (e.g. 0–2). Returns a continuous score, the distribution, and a confidence value.
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 One | LLM · System Two | |
|---|---|---|
| Output | Typed decisions (data) | Free-form text (strings) |
| Latency | ~70–500 ms | Seconds to minutes |
| Cost per decision | ~$0.00008 | ~$0.0139 |
| Sampling | Parallel — all answers at once | Sequential — one token at a time |
| Confidence | Calibrated probability on every answer | Often overconfident, inconsistent |
| Best at | Classify, route, score, gatekeep | Reasoning, 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.