Choosing a Question Type¶

Every question asks for one judgment about the state. Pick the type that matches the shape of the answer your code needs.

Decision guide¶

• Is it a yes/no condition? Use a Noul, for example "does this message ask for a refund?". Define the condition precisely: the answer is the probability that it holds.
• Is it one of several options? Use a Choice, for example routing to a team or classifying a document. Give the full list of options, and add other or none when the list might not cover every input.
• Is it a degree? Use a Score, for example severity, frustration, or skill. Describe what each level looks like.

A Noul value of 0.5 means yes and no are equally likely. It doesn't mean "medium". "Is the candidate strong in Python?" is a poor Noul, because "strong" isn't defined. Either ask a clean yes/no ("Does the resume state the candidate used Python at work?") or measure degree with a Score whose levels describe each amount of experience.

object PythonSkill : JevQuery() {
    // A Noul answers "is this true?": 0.5 means "equally likely", not "medium skill".
    val usedAtWork by noul("Does the resume state that the candidate has used Python at work?")

    // A Score measures degree along described levels.
    val depth by score("How much Python experience does the resume show?") {
        levels("No experience", "Some familiarity", "Daily use", "Deep expertise")
    }
}

One snap judgment per question¶

System One models are built for the judgment a knowledgeable person makes in a second, given the right context. "Does this message convey urgency?" is a good question. "Analyze this message and decide what to do" isn't: that needs deliberation, and it's a signal to split the task into narrow questions and combine their answers in code. Writing Good Questions covers this in depth.

Richer instructions and criteria¶

Instructions, option descriptions, levels and Noul criteria can all be structured JSON rather than plain strings, for example with contrastive what / not_for / examples fields. See Structured Criteria.

Limits¶

jev4k checks these locally and reports every problem at once, before anything is sent:

A request's state plus all its questions must also fit Jev's context window: 64k tokens in total, and 32k for the state plus the single longest question.