Carlo GiNotes

Choice vs Noul vs Score in Jev

3 min read

Jev is a decision model. You send it some state (the thing to judge) and a set of questions, and every question comes back as a typed answer, not text.

There are three question types for text. Picking the right one is most of the design work.

The three types

TypeAsksReturns
choiceWhich one of these options?the picked option, a probability per option, a confidence
noulIs this true?one probability of “yes”, from 0 to 1
scoreWhere on this ordered scale?a position between levels, a probability per level, a confidence

When to use which

  • choice: the options never overlap and exactly one is right. Ticket → billing / technical / sales. A choice always picks something, so add an other option if your list might not cover every input.
  • noul: a yes/no question where the probability itself is useful. One per label when an item can have zero, one, or many labels.
  • score: the answer is a level on a spectrum you can describe. Severity, frustration, skill level.

Tags are nouls, not a choice

Tagging a todo as groceries, chores, or work looks like a choice, but it is not:

  • “Buy dish soap” is groceries and chores.
  • “Renew passport” is none of them.

A choice cannot say either of those. One noul per tag can:

{
  "model": "typesafe/jev-1.14",
  "state": "Buy dish soap",
  "questions": {
    "groceries": {"type": "noul", "instructions": "Is this about buying food, drinks, or household supplies?"},
    "chores": {"type": "noul", "instructions": "Is this a task done in or around the home?"},
    "work": {"type": "noul", "instructions": "Is this related to a job or career?"}
  }
}
{"groceries": {"noul": 0.96}, "chores": {"noul": 0.92}, "work": {"noul": 0.09}}

Each tag is judged on its own, so adding a fourth tag does not change the scores of the other three. Then your code keeps every tag above a threshold.

Gotchas

  • Ask in one request. All questions about the same state go in one call. They run in parallel, so extra questions barely change the latency.
  • One snap judgment per question. “Is this about work?” is good. “Analyse this todo and decide what to do” is not; split it into several questions and combine the answers in code.
  • Question ids are not sent to the model. Put the full question in instructions, even if the id looks self-explanatory.
  • noul has no confidence. The probability already is the confidence. 0.5 means “no idea”, not “half true”.

Graph View