The three answers

The API takes one state and a set of questions, and answers each one independently. There are three kinds of question, and the integration uses the API's own names for them so what you pick in the UI is what goes over the wire.

Noul

A probability from 0 to 1. No confidence figure, because the number is the confidence.

In the UIIn YAMLAs an action

Pick Noul. Optionally set true_means and false_means to sharpen which way the answer points, and a threshold to also get a binary sensor.

- name: Laundry forgotten
  type: noul
  instructions: Is the laundry finished but still sitting in the machine?
  "true": Finished, and nobody has taken it out
  "false": Still running, or already emptied
  threshold: 0.7

- action: jev.noul
  response_variable: laundry
  target:
    entity_id: sensor.washing_machine_power
  data:
    instructions: Is the laundry finished but still in the machine?
    threshold: 0.7

Returns noul and is_true.

The key is spelled differently in YAML

A jev: block takes "true": and "false":, quoted, because unquoted YAML reads them as booleans and the schema never matches. The actions and the UI form take true_means and false_means. That inconsistency is historical and the docs say which is which rather than pretending otherwise.

Choice

One of your options, with the whole distribution and a confidence.

Between 2 and 255 options. The integration checks that before it sends anything, and the error names your number.

In the UIIn YAMLAs an action

Pick Choice, then write your options one per line as name: what it means. The description is optional but usually earns its place.

- name: Caller
  type: choice
  instructions: Who is at the door?
  criteria:
    delivery: A courier dropping off or collecting a parcel
    visitor: Someone the household knows
    sales: An unsolicited caller selling or canvassing

- action: jev.choice
  response_variable: caller
  data:
    instructions: Who is at the door?
    options:
      delivery: A courier dropping off a parcel
      visitor: Someone the household knows

Returns choice, probabilities and confidence.

The sensor takes the winning option as its state, with the full distribution in its attributes. Because the options are a fixed set, the sensor gets the enum device class, so the UI knows what values are possible.

Score

A rating against levels you describe, lowest first. The answer can land between levels, which is the point: 2.4 out of "Not at all, Worth a glance, Look today, Right now" says more than picking one.

Between 2 and 10 levels.

In the UIIn YAMLAs an action

Pick Score, then write your levels one per line, lowest first.

- name: Nudge urgency
  type: score
  instructions: How urgently should someone be reminded?
  criteria: [Not at all, When convenient, Soon, Right now]

- action: jev.score
  response_variable: urgency
  data:
    instructions: How urgently does this need a person?
    levels: [Never, Whenever, Within the hour, Right now]

Returns score, normalized, nearest_level, legend, probabilities and confidence.

Order matters

Levels are read lowest first. Writing them in the wrong order does not error, it just gives you answers that are backwards.

Picking between them

You want	Use
A yes or no, or something to threshold	Noul
To route between a handful of named outcomes	Choice
A dial rather than a switch	Score

When two options blur into each other, a Score with more levels often beats a Choice with more options, because it can answer "between these two" instead of being forced to pick.

On confidence

Choice and Score return a confidence. Noul does not, because its number already is one.

TypeSafe publishes no calibration evidence for confidence, and their own docs call it a convenient default. Treat 0.9 as higher than 0.6 rather than as right nine times in ten, until you have measured it on your own questions. See Limitations.