The three question types

Jev answers three kinds of question. Picking the right one is most of the work, because the answer's shape decides what your code can do with it.

Type	Ask it when	You get back
noul	A property either holds or it does not	The probability of yes, from 0 to 1
choice	Exactly one of up to 255 options must win	The winner, its confidence, and a probability per option
score	A described scale of 2 to 10 levels	A position on the scale, and a confidence

Each has a shortcut command for one question (jev noul, jev choice, jev score) and can be one of many questions in a request file run with jev eval.

noul: does this hold?

jev noul "Is the customer asking for a refund?" --state-file ticket.txt

The answer is a probability, not a label. Two things follow from that:

• Nouls are absolute, not competing. Several nouls about the same state can all be high or all be low. Ask one noul per label rather than expecting them to add up to 1.
• 0.5 means "cannot tell", not "medium". Send those cases to a person with --abstain-band 0.4,0.6, which exits 11 instead of guessing.

Gate on it with --fail-under or --fail-over:

jev noul "Does this commit message describe a user-facing change?" --state-file message.txt --fail-under 0.7

choice: which one of these?

jev choice "Which team should handle this?" --state-file ticket.txt \
  --option billing="Payments, invoicing, refunds" \
  --option technical="Bugs, outages, integrations" \
  --option other

The probabilities over your options add up to 1, so options compete. Two rules matter:

• Always give a way out, such as other, none_of_the_above or not_stated. The model cannot pick an option you did not offer, so without an escape it will pick the least wrong one. jev choice warns when there is none.
• Describe options that are easy to confuse. Say what each one covers and what it does not, in parallel terms, with examples that look like your real inputs.

Gate on it with --expect (repeatable) and --min-confidence:

jev choice "Which team should handle this?" --state-file ticket.txt \
  --option billing --option technical --option other --expect billing --min-confidence 0.6

score: where on this scale?

jev score "How frustrated is the customer?" --state-file ticket.txt \
  --level Calm --level Frustrated --level "Very angry"

Levels are counted from 0, so those three give a score between 0 and 2. Two rules matter:

• Describe each level as a recognisable situation. Levels are judged on their own, so "worse than the level before" tells the model nothing.
• One dimension per score. "How urgent and how angry" is two questions.

Gate on it with --fail-under, --fail-over and --min-confidence.

Reading the answer

• confidence is not accuracy. It says how peaked the distribution is. Two genuinely acceptable options give a low confidence, and that is not a problem.
• Answers are not bit-for-bit repeatable. Compare against thresholds, never for equality, and do not reuse a noul threshold for a choice or the other way round.
• Keep the raw probability. Store it and apply your policy in code, so the policy can change without asking the model again.
• cost_usd is an estimate. Only input tokens are billed, and an alias such as jev-latest has no published price, so the field can be null.

What Jev will not do

Jev reads literally and quickly. It cannot count, do arithmetic, compare dates or look anything up. Keep that in your code and ask Jev only for the meaning. There is a short list of habits that make a real difference to accuracy in Writing good questions.

Several questions in one call

Extra questions cost tokens, not round trips, and the state is only paid for once. So when you have more than one question about the same content, put them all in a request file and run jev eval instead of calling the shortcuts several times.

questions:
  is_urgent:
    type: noul
    instructions: Does the customer say the problem needs attention today?
    criteria:
      "true": A deadline, an outage, or money being lost right now
      "false": No time pressure is stated
  team:
    type: choice
    instructions: Which team should handle this ticket?
    criteria:
      billing: Charges, invoices, refunds, payouts
      technical: Errors, outages, integrations, bugs
      other: Anything else, or not enough information to tell
  frustration:
    type: score
    instructions: How frustrated does the customer sound?
    criteria:
      - Calm and neutral
      - Annoyed, but polite
      - Angry, or threatening to leave

jev eval -f triage.yaml --state-file ticket.txt -o json

Each answer comes back under its own question id, which makes --field answers.team.choice and --assert 'is_urgent >= 0.7' work. The full shape is in Exit codes and JSON contract.