Question types
Noul, choice and score, what each returns, and how to write them.
System One models answer three types of question. Each one comes back typed, with probabilities.
| Type | You give | You get back | Confidence |
|---|---|---|---|
noul | Instructions, plus optional true and false criteria | noul: the probability of yes, from 0 to 1 | No separate field |
choice | A map of option keys to descriptions (a description can be null) | The chosen option, a probability per option, and confidence | Yes |
score | An ordered list of 2 to 10 levels | A score weighted by probability, so it can land between levels, plus confidence | Yes |
A choice can have up to 255 options.
The question id is never sent to the model
In a spec, each question has an id such as real_person or cost_of_ignoring. That id is for you and your code. It is not sent to the model. Put the whole meaning of the question in instructions and criteria.
"real_person": {
"type": "noul",
"instructions": "Was `email` written by a real person to the recipient, rather than sent by an automated system, newsletter, or marketing tool?",
"criteria": {
"true": "A person wrote this message to the recipient.",
"false": "Automated, bulk, newsletter, receipt, or marketing mail."
},
"meta": { "label": "Real person wrote it" }
}instructions and criteria can be plain strings, or JSON objects and arrays when you want structure such as definitions or examples. Point at parts of the state with backticked paths like `email.subject`.
Reading a noul
A noul near 0.5 means yes and no are about equally likely. It is not a medium-strength yes. That is why a noul has its own thresholds (trueAt, falseAt, reviewMargin) instead of a confidence cutoff. See confidence bands.
Writing questions that work
- Ask one narrow judgment per question. Split a fuzzy ask like "is this urgent?" into separate checks.
- Give a choice a
none_of_theseoption when nothing may fit. Without it, the model has to spread probability across wrong answers. - Write score levels that describe concrete situations and make sense on their own.
- Word the
truecriterion of a noul positively. - Give each question the state it needs. Missing evidence shows up as low confidence, which is the model telling you the truth.
For more on how the models behave, see TypeSafe's docs at docs.typesafe.ai.
Bandwise is an independent product built on TypeSafe's System One models. It is not TypeSafe's documentation. For the System One models themselves, see docs.typesafe.ai.