Answers and results

decide() resolves to a DecisionResult with three fields.

type DecisionResult<Questions, RawResponse> = {
  readonly answers: DecisionAnswers<Questions>
  readonly usage?: Usage | undefined
  readonly rawResponse: RawResponse
}

1. Typed answers

answers has one entry per question name, and each entry's type is derived from its question.

const result = await decide({ model, state, questions })

result.answers.department.choice // 'billing' | 'technical' | 'general'
result.answers.topics.labels // readonly ('subscription' | 'duplicateCharge' | 'refund')[]
result.answers.urgency.score // number on the 0 to 3 rubric
result.answers.cancellation.probability // number from 0 to 1

Each answer has a type discriminator matching its question ('choice', 'multi-label', 'score', or 'check').

Answer	Fields
ChoiceAnswer	choice, optional probabilities keyed by option, optional confidence
MultiLabelAnswer	labels (selected), probabilities keyed by option
ScoreAnswer	score, rubric, optional probabilities in rubric order, optional confidence
CheckAnswer	probability

2. Probabilities and confidence

• Choice and score probabilities are exclusive distributions that sum to one, allowing for ordinary rounding.
• Multi-label probabilities are independent and do not need to sum to one.
• confidence is reported by the provider. It is not derived from the highest probability.
• Anvia does not calibrate these values. Treat them as estimates and set your own action thresholds, ideally from labeled examples.
• Relevance scores from native rerankers have different semantics and are not converted into probabilities by this contract.

3. Usage

result.usage is normalized Anvia token usage when the provider reports it, and is omitted otherwise. When present it passes the same non-negative-number validation as other usage.

4. Raw response

result.rawResponse is the original provider result. Its type flows from the model, so a model typed DecisionModel<MyRaw> returns MyRaw. Use it for diagnostics and provider-native fields. Do not return it to end users, and apply the same retention policy as other provider payloads.

5. Validation guarantee

Before decide() returns, core checks that:

• the answers have exactly the requested question names and matching types;
• a choice is one of the declared options and its probabilities cover exactly those options;
• multi-label probabilities cover exactly the options, labels are unique, and they match the threshold rule;
• a score is within the rubric bounds, the returned rubric equals the one you sent, and any distribution matches the rubric length;
• a check probability and any confidence are between zero and one;
• usage values are non-negative finite numbers.

A violation throws DecisionProviderOutputError, so downstream branching can trust the shapes.