Score question

A score question is Laya's ordinal primitive. Use it when the answer is a position on a ladder (urgency, severity, frustration, difficulty) rather than an unordered category.

Shape

criteria must be a list of level descriptions, index 0 first:

"frustration": {
  "type": "score",
  "instructions": "How frustrated does the customer sound in `message`?",
  "criteria": ["calm and neutral", "concerned but civil",
               "clearly annoyed", "very angry or using strong language"]
}

Each level is rendered as level i: description at its own option marker. The answer returns:

• score: the expected level, Σ i · p_i, so it is usually fractional (for example 1.84 on a 0–2 scale),
• legend: index to description,
• probabilities: one per level,
• confidence: normalised-entropy confidence.

Training and behaviour

During RLCD, score questions receive an extra ranked probability score penalty, so being one level off costs less than being three levels off. In practice uncertain answers spread over adjacent levels.

Using it well

• Describe levels with observable anchors ("blocking issue or hard deadline") rather than adjectives alone.
• Three to five levels is typical; more levels are harder to separate.
• Threshold on score for routing (for example, escalate when score >= 2.0), and check probabilities for the top level directly when the tail matters.

Be honest about limits: the model card calls score the weakest primitive (SST-5 accuracy 0.372), though on the fine-tuned typed-decisions benchmark it reaches 0.723. Validate score thresholds on your own data. See choice, score and noul.