Score

A score question rates the context on ordered semantic levels, lowest first. Krun returns a probability distribution over the levels and the expected score.

Complete example

curl https://api.krun.ai/v1/decide \
  -H "Authorization: Bearer $KRUN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "context": "Production is down: nobody can log in and there is no workaround.",
    "questions": {
      "severity": {
        "type": "score",
        "instructions": "How severe is the reported issue?",
        "levels": [
          "Cosmetic issue; no functionality affected",
          "Feature degraded; workaround available",
          "Critical functionality blocked; no workaround"
        ]
      }
    }
  }'

from krun import Krun, ScoreQuestion

client = Krun()  # reads KRUN_API_KEY

result = client.decide(
    context="Production is down: nobody can log in and there is no workaround.",
    questions={
        "severity": ScoreQuestion(
            "How severe is the reported issue?",
            [
                "Cosmetic issue; no functionality affected",
                "Feature degraded; workaround available",
                "Critical functionality blocked; no workaround",
            ],
        ),
    },
)

severity = result.score("severity")
print(severity.score)          # 1.88
print(severity.probabilities)  # {"0": 0.005, "1": 0.107, "2": 0.887}

Response:

{
  "model": "krun-one-v0.3",
  "answers": {
    "severity": {
      "type": "score",
      "score": 1.881956,
      "confidence": 0.885152,
      "legend": {
        "0": "Cosmetic issue; no functionality affected",
        "1": "Feature degraded; workaround available",
        "2": "Critical functionality blocked; no workaround"
      },
      "probabilities": {"0": 0.005369, "1": 0.107306, "2": 0.887325}
    }
  },
  "usage": {
    "input_tokens": 59
  }
}

The body always has a context and a questions map keyed by ids you choose (severity here); the answer comes back under the same id. A request can mix score questions with choice and noul questions, up to 16.

Expected score

The score is the expected level index under the distribution, not the most likely level:

score = Σ (index × probability)
      = 0 × 0.00 + 1 × 0.57 + 2 × 0.43 = 1.43

It lies between 0 and levels − 1. With a distribution like {"0": 0.0, "1": 0.57, "2": 0.43}, the score is 1.43: “between degraded and critical, closer to degraded”, which you can threshold, average across tickets or sort by. If you need a single level, use the most likely one from probabilities, or round the score.

Fields

Field	Description
`score`	Expected level, in [0, levels − 1].
`probabilities`	Level index (`"0"`, `"1"`, …) → probability, in level order.
`legend`	Level index → the level text you sent.
`confidence`	How concentrated the distribution is, in [0, 1].

confidence is 1 − Var / Varmax, where Var is the variance of the level index under probabilities and Varmax = ((levels − 1) / 2)² is the largest variance possible. It is 1 when all probability is on one level and 0 when it is split between the lowest and the highest level. It measures how sure the model is about the score, so a distribution split between two neighbouring levels has higher confidence than one split between the extremes.

Question fields

A score question is one entry of questions:

Field	Required	Description
`type`	Yes	`"score"`
`instructions`	Yes	What to rate (1–1,000 characters).
`levels`	Yes	2 to 16 level descriptions, lowest first (1–500 characters each).

Levels

Rule
2 to 16 levels	Each 1–500 characters.
Lowest first	Index `0` is the lowest level.
Order is meaning	Krun never reorders levels. Changing their order changes what the score means.

Write levels as descriptions, not bare numbers. In a request:

{
  "context": "The export button is slow but the file still downloads.",
  "questions": {
    "severity": {
      "type": "score",
      "instructions": "How severe is the reported issue?",
      "levels": ["Minor issue", "Feature degraded", "Blocking issue"]
    }
  }
}

is much better than ["0", "1", "2"] or ["low", "medium", "high"] with no context: the model reads the level text.

Krun One was trained and evaluated on 5-level scales (response quality, harm severity, sentiment). Other level counts work the same way, but have been validated less.

Send corrections with feedback: expected: {"type": "score", "value": 2} (the correct level index).

Overview

Concepts

Guides

SDKs

API Reference

Resources

Complete example

Expected score

Fields

Question fields

Levels

​Complete example

​Expected score

​Fields

​Question fields

​Levels

Complete example

Expected score

Fields

Question fields

Levels