Primitives· Last updated 2026-09-21

Jev Score questions

A Score rates the state along a rubric you write. criteria is an ordered array of level descriptions. The API wants at least two levels and accepts up to 10.

The answer includes:

• score — a probability-weighted position. It can fall between levels (for example 1.05 on a 0–2 rubric).
• legend — level index → description.
• probabilities — each level (string keys) → probability.
• confidence — how concentrated the distribution is.

Minimal shape

{
  "frustration": {
    "type": "score",
    "instructions": "How frustrated is the customer?",
    "criteria": ["Calm", "Frustrated", "Very angry"]
  }
}

Write levels as descriptions a person could apply, not as naked integers. The model is judging language against your rubric.

Do not interpolate into a physical quantity

The jaggedness note is specific: do not use Score expectations to reconstruct the exact magnitude of a number between two levels. jev-1.13 levels are “weak in numerical calibration.” You may threshold the expectation (“is this past level 1?”). You should not treat 1.37 as a precise measurement of anything except position on your rubric.

Split multi-factor scores

Ticket priority that mixes severity, frustration, and engineer-readiness is three Scores, then weights in code. That is the composite scoring pattern. Keep math in code.