Jev Score: Rubrics, Levels, Probabilities & Examples

What is Score?

Score rates state against an ordered rubric. TypeSafe accepts 2–10 level descriptions. Unlike Choice, the order matters: later levels represent progression along the dimension you define.

How levels work

{
  "type": "score",
  "instructions": "How urgent is the request?",
  "criteria": [
    "No time constraint",
    "Can wait this week",
    "Needs attention within a day",
    "Immediate outage or harm"
  ]
}

The HTTP result’s legend maps level indices back to descriptions. These indices begin at zero in the documented response. For this four-level rubric, the range is zero through three. A different number of levels changes the range.

Legend, probabilities, and confidence

Read legend, probabilities, score, and confidence together. A score can fall between levels because it is derived from their probability distribution. Two distributions can produce similar weighted averages while reflecting very different ambiguity.

For example, an illustrative distribution with half the probability on level 0 and half on level 2 averages to 1. It is not the same evidence as all probability on level 1. A policy based only on the average would miss that distinction.

Weighted Score explained

The conceptual computation is the sum of each level index multiplied by its probability. Your code should still use the returned value and legend rather than inventing a new scale. If you normalize several scores for a composite, normalize each against its own number of levels.

Do not treat Score as exact numerical regression. A semantic urgency rating is not a measured response deadline, refund amount, or exact probability of harm.

Good and bad rubrics

Bad: low, moderate, high, very high, with no operational definitions. Better: no deadline, this week, within one day, immediate ongoing harm. The better descriptions still require testing, but they tell reviewers what each level means.

Keep one dimension per rubric. A level that mixes company size, budget, urgency, and willingness to buy is difficult to interpret. Use composite scoring to separate those dimensions.

Urgency and risk examples

A support urgency Score can prioritize review. A tool-risk Score can summarize potential impact. Neither should replace exact checks such as “is this resource allowed?” or “does the user own this account?” Keep those checks deterministic and deny unauthorized operations before the model call.

Common mistakes

Do not compare raw scores from rubrics of different lengths. Do not rewrite the rubric without re-evaluating thresholds. Do not call 0.5 “50% likely urgent” unless you asked a Noul proposition that actually means that.