Score

Provide at least two and at most ten levels, ordered strictly increasingly by numeric ID. IDs are strings from "1" through "10"; they need not be consecutive. Define each level in terms a reviewer could apply.

{
  "model": "ren-zero-2026-11-12",
  "type": "score",
  "situation": "The draft reply acknowledges the issue and gives a specific next step, but omits the timeline.",
  "question": "How complete is the support reply?",
  "levels": [
    {"id": "1", "description": "Does not address the issue"},
    {"id": "2", "description": "Addresses the issue but misses an important detail"},
    {"id": "3", "description": "Addresses the issue with all required details"}
  ],
  "threshold": 0.9
}

choice is the most probable level. expected_score is the sum of each numeric level multiplied by its probability, so it may be fractional. The confidence used by the gate is the probability of the most probable level, not the expected score divided by the maximum level.

For illustration, probabilities {"1": 0.1, "2": 0.6, "3": 0.3} give an expected score of 2.2, a chosen level of "2", and confidence of 0.6. This arithmetic example is not a model result.

Keep arithmetic, totals, and fixed business rules in code. The model judges the stated rubric; your application decides how to use it.