Decision-1: ask three questions in one request before a Sume render

Yes: Microsoft-Decision-1 lets questions with the same state share one request, so you can run a brief through a gate, a format pick and a risk check in a single call. Microsoft says combining related questions reduces the number of requests and keeps the decision context consistent. Render on Sume only when every answer clears your thresholds.

The request

Each question has a name and the answers come back under the same names. The three types are noul (a probability from 0 to 1), choice (one option, plus a probability per option) and score (a position on an ordered scale).

{
  "model": "<deployment-name>",
  "state": {"brief": "9:16 teaser, 8 seconds, ceramic mug, slow push-in"},
  "questions": {
    "ready": {"type": "noul", "instructions": "Is the brief complete enough to render?"},
    "kind": {"type": "choice", "instructions": "What is it?",
      "criteria": {"product": "A product shot", "person": "Features a person", "other": "Anything else"}},
    "risk": {"type": "score", "instructions": "How risky is it to render unreviewed?",
      "criteria": ["None", "Low", "Medium", "High"]}
  }
}

Reading the answers

Define the rule in code before you look at any answer. A scale runs from index 0 up, and Microsoft says the returned score is the probability-weighted average of the level indexes, so it can fall between levels. Use scores for ordering and thresholds, not as calibrated ratings.

After the gate

Submit with the Videos API, which answers 202 with id and polling_url. Microsoft advises setting thresholds from the cost of false positives and negatives, and adding an abstention option such as cannot tell when a forced choice is wrong.

Testing the three-question gate

Write ten briefs you know are good and ten you know are bad, run both sets, and print each answer next to your label. Change one question's wording, run again and compare; Microsoft warns that scores can change with wording and option order. Keep the combination that separates the two sets best, and write the thresholds next to the questions.

If any answer is missing from the response, treat the whole brief as review.

• Ten good, ten bad.
• One change at a time.
• Missing answer means review.