Decision-1 as a gate for avatar requests: where a person must decide

Do not let Decision-1 alone approve a request that depicts a real person. Microsoft's page says the model may reflect biases from its base model and training data, and that its scores should not be the sole basis for decisions about individuals; for consequential decisions it asks for meaningful human review. A gate in front of an avatar or likeness render should therefore end in a person.

What Microsoft lists as limits

These are the page's own known limitations. They apply to any gate, but they bite hardest on briefs about people.

A split that works

Use the model for the dull questions (is the brief complete, is the aspect ratio present) and a person for the rest. Add a cannot tell option as Microsoft suggests. Send anything that names or depicts a person to review no matter what the probability says.

On the Sume side, creating avatars and avatar videos are paid write tools on hosted MCP (avatars_create, avatar-videos_create), so they need a write-scope session and an idempotency_key; see the tools and gates page. Those rules apply after your reviewer approves, not instead of it.

What to log

Keep a record of each gated request: the state you sent, the question wording, the options returned and the person who decided. The model gives no rationale, so your log is the only explanation you will have later. Remove anything that is not needed for the decision; Microsoft also advises avoiding sensitive attributes that the use case does not require and monitoring outcomes for errors and unfair disparities.

Tell affected users when AI contributed to a decision, as Microsoft's list says.

• No rationale from the model.
• Store the human's decision.
• Review outcomes on a schedule.