A gate is a routing rule.
A high probability is information. Deciding what to do with it is a separate job.
Separate prediction from permission
A model ranks the options it sees. Your application decides whether the result may trigger an action. Combining those two jobs hides a policy decision inside a prediction. Ren 1.1 keeps a gate field in the response so the routing policy is visible.
The threshold is supplied with the request. If confidence falls short, the result asks for human review. If it clears the threshold, the response allows the application's next step. Permissions, business rules and checks for harmful outcomes still belong in your code.
Measure on your own work
Calibration describes a relationship between reported probabilities and observed outcomes on a distribution. It does not promise that unfamiliar tasks will behave the same way. A deployment can change the wording, available options and costs of mistakes.
Collect representative development cases. Compare model decisions with reviewed answers. Look at disagreements and at cases where an appropriate option is missing. Choose a threshold for the consequences of your task, then keep monitoring after release.
Make escalation useful
A case sent to a human should retain the original situation, the available options and the reason the workflow paused. The API gives your application structured probabilities and a gate action. Build the review experience around the information the reviewer needs.
An explicit escalation option or a separate suitability check can help when the offered actions are incomplete. It still needs evaluation. Good routing is more than choosing a large confidence number; it is deciding where the model's job ends.