Guide
Accuracy has to be a number.
A model wired to a button is easy. The hard part is knowing whether it helps, what it costs, and who is allowed to let it act.
Name the decision
Write down what the model is allowed to decide, what data it can see, and where a person stays in the loop. If that page is vague, the feature is a demo.
Evaluate before launch
A baseline you can compare against. Not a feeling in a meeting. LobeStack stores rejected posts as preference so the next batch changes. Averil refuses vanity scores.
Keep a person on the expensive path
Where being wrong costs money or reputation, approval is a product feature. Nothing publishes because a model felt confident.