Our summary
Two AI answers can share a confidence number while disagreeing in very different ways. Bruno Coelho's explainer lets readers move probability bars and see the numbers change. It helps people planning a Jev feature understand what the returned numbers describe before choosing rules that act on them.
For a choice between categories, confidence compares the winner with an equal split across every option. For an ordered rating, it also accounts for how far other likely levels lie from the peak. A yes-or-no question returns a yes probability, with no separate confidence value.
These controls simulate answer patterns; moving them does not ask the model a new question. The example decision bands are not tested safety limits. An average rating can hide disagreement, and tied peaks raise an unresolved calculation question. Use labeled cases from your own work before adopting a cutoff.
Key takeaways
- Use unordered choices for categories and ordered ratings only when a near miss differs meaningfully from a far miss.
- Read the full probability pattern: the same winning probability can hide a close rival or several weaker alternatives.
- Ask separate yes-or-no questions for anger and a refund request, then combine the answers in code.