Determinism and confidence
Separate repeatable calculations from uncertainty in Jev evidence.
Same model and facts, same answer
Decision-only execution applies the supplied definition to matching facts, locally, with no provider involved. Repeating that model and state produces the same decision values and rule trace; each call still reports usage and incurs its own charge. Because you submit the definition with every case, nothing is generated or reinterpreted between runs.
Jev evidence remains distinct
In combined requests, Jev Noul answers become numeric probabilities, Choice answers become strings, and Score answers become numbers. The native model evaluates those values deterministically. Jev confidence does not automatically propagate through native execution as a distribution.
For example, a model may explicitly require damage_probability >= 0.8. The resulting boolean is an answer to that rule, not a claim that the customer’s report is certainly true. Jev answers and probabilities remain in the public answers; their input values and sources appear in receipt.read. Native answers own overlapping names. The full upstream response is console/internal only.
The same validated facts give the same calculation. The same raw message is not a guarantee of identical Jev answers on a later call. See Using Jev with aityx.
What confidence does and does not say
Native matching-fact answers have confidence 1 because they follow from the supplied model and facts. That does not certify the facts or the rules. Use the receipt to distinguish supplied facts from Jev evidence; use expected cases to check calculations.
Your application decides how to handle uncertainty, missing facts and any action based on the answer. Errors explains failed and partially completed calls.

