Keeping these jobs separate makes it easier to test failures. When a total is wrong, inspect the calculation. When a case is misrouted, inspect the evidence and question. When an explanation is misleading, inspect its source and wording.
Combining tools works best when each has an explicit responsibility. Business rules that can be written down stay in code. Jev answers the questions rules cannot, such as what a carrier note means or how urgent a complaint is.
For hard cases that need free-text reasoning, the orchestrator can escalate to a reasoning model—for example one deployed in Azure OpenAI—after policy has decided that escalation is warranted.
Jev does not write replies, produce code or generate explanations of its reasoning. Typed Choice, Score and Noul answers are the contract; your software decides what to do with them.