Jev vs asking an LLM for JSON

JSON mode on an LLM still samples tokens. It can emit a key you did not declare, skip a field, or invent a label. You then write a parser and a retry loop.

Jev's output space is the criteria you sent. A choice question with three options returns one of those three plus a distribution over all three.

Keep the LLM for generation. Put Jev on the line of code that currently says `if (await classify(text) === "billing")`.