Same ticket. Two ways to decide. One is a sentence, one is a value.
A working demo of the "typed decision" pattern TypeSafe's Jev model introduced: instead of generating a sentence you then have to parse, the model fills in a fixed schema — Choice, Score, Boolean — in a single pass. This runs the same support ticket down two lanes so you can see the shape of the difference, not just read about it.
How this demo works: Jev's public API isn't open yet (closed beta, launched Sept 2026), so the left lane is a local simulator built to the same I/O contract — instant, schema-locked output. The right lane simulates a conventional prose-generating LLM at realistic token speed, then parses its own sentence back into the same fields. Swap-in code for the real Jev endpoint is at the bottom.
JEV (SYSTEM 1)—
Waiting for a ticket…
CONVERSATIONAL LLM (SYSTEM 2)—
Waiting for a ticket…
Throughput at scale
Jev's pitch isn't "better," it's "cheap enough to call constantly." Run 50 simulated tickets through both lanes to see why that matters for tight loops — routing, moderation, agent tool-selection — where you'd never afford a full chat completion per decision.
The contract
Every field is pre-declared before the call. There's no string to parse and nothing off-schema to catch in a try/catch — the model can only ever land on one of the values you gave it.