Jev: Decisions, Not Text
Stop parsing LLM prose just to get a yes or no.
Why use it
Known answer set? Let Jev make the fast, cheap, typed call, and save the LLM for real reasoning.
Context
Jev, from TypeSafe AI (out of stealth on 15 September 2026), is a "System One" model in Kahneman's sense: fast judgement instead of slow written reasoning. It doesn't write text. You define the question and the answer space in advance, and it returns a typed decision: a Choice from your list, a Score on your scale, or a yes/no probability, each with a calibrated confidence.
The episode's system is a support pipeline that routes every incoming ticket. Today that's an LLM call per ticket whose prose answer gets parsed into a category: slow, comparatively expensive, and one odd phrasing away from a parse error. Ticket routing is high-volume and the answer space is fixed, which is exactly the workload a decision model targets.
Architecture
| Component | Role | Notes |
|---|---|---|
| Customer ticket | Input | "Where's my refund?" |
| Jev decision model | Choice(bill, bug, faq) + confidence | Answer is always one of the options; no free text |
| Confidence threshold | Plain code: if confidence >= 0.85 | 85% is the example threshold Correlation One gives; illustrative |
| Billing flow | Automated handling of confident, routine cases | Deterministic workflow, no LLM call |
| LLM agent | Uncertain or complex cases | Reasoning, tools, drafting replies; a human tier can sit behind it |
Request flow
- The ticket and a typed question ("Which queue: bill, bug or faq?") go to Jev.
- Jev returns one option from the list plus a confidence.
- Code branches on the result: at or above the threshold the matching workflow runs; below it, the ticket goes to the LLM agent (or a human).
The advantages
- Decisions, not text. The output is a typed value, so there is no prompt-for-JSON, no regex over prose and no "the model said something unexpected" branch.
- Guaranteed inside the set. The answer is always one of the options you defined (or a score on your scale), so every downstream branch is known in advance.
- A confidence you can threshold. Each decision carries a confidence, which gives a natural escalation rule: act automatically when sure, hand off when not.
- Fast and cheap per call (vendor-claimed). TypeSafe claims more than 100x lower cost and about 200x faster than frontier LLMs on classification. Not independently verified.
- Right-sized for the job. As a router, gate or classifier in an agent pipeline it takes the high-volume, closed questions, so the LLM only sees the cases that need open-ended reasoning or generation.
Where it fits (and where it doesn't)
- Good fit: routing, triage, moderation labels, screening, approvals, tool selection, anything whose answers you can list up front.
- Not a fit: open-ended writing, multi-step reasoning, code generation, or questions whose answer space you can't enumerate. Keep the LLM there.
- Still a model: it reads natural language, so treat its input as untrusted and keep the usual guardrails on high-impact actions (Check Point showed planted evidence can flip its decisions). This episode focuses on the benefits.
Trade-offs
- You must design the answer set up front. A missing category forces bad picks, so add an "other" option and watch its rate.
- Two models to operate instead of one: thresholds need calibration on your own data, and the escalation path must scale.
- Public accuracy benchmarks are still limited (noted by The Register), so evaluate it against your current LLM classifier before switching.
Numbers worth knowing
- Vendor-claimed, unverified: more than 100x lower cost and roughly 200x faster than frontier LLMs on classification, with responses in the tens to hundreds of milliseconds (TypeSafe, via Correlation One).
- The Register reports input pricing of $0.042 per million tokens with no output-token charge, and latency from about 150 ms to about 620 ms for complex decisions.
- The 85% threshold (Correlation One's example) is illustrative.
Coming next in the series: Model Cascades
