Jev use cases: what a decision model is actually good for

Twelve Jev use cases, measured side by side

Each row is the call printed on that page, made with the question definitions in its code sample and run against typesafe/jev-1.13 through openrouter.ai: the first eight on 2026-09-18, the last four on 2026-10-11. Some states are borderline on purpose, like a rude but not abusive post, so the probabilities have something to say. Latency is the median of three runs, network floor included (198.8 ms, then 260 ms); cost is input tokens at $0.042 per million, the only thing Jev bills.

Question text is billed as input, so the five-word routing state still billed 369 tokens and the priciest pattern costs only about 2.0 times the cheapest. More questions barely move latency: phishing asks five and sits among the one-question patterns. Chat-model comparisons on each page come from a 2026-09-20 head-to-head, re-runnable from jev-measured; the full tables are on the benchmarks page.

Which Jev use case fits your pipeline

Sort the Jev use cases by where the decision happens in your system, not by industry.

• When something arrives. Support ticket triage, content moderation, phishing detection and lead scoring label a ticket, post, email or form with several judgements from one call.
• When records come in from outside. Invoice classification, product categorization and CSV data validation check records before they reach the ledger, catalogue or warehouse; an AI news filter does it per reader.
• Before a model call. LLM model routing decides how much model a prompt deserves before you pay for it.
• Inside an agent loop. Agent tool selection picks the next tool and leaves only the arguments to the LLM.
• Between retrieval and generation. RAG reranking scores each retrieved passage and keeps the ones that answer the question.
• Between generation and the user. Agent output guardrails check a draft before it is sent.

If your step is not listed, apply the test all twelve pass: can every acceptable answer be written down before the request? Then it is a Choice, Score or Noul question; if the answer is a sentence, use a generative model.

The pattern underneath all of them

Every use case here makes the same move: find the step where the answer is a decision rather than a sentence, stop paying a generative model for it, and, since extra questions cost almost nothing, ask the adjacent ones too. In the phishing benchmark, five cheap signal questions beat one big question. The other shared habit is the confidence gate: act alone when the answer is decisive, and pass the rest to a person, a bigger model or a review queue.

Where none of these Jev use cases fit

Besides writing nothing (see the questions below), Jev chooses only among options you send, so open-ended extraction has no clean way in. It reads text only through a hosted API with closed weights, and it is weakest at counting, arithmetic and dates, so several pages compute those in code first. The full list is on the limitations page.

Common questions about Jev use cases