Frequently asked questions

Short answers to the questions we get most, with links to the long form.

Is Jev a replacement for LLMs?

No — and it isn't trying to be. It complements the generative layer: LLMs generate, Jev handles routing, verification, scoring and gating — the thousands-of-times-per-second micro-judgments. Confidence gating decides when to escalate up to an LLM or a human.

Why not just "LLM + JSON mode"?

Three words: fast, cheap, calibrated. A single forward pass returns in milliseconds; $0.042/Mtok undercuts the cheapest chat models by an order of magnitude; and the probabilities are the training objective, not a byproduct — so there's a ready-made, trustworthy confidence score for gating. See the cost math.

What is it good for?

Intent routing, content moderation, ticket triage, RAG re-ranking, risk scoring, agent action gating, extraction validation — any judgment point that is too fuzzy for rules and too expensive to send to an LLM.

Does it work in Chinese / other languages?

It works, but English is by far the most accurate and CJK precision is lower. For production in Chinese, run your own eval on real data and rely on the confidence threshold for fallback routing. See the 中文 edition.

How do I control cost?

Three levers: batch questions into one call (12.2× effect), shrink the state (input-only billing means state size is your cost), and gate so expensive downstream actions only fire on high confidence.

Can I fine-tune it for my domain?

No per-customer fine-tuning. Adaptation is via: richer domain content in the state, precise rules in the criteria, and weights composed in your code.

Is this site affiliated with TypeSafe?

No. OpenJev is an independent, unofficial community resource. See the terms and about pages.