Jev FAQ: Models, Primitives, SDKs and Integrations

What is Jev?

Jev is TypeSafe AI’s System One model. It evaluates state against typed questions and returns bounded decisions your software can interpret. It is useful for semantic classification, routing, scoring, and verification when the answer space is known.

Who makes Jev?

TypeSafe AI develops Jev. Jev Manual is an independent educational resource and is not owned, sponsored, or endorsed by TypeSafe AI. Official documentation and provider references are linked on each guide.

Is Jev an LLM?

TypeSafe presents Jev as a System One decision model. Its public API focuses on bounded Choice, Score, and Noul outputs rather than chat completions. This manual does not infer an unpublished internal architecture from that interface.

Does Jev generate text?

Jev does not provide open-ended text generation through this decision interface. Use an LLM or a template for prose. Jev can select a route or template, while code or another model produces the final message.

Can Jev write code?

Use a generative coding model for code creation. Jev can help classify a task or assess a proposed action, but its bounded output does not replace a code-generation interface.

Can Jev process images?

The documented direct model accepts text only. Extract text or structured descriptions from images before passing state to Jev. Evaluate the extraction stage separately because missing visual evidence cannot be recovered by the decision stage.

Can Jev process audio?

Not as native audio input on the documented direct API. Transcribe audio first, preserve useful context, and evaluate the resulting text. A transcription error can change the decision, so test the entire pipeline.

What is Noul?

Noul represents the probability that the answer to one yes/no question is yes. Values near zero favor no; values near one favor yes. A value near 0.5 means uncertainty, not a medium degree on a scale.

What is Choice?

Choice selects one item from a criteria map. It returns the selected label, probabilities over all options, and confidence. Departments, intent classes, and approved handlers are common examples.

What is Score?

Score evaluates an ordered rubric. It returns a weighted score, the rubric legend, a probability distribution, and confidence. It is a semantic rating, not exact measurement or numerical regression.

Does Noul have confidence?

No separate confidence field is returned for Noul. Use its yes probability according to your workflow’s policy. Choice and Score have their own confidence fields derived from their distributions.

What model is current?

The current direct model is jev-1.13.0, checked on 2026-09-21. The model guide tracks aliases and pinning advice. Provider-specific names and availability are checked separately.

What is jev-latest?

It is the direct stable alias and currently points to jev-1.13.0. It can change when a stable release ships. Pin a version when your production thresholds depend on tested probability behavior.

How much does Jev cost?

The documented direct input price is $0.042 per million tokens, with free output. Provider pricing may differ. Use the calculator as an estimate and compare it with actual usage.

How many Choice options are allowed?

The direct API accepts up to 255 options. That is a maximum, not a recommendation to create enormous taxonomies. Use distinct categories and evaluate ambiguous cases.

How many Score levels are allowed?

A Score should use 2–10 ordered descriptions. More levels do not automatically produce better judgments. Define clear boundaries and interpret the returned legend.

Can I ask multiple questions?

Yes. Independent questions can share one state in one request. A question cannot read another answer in that same batch; dependent stages require code between calls or a different decomposition.

Can I use Jev with Python?

Yes. Install typesafe-sdk and use TypeSafeClient or AsyncTypeSafeClient. The Python tutorial includes a complete support-ticket example, environment setup, response handling, and production notes.

Can I use Jev with TypeScript?

Yes. Install @typesafe-ai/sdk and use TypeSafeClient.systemOne with choice, score, and noul helpers. Node.js 20+ is required by the documented SDK.

Can I use Jev with OpenRouter?

OpenRouter documents a TypeSafe-compatible System One endpoint and a separate alpha Decisions surface. Use an OpenRouter key, its documented base URL, and its model identifiers rather than mixing direct-provider settings.

Can I use Jev with Vercel?

Yes. Vercel documents an evaluation interface using AI SDK 7+ and experimental_evaluate with typesafe-ai/jev. Its boolean primitive returns probability rather than the direct API’s noul field. Use the provider guide for the exact contract.

Can I use Jev with Cloudflare?

A Cloudflare Worker can call the direct TypeSafe API using a server-side secret. The plan also identifies a Workers AI Jev route; availability must be confirmed against the current Cloudflare model catalog before using the binding-specific variant.

Can Jev do math?

Do exact arithmetic in code. Jev can select the relevant record or interpret semantic intent, but a model answer should not substitute for a calculator, accounting function, or exact numeric comparison.

Can Jev compare dates?

Use deterministic date/time code for comparisons and arithmetic. Jev can identify bounded semantic components, then a timezone-aware application resolves them using an explicit reference date and convention.

Should I pin the model?

Pin when probability behavior affects production decisions. Log the actual response model, evaluate upgrades on fixed labeled examples, and move aliases only after checking the resulting error and review rates.

Is jevmanual.com official?

No. It is an independent developer manual. Official facts, provider-specific behavior, and educational examples are distinguished, and the relevant sources are linked so readers can verify details.