jev: typed ai decisions, evals and policies

What Jev Does

A Jev request has two main parts: state, which contains the relevant text or structured fields, and questions, which defines the decisions the application needs. Jev evaluates independent questions against the shared state and returns machine-readable answers. Code still owns control flow, exact rules, calculations, permissions and the final action.

This is narrower than a chat model by design. Jev does not produce an explanation, call a tool or invent a value outside the answer space. A support workflow might ask which team should receive a ticket, how urgent it is and whether it requests a refund. The application can route the ticket, request review or decline to act based on those separate results.

The Request Is a Decision Graph Boundary

The API boundary is deliberately small. State carries observations; questions describe judgments; the response carries answers and usage. Question IDs exist for the caller and are not part of the semantic instruction. Independent questions can be evaluated in parallel, but they cannot read one another’s answers. If one decision changes what evidence or candidates exist for the next, make a second request.

{
  "model": "jev-1.13.0",
  "state": {
    "message": "I was charged twice. Please fix this today.",
    "account_tier": "team"
  },
  "questions": {
    "queue": {
      "type": "choice",
      "instructions": "Which team should handle this request?",
      "criteria": { "billing": null, "technical": null, "other": null }
    },
    "refund_requested": {
      "type": "noul",
      "instructions": "Does the message request a refund?"
    }
  }
}

What TypeSafe Says About the Model Architecture

TypeSafe’s launch post says Jev uses a new model architecture, a parallel sampler and RLCD. It contrasts Jev with autoregressive LLMs that emit one token after another, saying Jev produces typed outputs and their probabilities in parallel. That public description explains the interface and latency thesis, but it does not disclose enough detail to infer parameter count, layers, training corpus, reward construction or reproduce the model.

TypeSafe reports 70-500 ms end-to-end response time from its early-access service and says its published evaluations were generally run from laptops on the US West Coast, where the service was based. Treat that range as a dated first-party measurement, not a latency guarantee. Network path, state size, question count, provider, retries and load all affect an application measurement.

The Three Jev Primitives

The types are not interchangeable. A Choice compares alternatives relative to one another. Separate Nouls test labels independently, so several labels may be plausible or none may fit. A Score is for an ordered semantic rubric, not arithmetic. RLCD and calibration explains what the returned probabilities do and do not mean.

Jev as an Eval and as a Policy Signal

As an agent evaluator, Jev can score a completed session against bounded semantic criteria: whether the task was completed, whether a tool choice fit the request, or whether cited evidence supports a claim. The result should retain the model version, question version and full distribution beside the trace so teams can measure it against observed outcomes.

As a policy signal, Jev can judge the smallest relevant state before an action. The policy layer, not Jev, decides allow, observe, review or block. Exact checks, permission boundaries and authorization remain deterministic. Teams should backtest and shadow the semantic check before enforcement, and define what happens on timeout, rate limiting or uncertainty.

Where Jev Fits and Where It Does Not

• Good fit: classification, routing, semantic verification, rubric scoring, candidate selection, trace evaluation and other repeated bounded judgments.
• Keep in code: arithmetic, counting, timestamps, schema validation, path and permission checks, exact allowlists and deterministic business rules.
• Use a generative model: writing, coding, explanations, open-ended extraction and multi-step reasoning.
• Test carefully: non-English data, adversarial text, long state with irrelevant detail and consequential automated actions.

Jev Docs, API Access and GitHub Code

If you searched for “Jev AI,” “AI Jev,” “Jev model” or “TypeSafe Jev,” these names refer to the same TypeSafe AI model described on this page. The Chinese query “Jev 模型” means “Jev model.” Jev is not an OpenAI product. TypeSafe’s documentation is the primary source for the model contract, primitives, current aliases, limits and direct price. The quickstart explains how to use Jev through the official SDKs and HTTP API.

TypeSafe’s public GitHub organization contains its Python SDK, JavaScript SDK and agent skill. Those repositories are client tools, not Jev’s model weights. The hosted model is not open source, and a similarly named community repository is not automatically official. See the open-source status, platform access map and launch tracker before choosing a route.

Public access conditions can change. TypeSafe publishes a Playground and direct API documentation, while Cloudflare has a verified provider route. An OpenRouter page exists, but catalog availability was unconfirmed when checked September 22, 2026. For Jev evals or policies in Failproof AI, beta users can email nikita@befailproof.ai.