how does jev work?

The End-to-End Decision Path

A useful mental model is “evidence in, bounded measurements out.” Code first gathers observations and projects them into a small, named state. It then defines questions whose answer spaces exist before inference. Jev evaluates the questions, and code consumes the resulting distributions. This observable contract is the architecture developers can rely on; an imagined diagram of Jev’s internal layers is not.

The boundary matters operationally. A database lookup, permission check, amount calculation or schema validation should happen outside Jev. A semantic judgment—whether a message requests a refund, which approved queue best fits it, or how urgent it is under a written rubric—can become a Jev question. The final refund, route or escalation remains an application action.

State Is Evidence; Questions Are the Measurement Instrument

State may be a string, object or array of text values. Field names are part of the evidence because they tell the model what a value represents. More context is not automatically better: TypeSafe’s Jev 1.13 notes describe degradation with large amounts of irrelevant state. Build and version a deliberate projection rather than serializing an entire trace or database record.

Each question has its own instructions and answer type. A Noul measures one yes/no proposition, a Choice selects among caller-defined alternatives, and a Score maps evidence onto ordered descriptive levels. IDs identify answers for code; TypeSafe says they are not semantic instructions. Put meaning in instructions and criteria, not in an opaque key such as q7.

{
  "model": "jev-1.13.0",
  "state": {
    "customer_message": "The export failed again; this blocks payroll.",
    "plan": "business"
  },
  "questions": {
    "owner": {
      "type": "choice",
      "instructions": "Which approved team should own this ticket?",
      "criteria": {
        "billing": "Charges, invoices or payment",
        "technical": "Product errors or unexpected behavior",
        "other": "Neither category"
      }
    },
    "blocks_critical_work": {
      "type": "noul",
      "instructions": "The customer says the issue blocks a critical business process"
    }
  }
}

Parallel Questions Are Not a Hidden Chain of Thought

TypeSafe’s launch post contrasts Jev’s parallel sampler with an autoregressive LLM that emits one token after another. Multiple Jev questions can share a state and be answered in one call, but they are independent. The blocks_critical_work answer above does not become an input to owner. If the first result changes the available candidates or evidence for a later decision, application code must issue a second request.

This creates a useful architecture rule: batch independent measurements; orchestrate dependencies explicitly. It makes the decision graph visible in code and logs. It also prevents a convenient one-call optimization from silently changing the meaning of a later question.

Keep the Distribution, Not Only the Winning Answer

A Choice or Score answer contains more information than its selected value. A 0.51/0.49 Choice and a 0.99/0.01 Choice may choose the same option but support different actions. Noul returns the probability of its condition rather than a Boolean. Store the raw answer with model, provider, question, criteria and state-projection versions so thresholds can be changed or audited without pretending the original result was binary.

Probability is not authorization and a high value is not an explanation. Fit act, review and decline bands from held-out outcomes and the cost of each error. The confidence guide separates distribution concentration from calibration; the RLCD guide explains why calibration is a population property rather than a guarantee about one call.

What the Public Architecture Does and Does Not Establish

The launch latency was generally measured from West Coast laptops near TypeSafe’s service, and TypeSafe notes that short dense state favored Jev in its demonstration. Benchmark state construction, network, provider, retries and fallback in the environment that will actually run. Review the limitations before using the result in a consequential path.

The Production Runtime Around Jev

Jev can feed an offline evaluation, a pre-action policy, or an ordinary application branch. Those are different operational contracts even if they reuse the same question. Places to experiment and deploy are compared in Jev platforms.

• Validate schemas, identity, permissions, candidates and exact constraints before inference.
• Set a total latency budget and explicit timeout, rate-limit and malformed-response behavior.
• Log the serving provider as well as the model ID; gateway behavior can differ from TypeSafe’s direct API.
• Shadow the decision against outcomes before it controls an action.
• Replay representative fixtures when model, question, state projection or provider changes.