jev·agent

API

Jev API reference

One endpoint, one state, as many typed questions as you like — evaluated in parallel and returned in a single response.

Endpoint

Endpoint
POST https://api.typesafe.ai/v1/systemone
Model routes
jev-latest jev-1.13
Auth
API key issued by TypeSafe (early access), or via a gateway
Context
~32,000 tokens per request
Modalities
Text only — no image or audio input at launch
Latency
70–500ms end-to-end
SDKs
Python, JavaScript, plus raw HTTP

The request model

A request has two parts. The state is what you are asking about: raw text, a transcript, a document, or serialised program state. The questions are what you want decided about it. Every question is evaluated independently, in parallel, against that same state — which is why asking ten questions costs roughly the latency of asking one.

Questions are passed as a dict keyed by an id you choose, and answers come back under those same keys. That is what makes fan-out ergonomic: name the question, read the answer.

python
from typesafe_sdk import Choice, Noul, NoulCriteria, Score, TypeSafeClient

response = client.system_one(
    model="jev-latest",
    state=the_text_or_program_state,
    questions={
        "route": Choice(
            instructions="Which team should handle this?",
            criteria={
                "billing":   "Payments, invoicing, refunds",
                "technical": "Bugs, outages, integrations",
                "sales":     "Pricing, upgrades, new accounts",
            },
        ),
        "severity": Score(
            instructions="How severe is this?",
            criteria=["Low", "Medium", "High", "Critical"],
        ),
        "on_topic": Noul(
            instructions="Is this about our product?",
            criteria=NoulCriteria(true="Clearly about our product",
                                  false="Unrelated"),
        ),
    },
)

The three primitives

Choice — Choose an option from a list.

You supply a state and a fixed set of options. Jev returns which option fits, a probability for every option, and a confidence value — so your code can branch on the winner or fall back when confidence is low.

Returns

choice, probabilities, confidence

Reach for it when

Routing, tool selection, intent classification, triage.

Score — Score the state on a rubric.

You define ordered descriptive levels (a rubric). Jev returns a score against that rubric plus the probability distribution across levels, which you can sort or threshold on.

Returns

score, probabilities, confidence

Reach for it when

Quality grading, ranking, rubric-based evaluation, content moderation tiers.

Noul — Is this statement true?

A single yes/no question that returns the probability of "yes" rather than a hard boolean — letting you pick your own threshold instead of trusting a coin flip.

Returns

a noul value in the 0–1 range

Reach for it when

Boolean gates, guardrails, verification checks, filters.

Confidence is the whole point

Choice and Score both return a full probability distribution plus a confidence value, and that is the feature that makes Jev safe to put on a hot path. A traditional classifier hands you a label and shrugs. Jev tells you how sure it is, so you can write the rule your product actually needs:

python
response = client.system_one(
    model="jev-latest", state=ticket, questions={"route": route_question}
)

answer = response.answers["route"]
if answer.confidence >= 0.85:
    dispatch(answer.choice)     # fast path: Jev decided
else:
    escalate_to_llm(ticket)     # slow path: not sure enough

Response fields follow the primitive: .choice and .confidence for Choice, .score for Score, .noul for Noul, plus response.usage.input_tokens for billing.

Pick the threshold empirically against your own labelled data — the right number depends entirely on what a wrong decision costs you.

Documented architectural patterns

TypeSafe's docs describe several composition patterns worth knowing: fan-out (many questions against one state in a single call), routing (a Choice that selects the next handler), and composite scoring (several Score questions combined into an overall judgement). Their cookbook section covers 20+ worked examples including legal analysis, RAG passage classification and entity alignment.