Introduction
Provide the evidence in state and define questions using noul, choice or score. Each question evaluates that state independently, and the task returns answers under the names you supplied. Keep the complete request within 64k tokens and state plus each question within 32k.
| Field | Value |
|---|---|
| Input | Text |
| Output | Structured data |
| Publisher and model | TypeSafe — Jev 1.14 (jev-1.14.0) |
| Release date | 14 October 2026; previous build Jev 1.13 |
| Public API name | typesafe/jev-1.14/decision |
| State input | String, JSON object or JSON array |
| Question types | noul, choice and score |
| Questions per call | Up to 64, independent and parallel |
| Context limits | 64k tokens per request; 32k for state plus one question |
| Choice options | 1–255 named options with descriptions |
| Score rubric | 2–10 ordered levels, from low to high |
| Task output | output.decision contains answers and inputTokens |
| Observed latency | About 0.6 s in our calls; earlier probe 0.62 s |
| Billing unit | Measured input tokens; output tokens free |
Key features & highlights
A probability, a category or a rubric
Select among 3 question types to match the value your code needs. Noul returns a yes probability from 0–1; choice returns a named option with confidence and option probabilities; score returns a probability-weighted level with confidence, a legend and level probabilities.
One state, independent judgments
Up to 64 questions evaluate the same state independently and in parallel, with little latency change as questions are added. Ask about a ticket’s team, urgency and refund intent together, then combine those answers in application logic.
Inspect uncertainty beside the answer
Our short-video review returned a proof score of 1.44 with confidence 0.28, while its offer probability was 0.98. Those values describe different judgments: a clear offer does not make the evidence for a product claim equally clear.
Keep escalation in your code
Define the probability and confidence thresholds that send an item to a person or a generative model. Jev 1.14 supplies structured judgments without a written rationale, so review the original evidence when an uncertain answer affects the next action.
Request parameters
| Parameter | Required | Description |
|---|---|---|
questions | Yes | Name up to 64 questions and give each a type and instructions; use criteria for 1–255 choice options or 2–10 score levels arranged from low to high, with optional true/false des… |
state | Yes | Supply a string, JSON object or array containing the evidence for every judgment; questions cannot rely on another question’s answer. at least 1 characters |
Pricing
$0.0000504 / 1,000 tokens
Billed per measured input token; output tokens are free. The hold at acceptance is sized from the 64k-token request limit, and the difference is released when the request settles.
Use cases
Screen short-video advertising copy
Provide a product-clip transcript, on-screen text and frame descriptions as state, then ask about hook style, proof and an offer. Our shown example returned problem as the hook, proof 1.44 and an offer probability of 0.98; it used 537 input tokens.
Route a billing support ticket
Submit a message such as “I was charged after cancelling; please refund the renewal” with relevant account metadata. Request team, urgency and refund intent; our support example returned billing, urgency 1.6 and refund_requested 0.96 across 422 input tokens.
Separate review sentiment from intent
Use a review such as “The finish chipped, but I would order again if that is fixed” to ask separate sentiment, issue and return-intent questions. Our review example produced sentiment 1.45, quality as the issue and would_return 0.97 using 435 input tokens.
Check a listing before publication
Send a listing title, description and required fields, then ask whether the title agrees with the description and whether a required detail appears. Use the resulting yes/no probabilities as publishing gates, with ambiguous listings held for a person.
Score a lead before escalation
Pass an enquiry such as “We need a team rollout next month” with its supplied company and timing fields. Define ordered qualification levels and ask whether information is missing, then use scores and uncertainty to decide which leads need a generative follow-up.
Prompt guide & pro tips
Prepare the shared state
Place the ticket, transcript or review in state as a string, JSON object or array. Include the evidence each question needs, since questions cannot read one another’s answers.
Define typed questions
Give each question a name, type and instructions; add criteria to describe choice options or ordered score levels. Choose noul for a yes/no probability, choice for a category or score for a rubric.
Call the decision endpoint
POST /api/v1/jobs/createTask?wait=30 with model set to typesafe/jev-1.14/decision and input containing state and questions. Read output.decision from the completed task, or create the task and poll recordInfo.
Validate and route answers
Compare probabilities and confidence with human labels on your own sample before choosing thresholds. Route uncertain cases to review and record inputTokens alongside the decisions.
Jev 1.14 API for typed decisions
Jev 1.14 is TypeSafe’s System One decision model for classifying and scoring text or JSON. This endpoint returns yes/no probabilities, named choices and rubric scores without generating prose or reasoning text. Apply the answers to support routing, review checks and escalation.
Send a decision request
Call POST /api/v1/jobs/createTask?wait=30 with the public model name typesafe/jev-1.14/decision. This illustrative billing request is separate from our measured example.
{
"model": "typesafe/jev-1.14/decision",
"input": {
"state": {
"message": "I cancelled before renewal but was charged again. Please refund it today."
},
"questions": {
"team": {
"type": "choice",
"instructions": "Which team should handle this message?",
"criteria": {
"billing": "Charges, renewals and refunds",
"technical": "Product errors and access problems"
}
},
"urgency": {
"type": "score",
"instructions": "How urgently does this message request action?",
"criteria": ["No time pressure", "Action requested soon", "Immediate action requested"]
},
"refund_requested": {
"type": "noul",
"instructions": "The customer is asking for money back",
"criteria": {
"true": "A refund or repayment is requested",
"false": "No refund or repayment is requested"
}
}
}
}
}
Read the task output
Completed results expose output.decision.answers and output.decision.inputTokens. You can also create the task and poll recordInfo. This excerpt contains selected fields from our measured support-triage response.
{
"output": {
"decision": {
"answers": {
"team": {"choice": "billing", "confidence": 1.0},
"urgency": {"score": 1.6, "confidence": 0.4},
"refund_requested": {"noul": 0.96}
},
"inputTokens": 422
}
}
}
Parameters and answer types
State supplies shared evidence for up to 64 questions, evaluated independently in parallel.
| Field | Accepted value | Practical use |
|---|---|---|
state | String, JSON object or array | Label the evidence fields |
questions | Object keyed by caller-chosen names | Map answers to checks |
type | noul, choice, score | Select a yes probability, category or ordered score |
instructions | Question instructions | Ask for one focused judgment |
criteria for noul | Optional true / false descriptions | Define both interpretations when wording is ambiguous |
criteria for choice | 1–255 named, described options | Define category boundaries |
criteria for score | 2–10 ordered level descriptions | Arrange the rubric from low to high |
A noul answer is the probability of yes from 0–1. Choice returns choice, confidence and per-option probabilities. Score returns a probability-weighted score, confidence, a level-to-text legend and level probabilities; fractional scores are expected, so read them against your rubric rather than as category labels.
Improve the decision criteria
Explicit evidence and criteria help define useful routing rules:
- Include every fact the judgment depends on in state; another question’s answer is unavailable to it.
- Describe category boundaries and score levels with concrete wording rather than unexplained labels.
- Ask separate questions for separate factors, such as refund intent and urgency, then combine them in code.
- Compare a labelled sample with the returned probabilities and confidence, and select review thresholds from those disagreements.
Low confidence signals uncertainty: escalate to a person or a generative model. Noul has no separate confidence field.
What our examples showed
Our calls returned in about 0.6 s, and an earlier 3-question probe recorded 0.62 s with 425 input tokens. TypeSafe says end-to-end latency is 70–500 ms in its launch announcement; that publisher statement is separate from our observations.
| Example | Selected results | Input tokens |
|---|---|---|
| Short-video review | hook: problem, confidence 1.0; proof: 1.44, confidence 0.28; has_offer: 0.98 | 537 |
| Support triage | team: billing, confidence 1.0; urgency: 1.6, confidence 0.4; refund_requested: 0.96 | 422 |
| Review check | sentiment: 1.45, confidence 0.32; issue: quality, confidence 1.0; would_return: 0.97 | 435 |
A confident category can coexist with an uncertain score; these examples do not establish accuracy on your dataset.
Pricing and usage
Billing uses measured input tokens, and output tokens are free. Consult the displayed price and track inputTokens. The acceptance hold is based on the 64k-token request limit, with the difference released when the task settles.
Limits and troubleshooting
This endpoint accepts up to 64 questions, 64k tokens per complete request and 32k tokens for state plus any single question. Shorten state or criteria when context is exceeded; fewer questions may still leave one over its limit.
Text and JSON are supported here; image, video and audio files and box questions are unavailable. Represent a clip with a transcript, on-screen text or frame descriptions. If a request is rejected, check the question type, criteria shape, option counts and both context limits before resubmitting. For wrong judgments, revisit evidence and criteria.
Choose Jev or an alternative
Choose Jev for typed judgments and a generative model for prose.
| Option | When it fits |
|---|---|
| Jev 1.14 | Probabilities, categories and rubric scores |
| Jev 1.13 | Preceding build; no measured sibling comparison here |
| Generative model | Written explanation or reply |
| Human review | Ambiguous evidence or uncertain judgments |
API quick start
# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
-d '{
"model": "typesafe/jev-1.14/decision",
"input": {
"state": "example",
"questions": {
"example": {
"type": "noul",
"instructions": "The customer asks for a refund after being billed twice."
}
}
}
}'
# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
-H "Authorization: Bearer $SPICY_API_KEY"
See the API reference for authentication, the complete schema, polling, and error handling.
