Compare GPT-6 Luna and other decision models in one playgroundExplore the playground

Developer documentation

Build with Decisions API

Start with one real decision: choose an available decision model, prepare state, define typed questions, and hand the probability-backed result to your code.

Meet Decisions API

A unified decision interface for software

Traditional LLMs mainly generate text for people to read. Decisions API provides one decision interface across available models, including Jev: send one state and typed questions, then get structured results your code can branch on, sort, and route.

Typed results

Parallel decisions

Probability and confidence

Supported models

Choose the right decision model for the job

Decisions API exposes the currently available decision models through one interface. Choose a model for the task, then define questions that match its supported types. Availability, speed, pricing, and result quality can change, so validate your own cases.

Catalog checked against OpenRouter on October 15, 2026: 21 model variants and 2 latest aliases. View the live OpenRouter catalog

GPT-6 Luna Decisions

OpenAI’s GPT-6 Luna decision model for classification, routing, moderation, and rubric scoring with Choice, Score, and Noul outputs.

Model ID

openai/gpt-6-luna-decisions

Supported question types

Choice · Score · Noul

Jev 1.14

TypeSafe’s Jev 1.14 decision model for typed choices, scores, and yes-or-no probabilities over text and images.

API alias: ~typesafe/jev-latest

Model ID

typesafe/jev-1.14

Supported question types

Choice · Score · Noul

Jev 1.13

The previous Jev release for workflows pinned to version 1.13 and comparisons with Jev 1.14.

Model ID

typesafe/jev-1.13

Supported question types

Choice · Score · Noul

Classify-1

A decision model for classifying text and documents with choice, score, and yes-or-no probabilities.

Model ID

scaledown/classify-1

Supported question types

Choice · Score · Noul

Drex v2.0

A 35B mixture-of-experts decision model with 3B active parameters for typed decisions over text and JSON.

Model ID

nace-ai/drex-v2.0

Supported question types

Choice · Score · Noul

Microsoft-Decision-1

A compact decision model that scores fixed answer options for classification, verification, and workflow control.

Model ID

microsoft/microsoft-decision-1

Supported question types

Choice · Score · Noul

Drex v1.5

A 9B decision model for choice, ordered-score, and yes-or-no questions over text and JSON.

Model ID

nace-ai/drex-v1.5

Supported question types

Choice · Score · Noul

Clef Omni

Cloudflare’s multimodal Clef model for typed decisions over text, images, audio, and video.

Model ID

cloudflare/clef-omni

Supported question types

Choice · Score · Noul

Clef

Cloudflare's open 27B multimodal decision model for typed answers.

Model ID

cloudflare/clef

Supported question types

Choice · Score · Noul

Clef Flash

Cloudflare's fast 9B Clef decision model for responsive structured decisions.

Model ID

cloudflare/clef-flash

Supported question types

Choice · Score · Noul

Decider V1.1 27B

Perplexity’s V1.1 multimodal decision model for typed answers over text and images, with a latest alias.

API alias: ~perplexity/pplx-decider-latest

Model ID

perplexity/pplx-decider-v1.1-27b

Supported question types

Choice · Score · Noul

Liquid d1

A fast decision model for structured decisions and probability estimates.

Model ID

liquid/d1

Supported question types

Choice · Score · Noul

Tev1 4B Experimental

An experimental decision model for bounded routing and classification with labeled options.

Model ID

togethercomputer/tev1-4b-experimental

Supported question types

Choice · Score · Noul

Mercury Decide

Inception’s paid Mercury Decide endpoint for structured decisions over text and images.

Model ID

inception/mercury-decide

Supported question types

Choice · Score · Noul

Mercury Decide (free)

Inception's free structured decision model for choices, scores, and yes-or-no judgments.

Model ID

inception/mercury-decide:free

Supported question types

Choice · Score · Noul

Solar Decide Flash

The low-latency Solar Decide variant for text classification, scoring, and yes-or-no judgments.

Model ID

upstage/solar-decide-flash

Supported question types

Choice · Score · Noul

Solar Decide

A decision model from Upstage for structured classification, scoring, and yes-or-no decisions.

Model ID

upstage/solar-decide

Supported question types

Choice · Score · Noul

Span-01

A specialized Noul behavior model. Use the Span playground to test behavior judgments.

Model ID

respan/span-01

Supported question types

Noul

Span-01 Lite

A lightweight Span behavior model for Noul yes-or-no judgments.

Model ID

respan/span-01-lite

Supported question types

Noul

Span-01 Lite Free

A free lightweight Span behavior model for Noul yes-or-no judgments.

Model ID

respan/span-01-lite:free

Supported question types

Noul

Kev 4B

A compact 4B decision model for typed answers and probability estimates.

Model ID

jaredpalmer/kev-4b

Supported question types

Choice · Score · Noul

Span models support Noul only. The Playground refreshes from OpenRouter and keeps the known models, including latest aliases, if loading fails. Jev 1.14 is the default; typesafe/jev-1.13 remains available for existing integrations. Latest aliases can change their target version; choose a versioned ID when you need a fixed release.

Discover current models

GET /api/playground/models is public and requires no API key. Successful responses set code to 0 and return model IDs and questionTypes in data.models. Send one of those IDs as model to POST /v1/systemone. List responses are cached for 5 minutes; provider failures return 502.

curl https://decisionapi.net/api/playground/models

Quickstart

Validate one decision in the playground

The Playground is the fastest way to understand Decisions API inputs and outputs. Choose an available model, validate the question, then create an API key and connect it to your product.

  1. 1

    Open the playground and choose a model

    Sign in, open the DecisionsApi playground, and choose a model from the available decision-model list.

  2. 2

    Prepare state

    Use text, a JSON object, or an array of text to provide the context the decision needs.

  3. 3

    Add questions

    Choose choice, score, or noul from the types supported by the selected model. You can mix them when the model supports all three.

  4. 4

    Connect your code

    Create an API key in your workspace, send the selected model ID, and call the production endpoint with an SDK or REST.

Input

Give state the context the decision needs

State is the content every question reads. Use a string for a simple case; use a JSON object when the decision needs a ticket, order, and policy together.

text

Natural language, a ticket, or a message

object

Structured records and nested fields

array

Context made of multiple text items

The Playground edits text and JSON state. Input modalities depend on the selected model: GPT-6 Luna Decisions accepts text and images. Image inputs must use the provider-supported state format and fit this platform’s 32 KiB request-body limit.

Question types

Compose decisions from small questions

Each question should ask one specific, well-scoped thing. Multiple questions are evaluated in parallel against the same state, so you do not need to chain calls just to split a decision. The available question types depend on the selected model.

TypeUse it forReturns
Choice
Classify or route from optionschoice · probabilities · confidence
Score
Rate state on an ordered rubricscore · legend · probabilities · confidence
Noul
Judge whether a statement is truenoul (yes probability)

Shared fields and structure

A Question uses choice, score, or noul; all types include type and instructions, then add criteria according to the type. The available types depend on the selected model. Instructions can be a string, object, or array; when a question needs extra context, put the question and data in a structured object and refer to the data by field name.

type

Required: noul, choice, or score.

instructions

Required: a string, object, or array describing the decision.

criteria

Type-specific: optional object for Noul, required map for Choice, required array for Score.

{
  "type": "noul",
  "instructions": "Does this message convey urgency?",
  "criteria": {
    "true": "Explicitly needs immediate attention",
    "false": "No urgency expressed"
  }
}

Structured instructions are useful for longer questions or questions that reference extra data: put the question in one field, put context in the others, and refer to those fields by name.

"instructions": {
  "potential_duplicate": {
    "name": "John Smith",
    "location": "Oakland, California",
    "last_employer": "Google"
  },
  "question": "Is the resume for the same person as `potential_duplicate`?"
}

Choice

Use Choice to select one answer from predefined options. type must be choice, instructions describes the decision, and criteria must map options to descriptions; a Choice can have up to 255 options, with each description as a string, object, array, or null.

{
  "state": "Help! My payouts have been failing for 3 days.",
  "model": "typesafe/jev-1.14",
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {
        "billing": "Payments, invoicing, refunds",
        "technical": "Bugs, outages, integrations",
        "sales": "Pricing, upgrades, new accounts"
      }
    }
  }
}

Score

Use Score for descriptive levels on a spectrum, such as severity or satisfaction. type must be score, instructions describes what to rate, and criteria is an ordered low-to-high array whose items can be strings, objects, or arrays; it needs at least 2 and at most 10 levels. The returned score is probability-weighted and can fall between levels.

{
  "state": "Help! My payouts have been failing for 3 days.",
  "model": "typesafe/jev-1.14",
  "questions": {
    "frustration": {
      "type": "score",
      "instructions": "How frustrated is the customer?",
      "criteria": ["Calm", "Frustrated", "Very angry"]
    }
  }
}

Noul

Use Noul for a yes / no judgment. type must be noul and instructions is the question to evaluate; criteria is optional and uses true and false to describe yes and no, with each value allowed to be a string, object, or array. Noul is the probability that the answer is yes, not a second confidence field.

{
  "state": "Help! My payouts have been failing for 3 days.",
  "model": "typesafe/jev-1.14",
  "questions": {
    "is_urgent": {
      "type": "noul",
      "instructions": "Does this convey urgency?",
      "criteria": {
        "true": "Explicitly time-sensitive",
        "false": "No urgency expressed"
      }
    }
  }
}

Output

What the response gives your code

result.answers uses the same question IDs you sent. Typed output guarantees the field shape, but your application should still set thresholds by risk and keep a human-review path where appropriate.

  • answers: Choice returns the selected choice, probabilities, and confidence; Score returns score, legend, per-level probabilities, and confidence; Noul returns noul.
  • usage: Includes input_tokens and output_tokens, and may include cost in USD.
  • elapsedMs: Time from sending the request to receiving the result, including validation—not pure model inference time.

Probability and confidence are signals for automation, not a guarantee of business accuracy. Use higher thresholds or human review for high-risk actions.

Response fields

modelThe model that performed the evaluation; this project returns answers and usage inside result.
answersOne Answer per question, keyed by the same question IDs from the request.
usageContains input_tokens and output_tokens.
elapsedAdditional request time returned by this project, in milliseconds.

Response example

{
  "model": "typesafe/jev-1.14",
  "answers": {
    "is_urgent": {
      "type": "noul",
      "noul": 0.95
    }
  },
  "usage": { "input_tokens": 296, "output_tokens": 20 }
}

Answer types

Every answer carries a type matching its question. Choice and Score answers also carry confidence from 0 to 1, derived from the answer probability distribution.

Choice

Returns the highest-probability choice, probabilities for every option, and confidence derived from the probability distribution.

type

Required; value is choice.

choice

Required string; the highest-probability option.

probabilities

Required map<string, number>; probabilities for all options sum to 1.

confidence

Required number; certainty derived from the probability distribution.

{
  "model": "typesafe/jev-1.14",
  "answers": {
    "department": {
      "type": "choice",
      "choice": "billing",
      "probabilities": { "billing": 0.88, "technical": 0.12, "sales": 0.0 },
      "confidence": 0.81
    }
  },
  "usage": { "input_tokens": 318, "output_tokens": 34 }
}

Score

Returns a probability-weighted score, a legend for each level, per-level probabilities, and confidence. The score can land between levels.

type

Required; value is score.

score

Required number; the probability-weighted score across levels.

legend

Required map<string, string>; maps each level number back to its description.

probabilities

Required map<string, number>; each level and its probability sum to 1.

confidence

Required number; certainty derived from the probability distribution.

{
  "model": "typesafe/jev-1.14",
  "answers": {
    "frustration": {
      "type": "score",
      "score": 1.05,
      "legend": { "0": "Calm", "1": "Frustrated", "2": "Very angry" },
      "probabilities": { "0": 0.0, "1": 0.95, "2": 0.05 },
      "confidence": 0.92
    }
  },
  "usage": { "input_tokens": 304, "output_tokens": 18 }
}

Noul

Returns noul on a 0 to 1 scale, representing the probability that the answer is yes.

type

Required; value is noul.

noul

Required number; 0 means no and 1 means yes.

{
  "model": "typesafe/jev-1.14",
  "answers": {
    "is_urgent": {
      "type": "noul",
      "noul": 0.95
    }
  },
  "usage": { "input_tokens": 307, "output_tokens": 20 }
}

Usage fields

input_tokens

integer · Number of input tokens used by the request.

output_tokens

integer · Number of output tokens generated by the request.

API reference

Evaluate state and return structured answers

Full HTTP API reference: choose an available decision model, evaluate a state against typed questions, and receive one structured answer for each question.

Evaluation endpoint

POST https://decisionapi.net/v1/systemone

Send an Authorization Bearer API key and an application/json content type with every request.

Authorization: Bearer <API_KEY>
Content-Type: application/json

Request body

Every request needs three top-level fields: the selected model, the state to evaluate, and a questions map. You choose each question key; those keys are reused in the response.

statestring | object | array · required: the text or structured data to evaluate.
modelstring · required: the model ID that handles the request. Use an ID from the currently available model list; supported question types may vary by model.
questionsmap<string, Question> · required: the questions to evaluate in parallel.

Use a model ID from the list above or GET /api/playground/models. Jev 1.14 and the other listed models use the same POST /v1/systemone endpoint, state, questions, and response envelope. Keep the leading ~ in ~typesafe/jev-latest and ~perplexity/pplx-decider-latest. The platform forwards the old ~perplexity/pplx-decider-27b-latest alias to ~perplexity/pplx-decider-latest. The examples use typesafe/jev-1.14; use typesafe/jev-1.13 to keep the previous release.

This platform accepts 1–8 questions per request and a JSON request body of at most 32 KiB, including state and questions. These limits also apply to Luna, regardless of the upstream model’s larger context or question limits.

You choose each key in questions; the matching Answer is returned under the same ID. The key is not sent to the underlying model and is not used in inference.

Request example

curl -X POST https://decisionapi.net/v1/systemone \
  -H "Authorization: Bearer <API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-1.14",
    "state": "Help! My payouts have been failing for 3 days.",
    "questions": {
      "is_urgent": {
        "type": "noul",
        "instructions": "Does this convey urgency?"
      }
    }
  }'

Request body example

{
  "state": "Help! My payouts have been failing for 3 days.",
  "model": "typesafe/jev-1.14",
  "questions": {
    "is_urgent": {
      "type": "noul",
      "instructions": "Does this convey urgency?"
    }
  }
}

Keep your API key in a server-side environment variable. Never put it in browser code or commit it to your repository. This page includes the request fields, question types, response shape, error codes, and retry behavior.

Agent usage

Use Jev inside a coding agent

The Jev Agent Skill teaches Codex, Claude Code, Cursor, and other compatible agents how to call this API for bounded decisions while keeping execution and permissions in your application.

Install and configure

Install the Skill, create a DecisionsApi API key, and choose the language for onboarding and examples.

Configure once

Use environment variables so the API key stays out of source code, logs, and Agent transcripts.

Ask a bounded question

Tell the Agent what decision is needed; it should choose Choice, Score, or Noul and send the smallest useful state.

Install and configure

npx skills add jev-ai/jev-agent-skill

export JEV_API_KEY="sk_your_key_here"
export JEV_LANGUAGE="en-US"

Create a key at https://decisionapi.net/settings/apikeys. English (en-US) is the default; set JEV_LANGUAGE=zh-CN for Simplified Chinese guidance. Never paste a real key into source code or a public prompt.

Five useful starting points

Copy one of these prompts after installation. They show how an Agent can use Jev for decisions without giving Jev permission to execute the final action.

1

Route a support ticket

Use Choice to select one approved team, then let application code route the ticket and send uncertain cases to review.

Use the Jev Agent skill. Classify this support ticket into exactly one team: billing, technical, account, or sales. Return the selected team, probabilities, and confidence. Do not contact the customer or modify any ticket yet.

Ticket: I was charged twice for my annual plan and need a refund.
2

Guard a tool call

Use Noul to judge whether a proposed action needs approval, while deterministic permissions and policy remain authoritative.

Use the Jev Agent skill before running this proposed tool call. Judge whether it is safe without human approval. Consider side effects, reversibility, scope, and policy. If risky or uncertain, do not execute it.

Tool: delete_customer_records
Arguments: {where: last_login < 2023-01-01}
Policy: destructive database operations require a backup and human approval.
3

Route to an approved model

Use Choice for the allowlisted candidates and a separate Noul question when no candidate is suitable.

Use the Jev Agent skill to choose one approved model for this task. Optimize for quality first, then context capacity and cost. Return the selected model, probabilities, and whether to escalate. Do not call any model yet.

Task: review a 100k-token customer dispute.
Candidates: fast-model (32k, low cost), reasoning-model (200k, high cost), fallback-model (128k, medium cost).
4

Verify research evidence

Use Noul to assess whether evidence is sufficient before an Agent publishes or cites a claim.

Use the Jev Agent skill to check whether the evidence is sufficient to publish this claim. Consider source quality, freshness, direct support, and contradictions. Return a yes probability and the missing verification work. Do not publish yet.

Claim: Our API reduced median processing time by 40%.
Evidence: an internal benchmark from last month with 120 cases; no production traffic data; an older report showing a 12% improvement.
5

Review task completion

Use Choice or Score to decide whether the work is complete, needs verification, or is incomplete before reporting success.

Use the Jev Agent skill to review whether this task is complete. Return one of complete, verify_more, or incomplete. Consider the objective, files changed, tests run, known gaps, and target-environment verification.

Objective: add API-key authentication to the production endpoint.
Completed: added the Authorization check and API-key lookup.
Verification: unit tests pass; production request and rate-limit behavior were not tested.
The Skill teaches an Agent when and how to ask Jev for a judgment. It does not create a tool, grant authority, intercept shell calls, or replace your permissions, deterministic rules, or human approval boundaries.

Error handling

Errors and retries

The endpoint uses standard HTTP status codes with a JSON response describing the error.

StatusMeaning
401Unauthorized: the API key is missing or invalid. Check the Authorization header.
422Unprocessable Entity: the request body failed validation, such as a missing field or malformed question; the response identifies the offending field.
429Too Many Requests: you exceeded the rate limit. Wait and retry.
529Overloaded: the service is temporarily overloaded. Wait and retry.

When you receive a 429 or 529, retry with exponential backoff instead of sending the same request immediately. SDKs with their default retry policy can handle these retries automatically.

What to do next

Start with one low-risk, well-scoped decision. Then connect it to routing, queues, guardrails, or an agent workflow as you learn where the signal is useful.