Open-source decision model  ·  by Singularity Compute

Super Decider

The fast decision layer for agents and apps. Ask about any situation and get a typed answer with a calibrated probability, so your code knows when to act and when to hand off.

Start on ASI:Cloud See the benchmarks
Open weights, Apache 2.0 GitHub Hugging Face
POSTllm.a.singularitycompute.com/v1/systemone
{
  "model": "super-decider",
  "state": {
    "ticket": "I was charged twice for my March invoice."
  },
  "questions": {
    "route": {
      "type": "choice",
      "instructions": "Which team should handle this ticket?",
      "criteria": {
        "billing": null, "technical": null,
        "account": null, "sales": null
      }
    }
  }
}
200 OKroute

Choice

billing

  • billing0.94
  • account0.04
  • technical0.01
  • sales0.01

Illustrative responses. The probabilities sum to 1 across the options you send.

Benchmarks

Measured on public decision benchmarks

Super Decider next to TypeSafe's Jev, Cloudflare's Clef and OpenAI Decisions on six public benchmarks, from agent tool calls to phishing. We ran Super Decider with each benchmark's public evaluation code; the other scores come from the public leaderboards, as of October 2026. Higher is better.

Decisions your software can act on

Super Decider is a System One model. It answers typed questions about a state, with a probability for every answer.

Available on ASI:Cloud

Pick Super Decider in the ASI:Cloud model catalogue, copy your key and start sending decisions.

ticket“The parcel arrived damaged. Can I get my money back?”

  • refund0.81
  • replacement0.15
  • other0.04

Escalate to the billing team Escalate to general support

Calibrated probabilities

Every answer carries a probability for each option. Set the threshold in code: when the model is sure, the ticket goes straight to billing; when it is not, to general support. Drag it to try.

Typed decisions

Pick one option, answer yes or no, score on your own scale or quote the input. Your code acts on the answer without parsing prose.

Over 150 options

Define routes, labels, actions or moves per request. We tested 151 options in one question and up to 64 questions in one call.

Policy in your code

Thresholds, fallbacks and escalation rules stay in your code, where you can test and change them.

Two APIs, one key

POST /v1/systemone in the System One request shape, or /v1/chat/completions with the OpenAI SDK you already use.

Reads images

Up to 8 images per request next to the text: invoices, receipts, screenshots and scanned forms.

Fast, with thinking on demand

About 0.3 s per request over the public internet. Turn on thinking for multi-step questions. The model writes a short thought first and scores 2 points higher on DecisionBench.

How it works

From state to decision in one request

  1. Describe the state01

    Put what the decision depends on into the request as JSON: a ticket, a record, a position on a board.

  2. List the options02

    Name the answers you accept and write the instructions once. Give any option a short description when it needs one.

  3. Act on the answer03

    Read the choice and its probability, then act, ask a person or try a larger model, by the threshold you set.

Spin it up on ASI:Cloud

Use cases

Where a typed decision beats a paragraph

Support and operations

Route tickets to the right team, flag urgent or sensitive requests, pick the next step in a runbook, and escalate when the probability is low.

Agents and automation

Call a tool, ask a follow-up or answer directly; approve or hold an action. 85.6 on NVIDIA's When2Call, ahead of Jev and Clef.

Documents and compliance

Check invoices, contracts and forms, as text or images, against your rules and flag the exceptions. 94.5% accurate on documents up to 16k tokens.

Try it on ASI:Cloud

F.A.Q

Frequently Asked Questions

What is a decision model?

You give a decision model the state of a situation, say a support message, and ask typed questions: is it urgent, which team should take it, how angry is the customer. It returns one answer per question with a probability for every option, and your code routes the ticket, escalates it or hands it to a person.

How is it different from an LLM?

An LLM writes open-ended text and tool calls, so its output has to be parsed and can drift from run to run. A decision model only picks from the answers you allow, with calibrated probabilities, in about 0.3 seconds. Put Super Decider in the hot path where an agent has to decide, and keep an LLM for the steps that need writing or long reasoning.

How does Super Decider compare with Jev and Clef?

It is ahead of TypeSafe's Jev on the benchmarks we compared, including NVIDIA's When2Call and PhishNChips. Against Cloudflare's Clef it is ahead on When2Call and PhishNChips and 1.6 points behind on BFCL tool selection. It also reads up to 8 images per request, takes up to 131,072 tokens of state, holds 94.5% accuracy on documents up to 16k tokens, and can think briefly before multi-step questions when you ask it to.

What happens when it isn't sure?

Every answer carries its probabilities, so you set the bar in code: act on its own above the threshold, ask a person or a larger model below it. The rule lives in your code, where you can test it and change it without retraining anything.

How do I call Super Decider?

Through the ASI:Cloud API. Create a key at asicloud.cudos.org and send your state and questions with "model": "super-decider", in the same request shape as the Jev API or through the OpenAI SDK. The ASI:Cloud documentation has the endpoints, question types and code examples.

How do I get access?

Super Decider runs on ASI:Cloud. Sign in at asicloud.cudos.org, pick Super Decider in the model catalogue, create a key and send your first decision. The model is also open source under Apache 2.0: the code is on GitHub and the weights on Hugging Face.

Singularity Compute

Put Super Decider
to work in your stack

Start on ASI:Cloud