What is a System One model?
A System One model makes a focused judgment about the information you give
it. You provide a state (such as a support ticket) and define the questions
to ask. It returns bounded, typed answers and probability estimates that your application
can use directly, rather than a paragraph of generated text.
For example, given a customer message, you might ask whether it requests a refund, which team should handle it, and how urgent it is. Your application can then route the ticket, prioritize a queue, or ask a person to review the result.
Four ways to automate a decision
| Approach | What you build | Tradeoff |
|---|---|---|
| Code rules | Explicit conditions and exceptions. | Easy to audit, but each new edge case may need another rule. |
| Classical ML | A classifier trained on labeled examples. | Learns patterns, but each task needs labels, training, and upkeep. |
| Generative AI | A prompt asking an LLM to make a judgment in text. | Flexible, but poorly calibrated and the result may need parsing and validation before code can use it |
| System One | Focused questions with defined answer types. | Returns decisions and probabilities directly; you still need to evaluate quality on your own cases. |
The difference is what the model is built to return. An LLM is useful for
drafting, explaining, coding, and planning. A System One model is for bounded
decisions within a larger workflow. mo-fast can optionally reason
before returning its typed answer; see
reasoning modes.
What does it return?
You define the possible answers and criteria for each question:
| Question type | Example | Answer |
|---|---|---|
noul | Does this message request a refund? | Probability of true, from 0 to 1. |
choice | Which team should handle it? | Selected option, probabilities for every option, and confidence. |
score | How urgent is it? | Score against a rubric, probabilities over its levels, and confidence. |
You can ask several questions about the same state in one decision. For a
choice question, an illustrative answer might be:
{
"type": "choice",
"choice": "billing",
"probabilities": { "billing": 0.94, "technical": 0.04, "account": 0.02 },
"confidence": 0.94
}
Probabilities help you decide when to act and when to request review; they do not guarantee a particular answer is correct. Measure agreement and calibration on examples from your use case before setting automated thresholds. See confidence estimation for guidance.
Make it yours
Start with a base model and typed questions. When you have a repeatable task, save the questions and criteria in a custom, versioned System One model. Review examples, add feedback, and align the model to improve it for your domain.
Explore use cases or follow the quickstart to make your first decision.