How System One Models Work
Most of what a business asks of AI is not writing. It is deciding: which team, how urgent, yes or no. A language model answers those by writing a paragraph and hoping the answer is in it. A System One model reads the input once and returns typed answers, in milliseconds. Pick an example, press Decide to watch the single pass in slow motion, then race it against a chatbot answering the same three questions.
What you are looking at
Fast judgement, not slow reasoning
The name comes from Daniel Kahneman: System 2 is slow, deliberate thinking, System 1 the instant judgement that tells you a face is angry. A language model is all System 2, reasoning and writing one word at a time. A System One model is built for the other half: it reads the input once and makes the call.
Nothing is generated, so nothing waits on itself
A chatbot runs the whole network again for every word it writes, so a short answer is a hundred passes in a row. A System One model reads the state in a single parallel pass and every declared question is answered from it at the same moment. Published figures put a decision at 70 to 500 milliseconds, where a frontier model takes seconds.
It cannot answer outside the question
The output is not text with an answer somewhere inside it. A Choice can only be one of the options you listed, a Score a point on your scale, a Noul a number between 0 and 1. There is nothing to parse and no way to invent a fifth option. The right shape can still hold the wrong answer, though: typed is not the same as correct.
Three question types
They are the whole interface. You do not write a prompt; you hand over the state, plain text or JSON, and declare what you want to know about it, each question with one of these types.
Choice
type: "choice"
options: ["billing", "technical", ...]
→ { choice, confidence, probabilities }Pick one from a fixed list, up to 255 options. Every option comes back with a probability, softmaxed so they sum to one, and the model answers with the most probable. Confidence is separate from the winning probability: it reads how the whole distribution is spread, so a 40% winner over three close rivals shows as unsure.
Score
type: "score"
levels: ["whenever", ..., "drop everything"]
→ { score, confidence }Place the state on an ordered scale of 2 to 10 levels, each described in words. The score is the probability-weighted position, so it can land between two levels: a 3.6 means somewhere past "today" and short of "within the hour". Confidence falls as the probability smears along the scale.
Noul
type: "noul"
"Is the customer threatening to leave?"
→ { noul: 0.93 }A yes or no, returned as one probability that the statement is true. There is no confidence field, because the number is the confidence: near 1 is a confident yes, near 0 a confident no, and near 0.5 is the model saying it cannot tell. A low number is an answer, not a weak one.
Where this gets used
Anywhere the same small question is asked thousands of times a day, and the answer has to be fast, cheap and in a shape a program can act on.
Routing and triage
Which queue, how urgent, does it need a person. Every ticket, email and claim gets the same three questions in a tenth of a second, at a price that makes it sensible to ask about all of them rather than a sample.
Checks inside an agent
An agent built on a language model makes dozens of small calls a task: is this tool result an error, is the task done, is this reply on brand. Each one is a System 1 question, and answering it with a typed value in milliseconds rather than a paragraph in seconds is what keeps the loop fast.
Guardrails and moderation
Is this message abusive, does this output leak personal data, is this payment the cardholder's usual pattern. Questions that have to be asked of every single item, and where a probability you can set a threshold on is worth more than an explanation.
Where it is the wrong tool
Anything that needs reasoning written out, a plan, or knowledge the input does not contain. On published benchmarks, System One models sit a few points behind frontier language models; they win on speed and cost, not on intelligence. Independent tests have also found their probabilities overconfident on hard knowledge questions, so thresholds want checking against your own labelled data.
A demonstration of the idea, not a production model: the scoring on this page is a small hand-written stand-in, and the timings in the race are illustrative. Pairing a fast decision model with a language model, each on the questions it is good at, is the kind of system we design and build for clients.
Talk to us about your own problem →