Concept
What is a System One model?
A model that answers fixed, typed questions in one pass, with a calibrated confidence, instead of generating text token by token. The name nods at Kahneman’s fast, intuitive System 1, as against the slow deliberate System 2 that an autoregressive LLM more closely resembles.
The problem it addresses
An enormous share of production LLM calls are not writing tasks at all. They are decisions wearing a writing task’s clothes: is this spam, which of these six categories applies, does this answer agree with that one, is this severe enough to page someone. For every one of those, the usual approach asks a text generator to produce a word, then parses the word back out, and pays for the whole autoregressive apparatus to do it.
That is slow, it costs more than it should, and it has an irritating failure mode: the model returns something outside your set of options. So you add a retry, or a parser, or a schema-constrained decoding layer, and now a yes-or-no question has infrastructure.
What changes
Typed by construction
You declare the questions and their answer types. An out-of-range answer is not possible, so there is no parsing step and no retry loop.
One parallel pass
No token-by-token generation, so latency is roughly flat in the number of questions rather than growing with the length of the answer.
Calibrated confidence
The score behaves like a probability. That is what lets you automate on it, and it is the part that is hard to build and easy to claim.
The shape of a call
A state, plus the questions you want answered about it. Three families cover most of what a triage layer needs: pick one of these, score this on a scale, and answer this yes or no with a calibrated confidence.
the idea, not a vendor API
{
"state": "Ticket: I was charged twice and cannot find a refund option.",
"questions": [
{ "type": "choice", "options": ["billing", "bug", "feature", "access"] },
{ "type": "score", "range": [0, 1], "of": "how complex is this" },
{ "type": "bool", "of": "does this need a human" }
]
}Where an LLM is still the right tool
Anything open-ended. If the answer is prose, code, a summary or a plan, a System One model has nothing to offer, because the whole trade is giving up generation in exchange for speed, cost and type safety. It is not a cheaper LLM. It is a different instrument, useful for a category of work that has been served badly by text generators because they were the only thing to hand.
Why this matters for verification
A verification layer answers structured questions all day. Which tier should handle this. Do these two answers conflict. Is this claim supported by this passage. Should this escalate. Every one of those is a typed decision, which makes this category of model close to purpose-built for the job, and makes always-on verification affordable at production volume for the first time.
The category is new enough to have exactly one product in it, which is a genuine risk rather than a footnote. How we handle that is a pluggable backend interface and a fallback that needs no vendor at all.
When one routes for you
Put a model like this in front of a pool of LLMs and you have a System One router: the choice of which model answers is itself a typed decision, made in one pass, for a fraction of a cent. That page compares it against the three other ways a router can decide.