Introduction
Instinct One reads text and images and answers your questions with probabilities, in one forward pass.
Instinct One is a "System One" model: it does not write text. You give it a state (text, structured data, images) and a set of questions, and it returns a probability for every possible answer. All questions of a request are answered in a single forward pass, so a decision takes about a hundred milliseconds of model time, not seconds.
{
"state": {
"channel": "email",
"message": "The export button on the reports page threw an error twice today. I just want it fixed."
},
"questions": {
"team": {
"type": "choice",
"instructions": "Which team should handle this ticket?",
"criteria": {
"billing": "Payments, invoices, refunds, subscriptions.",
"technical": "Bugs, errors, outages, integration problems.",
"sales": "Pricing questions, upgrades, new purchases."
}
},
"wants_refund": { "type": "noul", "instructions": "Is the customer asking for a refund?" }
}
}{
"model": "instinct-one-latest",
"answers": {
"team": {
"type": "choice",
"choice": "technical",
"confidence": 0.94,
"probabilities": { "billing": 0.03, "technical": 0.96, "sales": 0.01 }
},
"wants_refund": { "type": "noul", "noul": 0.02 }
},
"usage": { "input_tokens": 118, "output_tokens": 61 },
"latency_ms": 97
}What it is for
Use it wherever a program has to make a small decision many times, fast, and needs a number it can put a threshold on instead of prose it has to parse.
| Use | What you ask |
|---|---|
| Routing and triage | Which team, queue or handler does this message belong to? |
| Checks and guards | Does this text break the rule? Is the form on this screenshot showing an error? |
| Rating | How severe is this alert, on an ordered scale? |
| Reading screens | Which of these elements does "press Save" mean? Is the dialog still open? |
| Reading charts, tables, documents | Which series is larger? Is the invoice total above the limit? |
| Agent loops | A large model plans; Instinct One answers the small "which one / is it done" questions at every step |
What makes it different
- Probabilities, not text. Every answer is a distribution over the options you named. There is nothing to parse and nothing that can be worded differently next time.
- Images in the state. Screenshots, photos, document pages and charts sit in the state next to text, up to eight in one request.
- Many questions, one pass. Ten questions about one state cost one request and roughly the time of one.
- A familiar API. The request and the response follow TypeSafe's System One (Jev) format. A Jev or Kev client works unchanged; images are an addition on top.