Skip output decoding.
The model reads the input in a forward pass. It does not spend additional decoding steps writing an answer token by token.
decision-machine-1
Small model. Big decisions.
A label. A field. A yes or no. Meet decision-machine-1, the model for the little decisions that keep your app moving.
The model
Language in.
Application logic out.
In software, an AI response often becomes a branch, a field, or a score. The application needs the value to take its next step.
decision-machine-1 scores labels and extracts values directly from text. For these bounded tasks, it skips the token-by-token generation an LLM uses—even with structured outputs.
Read the model rationaleLanguage in
“Losing sales,
fix this today.”
Is the customer expressing urgency?
Decision outtrue
Probability of urgency 1.00
/yes-no response. Your code can use the boolean to raise a ticket’s priority.Define the task.
Supply labels to choose between, questions to answer from the source, or fields to extract.
See the alternatives.
A classification includes the full score distribution and confidence. Your code can see when two labels compete, instead of treating every winner alike.
Calls
Choose the shape
of your decision.
Choose the labels, questions, or fields your application needs.
POST /yes-noRecorded API example
Flag messages that need a faster response.
Request{
"text": "Losing sales, fix this today.",
"statement": "The customer expresses urgency."
}{ "answer": true, "probability": 1 }POST /classifyRecorded API example
Choose a destination from your support queues.
Request{
"text": "My card was charged twice this month.",
"labels": ["billing", "shipping", "technical", "other"]
}{
"label": "billing",
"probability": 0.74,
"confidence": 0.555,
"scores": {
"billing": 0.74,
"shipping": 0.001,
"technical": 0.008,
"other": 0.25
}
}POST /rateRecorded API example
Turn customer frustration into a sortable score.
Request{
"text": "Third time I write. Nobody answers.",
"scale": ["calm", "annoyed", "angry", "furious"]
}{
"score": 1.998,
"level": 1,
"confidence": 0.208,
"scores": [0, 0.334, 0.334, 0.332]
}POST /answerRecorded API example
Locate the shipment destination in a status update.
Request{
"text": "The shipment reached the Halifax warehouse on September 6, two days after the promised date.",
"questions": ["Where did the shipment arrive?"]
}{
"results": [
{
"answer": "Halifax warehouse",
"probability": 0.992,
"start": 25,
"end": 42
}
]
}POST /extractRecorded API example
Map invoice text to the fields in your records.
Request{
"text": "Invoice #A-1042 from Nordik Supply, due 2026-10-01, total 4,250.00 CAD, net 30.",
"schema": {
"type": "object",
"properties": {
"invoice_number": { "type": "string" },
"vendor": { "type": "string" },
"total": { "type": "number" },
"currency": { "type": "string" }
}
}
}{
"data": {
"invoice_number": "A-1042",
"vendor": "Nordik Supply",
"total": 4250,
"currency": "CAD"
}
}POST /entitiesRecorded API example
Identify people, organizations, and references in a claim note.
Request{
"text": "Marie Tremblay from Desjardins called about claim 44812 on March 3.",
"types": ["person", "organization", "claim_id", "date"]
}{
"entities": [
{ "type": "person", "text": "Marie Tremblay", "probability": 0.999 },
{
"type": "organization",
"text": "Desjardins",
"probability": 0.997
},
{ "type": "claim_id", "text": "claim 44812", "probability": 0.714 },
{ "type": "date", "text": "March 3", "probability": 0.997 }
]
}POST /verifyRecorded API example
Check a proposed deductible against the policy text.
Request{
"text": "Policy holder: Jean Roy. Deductible $500.",
"field": { "name": "deductible" },
"value": "$1000"
}{ "matches": false, "probability": 0, "found": ["$500"] }Need a nested taxonomy? Classify-tree walks your label hierarchy in one API request. Try the interactive demos.
Your application stays in controlSet the bar.
Set the bar.
Keep a review path.
Use a decision call first, then accept results that meet your thresholds, and send ambiguous cases to an LLM or a person.
That can reduce large-model calls. It also adds a first step, so measure the cost and latency of the whole path on your traffic.
Build an LLM cascadeRecorded classificationbilling
Label probability 0.74Confidence 0.555
Your application’s gate
probability >= 0.90
confidence >= 0.70Meets both thresholdsUse the result
Below either · this exampleLLM or human review
What to know before you build.
Does structured output mean the answer is correct?
No. A well-formed response can still contain a wrong value. Evaluate on your own data, validate business rules, and review consequential or ambiguous results. Verification can check a value against the source; it is not a guarantee of correctness.
When should I use an LLM instead?
Use an LLM for writing, conversation, multi-step reasoning, or judgments that need knowledge beyond the supplied text. Use milliseconds when the task has defined labels, questions, or fields.
What do probability and confidence tell me?
For classification, probability is the winning label’s share of the scores. Confidence describes how concentrated the whole distribution is. Neither is a guarantee that the answer is right. Other capabilities return different scoring fields; tune thresholds for the task and the cost of a mistake.
Can I keep my OpenAI SDK?
Yes, for JSON-schema extraction and tool calls. Set the base URL to https://api.milliseconds.ai/v1 and select decision-machine-1. This is not a drop-in chat model: plain chat is unsupported, and the compatibility layer reads user turns only.