Interfaces
Decisions
Call A8 through OpenAI’s Decisions API, the fields it reads, what the number on each answer promises, and where an abstention lands.
Preview. Applies to a8-1. Last reviewed 2026-10-31.
This page as Markdown: https://u22a8.ai/docs/a8-decisions.md. The index of every docs page: https://u22a8.ai/llms.txt.
A8 answers OpenAI’s Decisions API at POST https://api.u22a8.ai/v1/decisions. Any client that takes a custom base URL calls it this way: the OpenAI SDKs (decisions.create), LangSmith’s decision-model evaluators, promptfoo’s openai:decisions provider, MLflow judges with model_api="decisions", BAML and the gateways that serve /v1/decisions, LiteLLM among them.
§1Request
One request asks several questions of one input. A8 judges each question as a request of its own, against the same version of your organization’s model:
| Field | Type | How A8 reads it | Default |
|---|---|---|---|
model | string | a8 for your organization’s latest version, a pinned identifier, or a8-1-1@reference. Its tag sets the minimum accuracy: a8:95 | Required |
input | string or array | The subject. System and developer items open every question’s criterion | Required |
questions | array | 1 to 400 questions, answered in order | Required |
questions[].type | string | predicate for a yes or no, choice for one of choices, score for one of levels | Required |
questions[].instructions | string | The criterion | Required |
questions[].name | string | Echoed on the answer | null |
questions[].choices | array | 2 to 255 values, each a string or a boolean, with an optional description | None |
questions[].levels | array | 2 to 10 labels, lowest first, each with an optional description | None |
The fields A8 reads. Every other field is accepted and ignored, safety_identifier and the prompt-cache settings among them.
§2What is refused
Each of these returns 400, with an error that names the field and the value that passes:
| Request | Status | Send |
|---|---|---|
min_accuracy in the body | 400 | The minimum accuracy as the model’s tag, as a8:95 |
A model tag outside 0 to 99 | 400 | A whole number from 0 to 99 |
expected in the body | 400 | The answer as a teach, over MCP |
No questions, or more than 400 | 400 | 1 to 400 questions |
Two questions with one name | 400 | A name per question |
A choices or levels count outside its range | 400 | 2 to 255 choices, or 2 to 10 levels |
An image part, input_image | 400 | Text or JSON parts |
An input with no user message | 400 | The subject as a user message |
Refused fields and values. Error bodies are on Errors.
§3Minimum accuracy
The minimum accuracy is the accuracy A8 promises on the verdicts it returns. At 95, at least 95% of answered verdicts are correct. The model’s tag sets it for every question in the request: a8:95 asks for 95, and a pinned identifier carries it too, as a8-1-1:95@20260715t140322z.
- The tag is a whole number from
0to99. A higher level answers less often and is right more often. - By default, the minimum accuracy is 90.
- The response’s
modelnames the level that answered.
A question whose verdict falls short of the minimum accuracy comes back as a refusal, and the other questions still answer. At a8:0, A8 answers wherever a reading exists, and each answer’s number says which levels it clears, so one request lets you choose the threshold afterwards. What the minimum accuracy promises is on Accuracy and abstention.
§4The number on each answer
Every answer in this API carries a number. A8’s number is its promise: the strictest minimum accuracy the verdict clears, over 100. A verdict that clears 97 carries 0.97.
- A threshold is a minimum accuracy. Keeping the answers whose number is at least
0.95keeps exactly the answersa8:95returns, and at least 95% of them are correct. The threshold this API asks you to tune on labelled data is a level A8 has already earned on what you taught. - A predicate carries the number in
probabilityfor a yes, and one minus the number for the opposite answer, so it always sits on the answer’s side of0.5. - A choice carries the number on the chosen value, in its
probabilitiesentry and inconfidence. The rest is spread evenly over the other values, since A8’s promise covers the chosen value alone. The chosen value always holds at least an even share. - A score spreads its levels the same way, and
scoreis the mean level index of that spread, from0for the lowest level.
The promise covers all the answers at or above a threshold, taken together. Measured alone, a band between two thresholds can run below the lower one. Each answer also carries min_accuracy, the level it was answered at.
§5Reasoning
This API states which part is which, so A8 reads each question as written, as the other interfaces read a request with reasoning off:
| Part | Where it goes |
|---|---|
| Criterion | The input’s system and developer items, then the question’s instructions, then a line for each choice or level with a description |
| Subject | An input string, or its one user message. Any other input is its transcript: one role: text line per item, tool calls included |
| Labels | The question’s choices values, or its levels labels |
How a question is read. A JSON part, input_json, is read as its JSON.
A threshold written in instructions counts as part of the criterion. A layout read as written writes 0 reasoning tokens.
§6Response
A8 answers with a standard decision object. The answers follow the questions, in order:
{
"answers": [
{
"type": "predicate",
"name": "refund_requested",
"min_accuracy": 95, // the level this answer is promised at
"probability": 0.97 // the strictest level it clears
},
{
"type": "choice",
"name": "queue",
"min_accuracy": 95,
"choice": "shipping",
"probabilities": [
{"value": "billing", "probability": 0.02},
{"value": "shipping", "probability": 0.96},
{"value": "account", "probability": 0.02}
],
"confidence": 0.96
},
{
"type": "refusal",
"name": "severity",
"abstention": {
"kind": "too_close",
"asked": 95,
"answerable_at": 80,
"earned": 97,
"missing": null
}
}
],
"model": "a8-1-1:95@20260715t140322z", // the version and level that answered
"usage": {
"input_tokens": 1240,
"input_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 0},
"output_tokens": 0,
"total_tokens": 1240
},
"system_fingerprint": "fit-1a2b3c4d5e6f7890.9e2a1b3c"
}
modelnames the exact version that answered, with the minimum accuracy in its tag. Send it back to get the same answers. See Horizons.usage.input_tokenscounts each text the request reads once, so the input every question shares counts once. A verdict is scored, sooutput_tokensis0.min_accuracy,abstentionandsystem_fingerprintare A8’s fields beside the standard ones. Standard clients carry them through.
§7An abstention
An abstention is the question’s refusal, which this API returns for a question the model declines while the other questions still answer. Its abstention says which check the evidence failed, and at what level the same question is answered. The kinds, and what clears each, are on Accuracy and abstention.
Some clients score a refusal as a failure: promptfoo fails the assertion, and Langfuse fails the evaluator run. Ask at a8:0 and threshold on the number to keep every case in those tools, or send the refused cases to a person and teach their answers.
§8Precedents and recall
Precedents and recall travel on Chat Completions, Responses, Anthropic Messages and MCP, whose responses carry a verdict with room beside it. A Decisions answer holds the fields this API defines and A8’s min_accuracy and abstention. Teaching is over MCP, and a criterion taught there answers on this API too.
§9Streaming
A Decisions response arrives whole, in one body, as this API defines it.
§10Example
A ticket router at a minimum accuracy of 95:
curl https://api.u22a8.ai/v1/decisions \
-H "Authorization: Bearer $U22A8_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "a8:95",
"input": "It arrived crushed. I want a refund.",
"questions": [
{
"type": "predicate",
"name": "refund_requested",
"instructions": "Is the customer asking for a refund?"
},
{
"type": "choice",
"name": "queue",
"instructions": "Which queue does the ticket belong in?",
"choices": [
{"value": "billing"},
{"value": "shipping"},
{"value": "account"}
]
}
]
}'
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.u22a8.ai/v1",
api_key=os.environ["U22A8_API_KEY"],
)
decision = client.decisions.create(
model="a8:95", # minimum accuracy 95
input="It arrived crushed. I want a refund.",
questions=[
{
"type": "predicate",
"name": "refund_requested",
"instructions": "Is the customer asking for a refund?",
},
{
"type": "choice",
"name": "queue",
"instructions": "Which queue does the ticket belong in?",
"choices": [
{"value": "billing"},
{"value": "shipping"},
{"value": "account"},
],
},
],
)
for answer in decision.answers:
if answer.type == "refusal":
print(answer.name, "goes to a person")
else:
print(answer.name, answer)
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.u22a8.ai/v1",
apiKey: process.env.U22A8_API_KEY,
});
const decision = await client.decisions.create({
model: "a8:95", // minimum accuracy 95
input: "It arrived crushed. I want a refund.",
questions: [
{
type: "predicate",
name: "refund_requested",
instructions: "Is the customer asking for a refund?",
},
{
type: "choice",
name: "queue",
instructions: "Which queue does the ticket belong in?",
choices: [
{ value: "billing" },
{ value: "shipping" },
{ value: "account" },
],
},
],
});
for (const answer of decision.answers) {
const said = answer.type === "refusal" ? "goes to a person" : answer;
console.log(answer.name, said);
}
{
"answers": [
{
"type": "predicate",
"name": "refund_requested",
"min_accuracy": 95,
"probability": 0.97
},
{
"type": "choice",
"name": "queue",
"min_accuracy": 95,
"choice": "shipping",
"probabilities": [
{"value": "billing", "probability": 0.02},
{"value": "shipping", "probability": 0.96},
{"value": "account", "probability": 0.02}
],
"confidence": 0.96
}
],
"model": "a8-1-1:95@20260715t140322z",
"usage": {
"input_tokens": 46,
"input_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 0},
"output_tokens": 0,
"total_tokens": 46
},
"system_fingerprint": "fit-1a2b3c4d5e6f7890.9e2a1b3c"
}
The samples need Python 3.10 or later with openai 3.26 or later, or openai 7.30 or later for TypeScript. The minimum accuracy travels in model, which every SDK and gateway passes through.