Decision API · live now
Give the API the facts, the question, and the options. It returns one choice with probabilities — as JSON, every time. About a second. Half a cent.
How it works
Chat models write sentences your code then has to interpret. This API never writes a sentence. It scores the options you sent and returns them as numbers that sum to 1.
A ticket, a filing, a transcript, another model’s answer — whatever your software already knows.
“Approve this refund?” “Which class is this?” “Does the evidence support expanding?”
Two to sixteen options, in your own words. The API picks one and gives odds on all of them.
POST https://www.bluusun.com/api/v1/decide
Authorization: Bearer bsu_YOUR_KEY
{
"evidence": "Q3 revenue beat by 12%, churn at an all-time low.",
"criterion": "Should we expand the sales team?",
"options": ["Expand now", "Hold steady", "Cut back"]
}
{
"decision": 0,
"options": [
{"option": "Expand now", "probability": 0.72},
{"option": "Hold steady", "probability": 0.21},
{"option": "Cut back", "probability": 0.07}
]
}
Illustrative probabilities. decision is the index of the winning option. There is also an OpenAI-compatible POST /api/v1/chat/completions endpoint; the assistant content is the same JSON, never model prose.
Why this shape
The model never gets to write free text. Output is built from the probability of each option you sent, then renormalized so the odds sum to 1. Prompt injection in the evidence can only move weight between those options. It cannot invent a new format, leak the prompt, or return a paragraph for your parser to trip over.
That is the whole product: a decision, with odds, over your choices.
Use the probability as a rule. Auto-approve above 0.9, send the rest to a person. Most chat models can be right often and still never tell you when they are guessing. This one reports a number on every call.
Flat $0.005 from prepaid credit. Oversize requests are rejected and not billed. Typical latency is about a second.
Tickets, applications, content — anywhere hand-written rules start to crack.
Refunds, claims, reviews. Clear cases go through; the rest escalate.
Leads, candidates, risk. One consistent judge, thousands of times.
Verify another model’s answer against a criterion before a user sees it.
Measured on a public benchmark
openjev is a public set of 144 evidence-and-options decisions. We ran every item through this API and scored the run with the benchmark’s published evaluator — not a private test.
The benchmark asks for the same thing the API serves: pick one option from evidence, with probabilities. There is no extra “parse the chatbot” step in between, so the number is measuring the product. Fixtures: openjev on GitHub. In-house retake, September 2026.
Pricing
One price, paid from prepaid credit. No subscription. Credits never expire.
Text decisions only. Evidence cap is 3,000 characters; oversize requests return an error and are not billed. Keys are created, rotated, and revoked in your dashboard.
Top up with Stripe. Sign up is free and includes $0.01 — two decisions to try the API.
2,000 decisions
never expires
5,000 decisions
never expires
10,000 decisions
never expires
20,000 decisions
never expires
Bluusun Group
We build systems we actually operate. The Decision API is the public product; the other two divisions stay in-house.
Calibrated decision intelligence. The API on this page turns evidence into probabilities — no free text, no surprise format.
Systematic strategies and risk analytics, researched and operated in-house — not a public trading product.
Autonomous systems for construction and industry, tested on the work they are meant to do.