Switch to Curva in 5 minutes

This guide is for teams that already get structured answers from a model: another typed-decision or classification API, or a prompt that asks an LLM to “reply with one of: billing, technical, sales” and parses the text. Curva does the same job and adds the parts that make the answers safe to automate.

Map your concepts

What you have today	In Curva
Pick one label from a list (single choice, classification)	Choice: option key → description; answer is choice + probabilities + confidence
A rating or ordered scale (1–5, low/medium/high)	Score: levels, lowest first; answer is the expected level + a probability per level
Yes/no, true/false, a flag	Noul: answer is noul = P(yes)
Tags, any subset of labels	Multi: each option gets its own probability; selected are those over the threshold
“Parse the model’s text and hope it’s valid”	Never needed: every answer is typed and validated
A confidence the model wrote in its reply	Probabilities read from the model’s token probabilities, then calibrated on your feedback

See Questions and answers for every field.

If your client sends criteria

Many typed-decision clients send questions in a criteria shape. Curva accepts it as is at POST /v1/systemone, an alias of /v1/decide:

{
  "state": {"ticket": "I was charged twice for order A-104"},
  "model": "curva-latest",
  "questions": {
    "team": {"type": "choice", "instructions": "Which team?",
             "criteria": {"billing": "payments, refunds", "technical": "bugs", "sales": null}},
    "anger": {"type": "score", "instructions": "How frustrated?",
              "criteria": ["calm", "annoyed", "angry"]},
    "refund": {"type": "noul", "instructions": "Asks for a refund",
               "criteria": {"true": "explicitly asks for money back", "false": "anything else"}}
  }
}

• Choice criteria: key → description (text, null or JSON). There is no escape option unless the question sets "escape": true, so answers match what your client expects.
• Score criteria: level descriptions, lowest first. Noul criteria is optional.
• model: curva-latest or a pinned curva-x.y.z config; a name without a provider prefix means the server’s default model; anything else is a model id or a plan.
• The response is the /v1/decide response plus usage and a type on every answer.
• GET /v1/models lists the config names model accepts.

The Python SDK has the same shape: client.system_one(...) with Choice(..., criteria={...}), Score(..., criteria=[...]) and Noul(..., criteria={...}) (reference).

Change the base URL

Run Curva (Getting started or Self-hosting), create a key, and point your client at it:

curva serve                                    # http://127.0.0.1:7777
curva keys create --name my-app                # prints curva_… once

# before: https://<your current provider>/…
export BASE_URL=http://127.0.0.1:7777
curl -s $BASE_URL/v1/systemone -H "authorization: Bearer $CURVA_API_KEY" \
  -H 'content-type: application/json' -d @request.json

That’s the whole switch for a criteria client. For a prompt you parse yourself, replace the prompt with a /v1/decide call: the question list is your label list.

What you gain

• Calibration on your data. Send the true answer as feedback; from 30 labels per question, Curva calibrates when that helps, so a 0.9 means right about 90% of the time on your traffic.
• Debiasing. Options are asked in both orders and averaged, so an answer doesn’t depend on which label came first (details).
• Abstain. Set min_confidence and unsure answers come back abstain: true, for a person or a stronger model.
• Audit. Every decision is logged with model, config, answers, cost and a salted hash of the state (never the state itself).
• Self-host. One binary and a SQLite file; your data stays on your machines.
• Any model. Any OpenRouter or OpenAI-compatible model, free ones included, blended with council, cascade or race, or your own fine-tuned model.

Before you switch all traffic

Replay a day of logged traffic with shadow mode to see how often Curva agrees with your current system, and at what confidence it can take over.

Checklist

• Each question mapped to Choice, Score, Noul or Multi
• Curva running, API key created
• Base URL changed (/v1/systemone for criteria clients, /v1/decide otherwise)
• model set (a pinned config for reproducible answers, or a model id)
• Shadow run on logged traffic, agreement checked
• Decision ids stored, feedback flowing
• min_confidence set where a wrong answer is expensive