Calibrated LLM model routing: use the cheap model only when it's trustworthy

A calibrated gate that uses a cheaper model when it's trustworthy enough and escalates to the stronger one when it isn't — verified against real outcomes.

The problem

A cheap model is great until it quietly isn't, and you can't tell which calls those are. Always-cheap drops quality; always-expensive burns budget. You want to use the cheap path exactly where it holds.

The code

Zero-dependency SDK. Mint a key on your dashboard and drop this in.

from warmwinter import WarmWinter

ww = WarmWinter(api_key="ww_YOUR_KEY",
                base_url="https://api.warmwinter.io")

d = ww.decide(domain="compute", decision_type="model_route",
              stated_confidence=0.82, stakes="medium")

answer = cheap_model(prompt) if d.verdict == "act" else big_model(prompt)
ww.outcome(d.decision_id, "success" if ok else "failure")

import { WarmWinter } from "./warmwinter";

const ww = new WarmWinter({ apiKey: "ww_YOUR_KEY", baseUrl: "https://api.warmwinter.io" });

const d = await ww.decide({
  domain: "compute", decisionType: "model_route",
  statedConfidence: 0.82, stakes: "medium",
});
const answer = d.verdict === "act" ? await cheapModel(prompt) : await bigModel(prompt);
await ww.outcome(d.decisionId, ok ? "success" : "failure");

Why it works

The gate returns act only where it has verified — against your own resolved outcomes — that this kind of action is trustworthy; otherwise it escalates or abstains. Stakesraise the bar for riskier actions. Nothing is graded on volume — only on outcomes that actually resolved. There is no global accuracy number; trust is per cell. That calibrated humility is the point: the system tells you where it doesn't know yet.

More recipes