Introducing Judgment API

Most model APIs generate text. Sometimes you don't want text — you want a decision: is this urgent, which team owns it, how angry is this customer on a scale of three.

Today we're launching the Judgment API: structured decisions from decision models, through the same billable, logged gateway as everything else on Cencori. First up is TypeSafe's Jev — and any future decision model plugs into the same endpoint with no integration changes on your side.

Send a state (text, JSON, or an array) plus typed questions. Get back one answer per question with probabilities — never generated prose:

import { Cencori } from 'cencori';

const cencori = new Cencori({ apiKey: process.env.CENCORI_API_KEY });

const result = await cencori.judgment.judge({
  state: "I've been trying to connect my Stripe account for 3 days. Losing sales. Help ASAP.",
  questions: {
    department: {
      type: 'choice',
      instructions: 'Which team should handle this?',
      criteria: {
        billing: 'Payments, invoicing, refunds',
        technical: 'Bugs, outages, integrations',
      },
    },
    is_urgent: { type: 'noul', instructions: 'Does this convey urgency?' },
  },
});

if (result.answers.department.type === 'choice' && result.answers.department.confidence > 0.8) {
  routeTicket(result.answers.department.choice); // 'technical'
}

Or raw HTTP — POST /v1/judgments with your CENCORI_API_KEY, same envelope. GET /v1/judgments lists models with live pricing, no key required.

The classic pattern is asking a chat model to "reply with JSON" and praying the output parses. Judgment inverts it: the model never writes, so there is nothing to parse. confidence becomes a second axis for your code — act above your threshold, escalate to a person or a reasoning model below it:

• Route support tickets by department, urgency, and sentiment in one request (questions evaluate in parallel against the same state)
• Gate automation: refund under $50 when requests_refund > 0.9 and confidence is high, else queue for review
• Score content against your own rubric — toxicity, quality, completeness — with a full distribution, not a single label

jev-latest (default), jev-preview, and the pinned jev-1.13.0. Aliases resolve upstream, so the response always names the version that actually judged. And the endpoint is vendor-neutral: Cloudflare's Clef decision models are live on it too — clef-latest (27B, $0.24/MTok input), clef-omni ($0.15/MTok, the tier that adds audio and video states), and clef-flash (9B, $0.038/MTok, 24k budget). Same body, same envelope; switch by changing one string.

Jev bills input tokens only at $0.042 per million, as does Clef at $0.24 per million and Clef Flash at $0.038 — output tokens are unbilled everywhere. Bring your own TypeSafe or Cloudflare key in Project → Providers and calls aren't charged at all — managed-key calls bill to credits like every other gateway request, with the same logs, spend caps, and key allowlists.

Judgment classifies and rates content you already have. It won't write your reply, summarize your document, or chain multi-step reasoning — pair it with Chat or Responses: let generation produce, let judgment verify, route, and gate.

Full reference: Judgment API docs.