Judgment API

Get calibrated decisions instead of generated text. Send a state plus typed questions and get back one answer per question with probabilities your code can branch on — routing tickets, gating automation on confidence, scoring content against a rubric. Any judgment-type catalog model works through this same endpoint (TypeSafe's Jev today).

This is a proxy over TypeSafe's System One API (POST /v1/systemone): Cencori adds its own auth, per-project keys (BYOK), input-token billing, logging, and usage metering.

Authenticate with your Cencori key (CENCORI_API_KEY: csk_... header or Authorization: Bearer). A TypeSafe key is resolved per project — your own typesafe provider key first, the managed key as fallback. BYOK calls are not charged; managed calls bill input tokens at $0.042/MTok (output tokens are unbilled).

curl https://api.cencori.com/v1/judgments \
  -H "CENCORI_API_KEY: csk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-latest",
    "state": "I was charged twice. Please refund the duplicate.",
    "questions": {
      "department": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
          "billing": "Payments, invoicing, refunds",
          "technical": "Bugs, outages, integrations"
        }
      },
      "requests_refund": {
        "type": "noul",
        "instructions": "Does the customer request a refund?"
      }
    }
  }'

import { Cencori } from 'cencori';
 
const cencori = new Cencori({ apiKey: 'csk_...' });
const result = await cencori.judgment.judge({
  state: 'I was charged twice. Please refund the duplicate.',
  questions: {
    department: {
      type: 'choice',
      instructions: 'Which team should handle this?',
      criteria: { billing: 'Payments, invoicing, refunds', technical: 'Bugs, outages' },
    },
    requests_refund: { type: 'noul', instructions: 'Does the customer request a refund?' },
  },
});
 
if (result.answers.department.type === 'choice' && result.answers.department.confidence > 0.8) {
  routeTicket(result.answers.department.choice);
}

Omit model to use the default (jev-latest). Aliases resolve upstream — the response names the version that actually judged (e.g. jev-1.13.0).

Every question needs instructions. choice needs criteria as an option → rubric map (use null where an option needs no detail); score needs criteria as the ordered level array. noul accepts optional true/false criteria describing each side. Answers come back under the same ids you sent.

state is what the questions are about: a string, a JSON object, or an array of text. The 64k-token per-request budget covers the state plus all questions combined (state plus the single longest question must fit 32k) — over-budget requests are rejected before billing.

GET /v1/judgments is the authoritative list with live pricing and availability.

{
  "object": "judgment",
  "model": "jev-1.13.0",
  "answers": {
    "department": {
      "type": "choice",
      "choice": "billing",
      "confidence": 0.91,
      "probabilities": { "billing": 0.93, "technical": 0.07 }
    },
    "requests_refund": { "type": "noul", "noul": 0.97 }
  },
  "usage": { "input_tokens": 214, "output_tokens": 18 }
}

confidence (choice/score) says how peaked the distribution is — use it as a second axis: act above your threshold, escalate to a person or a reasoning model below it. (noul carries no separate confidence; max(noul, 1 - noul) is the equivalent.)

Judgment classifies and rates content you already have — it never writes replies, summaries, or code, and it answers one flat batch of independent questions per call (no agent loop, no tool use). For generation or multi-step reasoning, use Chat or Responses and let a judgment call gate or route the result.