Grade: call summary faithfulness

The job

Jev judge eval (1 score) over a fixed state : runnable via experimental_evaluate.

The prompt

Copy this into Muse and replace anything in [square brackets] with your own details.

// Grade: call summary faithfulness — Jev judge eval (typesafe-ai/jev).
// An agent summarizes a support call; the judge scores how faithfully the summary matches the call.
// Model: typesafe-ai/jev via the Vercel AI Gateway ($0.042/1M input tokens).
// Auth: set AI_GATEWAY_API_KEY (vck_…) in the environment — never paste the key here.
// For zero-data-retention / no-training workloads, add
// providerOptions: { gateway: { … } } to the evaluate call below.

import { experimental_evaluate } from 'ai';

const result = await experimental_evaluate({
  model: 'typesafe-ai/jev',
  state: "Call: customer reported ORDER #5520 missing for 9 days; agent promised a reshipment plus a $10 credit; customer accepted.\nSummary: Customer called about a late order. Agent offered a refund, which the customer declined.",
  questions: {
    "faithfulness": {
      "type": "score",
      "instructions": "Score how faithfully the summary reflects the call. Higher is better.",
      "criteria": [
        "Fabricates the outcome (wrong resolution entirely).",
        "Omits key facts (order number, delay length, or agreed remedy).",
        "Mostly accurate with one minor error or omission.",
        "Fully accurate: order, delay, remedy, and acceptance all correct."
      ]
    }
  },
});

console.log(JSON.stringify(result, null, 2));