/**
 * Frozen demo datasets for `respan demo`, hand-authored as typed TypeScript.
 *
 * Two datasets, numbered to match the demo prompt/evaluator they pair with:
 *   - "Demo dataset 1: support messages"  -> prompt 1 / evaluator 1
 *       Rows are customer messages that probe Lumen's refund/escalation policy
 *       edges. No `expected_output` — evaluator 1 scores compliance from the
 *       reply alone (it uses {{input}}/{{output}} only).
 *   - "Demo dataset 3: call transcripts"  -> prompt 3 / evaluator 3
 *       Rows are call transcripts, each with a gold-label classification in
 *       `expected_output` (evaluator 3 compares against {{expected_output}}).
 *
 * `input` is an object of the prompt's template variables (`customer_message` /
 * `call_transcript`); it is stored stringified and parsed into the row's mapped
 * fields. `expected_output` is authored here as an object for readability and
 * JSON-stringified at seed time — passing a raw object makes the platform coerce
 * it to single-quoted python-repr (invalid JSON the grader would then see).
 */
/** One dataset row: template variables plus an optional gold label. */
export interface DemoDatasetRow {
    /** The prompt's template variables for this row. */
    input: Record<string, unknown>;
    /** Gold label (object; stringified to JSON at seed time). */
    expected_output?: unknown;
}
/** One demo dataset: name/description plus its rows. */
export interface DemoDatasetFixture {
    name: string;
    description?: string;
    rows: DemoDatasetRow[];
}
export declare const DEMO_DATASETS: DemoDatasetFixture[];