/**
 * Turn a benchmark row into the exact question its decision site asks, the
 * site's heuristic answer, and the mapping from an answer back to a label
 * (T12495).
 *
 * The request builders are the sites' own (`buildDuplicateDecisionRequest`,
 * `buildObservationTypeRequest`, `buildContradictionRequest`), so the
 * benchmark measures the production wording. The heuristic is the rule each
 * site falls back to: Tier-1 lexical similarity for duplicates, whole-word
 * keywords for observation types, the word-Jaccard collision check for
 * decisions.
 *
 * @task T12495
 * @epic T12486
 */
import type { DecisionAnswer, DecisionRequest } from '@cleocode/contracts';
import type { BenchRow } from './types.js';
/** One row made ready to ask. */
export interface BenchQuestion {
    /** The site's request, with `cache: false` (no provider answer cache). */
    readonly req: DecisionRequest;
    /** The site heuristic's answers, in the request's shape (the fallback). */
    readonly heuristicAnswers: Readonly<Record<string, DecisionAnswer>>;
    /** The label the heuristic predicts. */
    readonly heuristicLabel: string;
    /**
     * Map answers back to a label.
     *
     * @returns The predicted label, or `null` when the answer is unusable.
     */
    readonly predict: (answers: Readonly<Record<string, DecisionAnswer>>) => string | null;
}
/**
 * Build the question for `row`.
 *
 * @param row - A dataset row.
 * @returns Request, heuristic answer and label mapping.
 */
export declare function benchQuestionFor(row: BenchRow): BenchQuestion;
//# sourceMappingURL=questions.d.ts.map