/**
 * Build the System One benchmark dataset from CLEO's own history (T12495).
 *
 * The control group is CLEO dogfooded: every row is a question a decision
 * site really asks, labelled from what the project's records already say.
 *
 * - `duplicateDetection` — task pairs. Positives: a `duplicates` relation, a
 *   task cancelled with a reason that calls it a duplicate of a named task,
 *   or a note that records a duplicate of a named task. Negatives: pairs of
 *   tasks under the same parent with no relation between them.
 * - `observationType` — observations whose stored type can only have come
 *   from a caller: it is one of the offered options and differs from BOTH
 *   the keyword default the writer applies today and the substring default
 *   it applied before T12494. A type that agrees with a keyword default may
 *   have been defaulted, so it is not used as a label.
 * - `decisionContradiction` — decision pairs. Positives: a supersedes edge
 *   (newer → older). Negatives: random pairs with no supersedes edge between
 *   them. The newer decision's declared `supersedes` is never sent.
 *
 * Pair rows whose text gives the label away (one side names the other's id,
 * or says "duplicate of T123" / "supersedes D12") are dropped
 * ({@link leaksBenchLabel}).
 *
 * Every text field is redacted with the memory redaction patterns System One
 * uses (`redactContent`) and clipped before it is stored, so the dataset file
 * never holds a secret the provider would not have been sent anyway.
 *
 * Pure: the store is read through a {@link BenchSource}; randomness is seeded.
 *
 * @task T12495
 * @epic T12486
 */
import { type BenchRow, type BenchSite, type BenchSource } from './types.js';
/** Characters kept per text field in the dataset (the site builders clip further). */
export declare const BENCH_TEXT_MAX_CHARS = 1000;
/** Default cap on rows per site. */
export declare const DEFAULT_BENCH_MAX_ROWS_PER_SITE = 400;
/** Default negatives sampled per positive. */
export declare const DEFAULT_BENCH_NEGATIVES_PER_POSITIVE = 1;
/** Negatives sampled even when a site has few or no positives. */
export declare const MIN_BENCH_NEGATIVES = 10;
/** Options for {@link buildBenchDataset}. */
export interface BuildBenchDatasetOptions {
    /** Sites to build. Default: all of {@link BENCH_SITES}. */
    readonly sites?: readonly BenchSite[];
    /** PRNG seed. Default {@link DEFAULT_BENCH_SEED}. */
    readonly seed?: number;
    /** Row cap per site. Default {@link DEFAULT_BENCH_MAX_ROWS_PER_SITE}. */
    readonly maxRowsPerSite?: number;
    /** Negatives per positive. Default {@link DEFAULT_BENCH_NEGATIVES_PER_POSITIVE}. */
    readonly negativesPerPositive?: number;
    /** Redaction. Default: the memory redaction patterns (`redactContent`). */
    readonly redact?: (text: string) => string;
}
/** Row counts of a dataset, per site and label. */
export interface BenchDatasetSizes {
    /** Total rows. */
    readonly total: number;
    /** Rows marked owner-verified. */
    readonly ownerVerified: number;
    /** Site → label → rows. */
    readonly bySite: Readonly<Record<string, Readonly<Record<string, number>>>>;
}
/**
 * The type the pre-T12494 writer assigned when the caller gave none: the
 * first keyword group with a SUBSTRING match, else `discovery`.
 *
 * @param text - Observation narrative.
 * @returns The legacy default type.
 */
export declare function legacyKeywordObservationType(text: string): string;
/**
 * Whether a pair row's text gives its label away: either side names the
 * other's id, or says "duplicate of T123" / "supersedes D12". Such rows are
 * dropped, so a provider is never scored on reading the answer off the text.
 *
 * @param row - A dataset row.
 * @returns True when the text leaks the label.
 */
export declare function leaksBenchLabel(row: BenchRow): boolean;
/**
 * Build the labelled dataset from a store.
 *
 * @param source - Store port (the accessors in production, records in tests).
 * @param opts - Sites, seed, caps and redaction.
 * @returns Rows, grouped by site in {@link BENCH_SITES} order.
 */
export declare function buildBenchDataset(source: BenchSource, opts?: BuildBenchDatasetOptions): Promise<BenchRow[]>;
/**
 * Count rows per site and label.
 *
 * @param rows - Dataset rows.
 * @returns The sizes.
 */
export declare function benchDatasetSizes(rows: readonly BenchRow[]): BenchDatasetSizes;
/**
 * Serialise rows as JSONL (one row per line, trailing newline).
 *
 * @param rows - Dataset rows.
 * @returns The JSONL text.
 */
export declare function serializeBenchDataset(rows: readonly BenchRow[]): string;
/**
 * Parse and validate JSONL rows.
 *
 * @param text - JSONL text.
 * @returns The rows.
 * @throws Error naming the first invalid line.
 */
export declare function parseBenchDataset(text: string): BenchRow[];
/**
 * Write rows to a JSONL file.
 *
 * @param path - Target file.
 * @param rows - Dataset rows.
 */
export declare function writeBenchDataset(path: string, rows: readonly BenchRow[]): Promise<void>;
/**
 * Read rows from a JSONL file.
 *
 * @param path - Source file.
 * @returns The rows.
 */
export declare function readBenchDataset(path: string): Promise<BenchRow[]>;
//# sourceMappingURL=dataset.d.ts.map