/**
 * testManyCombos — turns context pieces on and off in many combinations, re-runs each,
 * and learns which pieces drive the wrong answer (a linear-surrogate attribution), then
 * confirms the top candidate with one clean ablation. The sampling one: it estimates an
 * influence weight per piece from counterfactual data rather than ranking by similarity.
 *
 * Method: random-subset ablation + linear-surrogate attribution. Deterministic
 * pseudo-random masking (a reproducible reimplementation of ContextCite's random subsets).
 */
import type { Finder } from './types.js';
export declare const testManyCombos: Finder;