export type UrlShapeReason = 'homepage' | 'serp' | 'social-promo';
export type ContentGateReason = 'low-content' | 'low-overlap';
export type ScoreFloorReason = 'negative-score';
export interface UrlShapeVerdict {
    reject: boolean;
    reason?: UrlShapeReason;
}
export interface ContentGateVerdict {
    reject: boolean;
    reason?: ContentGateReason;
}
export interface ScoreFloorVerdict {
    reject: boolean;
    reason?: ScoreFloorReason;
}
export declare const SCORE_FLOOR = 0;
/**
 * Classify a candidate research source by its (post-rerank) relevance score.
 * Rejects strictly-negative scores — the cross-encoder's "not relevant"
 * verdict. Positive scores (the keyless passthrough path's engine/RRF values,
 * or a relevant cross-encoder logit) always survive, so this is a no-op when
 * no cross-encoder ran. The caller is responsible for never emptying the pool
 * (it keeps the single best source as a floor of last resort).
 *
 * `withinBreadthKeep` marks the top-N best-ranked candidates (by the reranker's
 * own ordering) and disables the negative-score reject for them entirely. The
 * cross-encoder's ABSOLUTE logits are miscalibrated per-query: on a niche query
 * it damps the WHOLE pool below zero — including genuinely canonical, on-topic
 * pages (live C1: sqlite.org/fts5.html, the sqlite-vec author's post, dev.to /
 * deepwiki / kentcdodds explainers all scored below even -0.35), which collapses
 * standard depth to ~1 source. Its RELATIVE ordering stays meaningful, so inside
 * the keep window the quality bar is rank position + the upstream url-shape and
 * content gates, NOT an absolute cutoff. Outside the window the strict `< 0`
 * rule applies unchanged, so a genuinely off-topic page that ranks past the pool
 * (it sorts below the on-topic survivors) is still rejected — that is what keeps
 * junk out of the long tail while the window stays wide.
 */
export declare function classifyScoreFloor(relevanceScore: number, withinBreadthKeep?: boolean): ScoreFloorVerdict;
/**
 * Classify a candidate URL by shape alone (no fetch). A bare-root homepage or a
 * search-engine results page is junk for research synthesis. `includeDomains`
 * roots are exempt — if the caller scoped research to a domain, its root is an
 * intentional target.
 */
export declare function classifyUrlShape(url: string, includeDomains?: string[]): UrlShapeVerdict;
/**
 * Distinct, stop-word-stripped content terms from the research question. Used
 * by the content gate to measure query overlap.
 */
export declare function queryContentTerms(question: string): string[];
/**
 * Gate fetched (or snippet) content. Rejects only when the page is BOTH thin
 * AND off-topic — a short on-topic doc survives (high overlap) and a long page
 * survives (high word count). Fails open when there are no query terms.
 */
export declare function gateContent(markdown: string, queryTerms: string[]): ContentGateVerdict;
//# sourceMappingURL=source-validation.d.ts.map