import type { InteractionCategory, Interaction, InteractionAudit, NecessityJudgment, CategoryScore } from "../types/scoring.js";
/**
 * How each audit dimension contributes to a category's raw score.
 *
 * Environment and service evaluate EXECUTION QUALITY only (success + speed).
 * The agent's choice of what to invoke, with what parameters, and whether
 * it was necessary is evaluated under the agent dimension.
 */
export declare const CATEGORY_DIMENSION_WEIGHTS: Record<InteractionCategory, Record<string, number>>;
/** Calibration parameters for the log-normal CDF mapping. */
export interface CalibrationParams {
    /** The raw score (0-1) that maps to 50/100. */
    median: number;
    /** Controls the spread — lower = steeper curve. */
    sigma: number;
}
/** Default calibration. median = raw score that maps to 50/100. */
export declare const DEFAULT_CALIBRATION: Record<InteractionCategory, CalibrationParams>;
export declare const DEFAULT_AUDIT_SCORES: {
    readonly success: 1;
    readonly speed: 1;
    readonly weight: 1;
    readonly contextRelevance: 1;
};
/**
 * Approximation of the standard normal CDF using Abramowitz & Stegun.
 */
export declare function normalCDF(x: number): number;
/**
 * Map a raw 0-1 score through a log-normal CDF to produce 0-100.
 *
 * The log-normal mapping ensures:
 * - Improving from bad (20) to mediocre (50) is "easier" (smaller raw improvement needed)
 * - Improving from good (80) to great (95) requires significant raw improvement
 * - The mapping is S-shaped, rewarding getting out of the "bad" zone
 */
export declare function logNormalScore(rawScore: number, median: number, sigma: number): number;
/**
 * Severity-weighted average: bad scores pull harder than good scores push.
 *
 * Each value's effective weight is `(1 - value)² + 1`. Perfect scores (1.0)
 * get weight 1, while worse scores get progressively heavier, making outlier
 * problems hard to hide behind many good results.
 *
 * @param values - Scores in the 0-1 range
 */
export declare function severityWeightedAverage(values: number[]): number;
/**
 * Aggregate a single audit dimension across interactions in a category,
 * weighted by each interaction's contextBytes.
 */
export declare function aggregateDimension(audits: InteractionAudit[], interactions: Interaction[], dimension: "success" | "speed" | "weight" | "contextRelevance"): number;
/**
 * Compute the full category score from audits, necessity judgment, and interactions.
 * Applies dimension weights and log-normal mapping.
 */
export declare function computeCategoryScore(category: InteractionCategory, audits: InteractionAudit[], necessity: NecessityJudgment, interactions: Interaction[], calibration?: CalibrationParams): CategoryScore;
//# sourceMappingURL=category-score.d.ts.map