import { type SemanticInjectionClassifierOptions } from './semantic-injection-classifier.js';
/** One labelled prompt for calibrating the injection classifier. */
export interface InjectionCalibrationExample {
    /** The prompt. */
    text: string;
    /** Whether it is an attack. */
    label: 'attack' | 'safe';
    /** What kind of prompt it is, such as `jailbreak`. */
    category?: string;
}
/** How the classifier performed at one threshold. */
export interface InjectionCalibrationResult {
    /** The threshold tested. */
    threshold: number;
    /** Examples tested. */
    total: number;
    /** Share classified correctly. */
    accuracy: number;
    /** Share of prompts flagged that were attacks. */
    precision: number;
    /** Share of attacks flagged. */
    recall: number;
    /** Share of safe prompts flagged. */
    falsePositiveRate: number;
    /** Share of attacks missed. */
    falseNegativeRate: number;
}
/** A small built-in calibration set: five attacks and five safe prompts that look like them. */
export declare const SEMANTIC_INJECTION_CALIBRATION_SET: InjectionCalibrationExample[];
/**
 * Scores the semantic injection classifier at several thresholds, so you can pick one for your own
 * traffic.
 */
export declare function calibrateSemanticInjectionClassifier(dataset?: InjectionCalibrationExample[], thresholds?: number[], options?: Omit<SemanticInjectionClassifierOptions, 'threshold'>): InjectionCalibrationResult[];