/**
 * Policy Comparative Evaluator
 *
 * Compares multiple AI models on Kubernetes organizational policy intent management scenarios
 * Groups by interaction_id (e.g., policy_create_workflow) and evaluates
 * quality of policy creation, validation, and enforcement recommendations across different models
 */
import { BaseComparativeEvaluator, ComparativeEvaluationScore } from './base-comparative.js';
import { ComparisonScenario } from '../dataset-analyzer.js';
export declare class PolicyComparativeEvaluator extends BaseComparativeEvaluator {
    readonly name = "policy-comparative";
    readonly description = "Compares AI models on Kubernetes organizational policy intent management quality";
    protected readonly promptFileName = "policy-comparative.md";
    protected readonly toolName = "policy";
    constructor(datasetDir?: string);
    evaluateAllScenarios(): Promise<ComparativeEvaluationScore[]>;
    /**
     * Build the evaluation prompt - uses base class reliability context with policy-specific template
     */
    protected buildEvaluationPrompt(scenario: ComparisonScenario, modelResponsesText: string, modelList: string): string;
    /**
     * Get detailed breakdown of evaluation phases available
     */
    getEvaluationPhases(): {
        phase: string;
        description: string;
        availableModels: string[];
        scenarioCount: number;
    }[];
}
//# sourceMappingURL=policy-comparative.d.ts.map