import type { EvalReport } from './EvalRunner.ts';
export interface MetricDelta {
    metric: string;
    current: number;
    baseline: number;
    delta: number;
}
export interface RegressionReport {
    overall: 'pass' | 'flag';
    regressions: MetricDelta[];
    improvements: MetricDelta[];
    unchanged: MetricDelta[];
    missingMetrics: string[];
}
/** Snapshot of an EvalReport's aggregates, persisted as a comparison target. */
export interface BaselineRecord {
    member: string;
    suiteName: string;
    timestamp: string;
    taskCount: number;
    aggregate: Record<string, number>;
}
export declare const DEFAULT_THRESHOLD = 0.1;
/**
 * Compares a fresh EvalReport against a stored baseline, flagging metrics
 * whose aggregate dropped by more than the threshold.
 */
export declare class BaselineComparator {
    private readonly threshold;
    constructor(threshold?: number);
    compare(report: EvalReport, baseline: BaselineRecord): RegressionReport;
    /** Persists the report's aggregates as the new baseline. */
    saveBaseline(report: EvalReport, path: string): BaselineRecord;
    /** Reads a baseline file, or null when missing or unparseable. */
    loadBaseline(path: string): BaselineRecord | null;
}
//# sourceMappingURL=BaselineComparator.d.ts.map