import type { EmpiricalEnv } from './empirical.js';
import type { LatencyResult, LatencyStats, LatencyTrial, Provider } from './types.js';
export type { LatencyResult, LatencyStats, LatencyTrial };
export interface MeasureLatencyOptions {
    modelId: string;
    prompt: string;
    trials: number;
    maxTokens?: number;
    env: EmpiricalEnv;
}
/**
 * Linear-interpolation percentile. `p` is 0-100. With p95 over n=3 trials this
 * just returns max(values), which is what we want for the "tail" reading.
 */
declare const nthPercentile: (values: readonly number[], p: number) => number;
/**
 * Test seam: callers (tests) can substitute a provider's trial runner. The
 * production `measureLatency` always uses the real implementations above.
 */
export interface LatencyDeps {
    fetchImpl?: typeof fetch;
    trialRunner?: (provider: Provider, modelId: string, prompt: string, maxTokens: number, env: EmpiricalEnv) => Promise<LatencyTrial>;
}
/**
 * Run `trials` real streaming generations and return percentile-summarized
 * latency data. Each trial is retried once on failure; if both attempts
 * fail, the original provider error is rethrown.
 */
export declare const measureLatency: (options: MeasureLatencyOptions, deps?: LatencyDeps) => Promise<LatencyResult>;
export { nthPercentile };
//# sourceMappingURL=latency.d.ts.map