export interface AutoResearchBenchOptions {
    taskSet?: string;
    experimentTag?: string;
    runs?: string;
    split?: string;
    model?: string;
    harnessRef?: string;
    benchmarkSource?: string;
    serverUrl?: string;
    json?: boolean;
    /** Seed a deficiency for every failing task into .cortex/research-backlog.jsonl
     *  (default true; holdout split never seeds). Pass false for candidate-worktree
     *  benches so a not-yet-fixed task doesn't re-stamp the discovery record. */
    seedBacklog?: boolean;
    /** Non-cortex target: grade a shell command per task instead of POSTing to a cortex
     *  server. `runCmd` is the per-task template ({prompt}/{case} substituted). Pair with
     *  `numeric` verifiers in the task set. */
    runCmd?: string;
    /** Optional one-shot build, run once in --cwd before benching the command target. */
    buildCmd?: string;
    /** Working dir for --run-cmd / --build-cmd (default: project root). */
    cwd?: string;
    /** Comma list of exit codes whose stdout is graded (default "0"). */
    acceptExit?: string;
    /** Effectiveness-arm labels recorded with each scored run, so the matrix can rank
     *  (model × temperature × strategy). Both fall back to the CORTEX_SUBAGENT_TEMPERATURE /
     *  CORTEX_ARM_STRATEGY env stamp when omitted. */
    temperature?: string;
    strategy?: string;
}
export declare function autoResearchBench(options: AutoResearchBenchOptions): Promise<void>;
//# sourceMappingURL=bench.d.ts.map