@remnic/bench 9.3.675 → 9.3.677

package/dist/index.d.ts

@@ -349,6 +349,13 @@ interface BenchmarkResult {
         estimatedCostUsd: number;
         totalLatencyMs: number;
         meanQueryLatencyMs: number;
+        /**
+         * Number of underlying judge model calls actually issued. When a content-
+         * keyed judge-result cache is enabled (#1573 PR1) and answers are
+         * unchanged, this equals the number of cache misses. Re-runs after the
+         * first put perform zero new judge model calls.
+         */
+        judgeModelCalls?: number;
     };
     results: {
         tasks: TaskResult[];
@@ -404,6 +411,18 @@ interface RunBenchmarkOptions {
     amaBenchJudgeProtocol?: AmaBenchJudgeProtocol;
     amaBenchCrossJudge?: BenchJudge;
     amaBenchCrossJudgeProvider?: ProviderConfig | null;
+    /**
+     * Force-disable the content-keyed judge-result cache (#1573 PR1). When
+     * true, every judge call reaches the underlying model regardless of
+     * whether `judgeCacheDir` is set. CLI flag: `--no-judge-cache`.
+     */
+    noJudgeCache?: boolean;
+    /**
+     * Override the on-disk directory used to persist judge verdicts. Defaults
+     * to `<outputDir>/judge-cache` when outputDir is supplied; ignored when
+     * `noJudgeCache` is true. The directory is created on demand.
+     */
+    judgeCacheDir?: string;
     /** Called after each task completes for progress logging and partial result tracking. */
     onTaskComplete?: (task: TaskResult, completedCount: number, totalCount?: number) => void;
 }