{ "total_examples": 160, "test_passing_below_50_pct": 0.65, "score_resolved_correlation": 1.0, "mean_score_resolved": 35.6, "mean_score_unresolved": 1.4166666666666667, "known_bad_detected": 50, "known_bad_total": 50, "known_bad_detection_rate": 1.0, "score_mean": 22.78125, "score_std": 26.889284268841482, "score_median": 0.0, "dimension_stats": { "correctness": { "mean": 2.69, "std": 3.15, "median": 0.0, "min": 0, "max": 9 }, "completeness": { "mean": 1.96, "std": 2.26, "median": 0.0, "min": 0, "max": 6 }, "code_quality": { "mean": 2.37, "std": 2.93, "median": 0.0, "min": 0, "max": 9 }, "non_regression_risk": { "mean": 2.36, "std": 2.88, "median": 0.0, "min": 0, "max": 9 }, "merge_readiness": { "mean": 1.67, "std": 2.21, "median": 0.0, "min": 0, "max": 8 } } }