{
  "regression": {
    "overall": {
      "n": 664,
      "correct": 617,
      "accuracy": 0.9292168674698795,
      "macro_f1": 0.8939250255772296,
      "tp": 117,
      "fn": 33,
      "fp": 14,
      "tn": 500,
      "deep_recall": 0.78,
      "false_deep_rate": 0.027237354085603113,
      "deep_recall_wilson95": [
        0.7071754568460243,
        0.8388407052323733
      ],
      "false_deep_wilson95": [
        0.016292663069040196,
        0.045196410037485035
      ],
      "brier": 0.05831906802495684,
      "nll": 0.19566529778461716
    },
    "by_language": {
      "en": {
        "n": 303,
        "correct": 274,
        "accuracy": 0.9042904290429042,
        "macro_f1": 0.8721053780656429,
        "tp": 61,
        "fn": 24,
        "fp": 5,
        "tn": 213,
        "deep_recall": 0.7176470588235294,
        "false_deep_rate": 0.022935779816513763,
        "deep_recall_wilson95": [
          0.6141587579262482,
          0.8023128049451995
        ],
        "false_deep_wilson95": [
          0.009835515458069499,
          0.05255855310256116
        ],
        "brier": 0.07394316489740126,
        "nll": 0.25100498741147925
      },
      "zh": {
        "n": 361,
        "correct": 343,
        "accuracy": 0.9501385041551247,
        "macro_f1": 0.9155665280665282,
        "tp": 56,
        "fn": 9,
        "fp": 9,
        "tn": 287,
        "deep_recall": 0.8615384615384616,
        "false_deep_rate": 0.030405405405405407,
        "deep_recall_wilson95": [
          0.7573167852534054,
          0.9254099381231576
        ],
        "false_deep_wilson95": [
          0.016077105119947862,
          0.056766689737069984
        ],
        "brier": 0.045205213863320645,
        "nll": 0.14921674942744484
      }
    },
    "by_category": {
      "completion": {
        "n": 10,
        "correct": 10,
        "accuracy": 1.0,
        "macro_f1": 1.0,
        "tp": 6,
        "fn": 0,
        "fp": 0,
        "tn": 4,
        "deep_recall": 1.0,
        "false_deep_rate": 0.0,
        "deep_recall_wilson95": [
          0.6096569663469354,
          0.9999999999999999
        ],
        "false_deep_wilson95": [
          0,
          0.48990002040399916
        ],
        "brier": 0.0010905407833151288,
        "nll": 0.019759182865158858
      },
      "context_completed": {
        "n": 26,
        "correct": 26,
        "accuracy": 1.0,
        "macro_f1": 0.5,
        "tp": 0,
        "fn": 0,
        "fp": 0,
        "tn": 26,
        "deep_recall": null,
        "false_deep_rate": 0.0,
        "deep_recall_wilson95": null,
        "false_deep_wilson95": [
          0,
          0.12873304380462175
        ],
        "brier": 0.00024423200617112355,
        "nll": 0.011072157124723887
      },
      "context_unfinished": {
        "n": 28,
        "correct": 16,
        "accuracy": 0.5714285714285714,
        "macro_f1": 0.5087719298245614,
        "tp": 13,
        "fn": 9,
        "fp": 3,
        "tn": 3,
        "deep_recall": 0.5909090909090909,
        "false_deep_rate": 0.5,
        "deep_recall_wilson95": [
          0.38734477382893806,
          0.7674444033117963
        ],
        "false_deep_wilson95": [
          0.18761280689940868,
          0.8123871931005913
        ],
        "brier": 0.31489699811230015,
        "nll": 0.9439517192864099
      },
      "emotion": {
        "n": 10,
        "correct": 10,
        "accuracy": 1.0,
        "macro_f1": 1.0,
        "tp": 4,
        "fn": 0,
        "fp": 0,
        "tn": 6,
        "deep_recall": 1.0,
        "false_deep_rate": 0.0,
        "deep_recall_wilson95": [
          0.5100999795960008,
          1
        ],
        "false_deep_wilson95": [
          0,
          0.3903430336530645
        ],
        "brier": 3.401839728294804e-05,
        "nll": 0.0028612694791609336
      },
      "factual": {
        "n": 10,
        "correct": 10,
        "accuracy": 1.0,
        "macro_f1": 1.0,
        "tp": 6,
        "fn": 0,
        "fp": 0,
        "tn": 4,
        "deep_recall": 1.0,
        "false_deep_rate": 0.0,
        "deep_recall_wilson95": [
          0.6096569663469354,
          0.9999999999999999
        ],
        "false_deep_wilson95": [
          0,
          0.48990002040399916
        ],
        "brier": 0.0026418461177944662,
        "nll": 0.04135682687840026
      },
      "memory_fact": {
        "n": 10,
        "correct": 9,
        "accuracy": 0.9,
        "macro_f1": 0.8901098901098901,
        "tp": 3,
        "fn": 1,
        "fp": 0,
        "tn": 6,
        "deep_recall": 0.75,
        "false_deep_rate": 0.0,
        "deep_recall_wilson95": [
          0.3006360524426366,
          0.9544139373553637
        ],
        "false_deep_wilson95": [
          0,
          0.3903430336530645
        ],
        "brier": 0.03388762151034709,
        "nll": 0.1221164888785166
      },
      "new_constraint_after_completion": {
        "n": 30,
        "correct": 30,
        "accuracy": 1.0,
        "macro_f1": 0.5,
        "tp": 30,
        "fn": 0,
        "fp": 0,
        "tn": 0,
        "deep_recall": 1.0,
        "false_deep_rate": null,
        "deep_recall_wilson95": [
          0.8864829086095221,
          1
        ],
        "false_deep_wilson95": null,
        "brier": 0.0024941061534353476,
        "nll": 0.03609175294310134
      },
      "public_conversation": {
        "n": 362,
        "correct": 333,
        "accuracy": 0.919889502762431,
        "macro_f1": 0.8358738645779591,
        "tp": 37,
        "fn": 20,
        "fp": 9,
        "tn": 296,
        "deep_recall": 0.6491228070175439,
        "false_deep_rate": 0.029508196721311476,
        "deep_recall_wilson95": [
          0.519416545881409,
          0.7599975392182597
        ],
        "false_deep_wilson95": [
          0.01560036886817662,
          0.05512067389290219
        ],
        "brier": 0.07042387677218072,
        "nll": 0.24027162211281072
      },
      "public_single_turn": {
        "n": 100,
        "correct": 95,
        "accuracy": 0.95,
        "macro_f1": 0.8669859005054535,
        "tp": 8,
        "fn": 3,
        "fp": 2,
        "tn": 87,
        "deep_recall": 0.7272727272727273,
        "false_deep_rate": 0.02247191011235955,
        "deep_recall_wilson95": [
          0.4343497069897537,
          0.9025411942607853
        ],
        "false_deep_wilson95": [
          0.006184269063094168,
          0.07827786859933303
        ],
        "brier": 0.03411899684260777,
        "nll": 0.11113511759895704
      },
      "topic_shift": {
        "n": 30,
        "correct": 30,
        "accuracy": 1.0,
        "macro_f1": 0.5,
        "tp": 0,
        "fn": 0,
        "fp": 0,
        "tn": 30,
        "deep_recall": null,
        "false_deep_rate": 0.0,
        "deep_recall_wilson95": null,
        "false_deep_wilson95": [
          0,
          0.113517091390478
        ],
        "brier": 1.7556884895978534e-09,
        "nll": 3.98404888866351e-05
      },
      "user_pause": {
        "n": 36,
        "correct": 36,
        "accuracy": 1.0,
        "macro_f1": 1.0,
        "tp": 4,
        "fn": 0,
        "fp": 0,
        "tn": 32,
        "deep_recall": 1.0,
        "false_deep_rate": 0.0,
        "deep_recall_wilson95": [
          0.5100999795960008,
          1
        ],
        "false_deep_wilson95": [
          0,
          0.1071827150573635
        ],
        "brier": 9.071688464884205e-06,
        "nll": 0.00122522356652892
      },
      "writing": {
        "n": 12,
        "correct": 12,
        "accuracy": 1.0,
        "macro_f1": 1.0,
        "tp": 6,
        "fn": 0,
        "fp": 0,
        "tn": 6,
        "deep_recall": 1.0,
        "false_deep_rate": 0.0,
        "deep_recall_wilson95": [
          0.6096569663469354,
          0.9999999999999999
        ],
        "false_deep_wilson95": [
          0,
          0.3903430336530645
        ],
        "brier": 0.045280168944203315,
        "nll": 0.17686675055194026
      }
    }
  },
  "regression_scope": "Reused machine-supervised regression; not independent human speech.",
  "latency_common_128_inputs_twice": {
    "gpu": {
      "n": 256,
      "p50_ms": 10.602879337966442,
      "p95_ms": 11.419005924835801,
      "threads": null
    },
    "cpu": {
      "n": 256,
      "p50_ms": 47.47643414884806,
      "p95_ms": 121.94657069630921,
      "threads": 2
    }
  },
  "latency_compiled_full_664_inputs_twice": {
    "n": 1328,
    "p50_ms": 2.2120927460491657,
    "p95_ms": 3.2203245908021927
  },
  "gpu": "NVIDIA L20",
  "tokenization_included": true,
  "result_cache_disabled": true,
  "compile_and_warm_seconds": 61.08646277990192,
  "compiled_decisions_checked": 1222,
  "onnx_decisions_checked": 1222,
  "tokenizer_parity_rows": 7224,
  "native_macos_tested": false,
  "native_windows_tested": false,
  "live_asr_tts_contention_measured": false,
  "model_id": "Sorbett/VoiceMem-Jev"
}