{
 "run_date_utc": "2026-06-12",
 "n_texts": 60,
 "n_measurements": 180,
 "detectors": {
  "walter": {
   "name": "Walter Writes",
   "n_total": 60,
   "n_human": 20,
   "n_ai": 20,
   "n_humanized": 20,
   "correct_total": 47,
   "accuracy_pct": 78.3,
   "false_positives": 0,
   "fp_rate_pct": 0.0,
   "false_negatives": 0,
   "fn_rate_pct": 0.0,
   "humanized_detected": 7,
   "humanized_detect_rate_pct": 35.0,
   "mean_score_human": 13.2,
   "mean_score_ai": 98.7,
   "mean_score_humanized": 49.3,
   "core_benchmark_correct_of_40": 40,
   "core_benchmark_pct": 100.0
  },
  "gptzero": {
   "name": "GPTZero",
   "n_total": 60,
   "n_human": 20,
   "n_ai": 20,
   "n_humanized": 20,
   "correct_total": 42,
   "accuracy_pct": 70.0,
   "false_positives": 0,
   "fp_rate_pct": 0.0,
   "false_negatives": 1,
   "fn_rate_pct": 5.0,
   "humanized_detected": 3,
   "humanized_detect_rate_pct": 15.0,
   "mean_score_human": 0.0,
   "mean_score_ai": 90.9,
   "mean_score_humanized": 30.2,
   "core_benchmark_correct_of_40": 39,
   "core_benchmark_pct": 97.5
  },
  "zerogpt": {
   "name": "ZeroGPT",
   "n_total": 60,
   "n_human": 20,
   "n_ai": 20,
   "n_humanized": 20,
   "correct_total": 23,
   "accuracy_pct": 38.3,
   "false_positives": 7,
   "fp_rate_pct": 35.0,
   "false_negatives": 11,
   "fn_rate_pct": 55.0,
   "humanized_detected": 1,
   "humanized_detect_rate_pct": 5.0,
   "mean_score_human": 34.3,
   "mean_score_ai": 35.4,
   "mean_score_humanized": 10.7,
   "core_benchmark_correct_of_40": 22,
   "core_benchmark_pct": 55.0
  }
 },
 "texts_with_disagreement": 25,
 "human_fp_texts": {
  "H01": [
   "zerogpt"
  ],
  "H03": [
   "zerogpt"
  ],
  "H04": [
   "zerogpt"
  ],
  "H07": [
   "zerogpt"
  ],
  "H08": [
   "zerogpt"
  ],
  "H10": [
   "zerogpt"
  ],
  "H16": [
   "zerogpt"
  ]
 },
 "humanized_passed_all_three": [
  "A01",
  "A02",
  "A03",
  "A05",
  "B03",
  "B04",
  "B05",
  "B06",
  "B07",
  "C01",
  "D01",
  "D02"
 ],
 "metric_note": "core_benchmark = 20 human (low signal correct) + 20 unmodified AI (high signal correct); stress test (20 twice-revised AI texts) reported separately; overall alignment includes stress test"
}