tests/test_summary.json

{
  "total_samples": 20,
  "overall_accuracy": 0.65,
  "human_samples": 10,
  "human_correct": 4,
  "human_accuracy": 0.4,
  "ai_samples": 10,
  "ai_correct": 9,
  "ai_accuracy": 0.9,
  "false_positives": 6,
  "false_negatives": 1,
  "time_seconds": 255.3
}