{
  "status": "awaiting_human_annotations",
  "reviewed_samples": 0,
  "evaluation_samples": 0,
  "evaluation_medicines": 0,
  "classification_accuracy_on_reviewed_holdout": null,
  "double_reviewed_samples": 0,
  "label_agreement": null,
  "detection_precision": null,
  "detection_recall": null,
  "clinical_performance": null,
  "protocol": "Medicine-disjoint deterministic development/evaluation split (hash modulo 5). Blank labels, conflicts and automatic predictions are excluded from gold labels.",
  "limitations": [
    "Control paragraphs are sampled, not exhaustive independently annotated document pairs.",
    "Classification accuracy only describes the reviewed held-out sample.",
    "Detection precision/recall require independent full-document review, including missed changes.",
    "Declared reviewer qualifications are recorded, not independently certified by this software."
  ]
}