{
  "total_cells": 96,
  "passed_cells": 81,
  "pass_rate": 0.8438,
  "token_budget_compliance_rate": 1.0,
  "mean_uuid_recall": 1.0,
  "mean_citation_resolution_rate": 1.0,
  "privacy_leak_count": 3,
  "abstention_accuracy": 0.875,
  "by_use_case": {
    "A_reconcile": {
      "total": 39,
      "passed": 39,
      "pass_rate": 1.0,
      "mean_uuid_recall": 1.0,
      "mean_citation_resolution_rate": 1.0,
      "privacy_leaks": 0
    },
    "B1_encounter_summary": {
      "total": 3,
      "passed": 3,
      "pass_rate": 1.0,
      "mean_uuid_recall": 1.0,
      "mean_citation_resolution_rate": 1.0,
      "privacy_leaks": 0
    },
    "B1_longitudinal_summary": {
      "total": 9,
      "passed": 9,
      "pass_rate": 1.0,
      "mean_uuid_recall": 1.0,
      "mean_citation_resolution_rate": 1.0,
      "privacy_leaks": 0
    },
    "B2_directed_summary": {
      "total": 9,
      "passed": 6,
      "pass_rate": 0.6667,
      "mean_uuid_recall": 1.0,
      "mean_citation_resolution_rate": 1.0,
      "privacy_leaks": 3
    },
    "B3_qa_abstention": {
      "total": 12,
      "passed": 0,
      "pass_rate": 0.0,
      "mean_uuid_recall": 1.0,
      "mean_citation_resolution_rate": 1.0,
      "privacy_leaks": 0
    },
    "B3_qa_answerable": {
      "total": 24,
      "passed": 24,
      "pass_rate": 1.0,
      "mean_uuid_recall": 1.0,
      "mean_citation_resolution_rate": 1.0,
      "privacy_leaks": 0
    }
  }
}
