{
  "direct_accuracy": 0.42,
  "direct_exact": 21,
  "direct_parse_ok": 39,
  "mean_correct_visible_programs": 0.92,
  "mean_visible_programs": 1.14,
  "n": 50,
  "oracle_union": 29,
  "oracle_union_accuracy": 0.58,
  "oracle_visible_program": 18,
  "policies": {
    "consensus_2": {
      "accuracy": 0.48,
      "direct_correct_losses": 1,
      "direct_miss_recoveries": 4,
      "exact": 24,
      "program_commit_correct": 11,
      "program_commit_precision": 0.7857142857142857,
      "program_commits": 14
    },
    "consensus_3": {
      "accuracy": 0.48,
      "direct_correct_losses": 0,
      "direct_miss_recoveries": 3,
      "exact": 24,
      "program_commit_correct": 10,
      "program_commit_precision": 0.9090909090909091,
      "program_commits": 11
    },
    "direct": {
      "accuracy": 0.42,
      "direct_correct_losses": 0,
      "direct_miss_recoveries": 0,
      "exact": 21,
      "program_commit_correct": 0,
      "program_commit_precision": null,
      "program_commits": 0
    },
    "first_visible_program": {
      "accuracy": 0.56,
      "direct_correct_losses": 1,
      "direct_miss_recoveries": 8,
      "exact": 28,
      "program_commit_correct": 18,
      "program_commit_precision": 0.782608695652174,
      "program_commits": 23
    }
  },
  "tasks_with_visible_program": 23,
  "total_forward_tokens": 890030,
  "variants": [
    "verified_structural",
    "cell_parser",
    "row_column_rule",
    "header_aware",
    "split_fold_unpivot"
  ]
}