{
  "by_family": [
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 5,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "agriculture",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 5,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "craigslist_data_wrangler",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 0,
      "direct_parse_rate": 0.0,
      "family": "crime_data_wrangler",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "potters_wheel_divide",
      "final_hidden_exact": 1,
      "final_program_accuracy": 0.2,
      "final_visible_pass": 1,
      "final_visible_pass_rate": 0.2,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.8,
      "n": 5,
      "neither": 4,
      "oracle_union": 1,
      "oracle_union_rate": 0.2,
      "program_only": 1,
      "repair_added_program_correct": 1,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 5,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "potters_wheel_fold",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 5,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "potters_wheel_fold_2",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "potters_wheel_merge_split",
      "final_hidden_exact": 3,
      "final_program_accuracy": 0.6,
      "final_visible_pass": 4,
      "final_visible_pass_rate": 0.8,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.6,
      "initial_hidden_exact": 3,
      "mean_round_count": 1.6,
      "n": 5,
      "neither": 2,
      "oracle_union": 3,
      "oracle_union_rate": 0.6,
      "program_only": 3,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 0.25
    },
    {
      "agreement": 3,
      "agreement_correct": 3,
      "agreement_precision": 1.0,
      "both_direct_and_program": 3,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 2,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "potters_wheel_split_fold",
      "final_hidden_exact": 3,
      "final_program_accuracy": 0.6,
      "final_visible_pass": 3,
      "final_visible_pass_rate": 0.6,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.6,
      "initial_hidden_exact": 3,
      "mean_round_count": 2.2,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.4,
      "direct_exact": 2,
      "direct_only": 2,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "potters_wheel_unfold",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.4,
      "hybrid_exact": 2,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 3,
      "oracle_union": 2,
      "oracle_union_rate": 0.4,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 1,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "potters_wheel_unfold2",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 1,
      "final_visible_pass_rate": 0.2,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.6,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 1.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 5,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "proactive_wrangling_complex",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 2,
      "agreement_correct": 2,
      "agreement_precision": 1.0,
      "both_direct_and_program": 2,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 3,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "proactive_wrangling_fold",
      "final_hidden_exact": 2,
      "final_program_accuracy": 0.4,
      "final_visible_pass": 2,
      "final_visible_pass_rate": 0.4,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.2,
      "initial_hidden_exact": 1,
      "mean_round_count": 3.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 1,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 5,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "reshape_table_structure_data_wrangler",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 4,
      "agreement_correct": 4,
      "agreement_precision": 1.0,
      "both_direct_and_program": 4,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 1,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_10",
      "final_hidden_exact": 4,
      "final_program_accuracy": 0.8,
      "final_visible_pass": 4,
      "final_visible_pass_rate": 0.8,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.8,
      "initial_hidden_exact": 4,
      "mean_round_count": 1.6,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 4,
      "agreement_correct": 4,
      "agreement_precision": 1.0,
      "both_direct_and_program": 4,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_11",
      "final_hidden_exact": 5,
      "final_program_accuracy": 1.0,
      "final_visible_pass": 5,
      "final_visible_pass_rate": 1.0,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.2,
      "initial_hidden_exact": 1,
      "mean_round_count": 1.8,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 1,
      "repair_added_program_correct": 4,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 5,
      "agreement_correct": 5,
      "agreement_precision": 1.0,
      "both_direct_and_program": 5,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_12",
      "final_hidden_exact": 5,
      "final_program_accuracy": 1.0,
      "final_visible_pass": 5,
      "final_visible_pass_rate": 1.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 1.0,
      "initial_hidden_exact": 5,
      "mean_round_count": 1.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_13",
      "final_hidden_exact": 3,
      "final_program_accuracy": 0.6,
      "final_visible_pass": 4,
      "final_visible_pass_rate": 0.8,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.6,
      "initial_hidden_exact": 3,
      "mean_round_count": 1.8,
      "n": 5,
      "neither": 2,
      "oracle_union": 3,
      "oracle_union_rate": 0.6,
      "program_only": 3,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 0.25
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_15",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.6,
      "direct_exact": 3,
      "direct_only": 3,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_17",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.6,
      "hybrid_exact": 3,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 2,
      "oracle_union": 3,
      "oracle_union_rate": 0.6,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_18",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 4,
      "agreement_correct": 4,
      "agreement_precision": 1.0,
      "both_direct_and_program": 4,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 1,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_19",
      "final_hidden_exact": 4,
      "final_program_accuracy": 0.8,
      "final_visible_pass": 4,
      "final_visible_pass_rate": 0.8,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.8,
      "initial_hidden_exact": 4,
      "mean_round_count": 1.6,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 4,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_2",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 3,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_22",
      "final_hidden_exact": 1,
      "final_program_accuracy": 0.2,
      "final_visible_pass": 4,
      "final_visible_pass_rate": 0.8,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 1.8,
      "n": 5,
      "neither": 4,
      "oracle_union": 1,
      "oracle_union_rate": 0.2,
      "program_only": 1,
      "repair_added_program_correct": 1,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 3,
      "visible_false_pass_rate_among_visible": 0.75
    },
    {
      "agreement": 2,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_24",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 2,
      "final_visible_pass_rate": 0.4,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.4,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 2,
      "visible_false_pass_rate_among_visible": 1.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 4,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_25",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 2,
      "agreement_correct": 1,
      "agreement_precision": 0.5,
      "both_direct_and_program": 1,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 3,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_26",
      "final_hidden_exact": 1,
      "final_program_accuracy": 0.2,
      "final_visible_pass": 2,
      "final_visible_pass_rate": 0.4,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.4,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 1,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 0.5
    },
    {
      "agreement": 3,
      "agreement_correct": 2,
      "agreement_precision": 0.6666666666666666,
      "both_direct_and_program": 2,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 2,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_27",
      "final_hidden_exact": 2,
      "final_program_accuracy": 0.4,
      "final_visible_pass": 3,
      "final_visible_pass_rate": 0.6,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.0,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 2,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 0.3333333333333333
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_28",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_29",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 1,
      "final_visible_pass_rate": 0.2,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.6,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 1.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_3",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 5,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_30",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 2,
      "agreement_correct": 2,
      "agreement_precision": 1.0,
      "both_direct_and_program": 2,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 2,
      "direct_parse_ok": 4,
      "direct_parse_rate": 0.8,
      "family": "synthetic_33",
      "final_hidden_exact": 2,
      "final_program_accuracy": 0.4,
      "final_visible_pass": 3,
      "final_visible_pass_rate": 0.6,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.2,
      "initial_hidden_exact": 1,
      "mean_round_count": 2.4,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 1,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 0.3333333333333333
    },
    {
      "agreement": 1,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_34",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 1,
      "final_visible_pass_rate": 0.2,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.4,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 1.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 4,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_36",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 4,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_37",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 4,
      "agreement_correct": 4,
      "agreement_precision": 1.0,
      "both_direct_and_program": 4,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 1,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_4",
      "final_hidden_exact": 4,
      "final_program_accuracy": 0.8,
      "final_visible_pass": 4,
      "final_visible_pass_rate": 0.8,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.8,
      "initial_hidden_exact": 4,
      "mean_round_count": 1.6,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 5,
      "agreement_correct": 4,
      "agreement_precision": 0.8,
      "both_direct_and_program": 4,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_40",
      "final_hidden_exact": 4,
      "final_program_accuracy": 0.8,
      "final_visible_pass": 5,
      "final_visible_pass_rate": 1.0,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 2.0,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 4,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 0.2
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 0,
      "direct_parse_rate": 0.0,
      "family": "synthetic_41",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 1.0,
      "direct_exact": 5,
      "direct_only": 5,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_43",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 1.0,
      "hybrid_exact": 5,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 1,
      "agreement_correct": 1,
      "agreement_precision": 1.0,
      "both_direct_and_program": 1,
      "direct_accuracy": 0.6,
      "direct_exact": 3,
      "direct_only": 2,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_44",
      "final_hidden_exact": 1,
      "final_program_accuracy": 0.2,
      "final_visible_pass": 1,
      "final_visible_pass_rate": 0.2,
      "hybrid_accuracy": 0.6,
      "hybrid_exact": 3,
      "initial_accuracy": 0.2,
      "initial_hidden_exact": 1,
      "mean_round_count": 3.4,
      "n": 5,
      "neither": 2,
      "oracle_union": 3,
      "oracle_union_rate": 0.6,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 3,
      "agreement_correct": 3,
      "agreement_precision": 1.0,
      "both_direct_and_program": 3,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 1,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_45",
      "final_hidden_exact": 4,
      "final_program_accuracy": 0.8,
      "final_visible_pass": 4,
      "final_visible_pass_rate": 0.8,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.8,
      "initial_hidden_exact": 4,
      "mean_round_count": 1.6,
      "n": 5,
      "neither": 0,
      "oracle_union": 5,
      "oracle_union_rate": 1.0,
      "program_only": 1,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 4,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_46",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 2,
      "agreement_correct": 2,
      "agreement_precision": 1.0,
      "both_direct_and_program": 2,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 2,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_47",
      "final_hidden_exact": 2,
      "final_program_accuracy": 0.4,
      "final_visible_pass": 2,
      "final_visible_pass_rate": 0.4,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.4,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 2,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 1,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.2,
      "direct_exact": 1,
      "direct_only": 1,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_48",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 1,
      "final_visible_pass_rate": 0.2,
      "hybrid_accuracy": 0.2,
      "hybrid_exact": 1,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 4,
      "oracle_union": 1,
      "oracle_union_rate": 0.2,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 1.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.6,
      "direct_exact": 3,
      "direct_only": 3,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_49",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.6,
      "hybrid_exact": 3,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 2,
      "oracle_union": 3,
      "oracle_union_rate": 0.6,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 2,
      "agreement_correct": 2,
      "agreement_precision": 1.0,
      "both_direct_and_program": 2,
      "direct_accuracy": 0.4,
      "direct_exact": 2,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_5",
      "final_hidden_exact": 4,
      "final_program_accuracy": 0.8,
      "final_visible_pass": 5,
      "final_visible_pass_rate": 1.0,
      "hybrid_accuracy": 0.4,
      "hybrid_exact": 2,
      "initial_accuracy": 0.4,
      "initial_hidden_exact": 2,
      "mean_round_count": 1.4,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 2,
      "repair_added_program_correct": 2,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 0.2
    },
    {
      "agreement": 1,
      "agreement_correct": 1,
      "agreement_precision": 1.0,
      "both_direct_and_program": 1,
      "direct_accuracy": 0.4,
      "direct_exact": 2,
      "direct_only": 1,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_51",
      "final_hidden_exact": 3,
      "final_program_accuracy": 0.6,
      "final_visible_pass": 3,
      "final_visible_pass_rate": 0.6,
      "hybrid_accuracy": 0.4,
      "hybrid_exact": 2,
      "initial_accuracy": 0.6,
      "initial_hidden_exact": 3,
      "mean_round_count": 2.2,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 2,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_6",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.0,
      "hybrid_exact": 0,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 5,
      "oracle_union": 0,
      "oracle_union_rate": 0.0,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.8,
      "direct_exact": 4,
      "direct_only": 4,
      "direct_parse_ok": 5,
      "direct_parse_rate": 1.0,
      "family": "synthetic_7",
      "final_hidden_exact": 0,
      "final_program_accuracy": 0.0,
      "final_visible_pass": 0,
      "final_visible_pass_rate": 0.0,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 4.0,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 0,
      "repair_added_program_correct": 0,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "agreement": 0,
      "agreement_correct": 0,
      "agreement_precision": 0.0,
      "both_direct_and_program": 0,
      "direct_accuracy": 0.0,
      "direct_exact": 0,
      "direct_only": 0,
      "direct_parse_ok": 0,
      "direct_parse_rate": 0.0,
      "family": "synthetic_8",
      "final_hidden_exact": 4,
      "final_program_accuracy": 0.8,
      "final_visible_pass": 4,
      "final_visible_pass_rate": 0.8,
      "hybrid_accuracy": 0.8,
      "hybrid_exact": 4,
      "initial_accuracy": 0.2,
      "initial_hidden_exact": 1,
      "mean_round_count": 2.4,
      "n": 5,
      "neither": 1,
      "oracle_union": 4,
      "oracle_union_rate": 0.8,
      "program_only": 4,
      "repair_added_program_correct": 3,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    }
  ],
  "by_num_samples": {
    "1": {
      "agreement": 8,
      "agreement_correct": 4,
      "agreement_precision": 0.5,
      "both_direct_and_program": 4,
      "direct_accuracy": 0.38,
      "direct_exact": 19,
      "direct_only": 15,
      "direct_parse_ok": 46,
      "direct_parse_rate": 0.92,
      "final_hidden_exact": 6,
      "final_program_accuracy": 0.12,
      "final_visible_pass": 15,
      "final_visible_pass_rate": 0.3,
      "hybrid_accuracy": 0.4,
      "hybrid_exact": 20,
      "initial_accuracy": 0.08,
      "initial_hidden_exact": 4,
      "mean_round_count": 3.18,
      "n": 50,
      "neither": 29,
      "oracle_union": 21,
      "oracle_union_rate": 0.42,
      "program_only": 2,
      "repair_added_program_correct": 2,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 9,
      "visible_false_pass_rate_among_visible": 0.6
    },
    "2": {
      "agreement": 9,
      "agreement_correct": 8,
      "agreement_precision": 0.8888888888888888,
      "both_direct_and_program": 8,
      "direct_accuracy": 0.58,
      "direct_exact": 29,
      "direct_only": 21,
      "direct_parse_ok": 47,
      "direct_parse_rate": 0.94,
      "final_hidden_exact": 10,
      "final_program_accuracy": 0.2,
      "final_visible_pass": 13,
      "final_visible_pass_rate": 0.26,
      "hybrid_accuracy": 0.6,
      "hybrid_exact": 30,
      "initial_accuracy": 0.14,
      "initial_hidden_exact": 7,
      "mean_round_count": 3.38,
      "n": 50,
      "neither": 19,
      "oracle_union": 31,
      "oracle_union_rate": 0.62,
      "program_only": 2,
      "repair_added_program_correct": 3,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 3,
      "visible_false_pass_rate_among_visible": 0.23076923076923078
    },
    "3": {
      "agreement": 16,
      "agreement_correct": 12,
      "agreement_precision": 0.75,
      "both_direct_and_program": 12,
      "direct_accuracy": 0.6,
      "direct_exact": 30,
      "direct_only": 18,
      "direct_parse_ok": 47,
      "direct_parse_rate": 0.94,
      "final_hidden_exact": 17,
      "final_program_accuracy": 0.34,
      "final_visible_pass": 19,
      "final_visible_pass_rate": 0.38,
      "hybrid_accuracy": 0.6,
      "hybrid_exact": 30,
      "initial_accuracy": 0.22,
      "initial_hidden_exact": 11,
      "mean_round_count": 3.08,
      "n": 50,
      "neither": 15,
      "oracle_union": 35,
      "oracle_union_rate": 0.7,
      "program_only": 5,
      "repair_added_program_correct": 6,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 2,
      "visible_false_pass_rate_among_visible": 0.10526315789473684
    },
    "4": {
      "agreement": 12,
      "agreement_correct": 11,
      "agreement_precision": 0.9166666666666666,
      "both_direct_and_program": 11,
      "direct_accuracy": 0.62,
      "direct_exact": 31,
      "direct_only": 20,
      "direct_parse_ok": 47,
      "direct_parse_rate": 0.94,
      "final_hidden_exact": 15,
      "final_program_accuracy": 0.3,
      "final_visible_pass": 17,
      "final_visible_pass_rate": 0.34,
      "hybrid_accuracy": 0.64,
      "hybrid_exact": 32,
      "initial_accuracy": 0.18,
      "initial_hidden_exact": 9,
      "mean_round_count": 3.16,
      "n": 50,
      "neither": 15,
      "oracle_union": 35,
      "oracle_union_rate": 0.7,
      "program_only": 4,
      "repair_added_program_correct": 6,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 2,
      "visible_false_pass_rate_among_visible": 0.11764705882352941
    },
    "5": {
      "agreement": 10,
      "agreement_correct": 9,
      "agreement_precision": 0.9,
      "both_direct_and_program": 9,
      "direct_accuracy": 0.58,
      "direct_exact": 29,
      "direct_only": 20,
      "direct_parse_ok": 47,
      "direct_parse_rate": 0.94,
      "final_hidden_exact": 14,
      "final_program_accuracy": 0.28,
      "final_visible_pass": 14,
      "final_visible_pass_rate": 0.28,
      "hybrid_accuracy": 0.6,
      "hybrid_exact": 30,
      "initial_accuracy": 0.18,
      "initial_hidden_exact": 9,
      "mean_round_count": 3.3,
      "n": 50,
      "neither": 16,
      "oracle_union": 34,
      "oracle_union_rate": 0.68,
      "program_only": 5,
      "repair_added_program_correct": 5,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    }
  },
  "final_visible_exec_errors": {
    "CodeParse": 1,
    "IndentationError": 1,
    "IndexError": 11,
    "NameError": 2,
    "None": 220,
    "SyntaxError": 3,
    "TimeoutExpired": 1,
    "TypeError": 2,
    "ValueError": 9
  },
  "overall": {
    "agreement": 55,
    "agreement_correct": 44,
    "agreement_precision": 0.8,
    "both_direct_and_program": 44,
    "direct_accuracy": 0.552,
    "direct_exact": 138,
    "direct_only": 94,
    "direct_parse_ok": 234,
    "direct_parse_rate": 0.936,
    "final_hidden_exact": 62,
    "final_program_accuracy": 0.248,
    "final_visible_pass": 78,
    "final_visible_pass_rate": 0.312,
    "hybrid_accuracy": 0.568,
    "hybrid_exact": 142,
    "initial_accuracy": 0.16,
    "initial_hidden_exact": 40,
    "mean_round_count": 3.22,
    "n": 250,
    "neither": 94,
    "oracle_union": 156,
    "oracle_union_rate": 0.624,
    "program_only": 18,
    "repair_added_program_correct": 22,
    "repair_lost_program_correct": 0,
    "visible_false_pass": 16,
    "visible_false_pass_rate_among_visible": 0.20512820512820512
  },
  "prompt_iteration": [
    {
      "direct_accuracy": 1.0,
      "final_program_accuracy": 0.8333333333333334,
      "initial_accuracy": 0.8333333333333334,
      "n": 6,
      "name": "prefix6 r2",
      "oracle_union_rate": 1.0,
      "program_only": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    {
      "direct_accuracy": 0.44,
      "final_program_accuracy": 0.08,
      "initial_accuracy": 0.08,
      "n": 25,
      "name": "standard spread25 r2",
      "oracle_union_rate": 0.44,
      "program_only": 0,
      "visible_false_pass_rate_among_visible": 0.7142857142857143
    },
    {
      "direct_accuracy": 0.44,
      "final_program_accuracy": 0.16,
      "initial_accuracy": 0.08,
      "n": 25,
      "name": "strict spread25 r2",
      "oracle_union_rate": 0.52,
      "program_only": 2,
      "visible_false_pass_rate_among_visible": 0.42857142857142855
    }
  ],
  "records": "/workspace/experiments/qwen35_4b_foofah_program_repair_agent/reports/eval_records.jsonl",
  "round_stats": {
    "0": {
      "attempted": 250,
      "hidden_exact": 41,
      "visible_pass": 50,
      "visible_verified_hidden_exact": 40
    },
    "1": {
      "attempted": 200,
      "hidden_exact": 17,
      "visible_pass": 20,
      "visible_verified_hidden_exact": 16
    },
    "2": {
      "attempted": 180,
      "hidden_exact": 6,
      "visible_pass": 5,
      "visible_verified_hidden_exact": 5
    },
    "3": {
      "attempted": 175,
      "hidden_exact": 1,
      "visible_pass": 3,
      "visible_verified_hidden_exact": 1
    }
  },
  "target_char_buckets": {
    "201-500": {
      "agreement": 22,
      "agreement_correct": 20,
      "agreement_precision": 0.9090909090909091,
      "both_direct_and_program": 20,
      "direct_accuracy": 0.6309523809523809,
      "direct_exact": 53,
      "direct_only": 33,
      "direct_parse_ok": 84,
      "direct_parse_rate": 1.0,
      "final_hidden_exact": 25,
      "final_program_accuracy": 0.2976190476190476,
      "final_visible_pass": 29,
      "final_visible_pass_rate": 0.34523809523809523,
      "hybrid_accuracy": 0.6309523809523809,
      "hybrid_exact": 53,
      "initial_accuracy": 0.23809523809523808,
      "initial_hidden_exact": 20,
      "mean_round_count": 3.0595238095238093,
      "n": 84,
      "neither": 26,
      "oracle_union": 58,
      "oracle_union_rate": 0.6904761904761905,
      "program_only": 5,
      "repair_added_program_correct": 5,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 4,
      "visible_false_pass_rate_among_visible": 0.13793103448275862
    },
    "501-1000": {
      "agreement": 1,
      "agreement_correct": 1,
      "agreement_precision": 1.0,
      "both_direct_and_program": 1,
      "direct_accuracy": 0.7777777777777778,
      "direct_exact": 14,
      "direct_only": 13,
      "direct_parse_ok": 18,
      "direct_parse_rate": 1.0,
      "final_hidden_exact": 2,
      "final_program_accuracy": 0.1111111111111111,
      "final_visible_pass": 2,
      "final_visible_pass_rate": 0.1111111111111111,
      "hybrid_accuracy": 0.7777777777777778,
      "hybrid_exact": 14,
      "initial_accuracy": 0.0,
      "initial_hidden_exact": 0,
      "mean_round_count": 3.7777777777777777,
      "n": 18,
      "neither": 3,
      "oracle_union": 15,
      "oracle_union_rate": 0.8333333333333334,
      "program_only": 1,
      "repair_added_program_correct": 2,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 0,
      "visible_false_pass_rate_among_visible": 0.0
    },
    "<=200": {
      "agreement": 31,
      "agreement_correct": 22,
      "agreement_precision": 0.7096774193548387,
      "both_direct_and_program": 22,
      "direct_accuracy": 0.5190839694656488,
      "direct_exact": 68,
      "direct_only": 46,
      "direct_parse_ok": 125,
      "direct_parse_rate": 0.9541984732824428,
      "final_hidden_exact": 29,
      "final_program_accuracy": 0.22137404580152673,
      "final_visible_pass": 40,
      "final_visible_pass_rate": 0.3053435114503817,
      "hybrid_accuracy": 0.5190839694656488,
      "hybrid_exact": 68,
      "initial_accuracy": 0.1297709923664122,
      "initial_hidden_exact": 17,
      "mean_round_count": 3.2748091603053435,
      "n": 131,
      "neither": 56,
      "oracle_union": 75,
      "oracle_union_rate": 0.5725190839694656,
      "program_only": 7,
      "repair_added_program_correct": 12,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 11,
      "visible_false_pass_rate_among_visible": 0.275
    },
    ">1000": {
      "agreement": 1,
      "agreement_correct": 1,
      "agreement_precision": 1.0,
      "both_direct_and_program": 1,
      "direct_accuracy": 0.17647058823529413,
      "direct_exact": 3,
      "direct_only": 2,
      "direct_parse_ok": 7,
      "direct_parse_rate": 0.4117647058823529,
      "final_hidden_exact": 6,
      "final_program_accuracy": 0.35294117647058826,
      "final_visible_pass": 7,
      "final_visible_pass_rate": 0.4117647058823529,
      "hybrid_accuracy": 0.4117647058823529,
      "hybrid_exact": 7,
      "initial_accuracy": 0.17647058823529413,
      "initial_hidden_exact": 3,
      "mean_round_count": 3.0,
      "n": 17,
      "neither": 9,
      "oracle_union": 8,
      "oracle_union_rate": 0.47058823529411764,
      "program_only": 5,
      "repair_added_program_correct": 3,
      "repair_lost_program_correct": 0,
      "visible_false_pass": 1,
      "visible_false_pass_rate_among_visible": 0.14285714285714285
    }
  }
}