{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:24:55.869931+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "tests.test_analysis.test_percent_with_character_count_combined",
      "tests.test_classification.test_default_xml_marker_detection",
      "tests.test_core_counting.test_multilanguage_aggregation_accuracy",
      "tests.test_core_counting.test_deeply_nested_directory_5_levels",
      "tests.test_classification.test_custom_marker_not_found",
      "tests.test_classification.test_no_gen_excludes_only_generated",
      "tests.test_analysis.test_no_duplicates_short_flag",
      "tests.test_analysis.test_character_count_long_lines",
      "tests.test_analysis.test_complexity_default_enabled",
      "tests.test_core_counting.test_directory_with_minified_files",
      "tests.test_analysis.test_dryness_short_flag",
      "tests.test_classification.test_case_insensitive_marker_detection_upper",
      "tests.test_core_counting.test_file_no_comments",
      "tests.test_core_counting.test_large_java_file_with_javadoc",
      "tests.test_analysis.test_sort_by_complexity",
      "tests.test_classification.test_generated_default_no_markers",
      "tests.test_core_counting.test_nested_directories_mixed_structure",
      "tests.test_classification.test_min_gen_line_length_threshold_100",
      "tests.test_analysis.test_no_duplicates_with_uloc",
      "tests.test_analysis.test_uloc_small_duplicate_set",
      "tests.test_classification.test_case_insensitive_marker_detection_lower",
      "tests.test_analysis.test_no_duplicates_long_flag",
      "tests.test_analysis.test_complexity_disabled_with_no_complexity_flag",
      "tests.test_core_counting.test_go_file_with_comments",
      "tests.test_analysis.test_dryness_long_flag",
      "tests.test_classification.test_generated_with_gen_flag_adds_markers",
      "tests.test_core_counting.test_inline_comments_counted_as_code",
      "tests.test_core_counting.test_two_line_file",
      "tests.test_core_counting.test_multiple_files_same_language",
      "tests.test_classification.test_generated_csv_format_includes_marker",
      "tests.test_core_counting.test_ambiguous_m_extension_detection",
      "tests.test_analysis.test_sort_by_blanks",
      "tests.test_analysis.test_dryness_small_duplicate_set",
      "tests.test_classification.test_min_gen_line_length_threshold_200",
      "tests.test_analysis.test_character_count_combined_with_percent",
      "tests.test_analysis.test_complexity_values_simple_code",
      "tests.test_classification.test_marker_in_middle_of_line",
      "tests.test_core_counting.test_by_file_flag_shows_individual_files",
      "tests.test_core_counting.test_complex_docstring_and_quote_handling",
      "tests.test_analysis.test_complexity_disabled_with_short_flag",
      "tests.test_core_counting.test_c_file_with_header_comments",
      "tests.test_core_counting.test_escaped_strings_in_code",
      "tests.test_classification.test_minified_wide_format_includes_marker",
      "tests.test_classification.test_min_gen_z_flag_detects_both",
      "tests.test_analysis.test_percent_short_flag",
      "tests.test_core_counting.test_file_with_mixed_content",
      "tests.test_core_counting.test_multilanguage_directory_comprehensive",
      "tests.test_analysis.test_sort_by_comments",
      "tests.test_classification.test_marker_beyond_1000_bytes_not_detected",
      "tests.test_classification.test_mixed_directory_with_z_flag",
      "tests.test_analysis.test_complexity_values_complex_code",
      "tests.test_core_counting.test_javascript_with_inline_comments",
      "tests.test_analysis.test_empty_file_percent_shows_nan",
      "tests.test_core_counting.test_python_with_comments",
      "tests.test_analysis.test_wide_with_no_complexity_still_shows_complexity",
      "tests.test_analysis.test_percent_long_flag",
      "tests.test_core_counting.test_minified_javascript_long_lines",
      "tests.test_analysis.test_character_count_short_flag",
      "tests.test_core_counting.test_no_complexity_flag_removes_complexity_column",
      "tests.test_classification.test_minified_by_file_format_includes_marker",
      "tests.test_cost.test_cocomo_project_type_organic",
      "tests.test_classification.test_custom_generated_marker_detection",
      "tests.test_core_counting.test_complexity_calculation_cpp",
      "tests.test_analysis.test_sort_by_files_default",
      "tests.test_classification.test_mixed_directory_no_min_gen_only_normal",
      "tests.test_classification.test_marker_within_1000_bytes_detected",
      "tests.test_core_counting.test_directory_with_multiple_languages",
      "tests.test_classification.test_minified_default_no_markers",
      "tests.test_analysis.test_character_count_long_flag",
      "tests.test_analysis.test_by_file_with_percent",
      "tests.test_analysis.test_uloc_and_dryness_combined",
      "tests.test_analysis.test_sort_by_name",
      "tests.test_core_counting.test_include_ext_filter_single_extension",
      "tests.test_cost.test_cost_comparison_shows_both",
      "tests.test_classification.test_no_min_excludes_only_minified",
      "tests.test_cost.test_cocomo_project_type_custom",
      "tests.test_cost.test_currency_symbol_changes",
      "tests.test_cost.test_large_project_cocomo",
      "tests.test_core_counting.test_complexity_calculation_java_loop",
      "tests.test_analysis.test_by_file_output"
    ],
    "failed_tests_omitted": 218,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 476,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "boyter__scc.515f91c",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "5961344",
    "calls": "87",
    "created_at": "2026-05-17T10:05:07.696213+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "5.8175",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "6219452",
    "instance_id": "boyter__scc.515f91c",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "176",
    "n_system_errors": "0",
    "n_tests": "464",
    "n_warnings": "0",
    "output_tokens": "51543",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "12950",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-boyter-scc",
    "run_version": "",
    "score": "0.3793103448275862",
    "session_ended_at": "2026-05-17T10:18:33.681Z",
    "session_started_at": "2026-05-17T10:05:10.401Z",
    "test_branch_errors": "{}",
    "total_tokens": "6270995",
    "wall_clock_seconds": "803"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./scc.py"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-boyter-scc",
  "run_version": "",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}