{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:24:57.477173+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "eval.tests.test_extraction_features.test_aws_long_flag",
      "tests.test_extraction_basic.test_no_extraction_flags_shows_first_match_all_types",
      "tests.test_extraction_extended.test_google_keys_from_file",
      "tests.test_extraction_extended.test_phone_numbers_from_file",
      "tests.test_extraction_basic.test_email_in_log_context_suppress_stdin_msg",
      "tests.test_extraction_basic.test_email_tld_length_restrictions",
      "tests.test_extraction_basic.test_email_from_file_input",
      "tests.test_extraction_extended.test_aws_keys_from_file",
      "tests.test_extraction_extended.test_aws_with_display_filename",
      "tests.test_extraction_basic.test_multiple_extractors_stops_at_first_without_thorough",
      "tests.test_extraction_extended.test_credit_card_with_drop_filter",
      "tests.test_filter_drop.test_invalid_regex_filter_panics",
      "tests.test_filter_drop.test_invalid_regex_drop_panics",
      "tests.test_gap_filling.test_csv_header_hide_display_flags",
      "tests.test_gap_filling.test_csv_header_display_line_flags",
      "tests.test_gap_filling.test_csv_multiple_extractors_with_display_line",
      "tests.test_gap_filling.test_csv_data_with_commas_in_output",
      "tests.test_extraction_extended.test_phone_numbers_clean_mode",
      "tests.test_extraction_extended.test_aws_keys_multiple_prefixes",
      "tests.test_extraction_extended.test_dns_srv_records_from_file",
      "tests.test_output_destinations.test_csv_header_generation_basic",
      "tests.test_filter_drop.test_filter_url_domain_clean_mode",
      "tests.test_output_destinations.test_csv_header_with_line_flag",
      "tests.test_gap_filling.test_csv_clean_mode_extracts_only_matches",
      "tests.test_output_destinations.test_csv_automatic_detection_from_extension",
      "tests.test_output_destinations.test_large_output_many_matches",
      "tests.test_output_destinations.test_csv_output_with_clean_flag",
      "tests.test_gap_filling.test_csv_header_only_line_numbers",
      "tests.test_gap_filling.test_csv_header_hide_line_flags",
      "tests.test_output_destinations.test_stdin_input_with_csv_output",
      "tests.test_gap_filling.test_write_error_disk_full_simulation",
      "tests.test_gap_filling.test_csv_with_hide_and_multiple_extractors",
      "tests.test_output_modifiers.test_thorough_mode_deduplicates_same_match",
      "tests.test_filter_drop.test_drop_http_urls_and_specific_domain",
      "tests.test_output_destinations.test_csv_multiple_extraction_types",
      "tests.test_output_destinations.test_csv_header_with_display_flag",
      "tests.test_plugins_basic.test_system_path_plugin_file_loads",
      "tests.test_plugins_basic.test_plugin_extraction_with_clean_flag",
      "tests.test_plugins_basic.test_local_plugin_file_loads_and_lists",
      "tests.test_output_destinations.test_csv_parsing_with_python_csv_module",
      "tests.test_plugins_basic.test_plugin_extraction_full_line",
      "tests.test_output_destinations.test_output_error_suppressed_with_ignore_flag",
      "tests.test_output_destinations.test_csv_multiple_dot_extensions",
      "tests.test_output_destinations.test_csv_append_behavior",
      "tests.test_gap_filling.test_csv_header_only_display",
      "tests.test_output_destinations.test_no_matches_csv_output",
      "tests.test_output_destinations.test_url_extraction_to_csv",
      "tests.test_gap_filling.test_directory_with_csv_output",
      "tests.test_plugins_basic.test_list_multiple_plugins_formatting",
      "tests.test_output_destinations.test_csv_header_all_flags_combined",
      "tests.test_output_destinations.test_csv_header_with_hide_type_flag",
      "tests.test_plugins_basic.test_local_plugin_used_when_system_path_missing",
      "tests.test_plugins_basic.test_plugin_with_invalid_regex_no_crash",
      "tests.test_plugins_basic.test_plugin_with_complex_regex_hex_extraction",
      "tests.test_plugins_basic.test_plugin_extraction_hide_prefix",
      "tests.test_plugins_basic.test_plugin_with_line_numbers",
      "tests.test_plugins_basic.test_multiple_plugins_simultaneous_extraction",
      "tests.test_plugins_basic.test_plugin_no_matches_empty_output",
      "tests.test_output_destinations.test_csv_with_thorough_flag",
      "tests.test_gap_filling.test_csv_header_only_hide_type",
      "tests.test_plugins_basic.test_plugin_output_to_file",
      "tests.test_output_destinations.test_hash_extraction_to_csv",
      "tests.test_plugins_errors.test_missing_required_fields_in_json",
      "tests.test_plugins_errors.test_malformed_json_continues_when_user_declines_exit",
      "tests.test_plugins_errors.test_plugin_extraction_with_malformed_json_shows_error",
      "tests.test_plugins_basic.test_plugin_extraction_large_input_multiple_matches_per_line",
      "tests.test_plugins_basic.test_plugin_system_path_priority_over_local",
      "tests.test_plugins_basic.test_plugin_content_type_used_as_prefix",
      "tests.test_plugins_basic.test_plugin_with_display_flag_shows_filename",
      "tests.test_plugins_basic.test_plugin_with_thorough_flag_gets_all_matches",
      "tests.test_plugins_basic.test_plugin_extraction_with_stdin_input",
      "tests.test_plugins_basic.test_plugin_with_filter_flag",
      "tests.test_plugins_basic.test_plugin_with_drop_flag",
      "tests.test_plugins_errors.test_utf8_bom_in_plugins_json_causes_parse_error",
      "tests.test_plugins_basic.test_plugin_version_displayed_correctly",
      "tests.test_output_destinations.test_empty_file_csv_output",
      "tests.test_gap_filling.test_csv_header_no_flags_default",
      "tests.test_plugins_errors.test_json_object_instead_of_array",
      "tests.test_plugins_errors.test_empty_json_file_is_valid",
      "tests.test_plugins_errors.test_malformed_json_syntax_error_with_exit"
    ],
    "failed_tests_omitted": 74,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 564,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "drew-alleman__datasurgeon.d257cee",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "4965504",
    "calls": "79",
    "created_at": "2026-05-17T22:53:39.714545+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "4.6128",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "5096850",
    "instance_id": "drew-alleman__datasurgeon.d257cee",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "366",
    "n_system_errors": "0",
    "n_tests": "502",
    "n_warnings": "0",
    "output_tokens": "49112",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "16565",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-drew-alleman-datasurgeon-attempt-3",
    "run_version": "20260517T094700Z",
    "score": "0.7290836653386454",
    "session_ended_at": "2026-05-17T23:06:30.427Z",
    "session_started_at": "2026-05-17T22:53:42.944Z",
    "test_branch_errors": "{}",
    "total_tokens": "5145962",
    "wall_clock_seconds": "767"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./datasurgeon.py",
      "./executable"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-drew-alleman-datasurgeon-attempt-3",
  "run_version": "20260517T094700Z",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}