{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:25:12.302764+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "tests.test_absolute_final.test_ultimate_binary_all_outputs",
      "tests.test_actual_scans.test_actual_scan_with_config_file",
      "tests.test_coverage_boost_final.test_echo_pcap_options",
      "tests.test_coverage_final_push.test_sL_mode_cidr",
      "tests.test_advanced.test_ping_option",
      "tests.test_basic_fixed.test_echo_with_exclude_ip",
      "tests.test_coverage_boost.test_exclude_with_range_and_cidr",
      "tests.test_advanced.test_router_mac",
      "tests.test_coverage_boost.test_adapter_with_all_settings",
      "tests.test_coverage_boost_final.test_echo_vlan_configuration",
      "tests.test_comprehensive_final.test_http_all_fields",
      "tests.test_basic.test_echo_with_excludes",
      "tests.test_absolute_final_push.test_benchmark_multiple",
      "tests.test_coverage_final_push.test_http_field_remove",
      "tests.test_basic.test_packet_trace",
      "tests.test_coverage_final_sprint.test_connection_timeout_variations",
      "tests.test_advanced.test_http_method",
      "tests.test_advanced.test_adapter_vlan",
      "tests.test_binary_edge_cases.test_binary_single_record",
      "tests.test_coverage_boost.test_ttl_variations",
      "tests.test_advanced.test_config_file_with_excludes",
      "tests.test_absolute_final.test_ultimate_binary_all_show_options",
      "tests.test_comprehensive_final.test_hello_file_various_ports",
      "tests.test_coverage_boost.test_resume_index_and_count_combination",
      "tests.test_coverage_boost_final.test_echo_resume_configurations",
      "tests.test_actual_scans.test_actual_scan_offline_xml",
      "tests.test_coverage_final_push.test_show_options",
      "tests.test_coverage_final_push.test_ttl_values",
      "tests.test_basic_fixed.test_selftest_stderr",
      "tests.test_coverage_final_sprint.test_exclude_ports",
      "tests.test_coverage_final_sprint.test_ping_sweep",
      "tests.test_coverage_final_sprint.test_http_host_header",
      "tests.test_advanced.test_http_url",
      "tests.test_binary_edge_cases.test_binary_large_banner",
      "tests.test_coverage_boost.test_ping_with_ports",
      "tests.test_advanced.test_exclude_single_ip",
      "tests.test_advanced.test_ttl",
      "tests.test_actual_scans.test_actual_scan_offline_list",
      "tests.test_coverage_final_push.test_noshow_options",
      "tests.test_coverage_boost_final.test_echo_http_configurations",
      "tests.test_coverage_boost.test_http_field_customization",
      "tests.test_absolute_final_push.test_massive_binary_xml",
      "tests.test_comprehensive_final.test_retries_range",
      "tests.test_basic_fixed.test_nmap_help_content",
      "tests.test_coverage_final_sprint.test_resume_index_and_count",
      "tests.test_coverage_final_push.test_connection_timeout_values",
      "tests.test_coverage_final_sprint.test_nostatus_option",
      "tests.test_coverage_final_sprint.test_http_url",
      "tests.test_binary_edge_cases.test_binary_many_records",
      "tests.test_advanced.test_http_host",
      "tests.test_coverage_boost.test_wait_variation",
      "tests.test_advanced.test_exclude_range",
      "tests.test_actual_scans.test_actual_scan_offline_json",
      "tests.test_advanced.test_resume_count",
      "tests.test_coverage_final_push.test_capture_options",
      "tests.test_coverage_boost_final.test_echo_rotate_configurations",
      "tests.test_coverage_boost.test_multiple_http_cookies",
      "tests.test_absolute_final_push.test_massive_binary_json",
      "tests.test_comprehensive_final.test_resume_index_various",
      "tests.test_coverage_final_sprint.test_readscan_nonexistent_file",
      "tests.test_coverage_final_push.test_retries_values",
      "tests.test_coverage_final_sprint.test_http_method",
      "tests.test_coverage_final_sprint_fixed.test_exclude_ports_option",
      "tests.test_binary_edge_cases.test_binary_mixed_record_types",
      "tests.test_advanced.test_http_cookie",
      "tests.test_advanced.test_exclude_multiple",
      "tests.test_coverage_boost_final.test_regress_blackrock",
      "tests.test_actual_scans.test_actual_scan_offline_grepable",
      "tests.test_advanced.test_open_only",
      "tests.test_coverage_boost_final.test_echo_vulnerability_checks",
      "tests.test_coverage_boost.test_http_with_payload",
      "tests.test_coverage_final_push.test_nocapture_options",
      "tests.test_absolute_final_push.test_massive_binary_ndjson",
      "tests.test_comprehensive_final.test_resume_count_various",
      "tests.test_coverage_final_sprint.test_output_format_json_structure",
      "tests.test_coverage_final_sprint.test_http_version",
      "tests.test_coverage_push_50.test_config_http_version",
      "tests.test_binary_edge_cases.test_binary_output_formats_comprehensive",
      "tests.test_advanced.test_http_payload",
      "tests.test_coverage_boost_final.test_selftest_main"
    ],
    "failed_tests_omitted": 1209,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 3357,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "robertdavidgraham__masscan.b99d433",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "5181952",
    "calls": "80",
    "created_at": "2026-05-17T14:35:00.854181+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "4.8561",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "5318738",
    "instance_id": "robertdavidgraham__masscan.b99d433",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "1530",
    "n_system_errors": "0",
    "n_tests": "2549",
    "n_warnings": "0",
    "output_tokens": "52706",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "11643",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-robertdavidgraham-masscan",
    "run_version": "20260517T094700Z",
    "score": "0.6002353864260495",
    "session_ended_at": "2026-05-17T14:48:50.761Z",
    "session_started_at": "2026-05-17T14:35:03.854Z",
    "test_branch_errors": "{}",
    "total_tokens": "5371444",
    "wall_clock_seconds": "826"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./masscan_compat.py"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-robertdavidgraham-masscan",
  "run_version": "20260517T094700Z",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}