{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:25:06.689952+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "eval.tests.test_dupl.test_conflicting_html_plumbing_flags",
      "tests.test_dupl.test_html_and_plumbing_conflict",
      "tests.test_errors.test_non_go_text_file",
      "tests.test_errors.test_malformed_go_file",
      "tests.test_errors.test_binary_file",
      "tests.test_go_constructs.test_parse_error_handling",
      "tests.test_input_modes.test_vendor_directory_skipped_by_default",
      "tests.test_input_modes.test_stdin_file_list_strips_dot_slash",
      "tests.test_go_constructs.test_empty_statements_and_declaration_statements",
      "tests.test_errors.test_conflicting_output_flags",
      "tests.test_input_modes.test_multiple_directories",
      "tests.test_integration.test_ci_pipeline_plumbing_threshold_check",
      "tests.test_integration.test_stdin_test_files_only",
      "tests.test_integration.test_vendor_directory_comparison",
      "tests.test_input_modes.test_html_and_plumbing_flags_mutually_exclusive",
      "tests.test_large_scale.test_multifile_webservice_project_finds_all_clone_groups",
      "tests.test_integration.test_error_recovery_mixed_valid_invalid",
      "tests.test_large_scale.test_large_file_500_plus_lines_with_many_clones",
      "tests.test_large_scale.test_very_permissive_threshold_finds_many_clones",
      "tests.test_large_scale.test_multiple_separate_directories_analyzed_together",
      "tests.test_input_modes.test_vendor_flag_includes_vendor_directories",
      "tests.test_input_modes.test_default_current_directory",
      "tests.test_integration.test_developer_workflow_verbose_then_html",
      "tests.test_integration.test_stdin_workflow_find_filter",
      "tests.test_integration.test_nested_directories_recursive_scan",
      "tests.test_main_gaps.test_stdout_write_error_text_mode",
      "tests.test_large_scale.test_performance_large_project_under_30_seconds",
      "tests.test_large_scale.test_large_threshold_filters_most_clones",
      "tests.test_main_gaps.test_single_clone_group_prints_correctly",
      "tests.test_main_gaps.test_stdout_write_error_html_mode",
      "tests.test_output_formats.test_multiple_directories",
      "tests.test_threshold.test_threshold_long_flag_identical_to_short",
      "tests.test_threshold.test_mixed_size_duplicates_default_threshold",
      "tests.test_threshold.test_realistic_code_with_various_thresholds",
      "tests.test_input_modes.test_explicit_directory",
      "tests.test_threshold.test_progressive_filtering_with_increasing_threshold",
      "tests.test_large_scale.test_verbose_mode_on_large_project_includes_progress",
      "tests.test_main_gaps.test_multiple_files_text_output",
      "tests.test_verbose.test_verbose_with_nonexistent_path",
      "tests.test_verbose.test_verbose_with_conflicting_output_flags",
      "tests.test_output_formats.test_format_conflict_html_and_plumbing",
      "tests.test_main_gaps.test_broken_pipe_sigpipe_text_mode",
      "eval.tests.test_subcommand_dispatch.TestFlagBehavior.test_html_and_plumbing_mutually_exclusive",
      "tests.test_dupl.TestBasicCloneDetection.test_single_file_no_duplicates",
      "eval.tests.test_core_outputs.test_html_output_has_structure_and_escapes",
      "eval.tests.test_core_outputs.test_text_output_single_group_and_summary",
      "eval.tests.test_core_outputs.test_plumbing_output_exact_relative",
      "eval.tests.test_dupl_externalized.test_ext_cli_rejects_html_and_plumbing_together"
    ],
    "failed_tests_omitted": 0,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 450,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "mibk__dupl.1bf052b",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "4792704",
    "calls": "85",
    "created_at": "2026-05-17T12:56:35.240762+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "4.8169",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "4981838",
    "instance_id": "mibk__dupl.1bf052b",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "326",
    "n_system_errors": "0",
    "n_tests": "373",
    "n_warnings": "0",
    "output_tokens": "49163",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "19786",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-mibk-dupl",
    "run_version": "20260517T094700Z",
    "score": "0.8739946380697051",
    "session_ended_at": "2026-05-17T13:10:27.678Z",
    "session_started_at": "2026-05-17T12:56:38.275Z",
    "test_branch_errors": "{}",
    "total_tokens": "5031001",
    "wall_clock_seconds": "829"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./dupl.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-mibk-dupl",
  "run_version": "20260517T094700Z",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}