{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:25:16.156763+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "eval.tests.test_basic_invocation.test_args_flag",
      "eval.tests.test_configuration.test_invalid_config_file",
      "eval.tests.test_utility.test_zsh_completion",
      "eval.tests.test_utility.test_powershell_completion",
      "eval.tests.test_utility.test_elvish_completion",
      "tests.test_cli.test_args_flag_displays_parsed_arguments",
      "tests.test_cli.test_check_mode_accepts_correct_formatting",
      "tests.test_cli.test_check_mode_detects_incorrect_formatting",
      "tests.test_cli.test_completion_bash_script_syntax",
      "tests.test_cli.test_completion_elvish_script_syntax",
      "tests.test_cli.test_completion_fish_script_syntax",
      "tests.test_cli.test_completion_powershell_script_syntax",
      "tests.test_cli.test_completion_zsh_script_syntax",
      "tests.test_cli.test_fail_on_change_mode_returns_nonzero_when_modified",
      "tests.test_cli.test_man_page_generation_troff_format",
      "tests.test_cli.test_nonexistent_file_error",
      "tests.test_cli.test_nowrap_disables_line_wrapping",
      "tests.test_cli.test_print_mode_does_not_modify_file",
      "tests.test_cli.test_print_mode_formats_to_stdout",
      "tests.test_cli.test_quiet_flag_suppresses_warnings",
      "tests.test_cli.test_stdin_mode_formats_from_stdin_to_stdout",
      "tests.test_cli.test_stdin_rejects_file_arguments",
      "tests.test_cli.test_tabsize_4_uses_4_space_indentation",
      "tests.test_cli.test_tabsize_out_of_range_error",
      "tests.test_cli.test_trace_flag_shows_trace_messages",
      "tests.test_cli.test_trace_overrides_quiet_flag",
      "tests.test_cli.test_unknown_flag_returns_error",
      "tests.test_cli.test_usetabs_indents_with_tabs",
      "tests.test_cli.test_verbose_flag_shows_info_messages",
      "tests.test_cli.test_verbose_overrides_quiet_flag",
      "tests.test_cli.test_version_flag",
      "tests.test_cli.test_wraplen_edge_case_1000_chars",
      "tests.test_cli.test_wraplen_edge_case_120_chars",
      "tests.test_cli.test_wraplen_edge_case_40_chars",
      "tests.test_comments_comprehensive.test_comment_after_backslash_space",
      "tests.test_comments_comprehensive.test_comment_indentation_normalized",
      "tests.test_comments_comprehensive.test_comment_not_wrapped_with_nowrap_flag",
      "tests.test_comments_comprehensive.test_comment_with_no_spaces_cannot_wrap",
      "tests.test_comments_comprehensive.test_comment_wrapping_respects_custom_wraplen",
      "tests.test_comments_comprehensive.test_comments_at_file_boundaries",
      "tests.test_comments_comprehensive.test_comments_between_begin_and_end",
      "tests.test_comments_comprehensive.test_comments_in_environments",
      "tests.test_comments_comprehensive.test_comments_in_math_mode",
      "tests.test_comments_comprehensive.test_comments_in_nested_environments",
      "tests.test_comments_comprehensive.test_comments_with_blank_lines_between",
      "tests.test_comments_comprehensive.test_comments_with_latex_commands",
      "tests.test_comments_comprehensive.test_comments_with_special_characters",
      "tests.test_comments_comprehensive.test_consecutive_comment_lines",
      "tests.test_comments_comprehensive.test_empty_comment_lines_preserved",
      "tests.test_comments_comprehensive.test_escaped_percent_not_treated_as_comment",
      "tests.test_comments_comprehensive.test_file_with_only_comments",
      "tests.test_comments_comprehensive.test_indented_comments_normalized",
      "tests.test_comments_comprehensive.test_inline_comment_wrapped_with_percent_continuation",
      "tests.test_comments_comprehensive.test_inline_comments_preserved",
      "tests.test_comments_comprehensive.test_long_comment_wrapped_with_percent_continuation",
      "tests.test_comments_comprehensive.test_multiple_consecutive_percent_signs",
      "tests.test_comments_comprehensive.test_percent_in_verb_not_treated_as_comment",
      "tests.test_comments_comprehensive.test_standalone_comment_lines_preserved",
      "tests.test_config.test_cli_nowrap_overrides_config_wrap",
      "tests.test_config.test_cli_overrides_config_tabsize",
      "tests.test_config.test_cli_overrides_config_wraplen",
      "tests.test_config.test_cli_usetabs_overrides_config_spaces",
      "tests.test_config.test_cli_verbosity_overrides_config",
      "tests.test_config.test_config_check_flag",
      "tests.test_config.test_config_current_directory",
      "tests.test_config.test_config_fail_on_change",
      "tests.test_config.test_config_file_explicit_path_tabsize",
      "tests.test_config.test_config_file_wrap_disabled",
      "tests.test_config.test_config_file_wraplen",
      "tests.test_config.test_config_multiple_options",
      "tests.test_config.test_config_only_when_cli_absent",
      "tests.test_config.test_config_print_flag",
      "tests.test_config.test_config_tabchar_tab",
      "tests.test_config.test_config_unknown_keys_ignored",
      "tests.test_config.test_config_verbosity_quiet",
      "tests.test_config.test_config_verbosity_verbose",
      "tests.test_config.test_config_with_file_mode",
      "tests.test_config.test_config_with_lists_option",
      "tests.test_config.test_config_with_no_indent_envs",
      "tests.test_config.test_config_with_verbatims_option"
    ],
    "failed_tests_omitted": 199,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {
      "15eca2a4caf9": [
        {
          "error_code": "results_read_failed",
          "error_details": "Error response from daemon: Could not find the file /workspace/eval/results.xml in container a221548435b2b21311bdde1fae978c97188ebb001a1ab1f367bf3a7089fc9e2d"
        }
      ]
    },
    "test_records": 495,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "wgunderwood__tex-fmt.3f1aef6",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "3714304",
    "calls": "82",
    "created_at": "2026-05-17T16:35:25.009710+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "3.6435",
    "evaluator_problem": "True",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "3818063",
    "instance_id": "wgunderwood__tex-fmt.3f1aef6",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "186",
    "n_system_errors": "0",
    "n_tests": "455",
    "n_warnings": "0",
    "output_tokens": "42253",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "15647",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-wgunderwood-tex-fmt",
    "run_version": "20260517T094700Z",
    "score": "0.4087912087912088",
    "session_ended_at": "2026-05-17T16:48:01.511Z",
    "session_started_at": "2026-05-17T16:35:28.049Z",
    "test_branch_errors": "{\"15eca2a4caf9\": [\"results_read_failed\"]}",
    "total_tokens": "3860316",
    "wall_clock_seconds": "753"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./tex_fmt.py"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-wgunderwood-tex-fmt",
  "run_version": "20260517T094700Z",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}