{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:24:53.929442+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "eval.tests.test_error_handling.test_max_password_length_zero_error",
      "eval.tests.test_error_handling.test_starting_password_too_long_error",
      "eval.tests.test_error_handling.test_starting_password_with_dictionary_error",
      "eval.tests.test_error_handling.test_starting_password_too_short_error",
      "eval.tests.test_error_handling.test_starting_password_out_of_charset_error",
      "eval.tests.test_password_length.test_max_password_len_zero",
      "eval.tests.test_starting_password.test_starting_password_outside_charset",
      "eval.tests.test_starting_password.test_starting_password_with_dictionary",
      "eval.tests.test_starting_password.test_starting_password_too_short",
      "eval.tests.test_starting_password.test_starting_password_too_long",
      "tests.test_bruteforce.test_max_password_len_zero",
      "tests.test_bruteforce.test_starting_password_too_long",
      "tests.test_cli.test_maxpasswordlen_zero",
      "tests.test_cli.test_inputfile_directory",
      "tests.test_cli.test_startingpassword_at_minlen_boundary",
      "tests.test_dictionary.test_corrupted_zip_file",
      "tests.test_harvest.test_startingPassword_with_dictionary",
      "tests.test_harvest.test_startingPassword_length_exceeds_maxPasswordLen",
      "tests.test_harvest.test_maxPasswordLen_zero",
      "tests.test_performance.test_resume_from_invalid_starting_password_wrong_charset",
      "tests.test_performance.test_resume_from_invalid_starting_password_wrong_length",
      "tests.test_cli.test_inputfile_not_a_zip",
      "tests.test_cli.test_startingpassword_with_dictionary",
      "tests.test_cli.test_startingpassword_at_maxlen_boundary",
      "tests.test_performance.test_starting_password_with_dictionary_error",
      "tests.test_bruteforce.test_starting_password_with_dictionary_error",
      "tests.test_harvest.test_startingPassword_out_of_charset",
      "tests.test_zip_files.test_truncated_zip_file",
      "tests.test_zip_files.test_starting_password_with_dictionary",
      "tests.test_zip_files.test_max_password_length_zero",
      "tests.test_performance.test_empty_charset_file_error",
      "tests.test_zip_files.test_corrupted_zip_fake_extension",
      "tests.test_bruteforce.test_starting_password_out_of_charset",
      "tests.test_bruteforce.test_starting_password_too_short",
      "tests.test_cli.test_startingpassword_out_of_charset",
      "tests.test_harvest.test_startingPassword_length_below_minPasswordLen",
      "tests.test_cli.test_startingpassword_exceeds_maxlen",
      "eval.tests.test_argument_parsing.TestPasswordDictionaryArgument.test_dictionary_is_directory",
      "eval.tests.test_argument_parsing.TestInputFileArgument.test_input_file_is_directory",
      "eval.tests.test_argument_parsing.TestCharsetFileArgument.test_charset_file_is_directory",
      "eval.tests.test_argument_parsing.TestPasswordLengthArguments.test_max_password_len_zero",
      "eval.tests.test_argument_parsing.TestStartingPasswordArgument.test_starting_password_too_long",
      "eval.tests.test_argument_parsing.TestStartingPasswordArgument.test_starting_password_with_dictionary",
      "eval.tests.test_argument_parsing.TestStartingPasswordArgument.test_starting_password_too_short",
      "eval.tests.test_argument_parsing.TestStartingPasswordArgument.test_starting_password_out_of_charset",
      "eval.tests.test_argparse_validation.test_integer_option_rejects_overflow",
      "eval.tests.test_argparse_validation.test_inputfile_accepts_space_and_equals_forms[args3]",
      "eval.tests.test_argparse_validation.test_single_dash_value_is_treated_as_path_and_rejected_by_validation",
      "eval.tests.test_argparse_validation.test_double_dash_separator_is_not_supported_without_positionals",
      "eval.tests.test_argparse_validation.test_duplicate_charset_flag_is_rejected_by_parser",
      "eval.tests.test_zip_password_finder.TestPasswordLength.test_min_password_length_constraint",
      "eval.tests.test_zip_password_finder.TestDictionaryAttack.test_dictionary_password_not_found"
    ],
    "failed_tests_omitted": 0,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 792,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "agourlay__zip-password-finder.704700d",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "1226112",
    "calls": "37",
    "created_at": "2026-05-17T09:47:12.134208+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "1.8627",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "1320621",
    "instance_id": "agourlay__zip-password-finder.704700d",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "629",
    "n_system_errors": "0",
    "n_tests": "680",
    "n_warnings": "0",
    "output_tokens": "25903",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "9615",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-agourlay-zip-password-finder",
    "run_version": "",
    "score": "0.925",
    "session_ended_at": "2026-05-17T10:00:17.438Z",
    "session_started_at": "2026-05-17T09:47:16.624Z",
    "test_branch_errors": "{}",
    "total_tokens": "1346524",
    "wall_clock_seconds": "780"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./zip_password_finder.py"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-agourlay-zip-password-finder",
  "run_version": "",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}