{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:24:59.268823+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "eval.tests.test_error_handling.test_corrupted_obj_file",
      "eval.tests.test_error_handling.test_invalid_file_extension_error",
      "eval.tests.test_input_files.test_unsupported_extension",
      "eval.tests.test_rotation_flags.test_pitch_rotation_short",
      "eval.tests.test_rotation_flags.test_negative_rotation_long_flag",
      "eval.tests.test_input_handling.test_invalid_extension",
      "tests.test_file_formats.test_rotation_y_axis",
      "tests.test_file_formats.test_no_color_flag",
      "tests.test_coverage_push.test_obj_with_vertex_colors",
      "tests.test_coverage_push.test_malformed_stl_parse_error",
      "tests.test_file_formats.test_multiple_rotation_axes",
      "tests.test_file_formats.test_stl_complex_part",
      "tests.test_coverage_push.test_malformed_obj_parse_error",
      "tests.test_file_formats.test_invalid_extension_error",
      "tests.test_file_formats.test_no_extension_file",
      "tests.test_coverage_push.test_stl_file_coverage",
      "tests.test_gaps_context_geometry.test_extreme_aspect_ratio_100x1_cube",
      "tests.test_file_formats.test_empty_stl_file",
      "tests.test_file_formats.test_stl_simple_cube",
      "tests.test_file_formats.test_obj_simple_cube",
      "tests.test_coverage_push.test_multiple_stl_obj_mixed",
      "tests.test_file_formats.test_custom_height",
      "tests.test_file_formats.test_empty_obj_file",
      "tests.test_gaps_context_geometry.test_degenerate_triangle_colinear_vertices",
      "tests.test_coverage_push.test_extreme_rotation_values",
      "tests.test_image_basic.test_tiny_dimensions_2x2",
      "tests.test_gaps_context_geometry.test_rotation_parameters_transform_application",
      "tests.test_gaps_context_geometry.test_webify_mode_shorthand_flag",
      "tests.test_gaps_context_geometry.test_extreme_aspect_ratio_1x100_cube",
      "tests.test_image_basic.test_bw_mode_width_only",
      "tests.test_image_basic.test_very_wide_aspect_ratio",
      "tests.test_file_formats.test_obj_complex_ferris",
      "tests.test_image_basic.test_multiple_input_files",
      "tests.test_image_basic.test_stl_format_support",
      "tests.test_gaps_context_geometry.test_model_with_materials_color_rendering",
      "tests.test_image_rotation.test_rotation_2pi_full_circle",
      "tests.test_gaps_context_geometry.test_overlapping_triangles_z_buffer",
      "tests.test_image_basic.test_part_stl_with_bw_mode",
      "tests.test_image_rotation.test_negative_rotation_value",
      "tests.test_gaps_context_geometry.test_multiple_input_meshes_combined_bounding_box",
      "tests.test_gaps_context_geometry.test_stl_format_mesh_conversion",
      "tests.test_image_basic.test_very_tall_aspect_ratio",
      "tests.test_image_basic.test_suzy_model_medium_complexity",
      "tests.test_image_rotation.test_single_axis_x_rotation_pi_half",
      "tests.test_image_basic.test_width_only_defaults_height",
      "tests.test_image_rotation.test_single_axis_z_rotation_pi_half",
      "tests.test_image_rotation.test_long_form_pitch_flag",
      "tests.test_gaps_context_geometry.test_large_dimension_buffer_allocation",
      "tests.test_image_basic.test_odd_dimensions",
      "tests.test_gaps_context_geometry.test_webify_mode_html_output",
      "tests.test_gaps_context_geometry.test_complex_model_large_mesh",
      "tests.test_image_rotation.test_combined_xyz_rotation",
      "tests.test_image_rotation.test_combined_xy_rotation",
      "tests.test_image_rotation.test_rotation_zero_value",
      "tests.test_image_rotation.test_very_small_rotation_value",
      "tests.test_image_basic.test_skull_model_larger_dimensions",
      "tests.test_image_basic.test_ferris_bw_medium_size",
      "tests.test_image_rotation.test_long_form_yaw_flag",
      "tests.test_image_rotation.test_single_axis_y_rotation_pi_half",
      "tests.test_image_basic.test_cube_basic_color_render",
      "tests.test_image_rotation.test_rotation_with_no_color_flag",
      "tests.test_image_rotation.test_missing_rotation_value",
      "tests.test_image_rotation.test_rotation_with_explicit_zero",
      "tests.test_image_rotation.test_stl_format_with_rotation",
      "tests.test_image_rotation.test_long_form_roll_flag",
      "tests.test_image_webify.test_webify_requires_width_flag",
      "tests.test_image_rotation.test_rotation_changes_output_from_baseline",
      "tests.test_image_rotation.test_rotation_with_complex_model",
      "tests.test_materials_colors.test_stl_file_renders_with_default_yellow_color",
      "tests.test_image_rotation.test_x_rotation_pi",
      "tests.test_image_webify.test_webify_small_dimensions_multiple_frames",
      "tests.test_image_basic.test_pikachu_high_poly_model",
      "tests.test_image_webify.test_webify_basic_single_frame",
      "tests.test_materials_colors.test_cube_obj_loads_materials_renders_colors",
      "tests.test_image_webify.test_webify_with_no_color_flag",
      "tests.test_materials_colors.test_ferris_multi_material_renders_distinct_colors",
      "tests.test_image_webify.test_webify_differs_from_normal_image_mode",
      "tests.test_materials_colors.test_obj_with_no_mtllib_directive_uses_default_white",
      "tests.test_image_webify.test_webify_frame_count_zero_generates_one_frame",
      "tests.test_image_webify.test_webify_missing_frame_count"
    ],
    "failed_tests_omitted": 61,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 454,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "ecumene__rust-sloth.051c559",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "6051072",
    "calls": "82",
    "created_at": "2026-05-17T10:46:28.011872+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "5.9332",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "6267930",
    "instance_id": "ecumene__rust-sloth.051c559",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "245",
    "n_system_errors": "0",
    "n_tests": "380",
    "n_warnings": "0",
    "output_tokens": "60778",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "27420",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-ecumene-rust-sloth",
    "run_version": "",
    "score": "0.6447368421052632",
    "session_ended_at": "2026-05-17T11:04:21.912Z",
    "session_started_at": "2026-05-17T10:46:30.877Z",
    "test_branch_errors": "{}",
    "total_tokens": "6328708",
    "wall_clock_seconds": "1071"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./sloth.py"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-ecumene-rust-sloth",
  "run_version": "",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}