{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:24:59.698907+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "tests.test_card_parsing.test_frontmatter_custom_deck_name",
      "tests.test_error_handling.test_unclosed_frontmatter",
      "tests.test_error_handling.test_malformed_frontmatter",
      "tests.test_error_handling.test_stats_with_broken_symlink",
      "tests.test_export_command.test_export_with_frontmatter",
      "tests.test_media_handling.test_image_with_title",
      "eval.tests.test_card_parsing.test_parse_frontmatter_name",
      "eval.tests.test_cli_edge_cases.test_stats_html_format_not_implemented",
      "eval.tests.test_database_coverage.test_orphans_with_no_database",
      "eval.tests.test_database_direct.test_export_multiple_sessions",
      "eval.tests.test_collection_operations.test_collection_initialization",
      "eval.tests.test_database_direct.test_export_with_reviewed_cards",
      "eval.tests.test_database_coverage.test_check_creates_database_if_missing",
      "eval.tests.test_edge_cases.test_malformed_frontmatter",
      "eval.tests.test_error_handling.test_invalid_toml_frontmatter",
      "eval.tests.test_export_command.test_export_with_frontmatter",
      "eval.tests.test_database_coverage.test_database_schema_initialization",
      "eval.tests.test_database_direct.test_export_file_output_with_performance",
      "eval.tests.test_database_coverage.test_database_file_location",
      "eval.tests.test_database_direct.test_export_with_new_and_reviewed_cards",
      "eval.tests.test_comprehensive_scenarios.test_collection_with_frontmatter_variations",
      "eval.tests.test_database_coverage.test_database_persists_across_check_runs",
      "eval.tests.test_stats_command.test_stats_html_format",
      "eval.tests.test_stats_command.test_stats_default_format",
      "eval.tests.test_database_direct.test_stats_with_reviewed_cards",
      "tests.test_db_gaps.test_stats_html_format_not_implemented",
      "tests.test_database.test_stats_html_format",
      "tests.test_card_rendering.test_basic_card_with_inline_markdown_mix",
      "tests.test_check.test_check_separator_in_question",
      "tests.test_card_rendering.test_basic_card_question_with_bold_markdown",
      "tests.test_collection.test_collection_macros_without_space",
      "tests.test_database.test_database_file_permissions",
      "tests.test_check.test_check_invalid_toml_frontmatter",
      "tests.test_database.test_database_with_unicode_in_directory_name",
      "tests.test_collection.test_collection_creates_database_file",
      "tests.test_database.test_database_location_in_collection_directory",
      "tests.test_database.test_database_created_on_first_command",
      "tests.test_drill_internals.test_post_action_forgot_causes_card_repeat",
      "tests.test_database.test_export_empty_output_file_not_created_on_error",
      "tests.test_db_gaps.test_database_created_with_correct_permissions",
      "tests.test_drill_internals.test_katex_css_endpoint",
      "tests.test_db_gaps.test_database_isolation_between_collections",
      "tests.test_database.test_stats_default_format_is_html",
      "tests.test_check.test_check_unclosed_frontmatter",
      "tests.test_card_rendering.test_cloze_card_with_empty_surrounding_text",
      "tests.test_card_rendering.test_cloze_card_front_shows_dots",
      "tests.test_collection.test_collection_reuses_existing_database",
      "tests.test_collection.test_export_collection_custom_deck_name",
      "tests.test_database.test_export_output_to_directory_instead_of_file",
      "tests.test_db_gaps2.test_export_handles_partial_performance_data",
      "tests.test_database.test_database_schema_tables_created",
      "tests.test_errors.test_error_message_format_consistency",
      "tests.test_drill_internals.test_latex_math_content_rendered_in_html",
      "tests.test_drill_internals.test_post_action_hard_causes_card_repeat",
      "tests.test_db_gaps2.test_export_retrieves_card_performance_data",
      "tests.test_drill_internals.test_katex_font_cache_headers",
      "tests.test_database.test_multiple_collections_separate_databases",
      "tests.test_check.test_check_cloze_after_question",
      "tests.test_card_rendering.test_basic_card_with_code_block",
      "tests.test_errors.test_parser_error_invalid_toml",
      "tests.test_db_gaps2.test_export_date_formats",
      "tests.test_db_gaps2.test_export_includes_review_sessions_v2",
      "tests.test_card_rendering.test_cloze_card_back_reveals_deletion",
      "tests.test_fsrs.test_initial_stability_good_grade",
      "tests.test_database.test_database_persists_across_commands",
      "tests.test_drill_http.test_session_completion",
      "tests.test_fsrs.test_interval_clamped_to_maximum",
      "tests.test_errors.test_database_error_includes_details",
      "tests.test_drill_internals.test_completion_page_structure",
      "tests.test_drill_internals.test_post_action_undo_reverts_to_previous_card",
      "tests.test_harvest.test_parser_cloze_with_image",
      "tests.test_errors.test_parser_error_unclosed_frontmatter",
      "tests.test_db_gaps2.test_export_session_with_no_reviews",
      "tests.test_database.test_database_survives_deck_parsing_errors",
      "tests.test_db_gaps2.test_export_multiple_sessions_chronological_order",
      "tests.test_fsrs.test_initial_stability_forgot_grade",
      "tests.test_drill_internals.test_katex_js_endpoint",
      "tests.test_errors.test_database_error_corrupted",
      "tests.test_card_rendering.test_cloze_with_markdown_in_deletion",
      "tests.test_fsrs.test_initial_stability_easy_grade"
    ],
    "failed_tests_omitted": 141,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 1293,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "eudoxia0__hashcards.48aa136",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "7586304",
    "calls": "128",
    "created_at": "2026-05-17T11:01:25.673264+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "7.1188",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "7871387",
    "instance_id": "eudoxia0__hashcards.48aa136",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "845",
    "n_system_errors": "0",
    "n_tests": "1019",
    "n_warnings": "0",
    "output_tokens": "63342",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "17038",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-eudoxia0-hashcards",
    "run_version": "",
    "score": "0.8292443572129539",
    "session_ended_at": "2026-05-17T11:19:30.095Z",
    "session_started_at": "2026-05-17T11:01:28.469Z",
    "test_branch_errors": "{}",
    "total_tokens": "7934729",
    "wall_clock_seconds": "1081"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./hashcards.py"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-eudoxia0-hashcards",
  "run_version": "",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}