{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:24:59.991824+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "tests.test_cmdutils_final.test_stream_specifier_program_id_no_value_error",
      "tests.test_advanced_video.test_interlacing_flags",
      "tests.test_ffmpeg_opt_depth.test_attach_file_with_metadata",
      "tests.test_decoder_edges.test_decoder_ignore_err_flag",
      "tests.test_encoder_remaining.test_interlaced_mpeg4_encoding",
      "tests.test_filter_graph_edges.test_empty_filter_string_error",
      "tests.test_filter_deepest.test_trim_filter_subset_frames",
      "tests.test_filters.test_invalid_filter_name",
      "tests.test_demuxer_edge_cases.test_map_by_stream_index",
      "tests.test_decoder_deepest.test_video_frame_process_best_effort_timestamp",
      "tests.test_ffmpeg_dec_final.test_video_framerate_forced_duration_estimation",
      "tests.test_audio_options.test_codec_pcm_f32le",
      "tests.test_concat_segment.test_segment_with_audio",
      "tests.test_cmdutils.test_help_bsf_h264_mp4toannexb",
      "tests.test_codec_format_breadth.test_multiple_audio_streams_opus_flac",
      "tests.test_demux_final.test_bsf_flush_at_eof",
      "tests.test_ffmpeg_opt_depth.test_timelimit_option_accepts_valid_value",
      "tests.test_cmdutils_final.test_stream_specifier_group_idx_no_value_error",
      "tests.test_advanced_video.test_aspect_ratio_flag",
      "tests.test_filter_deepest.test_select_filter_even_frames",
      "tests.test_decoder_edges.test_decoder_careful_flag",
      "tests.test_audio_options.test_ac_1_explicit_mono",
      "tests.test_encoder_remaining.test_recording_time_limit",
      "tests.test_demuxer_edge_cases.test_dts_delta_threshold_configures_discontinuity_detection",
      "tests.test_ffmpeg_dec_final.test_audio_simple_decode_frame_count",
      "tests.test_cmdutils.test_help_nonexistent_encoder_error",
      "tests.test_decoder_deepest.test_decoder_subtitle_bitmap_linesize_handling",
      "tests.test_filter_graph_edges.test_fps_zero_causes_encoder_error",
      "tests.test_filters.test_invalid_filter_params",
      "tests.test_concat_segment.test_concat_different_resolutions_reencode",
      "tests.test_codec_format_breadth.test_mpeg2video_encoding",
      "tests.test_demux_final.test_input_packet_process_with_ts_offset",
      "tests.test_encoder_remaining.test_psnr_calculation_flag",
      "tests.test_cmdutils_final.test_stream_specifier_stream_id_no_value_error",
      "tests.test_advanced_video.test_colorspace_flags",
      "tests.test_ffmpeg_opt_depth.test_vstats_file_option_creates_log",
      "tests.test_cmdutils.test_help_nonexistent_decoder_error",
      "tests.test_filter_deepest.test_framerate_conversion_fps_30",
      "tests.test_demuxer_edge_cases.test_stream_loop_zero_means_no_loop",
      "tests.test_decoder_edges.test_decoder_with_framemd5_output",
      "tests.test_concat_segment.test_segment_individual_header_trailer",
      "tests.test_filter_graph_edges.test_concat_filter_multiple_inputs",
      "tests.test_demux_final.test_demux_final_stats_logging",
      "tests.test_ffmpeg_dec_final.test_skip_frame_option_parsing",
      "tests.test_codec_format_breadth.test_ac3_audio_encoding",
      "tests.test_audio_options.test_ba_extreme_low_bitrate",
      "tests.test_decoder_deepest_v2.test_audio_sample_rate_change_overflow_boundary",
      "tests.test_filters.test_rotate_filter",
      "tests.test_ffmpeg_opt_depth.test_progress_to_stdout_dash_syntax",
      "tests.test_cmdutils_final.test_stream_specifier_multiple_disposition_error",
      "tests.test_encoder_remaining.test_vstats_file_output",
      "tests.test_advanced_video.test_gop_size_parameter",
      "tests.test_cmdutils.test_help_nonexistent_muxer_error",
      "tests.test_demuxer_edge_cases.test_disable_data_streams_with_dn_flag",
      "tests.test_filter_deepest.test_format_pixel_conversion_rgb24",
      "tests.test_concat_segment.test_segment_break_non_keyframes",
      "tests.test_decoder_edges.test_decoder_null_packet_handling",
      "tests.test_filter_graph_edges.test_blend_filter_with_mode",
      "tests.test_ffmpeg_dec_final.test_rawvideo_decoder_explicit_params",
      "tests.test_audio_options.test_ba_high_bitrate_320k",
      "tests.test_codec_format_breadth.test_aac_audio_encoding_m4a",
      "tests.test_demux_final.test_do_send_with_all_consumers_done",
      "tests.test_filters.test_filter_complex_split",
      "tests.test_decoder_deepest_v2.test_subtitle_heartbeat_signal_timing_precise",
      "tests.test_ffmpeg_opt_depth.test_complex_filter_chain_with_multiple_outputs",
      "tests.test_cmdutils_final.test_stream_specifier_multiple_stream_list_designators_error",
      "tests.test_decoder_edges.test_decoder_bitstream_flag",
      "tests.test_cmdutils.test_help_nonexistent_demuxer_error",
      "tests.test_ffmpeg_dec_final.test_decoder_flags_configuration",
      "tests.test_demux_final.test_ts_fixup_with_start_time_effective",
      "tests.test_filter_final.test_filter_nonexistent_name_error",
      "tests.test_advanced_video.test_bframes_parameter",
      "tests.test_encoder_remaining.test_audio_encoder_frame_size",
      "tests.test_demuxer_edge_cases.test_max_interleave_delta_for_packet_ordering",
      "tests.test_filter_graph_edges.test_pad_filter_with_color",
      "tests.test_audio_options.test_ar_and_af_aresample_conflict",
      "tests.test_codec_format_breadth.test_3gp_container_mpeg4_aac",
      "tests.test_concat_segment.test_segment_atclocktime",
      "tests.test_filters.test_filter_complex_overlay",
      "tests.test_decoder_deepest_v2.test_eof_handling_flush_frames"
    ],
    "failed_tests_omitted": 3687,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 4165,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "ffmpeg__ffmpeg.360a402",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "5644416",
    "calls": "83",
    "created_at": "2026-05-17T11:07:51.578214+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "5.3396",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "5838039",
    "instance_id": "ffmpeg__ffmpeg.360a402",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "302",
    "n_system_errors": "0",
    "n_tests": "3050",
    "n_warnings": "0",
    "output_tokens": "51642",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "17663",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-ffmpeg-ffmpeg",
    "run_version": "",
    "score": "0.09901639344262295",
    "session_ended_at": "2026-05-17T11:21:09.590Z",
    "session_started_at": "2026-05-17T11:07:54.905Z",
    "test_branch_errors": "{}",
    "total_tokens": "5889681",
    "wall_clock_seconds": "794"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./info_texts",
      "./info_texts/bsfs.txt",
      "./info_texts/buildconf.txt",
      "./info_texts/codecs.txt",
      "./info_texts/decoders.txt",
      "./info_texts/demuxers.txt",
      "./info_texts/devices.txt",
      "./info_texts/encoders.txt",
      "./info_texts/filters.txt",
      "./info_texts/formats.txt",
      "./info_texts/hwaccels.txt",
      "./info_texts/license.txt",
      "./info_texts/muxers.txt",
      "./info_texts/pix_fmts.txt",
      "./info_texts/protocols.txt",
      "./miniffmpeg.py"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-ffmpeg-ffmpeg",
  "run_version": "",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}