{
 "deepseek-v4.1-flash": {
  "failed_not_opened": [
   "T05",
   "T09",
   "T15",
   "T26",
   "T28",
   "T29"
  ],
  "passing_sample": [
   "T02",
   "T04",
   "T07",
   "T08",
   "T10",
   "T12",
   "T14",
   "T17",
   "T20",
   "T23"
  ],
  "failed_targets": [
   "codex_cli",
   "playwright_ajax_capture",
   "skill_firecrawl",
   "skill_image_generation",
   "skill_outlook",
   "video-scoring-audio"
  ],
  "failed_with_target_opened": []
 },
 "gemini-3.8-flash": {
  "failed_not_opened": [],
  "passing_sample": [
   "T02",
   "T06",
   "T08",
   "T10",
   "T13",
   "T15",
   "T20",
   "T23",
   "T26",
   "T30"
  ],
  "failed_targets": [],
  "failed_with_target_opened": []
 },
 "grok-4-fast": {
  "failed_not_opened": [
   "T01",
   "T02",
   "T03",
   "T04",
   "T05",
   "T06",
   "T07",
   "T09",
   "T14",
   "T15",
   "T17",
   "T19",
   "T20",
   "T21",
   "T23",
   "T25",
   "T26",
   "T27",
   "T28",
   "T29",
   "T30"
  ],
  "passing_sample": [
   "T08",
   "T10",
   "T11",
   "T12",
   "T13",
   "T16",
   "T18",
   "T22",
   "T24"
  ],
  "failed_targets": [
   "apple_calendar",
   "asr",
   "bestpractice_pdf_to_markdown",
   "codex_cli",
   "mac_universal_clipboard",
   "narwal",
   "openreview",
   "playwright_ajax_capture",
   "skill_firecrawl",
   "skill_image_generation",
   "skill_opencode_data",
   "skill_outlook",
   "skill_playwright_test",
   "skill_resend_email",
   "smart_home_skill",
   "speaker_notes",
   "video-scoring-audio",
   "watch_development",
   "workflow_bilibili_whisper_transcription",
   "workflow_parallel_subagents",
   "workflow_watchdog"
  ],
  "failed_with_target_opened": []
 }
}