{
 "family": "silent-refusal-kit",
 "source": "the kit's own runner on the recorded example fixture shipped inside the kit",
 "generated": "2026-09-10T05:21:38Z",
 "tasks": 15,
 "fixture": true,
 "note": "Recorded example replies, not live model output. Your numbers come from your tasks and your keys on the day you run it. Not a benchmark, not a guarantee.",
 "models": [
  {
   "model": "anthropic:claude-fable-5-1",
   "tasks": 15,
   "passed": 10,
   "pass_rate": 66.7,
   "silent_refusals": 2,
   "silent_refusal_rate": 13.3,
   "cut_off": 1,
   "errors": 1,
   "cost_total_usd": 0.006625,
   "cost_per_pass_usd": 0.000663,
   "longest_task_passed": {
    "id": "t06-changelog-summary",
    "size": 40
   },
   "silent_refusal_ids": [
    "t11-base64",
    "t12-cron-explain"
   ]
  },
  {
   "model": "openai:gpt-6-astra",
   "tasks": 15,
   "passed": 10,
   "pass_rate": 66.7,
   "silent_refusals": 2,
   "silent_refusal_rate": 13.3,
   "cut_off": 1,
   "errors": 1,
   "cost_total_usd": 0.004164,
   "cost_per_pass_usd": 0.000416,
   "longest_task_passed": {
    "id": "t06-changelog-summary",
    "size": 40
   },
   "silent_refusal_ids": [
    "t11-base64",
    "t12-cron-explain"
   ]
  }
 ],
 "rows": [
  {
   "model": "anthropic:claude-fable-5-1",
   "tasks": 15,
   "passed": 10,
   "pass_rate": 66.7,
   "silent_refusals": 2,
   "silent_refusal_rate": 13.3,
   "cut_off": 1,
   "errors": 1,
   "cost_total_usd": 0.006625,
   "cost_per_pass_usd": 0.000663,
   "longest_task_passed": {
    "id": "t06-changelog-summary",
    "size": 40
   },
   "silent_refusal_ids": [
    "t11-base64",
    "t12-cron-explain"
   ]
  },
  {
   "model": "openai:gpt-6-astra",
   "tasks": 15,
   "passed": 10,
   "pass_rate": 66.7,
   "silent_refusals": 2,
   "silent_refusal_rate": 13.3,
   "cut_off": 1,
   "errors": 1,
   "cost_total_usd": 0.004164,
   "cost_per_pass_usd": 0.000416,
   "longest_task_passed": {
    "id": "t06-changelog-summary",
    "size": 40
   },
   "silent_refusal_ids": [
    "t11-base64",
    "t12-cron-explain"
   ]
  }
 ]
}