{
 "task": {
  "onetTaskId": 21670,
  "soc": "15-1252.00",
  "statement": "Modify existing software to correct errors, adapt it to new hardware, or upgrade interfaces and improve performance."
 },
 "instanceSet": {
  "version": "v1",
  "n": 9,
  "source": "Minds (open-source social platform, gitlab.com/minds), merged bug tickets with runnable tests",
  "selection": "9 of 45 test-backed tickets from 2,704 merged; a fix counts only when the ticket's own test fails on the bug and passes on the fix with the existing suite still green"
 },
 "humanBaseline": {
  "success": "9/9 (the team shipped all nine)",
  "timePerFix": "~25 min (inferred from 2-4h @ $75-100/h, NOT measured)",
  "costPerFix": "~$225 (inferred, NOT measured)"
 },
 "configurations": [
  {
   "name": "Claude Opus (model alone, one pass)",
   "bugsFixed": "1/9",
   "resolveRate": "11%",
   "timePerFix": "6s",
   "costPerFix": "$0.04"
  },
  {
   "name": "Test-and-retry loop \u00b7 Claude Opus",
   "bugsFixed": "7/9",
   "resolveRate": "93%",
   "timePerFix": "19s",
   "costPerFix": "$0.29",
   "note": "flaky on tickets 2741 and 2725 (2/3 runs)"
  },
  {
   "name": "Diagnose \u2192 fix hand-off \u00b7 Claude Opus",
   "bugsFixed": "9/9",
   "resolveRate": "100% (27/27 runs)",
   "timePerFix": "19s",
   "costPerFix": "$0.29",
   "atHumanBar": true
  },
  {
   "name": "GPT-5.5 (model alone, one pass)",
   "bugsFixed": "2/9",
   "resolveRate": "22% (6/27 runs)",
   "timePerFix": "30s",
   "costPerFix": "$0.03",
   "receipts": "/data/runs/21670-oneshot-gpt55.json"
  },
  {
   "name": "Gemini 3.1 Pro (model alone, one pass)",
   "bugsFixed": "0/9",
   "resolveRate": "0% (0/27 runs)",
   "timePerFix": "97s",
   "costPerFix": "$0.01",
   "receipts": "/data/runs/21670-oneshot-gemini31.json"
  }
 ],
 "instances": [
  {
   "ticket": 2746,
   "what": "Audio pause on destroy",
   "oneShot": "0/3",
   "loop": "3/3",
   "workflow": "3/3"
  },
  {
   "ticket": 2721,
   "what": "Join group by invite",
   "oneShot": "0/3",
   "loop": "3/3",
   "workflow": "3/3"
  },
  {
   "ticket": 2698,
   "what": "Group invite form",
   "oneShot": "0/3",
   "loop": "3/3",
   "workflow": "3/3"
  },
  {
   "ticket": 2696,
   "what": "Chatwoot emails",
   "oneShot": "0/3",
   "loop": "3/3",
   "workflow": "3/3"
  },
  {
   "ticket": 2735,
   "what": "Tenant site name",
   "oneShot": "0/3",
   "loop": "3/3",
   "workflow": "3/3"
  },
  {
   "ticket": 2740,
   "what": "Muted modal video",
   "oneShot": "3/3",
   "loop": "3/3",
   "workflow": "3/3"
  },
  {
   "ticket": 2741,
   "what": "Audio og:image",
   "oneShot": "0/3",
   "loop": "2/3",
   "workflow": "3/3"
  },
  {
   "ticket": 2725,
   "what": "Add-card loop",
   "oneShot": "0/3",
   "loop": "2/3",
   "workflow": "3/3"
  },
  {
   "ticket": 2702,
   "what": "Misconfigured domain upgrade",
   "oneShot": "0/3",
   "loop": "3/3",
   "workflow": "3/3"
  }
 ],
 "scope": {
  "population": "well-specified, test-backed, buildable bugs in one Angular codebase; not software bugs in general",
  "resolve": "passes the team's own test; a strong proxy for correct, not a proof of it",
  "localization": "agents were handed the right files; real work includes finding the fault",
  "sample": "a 9-bug pilot"
 },
 "publisher": "SPARK AI Research at SDSC \u00b7 measurement by Recursiv Labs \u00b7 Claude Opus throughout, only the structure varied",
 "fullEvidence": "per-ticket diffs and run logs publish with instance set v1"
}