{
  "set": "public-v1",
  "tasks": 6,
  "method": "Wilson 95% CI; provider errors counted as failures; single-response; identical instructions",
  "rows": [
    {
      "model": "claude-opus-5-5",
      "provider": "Anthropic",
      "version": "claude-opus-5-5",
      "date_tested": "2026-10-15",
      "mode": "single-response",
      "pass_rate": 66.7,
      "ci_low": 50.3,
      "ci_high": 79.8,
      "passes": 24,
      "attempts": 36,
      "attempts_per_task": 6,
      "mean_partial_score": 84.9,
      "cost_per_attempt_usd": 0.22,
      "provider_errors": 0
    },
    {
      "model": "gpt-6-astra",
      "provider": "OpenAI",
      "version": "gpt-6-astra",
      "date_tested": "2026-10-15",
      "mode": "single-response",
      "pass_rate": 16.7,
      "ci_low": 7.9,
      "ci_high": 31.9,
      "passes": 6,
      "attempts": 36,
      "attempts_per_task": 6,
      "mean_partial_score": 69.7,
      "cost_per_attempt_usd": 0.45,
      "provider_errors": 1
    },
    {
      "model": "grok-4.7",
      "provider": "xAI",
      "version": "grok-4.7",
      "date_tested": "2026-10-15",
      "mode": "single-response",
      "pass_rate": 16.7,
      "ci_low": 7.9,
      "ci_high": 31.9,
      "passes": 6,
      "attempts": 36,
      "attempts_per_task": 6,
      "mean_partial_score": 64.0,
      "cost_per_attempt_usd": 0.58,
      "provider_errors": 2
    }
  ],
  "per_task": [
    {
      "model": "claude-opus-5-5",
      "pub_01": "5/6",
      "pub_02": "5/6",
      "pub_03": "3/6",
      "pub_04": "6/6",
      "pub_05": "5/6",
      "pub_06": "0/6"
    },
    {
      "model": "gpt-6-astra",
      "pub_01": "2/6",
      "pub_02": "0/6",
      "pub_03": "1/6",
      "pub_04": "2/6",
      "pub_05": "1/6",
      "pub_06": "0/6"
    },
    {
      "model": "grok-4.7",
      "pub_01": "1/6",
      "pub_02": "2/6",
      "pub_03": "1/6",
      "pub_04": "2/6",
      "pub_05": "0/6",
      "pub_06": "0/6"
    }
  ],
  "per_category": [
    {
      "model": "claude-opus-5-5",
      "survivorship": "3/30",
      "unadjusted_split": "4/30",
      "unadjusted_dividends": "9/24",
      "stale_prices": "2/30",
      "trading_halt": "4/36",
      "ticker_rename": "2/12",
      "currency_switch": "6/36"
    },
    {
      "model": "gpt-6-astra",
      "survivorship": "3/29",
      "unadjusted_split": "1/29",
      "unadjusted_dividends": "10/24",
      "stale_prices": "2/29",
      "trading_halt": "3/35",
      "ticker_rename": "11/12",
      "currency_switch": "27/35"
    },
    {
      "model": "grok-4.7",
      "survivorship": "10/29",
      "unadjusted_split": "2/28",
      "unadjusted_dividends": "8/23",
      "stale_prices": "7/28",
      "trading_halt": "7/34",
      "ticker_rename": "6/12",
      "currency_switch": "23/34"
    }
  ]
}