{
  "title": "Memory first, docs second: the same-model floors and the control for the model comparison",
  "note": "The two floors, the control and the second-run check come from waves the benchmark repo publishes under data/published/. Each entry names the command that recomputes it from those files on a clean clone, and the repo's test/fig6.test.ts pins every figure. The update row is computed on the page from runs.json.",
  "floors": [
    {
      "label": "Opus 5.5",
      "setup": "two passes, same setup",
      "waves": "wave 7 against wave 5",
      "agreeing": 28,
      "comparable": 32,
      "interval": {
        "lo": 0.77,
        "hi": 0.97
      },
      "reproduce": "npm run compare -- --wave data/published/hosting-s1-night2-opus55.json --against data/published/hosting-s1-opus55.json --exclude-frame open"
    },
    {
      "label": "Opus 5",
      "setup": "a Mac pass against a container pass",
      "waves": "wave 2 against wave 3",
      "agreeing": 27,
      "comparable": 31,
      "interval": {
        "lo": 0.77,
        "hi": 0.97
      },
      "caveat": "Wave 2 ran straight on the Mac, outside the container, on the untrimmed commits of three apps, with a BENCHMARK.md some runs quoted. It is a rougher floor than the Opus 5.5 pair.",
      "reproduce": "npm run compare -- --wave data/published/hosting-s0-wave3.json --against data/published/hosting-s0-full.json --exclude-frame open"
    }
  ],
  "control": {
    "label": "Opus 5 on Claude Code 2.1.280, one run per app and brief",
    "waves": "wave 6, read against wave 3 (Opus 5) and wave 5 (Opus 5.5)",
    "changed": 11,
    "likeOlder": 8,
    "likeNewer": 3,
    "likeNeither": 0,
    "compared": 32,
    "matchesOlder": 28,
    "matchesNewer": 23,
    "reproduce": "npm run compare -- --wave data/published/hosting-s1-opus55.json --against data/published/hosting-s0-wave3.json --control data/published/hosting-s1-control-opus5.json --exclude-frame open"
  },
  "updateAgain": {
    "label": "The second Opus 5.5 run against Opus 5",
    "waves": "wave 7 against wave 3",
    "agreeing": 21,
    "comparable": 32,
    "interval": {
      "lo": 0.43,
      "hi": 0.85
    },
    "reproduce": "npm run compare -- --wave data/published/hosting-s1-night2-opus55.json --against data/published/hosting-s0-wave3.json --exclude-frame open"
  },
  "interval": {
    "method": "Resample apps with replacement, paired across the two models. 2000 draws, mulberry32 seed 20261015. The 2.5th and 97.5th percentiles.",
    "iterations": 2000,
    "seed": 20261015
  }
}