{
  "battery_run": "battery-1.0-2026-W30",
  "canary_day": "2026-07-21",
  "canary_run": "canary-1.0-2026-07-21",
  "counts": {
    "AFFIRMED": 6,
    "CONFIRMED": 9,
    "NEW BASELINE": 1,
    "WATCH": 0
  },
  "gaps": [
    "openweight-a"
  ],
  "headline": "6 of 16 AFFIRMED; 9 confirmed changes, 1 new baseline.",
  "occasions": 5,
  "params": {
    "comparison_se_factor": 1.41421356,
    "cusum_h": 2.0,
    "cusum_watch_fraction": 0.5,
    "equivalence_margin": 0.5,
    "evalue_alternative": 0.5,
    "level_alpha": 0.05,
    "power": 0.8
  },
  "schema_version": "1",
  "seats": [
    {
      "build": "claude-opus-4-8",
      "canary": {
        "ok": 4,
        "total": 5
      },
      "detectable_shift_bound": 2.31,
      "detectable_shift_median": 0.97,
      "display_name": "Anthropic flagship (Opus 4.8)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 3.0,
        "canary_divergent_today": 1,
        "headline": 3.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "cusum-alarm",
        "alert-confirmed:canary",
        "canary-divergent-today"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "anthropic-a",
      "state": "CONFIRMED"
    },
    {
      "build": "gpt-5.6-sol",
      "canary": {
        "ok": 5,
        "total": 5
      },
      "detectable_shift_bound": 1.941,
      "detectable_shift_median": 0.97,
      "display_name": "OpenAI flagship (GPT-5.6 Sol)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 0,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "alert-confirmed:canary"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "openai-a",
      "state": "CONFIRMED"
    },
    {
      "build": "gemini-2.5-pro",
      "canary": {
        "ok": 0,
        "total": 5
      },
      "detectable_shift_bound": null,
      "detectable_shift_median": null,
      "display_name": "Google flagship (Gemini 2.5 Pro)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 5,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "canary-divergent-today"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "google-a",
      "state": "CONFIRMED"
    },
    {
      "build": "grok-4.5",
      "canary": {
        "ok": 5,
        "total": 5
      },
      "detectable_shift_bound": 2.911,
      "detectable_shift_median": 1.585,
      "display_name": "xAI flagship (Grok 4.5)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 0,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [],
      "equivalence": null,
      "evalue": null,
      "evalue_reject_at": 20.0,
      "n_items": 50,
      "seat": "xai-a",
      "state": "AFFIRMED"
    },
    {
      "build": "deepseek-v4-pro",
      "canary": {
        "ok": 4,
        "total": 5
      },
      "detectable_shift_bound": 3.962,
      "detectable_shift_median": 1.321,
      "display_name": "DeepSeek flagship (V4 Pro)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 3.0,
        "canary_divergent_today": 1,
        "headline": 3.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "cusum-alarm",
        "alert-confirmed:canary",
        "canary-divergent-today"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "deepseek-a",
      "state": "CONFIRMED"
    },
    {
      "build": "claude-fable-5",
      "canary": {
        "ok": 5,
        "total": 5
      },
      "detectable_shift_bound": 1.941,
      "detectable_shift_median": 1.057,
      "display_name": "Anthropic (Fable 5)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 0,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [],
      "equivalence": null,
      "evalue": null,
      "evalue_reject_at": 20.0,
      "n_items": 50,
      "seat": "anthropic-b",
      "state": "AFFIRMED"
    },
    {
      "build": "claude-sonnet-5",
      "canary": {
        "ok": 5,
        "total": 5
      },
      "detectable_shift_bound": 1.585,
      "detectable_shift_median": 0.792,
      "display_name": "Anthropic (Sonnet 5)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 0,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [],
      "equivalence": null,
      "evalue": null,
      "evalue_reject_at": 20.0,
      "n_items": 50,
      "seat": "anthropic-c",
      "state": "AFFIRMED"
    },
    {
      "build": "claude-haiku-4-5-20251001",
      "canary": {
        "ok": 4,
        "total": 5
      },
      "detectable_shift_bound": 1.941,
      "detectable_shift_median": 0.97,
      "display_name": "Anthropic (Haiku 4.5)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.5,
        "canary_divergent_today": 1,
        "headline": 0.5,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "alert-confirmed:canary",
        "canary-divergent-today"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "anthropic-d",
      "state": "CONFIRMED"
    },
    {
      "build": "gpt-5.6-terra",
      "canary": {
        "ok": 5,
        "total": 5
      },
      "detectable_shift_bound": 2.687,
      "detectable_shift_median": 0.792,
      "display_name": "OpenAI (GPT-5.6 Terra)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 0,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [],
      "equivalence": null,
      "evalue": null,
      "evalue_reject_at": 20.0,
      "n_items": 50,
      "seat": "openai-b",
      "state": "AFFIRMED"
    },
    {
      "build": "gpt-5.6-luna",
      "canary": {
        "ok": 5,
        "total": 5
      },
      "detectable_shift_bound": 2.911,
      "detectable_shift_median": 0.97,
      "display_name": "OpenAI (GPT-5.6 Luna)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 0,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [],
      "equivalence": null,
      "evalue": null,
      "evalue_reject_at": 20.0,
      "n_items": 50,
      "seat": "openai-c",
      "state": "AFFIRMED"
    },
    {
      "build": "gemini-3.5-flash",
      "canary": {
        "ok": 1,
        "total": 5
      },
      "detectable_shift_bound": 1.483,
      "detectable_shift_median": 1.483,
      "display_name": "Google (Gemini 3.5 Flash)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.75,
        "canary_divergent_today": 4,
        "headline": 0.75,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "canary-divergent-today"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "google-b",
      "state": "CONFIRMED"
    },
    {
      "build": "gemini-2.5-flash",
      "canary": {
        "ok": 0,
        "total": 5
      },
      "detectable_shift_bound": null,
      "detectable_shift_median": null,
      "display_name": "Google (Gemini 2.5 Flash)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 5,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "canary-divergent-today"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "google-c",
      "state": "CONFIRMED"
    },
    {
      "build": "gemini-2.5-flash-lite",
      "canary": {
        "ok": 0,
        "total": 5
      },
      "detectable_shift_bound": null,
      "detectable_shift_median": null,
      "display_name": "Google (Gemini 2.5 Flash-Lite)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 5,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "canary-divergent-today"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "google-d",
      "state": "CONFIRMED"
    },
    {
      "build": "grok-4.5-fast",
      "canary": {
        "ok": 0,
        "total": 5
      },
      "detectable_shift_bound": null,
      "detectable_shift_median": null,
      "display_name": "xAI (Grok 4.5 Fast)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 5,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "alert-confirmed:canary",
        "canary-divergent-today"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "xai-b",
      "state": "CONFIRMED"
    },
    {
      "build": "deepseek-v4-flash",
      "canary": {
        "ok": 5,
        "total": 5
      },
      "detectable_shift_bound": 2.911,
      "detectable_shift_median": 0.991,
      "display_name": "DeepSeek (V4 Flash)",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.0,
        "canary_divergent_today": 0,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "cusum_s"
      },
      "drivers": [],
      "equivalence": null,
      "evalue": null,
      "evalue_reject_at": 20.0,
      "n_items": 50,
      "seat": "deepseek-b",
      "state": "AFFIRMED"
    },
    {
      "build": "simulant-uniform",
      "canary": {
        "ok": 4,
        "total": 5
      },
      "detectable_shift_bound": 3.631,
      "detectable_shift_median": 2.31,
      "display_name": "Simulant · dice ruler",
      "divergence": {
        "canary_cusum_h": 2.0,
        "canary_cusum_max": 0.75,
        "canary_divergent_today": 1,
        "headline": 0.0,
        "stance_delta_max": null,
        "stance_items_moved": 0,
        "unit": "none"
      },
      "drivers": [
        "first-baseline"
      ],
      "equivalence": null,
      "evalue": null,
      "n_items": 50,
      "seat": "simulant-a",
      "state": "NEW BASELINE"
    }
  ]
}