{
  "title": "AI Models vs Optimal Human Diet",
  "subtitle": "Same first-principles prompt. Many models. Side-by-side metabolically optimal plates.",
  "disclaimer": "Reasoning experiment only. Not medical advice. Every plate is a hypothesis from an AI run or joint debate — not a prescription.",
  "intro": {
    "eyebrow": "Public experiment · not medical advice",
    "heading": "What is this?",
    "body": [
      "This site is a small experiment: give several frontier language models the **same hard question** about food — what should a healthy young adult eat by default if the only goal is **metabolic optimum**? Each answer must lead with a clear **Plate description**, then **Substantiate plate choice** — that two-part outcome is the core of the lab.",
      "It is **not** a product, a medical service, or a popularity contest. Models must reason from **logic and first principles** only. Scientific studies, guidelines, and authority quotes are out of bounds: they are associative, confounded, and endlessly re-interpretable in debate.",
      "**Metabolic optimal** is the **sole outcome** this lab optimizes for: the intake pattern that best supports a healthy young adult’s core metabolic function — fuel handling, cellular repair, hormone and mineral balance, low chronic inflammatory / toxic noise — treated as the master hypothesis from which many other wellness goods tend to follow. It is a **peak** claim, not a survival diet that happens to work well: at the optimum, adding or subtracting anything makes it less optimal. It is **not** “whatever guidelines say,” not culinary variety, not population averages, and not “absence of frank deficiency.” Models must reason *for* that target, not trade it away for other scorecards.",
      "Every answer is a **hypothesis**. A good hypothesis is not instantly wrecked by healthy people on another pattern; nor may it demand that those people must be deficient when observation allows them not to be.",
      "The first card is a **joint debate fable** (Grok 4.5 + Claude Sonnet Fable 5) — one combined write-up, not the answer key. Other cards are single-model outcomes. Intelligence scores link to [Artificial Analysis](https://artificialanalysis.ai/); they measure model capability, not nutrition truth."
    ]
  },
  "prompt": {
    "id": "core-metabolic-optimal-v9",
    "label": "Core prompt (metabolic optimal · plate → substantiate)",
    "system": "You are a careful nutrition reasoner. Be concrete and opinionated. No medical diagnoses or treatment plans.\n\n## Sole objective\n**Metabolic optimal** for a healthy young adult — the only target.\n\n**Meaning:** the default intake that best supports core metabolic function (energy handling / fuel partitioning, structural repair, hormone and mineral balance, low needless metabolic/toxic load) so the organism runs high-function on food alone.\n\n**Peak, not survival:** not a subsistence or “people do fine” plate. A strong-outcome pattern can still be sub-optimal. **Optimal** means a peak hypothesis: default inclusions and exclusions such that **adding or subtracting anything makes it less optimal** for this target.\n\n**Downstream only:** clarity, resilience, fertility, long disease-free stretch, general wellness — treat as following from the metabolic state, not rival scorecards.\n\n**Out of scope as goals:** guidelines / “balanced plate” aesthetics; variety, tradition, popularity, adherence theater; “better than SAD”; mere non-deficiency; longevity, weight loss, sport, or ethics as primary aims (mention only if they fall out of the metabolic case).\n\nEvery food must earn its seat for metabolic optimal. If it mainly serves variety, culture, guidelines, or another goal, it is not default.\n\n## Method\n- Logic, first principles, food chemistry, human physiology, internal consistency only.\n- No scientific studies, trials, meta-analyses, guidelines, authorities, or “the literature” (associative, confounded, endlessly debatable).\n- Recognizable whole foods only — no supplements, isolates, or fortification as the answer.\n- Plate = **hypothesis** for metabolic optimal. Cohorts who thrive (or lack frank deficiency) on another pattern do not by themselves falsify yours; do not require them to be clinically sick.\n- Sun assumed for vitamin D unless you argue otherwise. Free amounts to satiety OK; no heroic calorie counting. Numbers only when they carry the logic (esp. calorie shares).\n- Follow the user’s output format exactly. Do not restate these rules in your answer.",
    "user": "Design the metabolically optimal **default plate** for a healthy young adult (sole objective as defined in the system message).\n\nOutput exactly this order — no text before the first heading; no extra headings:\n\n## Plate description\nBullets only (4–10). One food/group per line. Use `rare:` / `out:` where needed. No argument.\n\n## Substantiate plate choice\nOpen with one line on metabolic optimal (peak sole objective), then why **this** plate is the best peak hypothesis (essentials, ceilings, outs, energy). Where useful, why a notable add/remove would be less optimal. ≤**280 words**. Do not restate method rules.\n\n```\nCALORIE_SHARES\nanimal: <0-100>\nplant: <0-100>\nconstituents:\n- <item from plate>: <0-100>\n- <item from plate>: <0-100>\n```\n\nFooter only (not a third essay section). Integers 0–100; animal + plant = 100; constituents = % of total calories and match the plate.\n\nShape:\n## Plate description\n- …\n- rare: …\n- out: …\n\n## Substantiate plate choice\n…\n\n```\nCALORIE_SHARES\n…\n```\n"
  },
  "scoreSources": {
    "artificialAnalysis": {
      "label": "Artificial Analysis",
      "url": "https://artificialanalysis.ai/leaderboards/models",
      "note": "Intelligence / Coding / Agentic composite indices"
    }
  },
  "models": [
    {
      "id": "meta-llama/llama-3.3-70b-instruct",
      "name": "Llama 3.3 70B",
      "vendor": "Meta",
      "color": "#be185d",
      "note": "Plant-prior pick: older instruct model, strong public-health / guideline echo in training.",
      "scores": [
        {
          "label": "Intelligence",
          "value": 9,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "meta-llama/llama-4-scout",
      "name": "Llama 4 Scout",
      "vendor": "Meta",
      "color": "#db2777",
      "scores": [
        {
          "label": "Intelligence",
          "value": 10,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "meta-llama/llama-4-maverick",
      "name": "Llama 4 Maverick",
      "vendor": "Meta",
      "color": "#f472b6",
      "scores": [
        {
          "label": "Intelligence",
          "value": 14,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "google/gemma-4-31b-it",
      "name": "Gemma 4 31B",
      "vendor": "Google (open)",
      "color": "#94a3b8",
      "scores": [
        {
          "label": "Intelligence",
          "value": 22,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "anthropic/claude-haiku-4.5",
      "name": "Claude Haiku 4.5",
      "vendor": "Anthropic",
      "color": "#fb923c",
      "note": "Plant-prior pick: light assistant tier; often defaults to conventional healthy-plate priors.",
      "scores": [
        {
          "label": "Intelligence",
          "value": 24,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "openai/gpt-oss-120b",
      "name": "gpt-oss-120b",
      "vendor": "OpenAI (open)",
      "color": "#34d399",
      "scores": [
        {
          "label": "Intelligence",
          "value": 24,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "google/gemini-2.5-pro",
      "name": "Gemini 2.5 Pro",
      "vendor": "Google",
      "color": "#3b82f6",
      "scores": [
        {
          "label": "Intelligence",
          "value": 26,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "openai/gpt-5.6-luna",
      "name": "GPT-5.6 Luna",
      "vendor": "OpenAI",
      "color": "#059669",
      "note": "Plant-prior pick: mass-market OpenAI tier; productized healthy-diet priors.",
      "scores": [
        {
          "label": "Intelligence",
          "value": 38,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "mistralai/mistral-medium-3-5",
      "name": "Mistral Medium 3.5",
      "vendor": "Mistral",
      "color": "#f59e0b",
      "scores": [
        {
          "label": "Intelligence",
          "value": 30,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "deepseek/deepseek-v4-flash",
      "name": "DeepSeek V4 Flash",
      "vendor": "DeepSeek",
      "color": "#22d3ee",
      "scores": [
        {
          "label": "Intelligence",
          "value": 40,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "deepseek/deepseek-v4-pro",
      "name": "DeepSeek V4 Pro",
      "vendor": "DeepSeek",
      "color": "#06b6d4",
      "scores": [
        {
          "label": "Intelligence",
          "value": 44,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "qwen/qwen3.7-max",
      "name": "Qwen3.7 Max",
      "vendor": "Alibaba",
      "color": "#a855f7",
      "scores": [
        {
          "label": "Intelligence",
          "value": 46,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "google/gemini-3.5-flash",
      "name": "Gemini 3.5 Flash",
      "vendor": "Google",
      "color": "#60a5fa",
      "scores": [
        {
          "label": "Intelligence",
          "value": 50,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "anthropic/claude-sonnet-5",
      "name": "Claude Sonnet 5",
      "vendor": "Anthropic",
      "color": "#ea580c",
      "scores": [
        {
          "label": "Intelligence",
          "value": 53,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "x-ai/grok-4.5",
      "name": "Grok 4.5",
      "vendor": "xAI",
      "color": "#a78bfa",
      "scores": [
        {
          "label": "Intelligence",
          "value": 54,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "moonshotai/kimi-k3",
      "name": "Kimi K3",
      "vendor": "Moonshot",
      "color": "#e11d48",
      "scores": [
        {
          "label": "Intelligence",
          "value": 57,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "openai/gpt-5.6-sol",
      "name": "GPT-5.6 Sol",
      "vendor": "OpenAI",
      "color": "#10b981",
      "scores": [
        {
          "label": "Intelligence",
          "value": 59,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    },
    {
      "id": "anthropic/claude-opus-5",
      "name": "Claude Opus 5",
      "vendor": "Anthropic",
      "color": "#d97706",
      "scores": [
        {
          "label": "Intelligence",
          "value": 61,
          "source": "Artificial Analysis",
          "url": "https://artificialanalysis.ai/leaderboards/models"
        }
      ]
    }
  ],
  "joint": {
    "id": "joint-grok-claude-fable",
    "name": "Grok 4.5 × Claude Sonnet Fable 5",
    "vendor": "Joint debate fable · not a single API run",
    "color": "#c9a227",
    "note": "Combined debate fable from Grok 4.5 and Claude Sonnet Fable 5 under the same method. One hypothesis among others — not an answer key.",
    "scores": [
      {
        "label": "Grok 4.5",
        "value": 53.8,
        "source": "Artificial Analysis",
        "url": "https://artificialanalysis.ai/leaderboards/models"
      },
      {
        "label": "Sonnet Fable 5",
        "value": 53.4,
        "source": "Artificial Analysis",
        "url": "https://artificialanalysis.ai/leaderboards/models"
      }
    ],
    "calorieShares": {
      "animal": 92,
      "plant": 8,
      "constituents": [
        {
          "name": "Ruminant muscle & fat",
          "pct": 62
        },
        {
          "name": "Liver (small)",
          "pct": 3
        },
        {
          "name": "Bone-in / marrow",
          "pct": 6
        },
        {
          "name": "Marine fish & shellfish",
          "pct": 18
        },
        {
          "name": "Eggs / residual animal",
          "pct": 3
        },
        {
          "name": "Citrus & minor plant",
          "pct": 8
        }
      ]
    },
    "content": "## Plate description\n- Ruminant muscle and fat (main energy and protein)\n- Small regular liver\n- Bone-in cuts or marrow\n- Oily fish or shellfish\n- Salt and water\n- Small citrus (margin)\n- out: plant staples as the scaffold (grains, seed oils, juices, sugar loads)\n- out: supplements or fortification powders as the answer\n\n## Substantiate plate choice\n\n**Metabolic optimal** means the default intake that best keeps a healthy young adult's core metabolism high-functioning: reliable fuel, repair, mineral and vitamin coverage without overload, low needless toxin load, and real-food satiety. Secondary goods ride largely downstream.\n\nDay to day: close every non-negotiable nutrient bill in usable forms, supply energy without constant cleanup, and default off loads that cost more than they return. Other patterns may still be enough without being best.\n\nAnimal muscle and fat carry complete protein, dense energy, and bioavailable iron, zinc, and B-vitamins—the load-bearing center. Muscle-plus-fat alone under-explains calcium, long-chain n-3, iodine/selenium, and organ-concentrated micronutrients. Small liver covers preformed A, copper, folate, and B12; kept small because A and copper have ceilings. Bone-in or marrow closes calcium. Marine fish or shellfish closes marine gaps. Salt and water replace losses; sun covers vitamin D.\n\nPlants are not default staples: defense chemistry means large plant mass needs a net metabolic gain if the animal core already covers essentials. Small citrus is a cautious ascorbate margin under hard cooking or aging—not a plant-based plate. Seed oils, refined sugars, and juices fail harder.\n\nThis default leads on metabolic optimum until a stronger alternative is argued the same way. Thriving people on other patterns need not be cast as sick. Nearly all calories are animal; a single-digit plant share is a margin, not a pillar.\n\n```\nCALORIE_SHARES\nanimal: 92\nplant: 8\nconstituents:\n- Ruminant muscle & fat: 62\n- Liver (small): 3\n- Bone-in / marrow: 6\n- Marine fish & shellfish: 18\n- Eggs / residual animal: 3\n- Citrus & minor plant: 8\n```\n"
  },
  "scoresNote": "Scores are Artificial Analysis indices. Click any score to open the source leaderboard. Scores measure model capability — not whether a plate is true.",
  "openrouter": {
    "baseUrl": "https://openrouter.ai/api/v1",
    "maxTokens": 8192,
    "temperature": 0.4,
    "reasoning": {
      "effort": "low",
      "exclude": true
    }
  }
}