{
  "round": "Fresh four-way review · 28 October 2026 · frozen original submissions",
  "models": [
    {
      "id": "sol",
      "name": "GPT 6.1 Sol",
      "color": "#e1bf76",
      "model": "gpt-6.1-sol",
      "anonymousLabel": "C",
      "score": 93.81818181818181,
      "elapsedSeconds": 1097.906772,
      "input": 3046820,
      "cached": 2894720,
      "uncached": 152100,
      "output": 41141,
      "reasoning": 5500,
      "total": 3087961,
      "cacheShare": 0.9500790988637333,
      "estimatedUsd": null
    },
    {
      "id": "terra",
      "name": "GPT 6.1 Terra",
      "color": "#8cc9ff",
      "model": "gpt-6.1-terra",
      "anonymousLabel": "A",
      "score": 81,
      "elapsedSeconds": 348.139854,
      "input": 791539,
      "cached": 728192,
      "uncached": 63347,
      "output": 24021,
      "reasoning": 601,
      "total": 815560,
      "cacheShare": 0.9199698309243133,
      "estimatedUsd": null
    },
    {
      "id": "luna",
      "name": "GPT 6 Luna",
      "color": "#72d3c0",
      "model": "gpt-6-luna",
      "anonymousLabel": "D",
      "score": 80.63636363636364,
      "elapsedSeconds": 567.973218,
      "input": 3110420,
      "cached": 3009152,
      "uncached": 101268,
      "output": 47354,
      "reasoning": 9353,
      "total": 3157774,
      "cacheShare": 0.9674423389767298,
      "estimatedUsd": null
    },
    {
      "id": "deepseek",
      "name": "DeepSeek V4.1 Flash",
      "color": "#baadff",
      "model": "deepseek/deepseek-v4.1-flash (latest, no pinned 1071 snapshot)",
      "anonymousLabel": "B",
      "score": 78.18181818181819,
      "elapsedSeconds": 3518.355932,
      "input": 31878376,
      "cached": 31187584,
      "uncached": 690792,
      "output": 177989,
      "reasoning": null,
      "total": 32056365,
      "cacheShare": 0.978330389226854,
      "estimatedUsd": null
    }
  ],
  "sources": {
    "cluster-scores.json": {
      "sha256": "9378ca92390765db8b5a0fafe919b697fcdae365a6026a79b4b06ce267e7e8cd",
      "url": "/assets/evidence/ai-editorial-terra-20261028/cluster-scores.json"
    },
    "identity-efficiency-release.json": {
      "sha256": "7c6d76a97a5e16cc545a05192e0e625b43f1cbd074bb9cfe53a4e9c3860e4f06",
      "url": "/assets/evidence/ai-editorial-terra-20261028/identity-efficiency-release.json"
    }
  },
  "cost": {
    "status": "unavailable",
    "reason": "Actual task charges not measured; no comparable current rate set verified for this round."
  },
  "limitations": [
    "One production workflow per candidate and one AI judge; no statistical or general model superiority inference.",
    "Twelve directions are internal order checks by the same judge, not twelve independent judges.",
    "These are fresh four-way scores. Preserve the earlier two-way and three-way scores as separate historical reviews.",
    "Blinding is attempted: original prose, styles and research dates can contain identifying clues.",
    "Same exact production brief/rubric and source baseline; dates, model, provider, CLI, tools and context/goal supervision differ across workflows. Nominal high effort is not matched compute.",
    "DeepSeek was CommandCode latest, not proven pinned to 1017 or 1071. Its wall time includes approval waits; unavailable separate reasoning stays null.",
    "Cached input is a subset of input; reasoning is a subset of output. No double counting. Tokenizers/reporting differ; billed cost was not measured.",
    "Quality evidence is targeted fact verification, not exhaustive checking or human publication acceptance."
  ]
}