{
  "production": [
    {
      "model": "gpt-6.1-sol",
      "provider": "OpenAI",
      "harness": "Codex CLI 0.163.0",
      "mode": "native /goal",
      "reasoning_effort": "high",
      "production_elapsed_wall_seconds": 1097.906772,
      "tokens": {
        "input_tokens": 3046820,
        "cached_input_tokens": 2894720,
        "cache_write_input_tokens": 0,
        "output_tokens": 41141,
        "reasoning_output_tokens": 5500,
        "total_tokens": 3087961,
        "uncached_input_tokens": 152100
      },
      "common_method_body_words": 12976,
      "time_limitations": "Elapsed includes tool time and waiting; no separate inference-time measurement.",
      "single_run": true,
      "billing": "Actual task cost not measured; do not substitute rate-table estimates or flat subscription allowances.",
      "anonymous_label": "C",
      "new_four_way_score": 93.81818181818181,
      "elapsed_minutes": 18.2984462,
      "cache_read_fraction_of_input": 0.9500790988637333
    },
    {
      "model": "gpt-6.1-terra",
      "provider": "OpenAI",
      "harness": "Codex CLI 0.164.0",
      "mode": "native /goal",
      "reasoning_effort": "high",
      "production_elapsed_wall_seconds": 348.139854,
      "tokens": {
        "input_tokens": 791539,
        "cached_input_tokens": 728192,
        "cache_write_input_tokens": 0,
        "output_tokens": 24021,
        "reasoning_output_tokens": 601,
        "total_tokens": 815560,
        "uncached_input_tokens": 63347
      },
      "single_run": true,
      "time_limitations": "Wall time includes tools, not pure inference latency. Initial empty CLI cwd was sanctum-chat; every production tool call was instructed to use the isolated historical-baseline worktree.",
      "billing": "No actual per-task invoice or subscription charge measured.",
      "anonymous_label": "A",
      "new_four_way_score": 81.0,
      "elapsed_minutes": 5.8023309,
      "cache_read_fraction_of_input": 0.9199698309243133
    },
    {
      "model": "gpt-6-luna",
      "provider": "OpenAI",
      "harness": "Codex CLI 0.163.0",
      "mode": "native /goal",
      "reasoning_effort": "high",
      "production_elapsed_wall_seconds": 567.973218,
      "tokens": {
        "input_tokens": 3110420,
        "cached_input_tokens": 3009152,
        "cache_write_input_tokens": 0,
        "output_tokens": 47354,
        "reasoning_output_tokens": 9353,
        "total_tokens": 3157774,
        "uncached_input_tokens": 101268
      },
      "common_method_body_words": 14249,
      "time_limitations": "Elapsed includes tool time and waiting; no separate inference-time measurement.",
      "single_run": true,
      "billing": "Actual task cost not measured; do not substitute rate-table estimates or flat subscription allowances.",
      "anonymous_label": "D",
      "new_four_way_score": 80.63636363636364,
      "elapsed_minutes": 9.4662203,
      "cache_read_fraction_of_input": 0.9674423389767298
    },
    {
      "model": "deepseek/deepseek-v4.1-flash (latest, no pinned 1071 snapshot)",
      "provider": "CommandCode",
      "harness": "CommandCode CLI 1.88.0",
      "mode": "native /goal",
      "reasoning_effort": "high",
      "production_elapsed_wall_seconds": 3518.355932,
      "native_goal_elapsed_seconds": 3514.968,
      "tokens": {
        "input_tokens": 31878376,
        "cached_input_tokens": 31187584,
        "cache_write_input_tokens": 0,
        "output_tokens": 177989,
        "reasoning_output_tokens": null,
        "total_tokens": 32056365,
        "uncached_input_tokens": 690792
      },
      "common_method_body_words": 15984,
      "time_limitations": "Includes user/coordinator permission approval waits, not separately timed. Switched from accept-edits to YOLO during run at user request. Do not treat elapsed as pure model latency.",
      "single_run": true,
      "token_coverage": "105 unique persisted assistant requests; separate reasoning count and unpersisted verifier/background overhead unavailable. Setup/preflight excluded.",
      "native_goal_verifier": "Unavailable; harness accepted producer self-claim. Structural checks are separate, and blind quality review is the independent evaluation.",
      "billing": "Actual task cost not measured; do not substitute rate-table estimates or flat subscription allowances.",
      "anonymous_label": "B",
      "new_four_way_score": 78.18181818181819,
      "elapsed_minutes": 58.63926553333333,
      "cache_read_fraction_of_input": 0.978330389226854
    }
  ],
  "judge": {
    "model": "gpt-6.1-sol",
    "input_tokens": 18407853,
    "cached_input_tokens": 18001792,
    "output_tokens": 59262,
    "reasoning_output_tokens": 13881,
    "total_tokens": 18467115,
    "elapsed_wall_seconds": 2523.402688,
    "native_goal_active_seconds": 2148,
    "excluded_from_production_rows": true
  },
  "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."
  ],
  "export_scope": "Curated frozen usage view. Private runtime paths, session IDs and diagnostic transport records omitted; numerical usage and timing unchanged."
}