{
  "release_gate": "Passed: native goal achieved, phase1 complete and hashes reconciled, article score arithmetic independently checked, external quality archive frozen before coordinator combined scores with identities/efficiency.",
  "created_at_utc": "2026-10-25T14:24:43.182156+00:00",
  "labels": {
    "A": {
      "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."
    },
    "B": {
      "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."
    },
    "C": {
      "model": "deepseek/deepseek-v4.1-flash (latest; served snapshot unverified)",
      "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."
    }
  },
  "limitations": [
    "Same shared brief/rubric and repository baseline; one run per candidate.",
    "This is a new three-way review; historical two-way scores are preserved and must not be overwritten.",
    "Producer model, provider, CLI, system prompt, tools, context management, goal supervisor and research date changed together for the new run. It is a workflow comparison, not an isolated model-only causal test.",
    "High reasoning effort is a provider-specific label, not a matched compute budget.",
    "Different provider tokenizers and caching/reporting schemas limit direct token comparisons. Cached input is a subset of input; reasoning when reported is a subset of output. Do not add either twice.",
    "Preflight/setup and reviewer usage are separate overhead. Null means unreported, not zero.",
    "New production began 25 October UTC / 26 October Melbourne; historical runs were on 22 October. Extra anti-contamination/date instructions were appended to the original common goal.",
    "Fresh judge is gpt-6.1-sol high in Codex 0.163.0, the same exact designation as one producer. Blinding and reversed pair ordering do not eliminate model-family or stylistic bias.",
    "One judge and three single runs do not support significance tests, general superiority, manufactured confidence intervals or measured productivity savings.",
    "Review/analysis does not authorise publishing or replacing website articles."
  ],
  "released_at_utc": "2026-10-25T14:54:40.409487+00:00",
  "quality_freeze_at_utc": "2026-10-25T14:54:40.406052+00:00",
  "body_word_method": "Coordinator common regex convention with terminal reference exclusion; reproduces all 22 historical counts. Judge used a separate disclosed common convention. All ranges pass both; word totals are convention dependent."
}