{
  "as_of": "2026-09-26",
  "inventory_note": "Twelve models documented across different local experiments; not a common twelve-model benchmark.",
  "models": [
    {
      "name": "Qwen 3.6 35B A3B",
      "variant": "MLX · 4-bit"
    },
    {
      "name": "Gemma 4 31B QAT",
      "variant": "GGUF · Q4_0"
    },
    {
      "name": "Qwen 3.8 27B",
      "variant": "MLX · 4-bit"
    },
    {
      "name": "Gemma 4 26B A4B",
      "variant": "MLX · 4-bit"
    },
    {
      "name": "Qwen 3 30B A3B Thinking",
      "variant": "MLX · 4-bit"
    },
    {
      "name": "Ministral 3 14B Instruct",
      "variant": "GGUF · Q4_K_M"
    },
    {
      "name": "NVIDIA Nemotron 3 Nano 30B A3B",
      "variant": "GGUF · Q4_K_M"
    },
    {
      "name": "Liquid LFM2 24B A2B",
      "variant": "GGUF · Q6_K"
    },
    {
      "name": "IBM Granite 4.1 8B",
      "variant": "GGUF · Q8_0"
    },
    {
      "name": "GLM-4.7 Flash 30B A3B",
      "variant": "MLX · 6-bit"
    },
    {
      "name": "Phi-4 Mini Instruct",
      "variant": "GGUF · Q8_0"
    },
    {
      "name": "Muse Glimmer 30B",
      "variant": "GGUF · Q4_K_M · Archiv"
    }
  ],
  "task_test": {
    "date": "2026-09-24",
    "synthetic": true,
    "cases": 14,
    "model_calls": 84,
    "technically_valid": 84,
    "fields": [
      "model",
      "prompt",
      "original_correct",
      "corrected_reference_correct",
      "median_seconds",
      "false_positives",
      "missed_tasks_original"
    ],
    "rows": [
      ["Gemma 4 26B A4B", "v1", 13, 14, 1.3, 0, 0],
      ["Gemma 4 26B A4B", "v1-strict", 13, 14, 1.3, 0, 0],
      ["Qwen 3.6 35B A3B", "v1", 12, 13, 22.5, 0, 0],
      ["Qwen 3.6 35B A3B", "v1-strict", 11, 12, 23.2, 0, 1],
      ["Qwen 3 30B A3B Thinking", "v1", 11, 12, 8.2, 0, 0],
      ["Qwen 3 30B A3B Thinking", "v1-strict", 12, 13, 9.2, 0, 0]
    ],
    "correction": "One meeting-notes reference label corrected by task definition. Same stored responses, no new inference.",
    "recorded_run_sha256": "5c2d1201a91a1365d5c4c2edfe6f8baf2b37465ee543cf1fd412a2174b7dd8d3"
  },
  "ledger": {
    "date": "2026-09-22",
    "model": "Qwen 3.8 27B",
    "planner_passed": 28,
    "planner_total": 30,
    "integration_passed": 18,
    "integration_total": 18,
    "inference_calls": 9,
    "calculation_cases_passed": 28,
    "calculation_model_inference": false
  },
  "scope": "Recorded local tests, not a hardware-controlled general benchmark. Public figures use synthetic data only."
}