{
  "leg": "C3",
  "started_at": "2026-08-28T16:41:17.258272Z",
  "finished_at": "2026-08-28T16:44:06.112125Z",
  "host_facts": {
    "host": "the bench endpoint",
    "ollama_version": "0.32.14",
    "ps_before": {
      "models": []
    },
    "pin_before": {
      "checked_at": "2026-08-28T16:41:17.262809Z",
      "mode": "bench",
      "ok": true,
      "reachable": true,
      "failures": [],
      "checks": {
        "mainpid": {
          "ok": null,
          "got": null,
          "want": null,
          "note": "bench mode: bench daemon MainPID is asserted by core.bench_assert(), not by this function"
        }
      },
      "prod_untouched": {
        "at": "2026-08-28T16:41:17Z",
        "checked": false,
        "watch": "disabled",
        "reason": "PROD_WATCH=0: cloud-only reference arm. This run loads no weights on any estate box, so it cannot disturb a seat, and it addresses no production port at all.",
        "seats": [],
        "pin": {
          "checked": false
        },
        "ok": true
      },
      "residents": []
    },
    "pin_after": {
      "checked_at": "2026-08-28T16:44:06.112090Z",
      "mode": "bench",
      "ok": true,
      "reachable": true,
      "failures": [],
      "checks": {
        "mainpid": {
          "ok": null,
          "got": null,
          "want": null,
          "note": "bench mode: bench daemon MainPID is asserted by core.bench_assert(), not by this function"
        }
      },
      "prod_untouched": {
        "at": "2026-08-28T16:44:06Z",
        "checked": false,
        "watch": "disabled",
        "reason": "PROD_WATCH=0: cloud-only reference arm. This run loads no weights on any estate box, so it cannot disturb a seat, and it addresses no production port at all.",
        "seats": [],
        "pin": {
          "checked": false
        },
        "ok": true
      },
      "residents": []
    }
  },
  "runner_argv": [
    "tools/c3_run.py",
    "--host",
    "the bench endpoint",
    "glm-5.3-flash:cloud"
  ],
  "protocol": {
    "repeats": 2,
    "temperature": 0,
    "num_predict": 2048,
    "timeout_s": 300,
    "max_rounds": 8,
    "num_ctx": 32768,
    "glimmer_reasoning_strength": "high",
    "reasoning_strength_mechanism": "system-prompt line 'Reasoning strength: {strength}' appended by the house-law client (RULINGS-0812 core ruling 1); the think field stays boolean",
    "reasoning_strength_assignment": {
      "muse-glimmer:30b-q8_0-dflash": {
        "think_true": "high"
      },
      "muse-glimmer:30b-q4_K_M-dflash": {},
      "gemma4:26b": {},
      "gemma4:12b": {},
      "qwen3.5:27b": {},
      "qwen3.6:27b": {},
      "nemotron-3.5-lightning:30b-a3b": {},
      "glm-5.3-flash:cloud": {}
    },
    "posture_assignment": {
      "muse-glimmer:30b-q8_0-dflash": {
        "think_false": false,
        "think_true": true
      },
      "muse-glimmer:30b-q4_K_M-dflash": {
        "think_false": false
      },
      "gemma4:26b": {
        "think_false": false,
        "think_true": true
      },
      "gemma4:12b": {
        "think_false": false
      },
      "qwen3.5:27b": {
        "think_false": false
      },
      "qwen3.6:27b": {
        "think_false": false
      },
      "nemotron-3.5-lightning:30b-a3b": {
        "think_false": false
      },
      "glm-5.3-flash:cloud": {
        "think_true": true
      }
    }
  },
  "pins": {
    "task_file": {
      "name": "tasks.json",
      "sha256": "d172e9c326735df3848389e24f0a82d67b11d97df62805285b2a0b5d527ee140",
      "counts": {
        "total": 19,
        "single": 6,
        "chain": 5,
        "distractor": 2,
        "honesty": 5,
        "error_recovery": 1
      }
    },
    "tools_sha256": "37a3d20f99a366413a6be7b69be7d80ff2a7cb642c8abf8dcd90d7f81189941b",
    "tool_count": 10,
    "system_prompt_sha256": "7b38278b7435b84b2e872b7f61281c24288ba8148ae2b18030d08e399486294c",
    "fixed_clock": "2026-08-12T09:00:00-06:00",
    "http_allowlist": [
      "http://127.0.0.1:8971",
      "https://api.open-meteo.com"
    ]
  },
  "runs": [
    {
      "model": "glm-5.3-flash:cloud",
      "outcome": "RUN",
      "preprobe": "raw/c3/glm-5.3-flash_cloud/preprobe.json",
      "postures": [
        {
          "posture": "think_true",
          "think": true,
          "reasoning_strength": null,
          "trace": "raw/c3/glm-5.3-flash_cloud/think_true/trace.jsonl",
          "attempts": 38,
          "calls": 88,
          "rounds": 100,
          "transport_failures": 0
        }
      ],
      "pin_findings": [
        {
          "checked_at": "2026-08-28T16:42:45.635053Z",
          "mode": "bench",
          "ok": true,
          "reachable": true,
          "failures": [],
          "checks": {
            "mainpid": {
              "ok": null,
              "got": null,
              "want": null,
              "note": "bench mode: bench daemon MainPID is asserted by core.bench_assert(), not by this function"
            }
          },
          "prod_untouched": {
            "at": "2026-08-28T16:42:45Z",
            "checked": false,
            "watch": "disabled",
            "reason": "PROD_WATCH=0: cloud-only reference arm. This run loads no weights on any estate box, so it cannot disturb a seat, and it addresses no production port at all.",
            "seats": [],
            "pin": {
              "checked": false
            },
            "ok": true
          },
          "residents": []
        }
      ]
    }
  ]
}