{
  "format": "coderred-strata-local-benchmark",
  "schema_version": 1,
  "gpu_unlock": {
    "enabled": true,
    "module_version": "1.0.0",
    "kernel": "6.12.116+deb13-amd64",
    "verification_timing": "Collected after measurement from the current boot kernel journal; the benchmark ran in this same boot.",
    "boot_parameters": {
      "gpu_unlock.targets": "0000:09:00.0,0000:0a:00.0",
      "gpu_unlock.resume": 1
    },
    "devices": [
      {
        "name": "NVIDIA GeForce RTX 3090",
        "device_index": 0,
        "pci_bus_id": "0000:09:00.0",
        "unlocked": true,
        "used_for_strata_inference": false,
        "kernel_event": "[5.500805] pci 0000:09:00.0: gpu_unlock: unlocked"
      },
      {
        "name": "NVIDIA GeForce RTX 3060",
        "device_index": 1,
        "pci_bus_id": "0000:0a:00.0",
        "unlocked": true,
        "used_for_strata_inference": true,
        "kernel_event": "[5.499107] pci 0000:0a:00.0: gpu_unlock: unlocked"
      }
    ],
    "sources": [
      "/proc/modules",
      "/sys/module/gpu_unlock/version",
      "/proc/cmdline",
      "current-boot kernel journal",
      "nvidia-smi PCI device mapping"
    ]
  },
  "selected_run": {
    "id": "20261109T061124Z-68344736",
    "started_at": "2026-11-09T06:11:24Z",
    "status": "complete",
    "required_duration_s": 222.136,
    "optional_duration_s": null,
    "optional_concurrency": false,
    "command": "strata bench run --profile <tmp>/strata-rtx3060-controlled-20261109-uilke4xi/benchmark-profile.json --suite standard-v1 --track custom --track-notes Strata 0.2.12 Control/engine 0.1.58 after the approved signed-feed update; existing IQ3_XXS pack, int8 KV, automatic drafting resolved as the existing standard DFlash2 drafter (not the tuned strata1 drafter). Dedicated loopback server; context raised only in a copied profile from 16384 to 32768 for the frozen 16K input plus output. RTX 3060 only; RTX 3090 off. Both physical GPUs exclusively reserved against other generator jobs, but only RTX 3060 used for inference. No tuning or model download; no before/after comparison. Default route-tail skip optimization remains enabled; raw engine log retained. --quantization IQ3_XXS --custom-pack Existing setup-prepared IQ3_XXS pack; shipped expert-profile.bin; no repacking --out <tmp>/strata-rtx3060-controlled-20261109-uilke4xi/standard-results --output json"
  },
  "protocol": {
    "id": "strata-qwen38-thinking-v1-draft.1",
    "git_commit": "9780a8c27d84f408240ed4b07062ab7275f02f09",
    "profile": "qwen3.8-flash-next-thinking",
    "corpus_sha256": "e75eff8b9feea04f745520b392c4c17d97aefc1ba185e222aa6740efa8f68407",
    "model_card_revision": null
  },
  "hardware": {
    "gpus": [
      {
        "device_index": 1,
        "name": "NVIDIA GeForce RTX 3060",
        "memory_bytes": 12884901888,
        "memory_kind": "dedicated",
        "power_limit_w": 170.0
      }
    ],
    "cpu": "AMD Ryzen 9 5950X 16-Core Processor",
    "ram_bytes": 67347894272,
    "storage": [
      {
        "kind": "nvme",
        "model": "CSSD-M2RF72LT2000GB",
        "bus": "nvme"
      }
    ],
    "background_workloads": null
  },
  "software": {
    "os": {
      "name": "Debian GNU/Linux 13 (trixie)",
      "version": "6.12.116+deb13-amd64"
    },
    "drivers": [
      {
        "name": "gpu-driver",
        "version": "620.70.06"
      }
    ],
    "runtime": [
      {
        "name": "cuda",
        "version": "13.5"
      }
    ],
    "strata_commit": "9780a8c27d84f408240ed4b07062ab7275f02f09",
    "engine_version": "0.1.58",
    "build_kind": "release",
    "runner_commit": "9780a8c27d84f408240ed4b07062ab7275f02f09",
    "build_options": null
  },
  "model": {
    "id": "qwen3.8-flash-next-iq3_xxs",
    "revision": null,
    "source_revision": null,
    "quantization": "IQ3_XXS",
    "files": [
      {
        "path": "Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
        "sha256": "219ea929900dfa9ef091f3aa473fdba6874b65fcb36526d7d851ac9e95856d15"
      },
      {
        "path": "Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf",
        "sha256": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"
      },
      {
        "path": "draft_vocab.bin",
        "sha256": "b1e1d3a7a9e4bf862dcd5923ce661fb59bbd07907e594df5cf86a62ac235cb91"
      },
      {
        "path": "expert-profile.bin",
        "sha256": "8f59b4aa8873209dff11c11e37bcda9529a1335b724a1afeea37bf6388975baf"
      },
      {
        "path": "qwen3.8-flash-next-dflash2.q4_0.pack",
        "sha256": "cda40d62d0eb8f37ed6e272142641a1d2f0de4a32f9e14d0416e7a71534849f1"
      }
    ],
    "tokenizer_revision": null,
    "custom_pack": "Existing setup-prepared IQ3_XXS pack; shipped expert-profile.bin; no repacking"
  },
  "configuration": {
    "track": "custom",
    "track_notes": "Strata 0.2.12 Control/engine 0.1.58 after the approved signed-feed update; existing IQ3_XXS pack, int8 KV, automatic drafting resolved as the existing standard DFlash2 drafter (not the tuned strata1 drafter). Dedicated loopback server; context raised only in a copied profile from 16384 to 32768 for the frozen 16K input plus output. RTX 3060 only; RTX 3090 off. Both physical GPUs exclusively reserved against other generator jobs, but only RTX 3060 used for inference. No tuning or model download; no before/after comparison. Default route-tail skip optimization remains enabled; raw engine log retained.",
    "context_tokens": 32768,
    "server_capacity": 1,
    "kv_cache": {
      "--kv": "int8"
    },
    "expert_cache": {
      "--expert-profile": "expert-profile.bin",
      "--expert-cache": "auto"
    },
    "offloading": {},
    "prefill": {
      "--prefill": "auto"
    },
    "mtp": {
      "--draft": "dflash2",
      "--draft-model": "qwen3.8-flash-next-dflash2.q4_0.pack",
      "--draft-fallback": "none",
      "--mtp-draft-vocab": "draft_vocab.bin",
      "--spec": "6",
      "--draft-conf": "0.8"
    },
    "calibration": {},
    "experimental_options": {},
    "template": {
      "path": "provenance/chat_template.jinja",
      "sha256": "12827f24b742ea4e80cdc12dbcf9622227056b9f797252a3149263d4f9aaadce"
    },
    "sampling": {
      "temperature": 1.0,
      "top_p": 0.95,
      "top_k": 20,
      "min_p": 0.0,
      "presence_penalty": 0.0,
      "repetition_penalty": 1.0,
      "reasoning_effort": "xhigh"
    },
    "chat_template_kwargs": {
      "enable_thinking": true,
      "preserve_thinking": true
    }
  },
  "request_counts": {
    "warmup": {
      "planned": 1,
      "attempted": 1,
      "completed": 0,
      "truncated": 1,
      "failed": 0,
      "skipped": 0
    },
    "performance": {
      "planned": 9,
      "attempted": 9,
      "completed": 2,
      "truncated": 7,
      "failed": 0,
      "skipped": 0
    },
    "sanity": {
      "planned": 12,
      "attempted": 12,
      "completed": 12,
      "truncated": 0,
      "failed": 0,
      "skipped": 0
    },
    "concurrency": {
      "planned": 0,
      "attempted": 0,
      "completed": 0,
      "truncated": 0,
      "failed": 0,
      "skipped": 0
    }
  },
  "summaries": [
    {
      "phase": "performance",
      "case_id": "short",
      "concurrency": 1,
      "metric": "prompt_tps",
      "unit": "tokens/s",
      "median": 101.663586,
      "min": 101.023353,
      "max": 124.979711,
      "n": 3,
      "included_request_ids": [
        "performance-short-r1",
        "performance-short-r2",
        "performance-short-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "short",
      "concurrency": 1,
      "metric": "decode_tps",
      "unit": "tokens/s",
      "median": 56.911101,
      "min": 50.336725,
      "max": 61.898955,
      "n": 3,
      "included_request_ids": [
        "performance-short-r1",
        "performance-short-r2",
        "performance-short-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "short",
      "concurrency": 1,
      "metric": "ttft_s",
      "unit": "s",
      "median": 0.787769,
      "min": 0.64291,
      "max": 0.787918,
      "n": 3,
      "included_request_ids": [
        "performance-short-r1",
        "performance-short-r2",
        "performance-short-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "short",
      "concurrency": 1,
      "metric": "first_answer_s",
      "unit": "s",
      "median": 2.607672,
      "min": 2.12019,
      "max": 7.846768,
      "n": 3,
      "included_request_ids": [
        "performance-short-r1",
        "performance-short-r2",
        "performance-short-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "short",
      "concurrency": 1,
      "metric": "total_s",
      "unit": "s",
      "median": 7.237697,
      "min": 5.350388,
      "max": 10.793635,
      "n": 3,
      "included_request_ids": [
        "performance-short-r1",
        "performance-short-r2",
        "performance-short-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "prompt-4k",
      "concurrency": 1,
      "metric": "prompt_tps",
      "unit": "tokens/s",
      "median": 1010.11846,
      "min": 983.87058,
      "max": 1010.417695,
      "n": 3,
      "included_request_ids": [
        "performance-prompt-4k-r1",
        "performance-prompt-4k-r2",
        "performance-prompt-4k-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "prompt-4k",
      "concurrency": 1,
      "metric": "decode_tps",
      "unit": "tokens/s",
      "median": 51.596779,
      "min": 50.118444,
      "max": 58.285805,
      "n": 3,
      "included_request_ids": [
        "performance-prompt-4k-r1",
        "performance-prompt-4k-r2",
        "performance-prompt-4k-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "prompt-4k",
      "concurrency": 1,
      "metric": "ttft_s",
      "unit": "s",
      "median": 4.080739,
      "min": 4.079926,
      "max": 4.209126,
      "n": 3,
      "included_request_ids": [
        "performance-prompt-4k-r1",
        "performance-prompt-4k-r2",
        "performance-prompt-4k-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "prompt-4k",
      "concurrency": 1,
      "metric": "first_answer_s",
      "unit": "s",
      "median": 11.48394,
      "min": 9.674501,
      "max": 13.29338,
      "n": 2,
      "included_request_ids": [
        "performance-prompt-4k-r1",
        "performance-prompt-4k-r3"
      ],
      "excluded": [
        {
          "request_id": "performance-prompt-4k-r2",
          "reason": "first_answer_s is not available for this request"
        }
      ]
    },
    {
      "phase": "performance",
      "case_id": "prompt-4k",
      "concurrency": 1,
      "metric": "total_s",
      "unit": "s",
      "median": 13.980773,
      "min": 12.838881,
      "max": 14.399848,
      "n": 3,
      "included_request_ids": [
        "performance-prompt-4k-r1",
        "performance-prompt-4k-r2",
        "performance-prompt-4k-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "prompt-16k",
      "concurrency": 1,
      "metric": "prompt_tps",
      "unit": "tokens/s",
      "median": 1228.594992,
      "min": 1227.959745,
      "max": 1233.849351,
      "n": 3,
      "included_request_ids": [
        "performance-prompt-16k-r1",
        "performance-prompt-16k-r2",
        "performance-prompt-16k-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "prompt-16k",
      "concurrency": 1,
      "metric": "decode_tps",
      "unit": "tokens/s",
      "median": 49.102819,
      "min": 49.100935,
      "max": 62.024519,
      "n": 3,
      "included_request_ids": [
        "performance-prompt-16k-r1",
        "performance-prompt-16k-r2",
        "performance-prompt-16k-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "prompt-16k",
      "concurrency": 1,
      "metric": "ttft_s",
      "unit": "s",
      "median": 13.372609,
      "min": 13.318078,
      "max": 13.37857,
      "n": 3,
      "included_request_ids": [
        "performance-prompt-16k-r1",
        "performance-prompt-16k-r2",
        "performance-prompt-16k-r3"
      ],
      "excluded": []
    },
    {
      "phase": "performance",
      "case_id": "prompt-16k",
      "concurrency": 1,
      "metric": "first_answer_s",
      "unit": "s",
      "median": 17.184643,
      "min": 17.184643,
      "max": 17.184643,
      "n": 1,
      "included_request_ids": [
        "performance-prompt-16k-r1"
      ],
      "excluded": [
        {
          "request_id": "performance-prompt-16k-r2",
          "reason": "first_answer_s is not available for this request"
        },
        {
          "request_id": "performance-prompt-16k-r3",
          "reason": "first_answer_s is not available for this request"
        }
      ]
    },
    {
      "phase": "performance",
      "case_id": "prompt-16k",
      "concurrency": 1,
      "metric": "total_s",
      "unit": "s",
      "median": 23.719829,
      "min": 21.606041,
      "max": 23.781641,
      "n": 3,
      "included_request_ids": [
        "performance-prompt-16k-r1",
        "performance-prompt-16k-r2",
        "performance-prompt-16k-r3"
      ],
      "excluded": []
    }
  ],
  "quality": {
    "passed": 12,
    "total": 12,
    "categories": {
      "instruction": {
        "passed": 4,
        "total": 4
      },
      "retrieval": {
        "passed": 4,
        "total": 4
      },
      "tool_call": {
        "passed": 4,
        "total": 4
      }
    }
  },
  "memory_observation": {
    "window": "required request window, including warmup; excluding model load and hashing",
    "sampling": "approximately 2 seconds; nvidia-smi and psutil.virtual_memory",
    "sample_count": 108,
    "rtx3060_peak_used_mib": 11300.0,
    "rtx3090_peak_used_mib": 11.0,
    "system_used_peak_bytes": 53346893824,
    "system_available_min_bytes": 14001000448,
    "swap_used_min_bytes": 8913633280,
    "swap_used_max_bytes": 9167650816,
    "swap_in_bytes": 342142976,
    "swap_out_bytes": 0,
    "baseline_system_used_bytes": 6207856640,
    "baseline_rtx3060_used_mib": 64.0
  },
  "excluded_runs": [
    {
      "run_id": "20261109T060203Z-e7d63495",
      "selected_for_headline": false,
      "reason": "Concurrent non-benchmark RTX 3090 workload and active system swapping were observed. Do not attribute these timings or whole-system RAM peaks to isolated RTX 3060 inference.",
      "declaration_correction": "The original operator track_notes incorrectly describe MTP. The effective engine configuration and engine.log show the standard DFlash2 drafter, spec=6, draft-conf=0.8, with no MTP layer loaded. Original generated files are retained byte-for-byte rather than rewritten.",
      "resource_observation": {
        "window": "required request window, including warmup; excluding model load and hashing",
        "sampling": "approximately 2 seconds; nvidia-smi and psutil.virtual_memory",
        "sample_count": 110,
        "rtx3060_peak_used_mib": 11300.0,
        "rtx3090_peak_used_mib": 13421.0,
        "system_used_peak_bytes": 60758794240,
        "system_available_min_bytes": 6589100032,
        "swap_used_min_bytes": 8435798016,
        "swap_used_max_bytes": 15757467648,
        "swap_in_bytes": 5591121920,
        "swap_out_bytes": 7922106368,
        "baseline_system_used_bytes": 7657209856,
        "baseline_rtx3060_used_mib": 64.0
      }
    }
  ],
  "evidence": {
    "archive": "strata-benchmark-evidence.zip",
    "selected_report_path": "primary-run/results.json",
    "selected_report_sha256": "93fa29847333d44857348e1b4185dacb3897fc3529bc8b43c4380f4e654d0fd4"
  }
}