{
  "title": "Historical configured candidate inputs for the August 2026 six-configuration comparison",
  "evidence_boundary": "These are exact values from retained Git configuration before production promotion. They are not a run manifest and do not prove the effective settings, loaded checkpoint or cache state for any measured request. No current engine versions were substituted.",
  "source_snapshot": {
    "repository": "argocd",
    "git_revision": "dbffb69dc3b860946956286ffed30fab94fb4647",
    "commit_time": "2026-08-09T17:09:56+02:00",
    "files": [
      {
        "path": "sys/vllm/model-lock.yaml",
        "sha256": "baad607a75b6ce9ce8cf6dfcfb35644510c096603338b3254c07601ff87e56d3"
      },
      {
        "path": "sys/llamacpp/model-lock.yaml",
        "sha256": "48b60f1806838d5c72e81f8c1733abcc62be320df0f77fcfc908d9ea96130718"
      },
      {
        "path": "sys/vllm/second-brain-brain-canary.values.yaml",
        "sha256": "83ce27537a87001a6fd32ae768d2177e808d19e4eeeaa37e1d49f9f94fe1378a"
      },
      {
        "path": "sys/llamacpp/second-brain-brain-lanes.values.yaml",
        "sha256": "1953b4297f4e46afa5930240e6a4943517f71ab200fd1b2b516f69ab39be5af9"
      },
      {
        "path": "sys/llamacpp/values.yaml",
        "sha256": "de40b5d0d8d7daf2050b6a5794683b93407dbf63bb7a0de944b451fd67336a82"
      },
      {
        "path": "sys/llamacpp/prod.yaml",
        "sha256": "b1e98f8dea669c7663f0b796404c707591ce3d71a1906810fbf68853f700f284"
      },
      {
        "path": ".argo/sys/templates/vllm.yaml",
        "sha256": "839fd65f9a6f169e3a705326e943ca90950ac63f443a28732a8349b9d376ed7a"
      },
      {
        "path": ".argo/sys/templates/llamacpp.yaml",
        "sha256": "5af641c831965afb80e88de7dd50fb21f5bf15d0b2c2ea102741f746148f5767"
      }
    ]
  },
  "unavailable_for_this_review": {
    "exact_prompt": null,
    "original_harness": null,
    "per_request_results": null,
    "per_request_effective_launch_settings": null,
    "warmup_procedure": null,
    "cache_contents_or_resets_between_requests": null,
    "explanation": "The article evidence ledger records an earlier check of the original harness and six result files. Those artifacts were not available in this checkout for export; absence here does not establish that they were never retained."
  },
  "configured_engines": {
    "vllm": {
      "image_tag": "v0.26.0",
      "image_digest": "sha256:ffb2d59b1c059a5bd8d781320c9f5189de8293693b7d95da54befddaa54abf52",
      "image_repository": "Private mirror repository omitted."
    },
    "llama.cpp": {
      "image_tag": "server-cuda",
      "image_digest": "sha256:9b17daa3579c05543a0e7c87f3af92f05434b9f3d1608b45ae9ec855130eb8ac",
      "image_repository": "Private mirror repository omitted.",
      "source_commit": null
    }
  },
  "configured_candidates": [
    {
      "engine": "vllm",
      "checkpoint": {
        "repository": "nvidia/Qwen3.6-35B-A3B-NVFP4",
        "revision": "491c2f1ea524c639598bf8fa787a93fed5a6fbce"
      },
      "helm_values": {
        "maxModelLen": 32768,
        "gpuMemoryUtilization": "0.82",
        "dtype": "auto",
        "extraArgs": [
          "--quantization",
          "modelopt",
          "--kv-cache-dtype",
          "fp8",
          "--max-num-seqs",
          "1",
          "--enable-chunked-prefill",
          "--enable-prefix-caching",
          "--reasoning-parser",
          "qwen3",
          "--enable-auto-tool-choice",
          "--tool-call-parser",
          "qwen3_coder",
          "--language-model-only"
        ]
      }
    },
    {
      "engine": "llama.cpp",
      "checkpoint": {
        "repository": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "filename": "Qwen3.6-35B-A3B-Q8_0.gguf",
        "sha256": "d1a395809f65a43a13ad119eb4e7acdef1ac6d68120f39902c8ab96e72794a59",
        "size_bytes": 36903140320
      },
      "helm_values": {
        "ctxSize": 65536,
        "commonArgs_performance_subset": [
          "--cache-type-k",
          "q8_0",
          "--cache-type-v",
          "q8_0",
          "--n-gpu-layers",
          "999",
          "--flash-attn",
          "on",
          "--parallel",
          "1",
          "--jinja"
        ],
        "extraArgs": []
      }
    },
    {
      "engine": "vllm",
      "checkpoint": {
        "repository": "zai-org/GLM-4.7-Flash",
        "revision": "7dd20894a642a0aa287e9827cb1a1f7f91386b67"
      },
      "helm_values": {
        "maxModelLen": 32768,
        "gpuMemoryUtilization": "0.80",
        "dtype": "auto",
        "extraArgs": [
          "--max-num-seqs",
          "1",
          "--enable-chunked-prefill",
          "--enable-prefix-caching",
          "--enable-auto-tool-choice",
          "--tool-call-parser",
          "glm47",
          "--reasoning-parser",
          "glm47"
        ]
      }
    },
    {
      "engine": "llama.cpp",
      "checkpoint": {
        "repository": "t-tech/T-pro-it-2.1-GGUF",
        "revision": "cf18787d8901c26ca9334812e7ff8070f2ddab4a",
        "filename": "T-pro-it-2.1-Q8_0.gguf",
        "sha256": "af2b3a761473a95ca31ab9b8647b9dc4288604a4a71f795ffd9402d2938c7621",
        "size_bytes": 34816397344
      },
      "helm_values": {
        "ctxSize": 40960,
        "commonArgs_performance_subset": [
          "--cache-type-k",
          "q8_0",
          "--cache-type-v",
          "q8_0",
          "--n-gpu-layers",
          "999",
          "--flash-attn",
          "on",
          "--parallel",
          "1",
          "--jinja"
        ],
        "extraArgs": []
      }
    },
    {
      "engine": "vllm",
      "checkpoint": {
        "repository": "google/gemma-4-31B-it",
        "revision": "842da3794eaa0b77d5f08bae87a17459d91ff475"
      },
      "helm_values": {
        "maxModelLen": 32768,
        "gpuMemoryUtilization": "0.88",
        "dtype": "auto",
        "extraArgs": [
          "--max-num-seqs",
          "1",
          "--enable-chunked-prefill",
          "--enable-prefix-caching",
          "--enable-auto-tool-choice",
          "--tool-call-parser",
          "gemma4"
        ]
      }
    },
    {
      "engine": "vllm",
      "checkpoint": {
        "repository": "nvidia/Qwen3.5-122B-A10B-NVFP4",
        "revision": "98915d837c4e7c87ac8296d02e89de19b3207e6d"
      },
      "helm_values": {
        "maxModelLen": 32768,
        "gpuMemoryUtilization": "0.90",
        "dtype": "auto",
        "extraArgs": [
          "--quantization",
          "modelopt",
          "--kv-cache-dtype",
          "fp8",
          "--max-num-seqs",
          "1",
          "--enable-chunked-prefill",
          "--enable-prefix-caching",
          "--reasoning-parser",
          "qwen3",
          "--enable-auto-tool-choice",
          "--tool-call-parser",
          "qwen3_coder",
          "--language-model-only"
        ]
      }
    }
  ],
  "interpretation": [
    "The article reports small sequential smoke measurements from the earlier host configuration, not concurrent capacity or results for the consolidated ASUS host.",
    "Its rate is completion tokens divided by the whole streaming request duration, including time before the first token.",
    "The vLLM candidate definitions explicitly enable prefix caching. With a repeated prompt, reuse could affect the prefill contribution to the reported rate; actual cache state and per-request cache hits are unavailable here.",
    "The llama.cpp definitions do not explicitly establish comparable prefix-cache conditions or a warm-up procedure.",
    "Configuration values are transcribed without infrastructure endpoints, workload aliases or filesystem paths. Git source paths above identify the retained configuration files."
  ]
}
