{
  "schema": "anvil-serving.recipe-feasibility-input/v1",
  "campaign": "glm53-k3-dflash2-dual-pro6000-wsl2-default",
  "requirements": {
    "tokens": {
      "prompt": {
        "value": 250000,
        "unit": "tokens",
        "status": "assumed",
        "source": "one-week default long-context requirement",
        "observed_at": "2026-08-29"
      },
      "output_reserve": {
        "value": 8192,
        "unit": "tokens",
        "status": "assumed",
        "source": "router-visible reasoning and answer reserve",
        "observed_at": "2026-08-29"
      }
    },
    "concurrency": {
      "value": 2,
      "unit": "sequences",
      "status": "assumed",
      "source": "minimum useful Hermes default concurrency",
      "observed_at": "2026-08-29"
    },
    "physical_vram_bytes": {
      "value": 205283917824,
      "unit": "bytes",
      "status": "measured",
      "source": "local two-card inventory: 97887 MiB per RTX PRO 6000 Blackwell Max-Q",
      "observed_at": "2026-08-29"
    },
    "vram_reserves": {
      "idle_gpu_allocations": {
        "kind": "physical",
        "variable": {
          "value": 39845888,
          "unit": "bytes",
          "status": "measured",
          "source": "local pre-load idle allocation: 19 MiB per card",
          "observed_at": "2026-08-29"
        }
      },
      "five_percent_runtime_headroom": {
        "kind": "policy",
        "variable": {
          "value": 10224350003,
          "unit": "bytes",
          "status": "assumed",
          "source": "remainder of the pinned 0.950 GPU-memory ceiling after idle allocation",
          "observed_at": "2026-08-29"
        }
      }
    },
    "thresholds": {
      "min_deterministic_pass_rate": {
        "value": 1.0,
        "unit": "ratio",
        "status": "assumed",
        "source": "identity, retrieval, tools, image/OCR, and stability hard gate",
        "observed_at": "2026-08-29"
      },
      "max_relative_quality_loss": {
        "value": 0.03,
        "unit": "ratio",
        "status": "assumed",
        "source": "maximum acceptable loss against the current locally qualified 4 bpw GLM control",
        "observed_at": "2026-08-29"
      },
      "min_warm_e2e_gain": {
        "value": 0.25,
        "unit": "ratio",
        "status": "assumed",
        "source": "minimum worthwhile speed gain for changing the default quant/runtime",
        "observed_at": "2026-08-29"
      },
      "min_tasks_per_hour_ratio": {
        "value": 1.0,
        "unit": "ratio",
        "status": "assumed",
        "source": "new default may not reduce successful agent work per hour",
        "observed_at": "2026-08-29"
      }
    }
  },
  "tracked_variables": {
    "k3_checkpoint_payload_bytes": {
      "value": 137123823211,
      "unit": "bytes",
      "status": "confirmed",
      "source": "Hugging Face blob metadata at target revision 319d66a8b53092b491f698440ecea781e4ddd4e4",
      "observed_at": "2026-08-29",
      "notes": "Storage payload is not treated as resident VRAM."
    },
    "dflash2_checkpoint_payload_bytes": {
      "value": 2342169800,
      "unit": "bytes",
      "status": "confirmed",
      "source": "Hugging Face blob metadata at draft revision dc77ff1c99eeb2df044ee3d4f0094eb033fee410",
      "observed_at": "2026-08-29"
    },
    "wsl2_collective_compatibility": {
      "status": "unknown",
      "unit": "boolean",
      "notes": "Upstream qualification is native Linux; the managed local load must prove the PCIe collective paths under WSL2."
    }
  },
  "candidates": [
    {
      "id": "purtell-k3-dflash2-k5-fp8-1m-vision",
      "runtime_context_limit_tokens": {
        "value": 1048576,
        "unit": "tokens",
        "status": "confirmed",
        "source": "Purtell release d46fdeddf8c6fec2d4595b65535a32d80a5af787",
        "observed_at": "2026-08-29"
      },
      "measured_max_stable_context_tokens": {
        "value": 1000000,
        "unit": "tokens",
        "status": "confirmed",
        "source": "external cold six-needle result on two RTX PRO 6000 workstation GPUs",
        "observed_at": "2026-08-29",
        "notes": "External hardware-matched prior; local qualification remains required."
      },
      "resident_components": {
        "target_draft_runtime_and_graph_proxy": {
          "min": 158000000000,
          "max": 165000000000,
          "unit": "bytes",
          "status": "estimated",
          "source": "derived from the external 15.88 GiB per-card cache remainder at the 0.950 ceiling, widened for Max-Q and WSL2 runtime variance",
          "observed_at": "2026-08-29"
        }
      },
      "per_sequence_components": {},
      "per_token_components": {
        "heterogeneous_target_draft_request_cache_proxy": {
          "min": 12000,
          "max": 13000,
          "unit": "bytes/token",
          "status": "estimated",
          "source": "external 31.76 GiB aggregate cache divided by 2786881 full-request-equivalent tokens, widened for local variance",
          "observed_at": "2026-08-29"
        }
      },
      "kv_token_multiplier": {
        "value": 2,
        "unit": "copies",
        "status": "assumed",
        "source": "two independent maximum-requirement sequences",
        "observed_at": "2026-08-29"
      },
      "metrics": {
        "deterministic_pass_rate": {"status": "unknown", "unit": "ratio", "notes": "Run the local full gate."},
        "quality_score": {"status": "unknown", "unit": "score", "notes": "Run the matched local high-reasoning suite."},
        "reference_quality_score": {"status": "unknown", "unit": "score", "notes": "Populate from the current 4 bpw control."},
        "warm_e2e_seconds": {"status": "unknown", "unit": "seconds", "notes": "Measure locally."},
        "no_spec_warm_e2e_seconds": {"status": "unknown", "unit": "seconds", "notes": "Use the current fixed-K5 control for the selection comparison."},
        "successful_tasks_per_hour": {"status": "unknown", "unit": "tasks/hour", "notes": "Measure local successful work."},
        "reference_tasks_per_hour": {"status": "unknown", "unit": "tasks/hour", "notes": "Populate from the current control."}
      },
      "hard_failures": []
    },
    {
      "id": "brandon-v84-k4-dflash2-vision-98k",
      "runtime_context_limit_tokens": {
        "value": 98304,
        "unit": "tokens",
        "status": "confirmed",
        "source": "Brandon runtime release 24784d718cb7bbb99feaee2050eaddcfbc913386",
        "observed_at": "2026-08-29"
      },
      "resident_components": {},
      "per_sequence_components": {},
      "per_token_components": {
        "unmodeled_kv": {"status": "unknown", "unit": "bytes/token", "notes": "Context-limit rejection does not require a KV estimate."}
      },
      "kv_token_multiplier": {"value": 2, "unit": "copies", "status": "assumed", "source": "two independent sequences", "observed_at": "2026-08-29"},
      "metrics": {},
      "hard_failures": []
    },
    {
      "id": "cardillo-k4-adaptive-mtp-replayssm-262k",
      "runtime_context_limit_tokens": {
        "value": 262144,
        "unit": "tokens",
        "status": "confirmed",
        "source": "Cardillo release 5b5623ea07f48683f37f3774d8d5b8bf5b04fdf0",
        "observed_at": "2026-08-29"
      },
      "resident_components": {},
      "per_sequence_components": {},
      "per_token_components": {
        "unmodeled_kv": {"status": "unknown", "unit": "bytes/token", "notes": "The measured behavioral failure prunes this profile before memory modeling."}
      },
      "kv_token_multiplier": {"value": 2, "unit": "copies", "status": "assumed", "source": "two independent sequences", "observed_at": "2026-08-29"},
      "metrics": {},
      "hard_failures": ["Local repeated structured-tool gate passed only 12/20 and produced degenerate repeated output."]
    },
    {
      "id": "0xsero-glm53-flash-exl3-3bpw",
      "runtime_context_limit_tokens": {
        "status": "unknown",
        "unit": "tokens",
        "notes": "The publisher did not provide a complete supported TP=2 serve path."
      },
      "resident_components": {},
      "per_sequence_components": {},
      "per_token_components": {
        "unmodeled_kv": {"status": "unknown", "unit": "bytes/token", "notes": "The measured quality/runtime-path failures prune this quant before memory modeling."}
      },
      "kv_token_multiplier": {"value": 2, "unit": "copies", "status": "assumed", "source": "two independent sequences", "observed_at": "2026-08-29"},
      "metrics": {},
      "hard_failures": ["Publisher-held-out quality evaluation rejected the 3.0 bpw quant; no complete supported dual-GPU server path was published."]
    }
  ]
}