{
  "schema": "anvil-serving.recipe-feasibility-input/v1",
  "campaign": "2026-09-19-qwen38-huihui-ninfer",
  "requirements": {
    "tokens": {
      "prompt": {
        "value": 5888,
        "unit": "tokens",
        "status": "assumed",
        "source": "scout requirement",
        "observed_at": "2026-09-19",
        "notes": ""
      },
      "output_reserve": {
        "value": 2048,
        "unit": "tokens",
        "status": "assumed",
        "source": "scout requirement",
        "observed_at": "2026-09-19",
        "notes": ""
      },
      "protocol": {
        "value": 256,
        "unit": "tokens",
        "status": "assumed",
        "source": "scout requirement",
        "observed_at": "2026-09-19",
        "notes": ""
      }
    },
    "concurrency": {
      "value": 1,
      "unit": "requests",
      "status": "assumed",
      "source": "scout requirement",
      "observed_at": "2026-09-19",
      "notes": ""
    },
    "physical_vram_bytes": {
      "value": 34190917632,
      "unit": "bytes",
      "status": "measured",
      "source": "local nvidia-smi",
      "observed_at": "2026-09-19",
      "notes": ""
    },
    "vram_reserves": {
      "desktop": {
        "kind": "physical",
        "variable": {
          "value": 1175453696,
          "unit": "bytes",
          "status": "measured",
          "source": "total minus free VRAM before stop",
          "observed_at": "2026-09-19",
          "notes": ""
        }
      },
      "headroom": {
        "kind": "policy",
        "variable": {
          "value": 3221225472,
          "unit": "bytes",
          "status": "assumed",
          "source": "campaign default reserve",
          "observed_at": "2026-09-19",
          "notes": ""
        }
      }
    },
    "thresholds": {
      "min_deterministic_pass_rate": {
        "value": 1,
        "unit": "ratio",
        "status": "assumed",
        "source": "hard gate",
        "observed_at": "2026-09-19",
        "notes": ""
      }
    }
  },
  "candidates": [
    {
      "id": "huihui-8k-c1-scout",
      "runtime_context_limit_tokens": {
        "value": 8192,
        "unit": "tokens",
        "status": "assumed",
        "source": "scout recipe",
        "observed_at": "2026-09-19",
        "notes": ""
      },
      "resident_components": {
        "target": {
          "min": 19327352832,
          "max": 22548578304,
          "unit": "bytes",
          "status": "estimated",
          "source": "producer GPU weights 18.98 GiB text / 19.25 GiB vision; conservative interval",
          "observed_at": "2026-09-19"
        },
        "runtime": {
          "min": 1073741824,
          "max": 5368709120,
          "unit": "bytes",
          "status": "estimated",
          "source": "prior same-GPU NInfer total at much larger context; planning bound requires local confirmation",
          "observed_at": "2026-09-19"
        }
      },
      "per_token_components": {
        "kv": {
          "min": 8192,
          "max": 16384,
          "unit": "bytes/token",
          "status": "estimated",
          "source": "prior NInfer hybrid attention planning interval",
          "observed_at": "2026-09-19"
        }
      },
      "per_sequence_components": {},
      "kv_token_multiplier": {
        "value": 1,
        "unit": "copies",
        "status": "assumed",
        "source": "C1",
        "observed_at": "2026-09-19",
        "notes": ""
      },
      "metrics": {},
      "hard_failures": []
    }
  ]
}
