{
  "schema": "anvil-serving.recipe-feasibility-input/v1",
  "campaign": "glm53-k3-xgrammar-524k-dual-pro-c2",
  "requirements": {
    "tokens": {
      "prompt": {
        "value": 250000,
        "unit": "tokens",
        "status": "assumed",
        "source": "operator-approved 250K through 600K replacement envelope",
        "observed_at": "2026-08-31"
      },
      "output_reserve": {
        "value": 8192,
        "unit": "tokens",
        "status": "assumed",
        "source": "router-visible reasoning and answer reserve",
        "observed_at": "2026-08-31"
      }
    },
    "concurrency": {
      "value": 2,
      "unit": "sequences",
      "status": "assumed",
      "source": "operator-requested minimum useful concurrency",
      "observed_at": "2026-08-31"
    },
    "physical_vram_bytes": {
      "value": 205283917824,
      "unit": "bytes",
      "status": "measured",
      "source": "two local RTX PRO 6000 Blackwell Max-Q cards at 97887 MiB each",
      "observed_at": "2026-08-31"
    },
    "vram_reserves": {
      "idle_gpu_allocations": {
        "kind": "physical",
        "variable": {
          "value": 39845888,
          "unit": "bytes",
          "status": "measured",
          "source": "retained isolated-lane 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": "2026-08-31"
        }
      }
    },
    "thresholds": {
      "min_deterministic_pass_rate": {
        "value": 1.0,
        "unit": "ratio",
        "status": "assumed",
        "source": "identity, structured output, tools, context, quality, and stability hard gate",
        "observed_at": "2026-08-31"
      },
      "max_relative_quality_loss": {
        "value": 0.03,
        "unit": "ratio",
        "status": "assumed",
        "source": "maximum acceptable loss against the current qualified K3 control",
        "observed_at": "2026-08-31"
      },
      "min_warm_e2e_gain": {
        "value": 0.0,
        "unit": "ratio",
        "status": "assumed",
        "source": "operator requires a positive matched throughput improvement; absolute lane gate is greater than 67.9 output tokens per second near 240K",
        "observed_at": "2026-08-31"
      },
      "min_tasks_per_hour_ratio": {
        "value": 1.0,
        "unit": "ratio",
        "status": "assumed",
        "source": "replacement may not reduce successful agent work per hour",
        "observed_at": "2026-08-31"
      }
    }
  },
  "tracked_variables": {
    "current_spec_kv_capacity_tokens": {
      "value": 2917371,
      "unit": "tokens",
      "status": "measured",
      "source": "current exact K3 plus DFlash2 K5 startup telemetry",
      "observed_at": "2026-08-30",
      "notes": "The corrected image changes only structured-output control flow, not model, draft, KV dtype, or allocation geometry."
    },
    "runtime_patch_source_hashes": {
      "status": "confirmed",
      "unit": "sha256",
      "source": "hash-gated derived-image Dockerfile",
      "observed_at": "2026-08-31",
      "notes": "Pre-patch hashes are enforced in the derived-image Dockerfile; post-patch hashes are enforced after application."
    }
  },
  "candidates": [
    {
      "id": "glm53-k3-524k-nospec-xgrammar-fixed",
      "runtime_context_limit_tokens": {
        "value": 524288,
        "unit": "tokens",
        "status": "confirmed",
        "source": "managed candidate recipe",
        "observed_at": "2026-08-31"
      },
      "resident_components": {
        "target_runtime_and_graph_conservative_proxy": {
          "min": 158000000000,
          "max": 165000000000,
          "unit": "bytes",
          "status": "estimated",
          "source": "conservative carry-forward of the larger locally proven K3 plus DFlash2 resident envelope",
          "observed_at": "2026-08-31"
        }
      },
      "per_sequence_components": {},
      "per_token_components": {
        "request_cache_conservative_proxy": {
          "min": 12000,
          "max": 13000,
          "unit": "bytes/token",
          "status": "estimated",
          "source": "conservative carry-forward of the locally proven target-plus-draft request-cache density",
          "observed_at": "2026-08-31"
        }
      },
      "kv_token_multiplier": {
        "value": 2,
        "unit": "copies",
        "status": "assumed",
        "source": "two independent required sequences",
        "observed_at": "2026-08-31"
      },
      "metrics": {},
      "hard_failures": []
    },
    {
      "id": "glm53-k3-524k-dflash2-k5-xgrammar-fixed",
      "runtime_context_limit_tokens": {
        "value": 524288,
        "unit": "tokens",
        "status": "confirmed",
        "source": "managed candidate recipe",
        "observed_at": "2026-08-31"
      },
      "measured_max_stable_context_tokens": {
        "value": 950000,
        "unit": "tokens",
        "status": "measured",
        "source": "current exact K3 plus DFlash2 K5 local retrieval gate",
        "observed_at": "2026-08-30",
        "notes": "Same target, draft, KV dtypes, and allocation geometry; corrected runtime still requires local requalification."
      },
      "resident_components": {
        "target_draft_runtime_and_graph_proxy": {
          "min": 158000000000,
          "max": 165000000000,
          "unit": "bytes",
          "status": "estimated",
          "source": "locally proven K3 plus DFlash2 resident envelope",
          "observed_at": "2026-08-30"
        }
      },
      "per_sequence_components": {},
      "per_token_components": {
        "target_draft_request_cache_proxy": {
          "min": 12000,
          "max": 13000,
          "unit": "bytes/token",
          "status": "estimated",
          "source": "local K3 plus DFlash2 request-cache density proxy",
          "observed_at": "2026-08-30"
        }
      },
      "kv_token_multiplier": {
        "value": 2,
        "unit": "copies",
        "status": "assumed",
        "source": "two independent required sequences",
        "observed_at": "2026-08-31"
      },
      "metrics": {},
      "hard_failures": []
    }
  ]
}