{
  "schema": "anvil-serving.recipe-feasibility-input/v1",
  "campaign": "glm53-cardillo-adaptive-mtp-dual-sm120-wsl2",
  "requirements": {
    "tokens": {
      "prompt": {
        "value": 250000,
        "unit": "tokens",
        "status": "assumed",
        "source": "campaign long-context requirement",
        "observed_at": "2026-08-29",
        "notes": "Usable prompt budget; protocol, system, tool, and media tokens must fit inside this total."
      },
      "output_reserve": {
        "value": 8192,
        "unit": "tokens",
        "status": "assumed",
        "source": "campaign reasoning and answer reserve",
        "observed_at": "2026-08-29"
      }
    },
    "concurrency": {
      "value": 1,
      "unit": "sequences",
      "status": "assumed",
      "source": "single-agent qualification lane",
      "observed_at": "2026-08-29"
    },
    "physical_vram_bytes": {
      "value": 205283917824,
      "unit": "bytes",
      "status": "measured",
      "source": "local nvidia-smi: two 97887 MiB RTX PRO 6000 Blackwell Max-Q GPUs",
      "observed_at": "2026-08-29"
    },
    "vram_reserves": {
      "idle_gpu_allocations": {
        "kind": "physical",
        "variable": {
          "value": 39845888,
          "unit": "bytes",
          "status": "measured",
          "source": "local pre-mutation nvidia-smi: 19 MiB used on each GPU",
          "observed_at": "2026-08-29"
        }
      },
      "driver_and_desktop_headroom": {
        "kind": "policy",
        "variable": {
          "value": 6442450944,
          "unit": "bytes",
          "status": "assumed",
          "source": "campaign safe-operating policy: 3 GiB per GPU",
          "observed_at": "2026-08-29"
        }
      },
      "allocator_and_wsl2_uncertainty": {
        "kind": "policy",
        "variable": {
          "value": 3781899059,
          "unit": "bytes",
          "status": "assumed",
          "source": "remainder of the upstream 0.950 GPU-memory ceiling after idle allocation and 3 GiB per-card headroom",
          "observed_at": "2026-08-29",
          "notes": "The three reserves total five percent of physical VRAM. The runtime must not be raised above its externally qualified 0.950 ceiling."
        }
      }
    },
    "thresholds": {
      "min_deterministic_pass_rate": {
        "value": 1.0,
        "unit": "ratio",
        "status": "assumed",
        "source": "qualification hard gate for identity, exact retrieval, tools, JSON, and stability",
        "observed_at": "2026-08-29"
      },
      "max_relative_quality_loss": {
        "value": 0.03,
        "unit": "ratio",
        "status": "assumed",
        "source": "campaign intelligence-loss budget versus the retained Qwen3.8 Flash-Next baseline",
        "observed_at": "2026-08-29"
      },
      "min_warm_e2e_gain": {
        "value": 0.15,
        "unit": "ratio",
        "status": "assumed",
        "source": "minimum adaptive-MTP gain over an otherwise matched MTP-off control",
        "observed_at": "2026-08-29"
      },
      "min_tasks_per_hour_ratio": {
        "value": 0.65,
        "unit": "ratio",
        "status": "assumed",
        "source": "minimum useful coding-agent speed ratio versus the retained Qwen3.8 Flash-Next baseline",
        "observed_at": "2026-08-29",
        "notes": "The external adaptive decode claim is about 0.69 of the local Qwen short-context decode baseline; local tasks per hour remain unknown."
      }
    }
  },
  "tracked_variables": {
    "checkpoint_weight_bytes": {
      "value": 175642157752,
      "unit": "bytes",
      "status": "confirmed",
      "source": "Hugging Face LFS metadata for 120 safetensors at revision 5ab363a8dcf6405955fd5f99671e01a1c9fb124b",
      "observed_at": "2026-08-29",
      "notes": "Artifact bytes are not treated as resident GPU allocation."
    },
    "wsl2_runtime_compatibility": {
      "status": "unknown",
      "unit": "boolean",
      "notes": "The pinned runtime is externally qualified on native Linux, not this WSL2/Docker Desktop stack."
    },
    "system_ram_peak_bytes": {
      "status": "unknown",
      "unit": "bytes",
      "notes": "Measure during load, graph capture, 250K prefill, and unload. No CPU KV offload is planned."
    }
  },
  "candidates": [
    {
      "id": "cardillo-exact-vision-adaptive-mtp5-replayssm10-262k",
      "runtime_context_limit_tokens": {
        "value": 262144,
        "unit": "tokens",
        "status": "confirmed",
        "source": "samuelcardillo integration commit 5b5623ea07f48683f37f3774d8d5b8bf5b04fdf0",
        "observed_at": "2026-08-29"
      },
      "resident_components": {
        "native_linux_non_kv_allocation_proxy": {
          "min": 188505512660,
          "max": 191762889926,
          "unit": "bytes",
          "status": "estimated",
          "source": "derived from the external 545259-token vision KV pool at the 0.950 ceiling and the bounded KV-density proxy",
          "observed_at": "2026-08-29",
          "notes": "Aggregates weights, vision tensors, graphs, ReplaySSM state, and runtime workspace because the external logs do not decompose them. Replace with local startup evidence."
        }
      },
      "per_sequence_components": {},
      "per_token_components": {
        "nvfp4_mla_and_mtp_kv_proxy": {
          "min": 5973,
          "max": 11947,
          "unit": "bytes/token",
          "status": "estimated",
          "source": "bounded from the reported approximately 1.05 GiB vision allocation and 188744-token difference between external text and vision KV pools",
          "observed_at": "2026-08-29",
          "notes": "The lower bound treats 1.05 GiB as aggregate; the upper bound treats it as per-GPU replicated allocation."
        }
      },
      "kv_token_multiplier": {
        "value": 1,
        "unit": "copies",
        "status": "assumed",
        "source": "concurrency-one independent sequence",
        "observed_at": "2026-08-29"
      },
      "metrics": {
        "deterministic_pass_rate": {
          "status": "unknown",
          "unit": "ratio",
          "notes": "External tests are priors; run local independent gates."
        },
        "quality_score": {
          "status": "unknown",
          "unit": "score",
          "notes": "Run the retained local coding-agent suite."
        },
        "reference_quality_score": {
          "status": "unknown",
          "unit": "score",
          "notes": "Populate from the matched retained Qwen baseline metric."
        },
        "warm_e2e_seconds": {
          "status": "unknown",
          "unit": "seconds",
          "notes": "Measure locally after warmup."
        },
        "no_spec_warm_e2e_seconds": {
          "status": "unknown",
          "unit": "seconds",
          "notes": "Requires an otherwise matched MTP-off control."
        },
        "successful_tasks_per_hour": {
          "status": "unknown",
          "unit": "tasks/hour",
          "notes": "Compute from successful end-to-end local tasks."
        },
        "reference_tasks_per_hour": {
          "status": "unknown",
          "unit": "tasks/hour",
          "notes": "Populate from the retained Qwen baseline."
        }
      },
      "hard_failures": []
    },
    {
      "id": "tpurtell-text-only-adaptive-mtp5-replayssm10-262k",
      "runtime_context_limit_tokens": {
        "value": 262144,
        "unit": "tokens",
        "status": "confirmed",
        "source": "pinned tpurtell runtime and Cardillo language-model-only comparison profile",
        "observed_at": "2026-08-29"
      },
      "resident_components": {
        "native_linux_non_kv_allocation_proxy": {
          "min": 186250588092,
          "max": 190635522014,
          "unit": "bytes",
          "status": "estimated",
          "source": "derived from the external 734003-token text-only KV pool at the 0.950 ceiling and the bounded KV-density proxy",
          "observed_at": "2026-08-29",
          "notes": "Aggregates weights, graphs, ReplaySSM state, and runtime workspace; replace with local startup evidence."
        }
      },
      "per_sequence_components": {},
      "per_token_components": {
        "nvfp4_mla_and_mtp_kv_proxy": {
          "min": 5973,
          "max": 11947,
          "unit": "bytes/token",
          "status": "estimated",
          "source": "bounded from the external text-versus-vision KV-pool delta",
          "observed_at": "2026-08-29"
        }
      },
      "kv_token_multiplier": {
        "value": 1,
        "unit": "copies",
        "status": "assumed",
        "source": "concurrency-one independent sequence",
        "observed_at": "2026-08-29"
      },
      "metrics": {},
      "hard_failures": []
    }
  ]
}