{
  "schema": "anvil-serving.recipe-feasibility-input/v1",
  "campaign": "2026-09-19-swift-qwen38-apple-feasibility-evidence",
  "requirements": {
    "tokens": {
      "prompt": {
        "value": 24000,
        "unit": "tokens",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "output_reserve": {
        "value": 8192,
        "unit": "tokens",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "media_protocol_overhead": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "Exact tokenization/image expansion not measured; capacity request cannot assume zero."
      }
    },
    "concurrency": {
      "value": 1,
      "unit": "requests",
      "status": "assumed",
      "source": "run-plan.md",
      "observed_at": "2026-09-19T11:02:17.939934+00:00"
    },
    "physical_vram_bytes": {
      "value": 51539607552,
      "unit": "bytes",
      "status": "measured",
      "source": "baseline/hardware-filtered.json",
      "observed_at": "2026-09-19T11:02:17.939934+00:00"
    },
    "vram_reserves": {
      "os_desktop_voice": {
        "kind": "policy",
        "variable": {
          "value": 17179869184,
          "unit": "bytes",
          "status": "assumed",
          "source": "run-plan.md",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      }
    },
    "thresholds": {
      "min_deterministic_pass_rate": {
        "value": 1,
        "unit": "fraction",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "max_relative_quality_loss": {
        "value": 0.05,
        "unit": "fraction",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "min_warm_e2e_gain": {
        "value": 0.2,
        "unit": "fraction",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "min_tasks_per_hour_ratio": {
        "value": 1.2,
        "unit": "ratio",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      }
    }
  },
  "capacity_interpretation": "Physical shared RAM is an upper envelope, not dedicated VRAM or an observed Metal budget. Engine working-set ceiling unknown. Reserve includes OS/applications/voice once; observed used RAM is not added again.",
  "candidates": [
    {
      "id": "swift-q6-k",
      "runtime_context_limit_tokens": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "No selected supported managed runtime pinned; source advertised 262144 does not prove engine context."
      },
      "measured_max_stable_context_tokens": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "No candidate requests permitted before safety gate."
      },
      "resident_components": {
        "target_weights": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "projector": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "runtime_fixed": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "graphs_workspace_recurrent_state": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        }
      },
      "per_sequence_components": {
        "sequence_state": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "Not measured."
        }
      },
      "per_token_components": {
        "target_kv": {
          "unit": "bytes/token",
          "status": "unknown",
          "notes": "Hybrid KV runtime representation not measured; architectural FP16 full-attention payload is 65536 bytes/token, excluding recurrent state and workspace."
        }
      },
      "kv_token_multiplier": {
        "value": 1,
        "unit": "copies",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "metrics": {},
      "hard_failures": [],
      "artifact_bytes": {
        "weight": {
          "value": 22884407456,
          "unit": "bytes",
          "status": "confirmed",
          "source": "identity-summary.json",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        },
        "projector": {
          "value": 927606976,
          "unit": "bytes",
          "status": "confirmed",
          "source": "identity-summary.json",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      },
      "speculation": {
        "enabled": false,
        "draft_resident_bytes": {
          "value": 0,
          "unit": "bytes",
          "status": "assumed",
          "source": "Disabled campaign arm, not a claim about packaged MTP tensor allocation.",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      }
    },
    {
      "id": "stock-q6-k",
      "runtime_context_limit_tokens": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "No selected supported managed runtime pinned; source advertised 262144 does not prove engine context."
      },
      "measured_max_stable_context_tokens": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "No candidate requests permitted before safety gate."
      },
      "resident_components": {
        "target_weights": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "projector": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "runtime_fixed": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "graphs_workspace_recurrent_state": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        }
      },
      "per_sequence_components": {
        "sequence_state": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "Not measured."
        }
      },
      "per_token_components": {
        "target_kv": {
          "unit": "bytes/token",
          "status": "unknown",
          "notes": "Hybrid KV runtime representation not measured; architectural FP16 full-attention payload is 65536 bytes/token, excluding recurrent state and workspace."
        }
      },
      "kv_token_multiplier": {
        "value": 1,
        "unit": "copies",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "metrics": {},
      "hard_failures": [],
      "artifact_bytes": {
        "weight": {
          "value": 23860565728,
          "unit": "bytes",
          "status": "confirmed",
          "source": "identity-summary.json",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        },
        "projector": {
          "value": 927607008,
          "unit": "bytes",
          "status": "confirmed",
          "source": "identity-summary.json",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      },
      "speculation": {
        "enabled": false,
        "draft_resident_bytes": {
          "value": 0,
          "unit": "bytes",
          "status": "assumed",
          "source": "Disabled campaign arm, not a claim about packaged MTP tensor allocation.",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      }
    },
    {
      "id": "swift-q4-k-m",
      "runtime_context_limit_tokens": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "No selected supported managed runtime pinned; source advertised 262144 does not prove engine context."
      },
      "measured_max_stable_context_tokens": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "No candidate requests permitted before safety gate."
      },
      "resident_components": {
        "target_weights": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "projector": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "runtime_fixed": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "graphs_workspace_recurrent_state": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        }
      },
      "per_sequence_components": {
        "sequence_state": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "Not measured."
        }
      },
      "per_token_components": {
        "target_kv": {
          "unit": "bytes/token",
          "status": "unknown",
          "notes": "Hybrid KV runtime representation not measured; architectural FP16 full-attention payload is 65536 bytes/token, excluding recurrent state and workspace."
        }
      },
      "kv_token_multiplier": {
        "value": 1,
        "unit": "copies",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "metrics": {},
      "hard_failures": [],
      "artifact_bytes": {
        "weight": {
          "value": 18024380576,
          "unit": "bytes",
          "status": "confirmed",
          "source": "identity-summary.json",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        },
        "projector": {
          "value": 927606976,
          "unit": "bytes",
          "status": "confirmed",
          "source": "identity-summary.json",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      },
      "speculation": {
        "enabled": false,
        "draft_resident_bytes": {
          "value": 0,
          "unit": "bytes",
          "status": "assumed",
          "source": "Disabled campaign arm, not a claim about packaged MTP tensor allocation.",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      }
    },
    {
      "id": "stock-q4-k-m",
      "runtime_context_limit_tokens": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "No selected supported managed runtime pinned; source advertised 262144 does not prove engine context."
      },
      "measured_max_stable_context_tokens": {
        "unit": "tokens",
        "status": "unknown",
        "notes": "No candidate requests permitted before safety gate."
      },
      "resident_components": {
        "target_weights": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "projector": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "runtime_fixed": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        },
        "graphs_workspace_recurrent_state": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "No exact managed native peak measurement or enforceable host-RAM ceiling."
        }
      },
      "per_sequence_components": {
        "sequence_state": {
          "unit": "bytes",
          "status": "unknown",
          "notes": "Not measured."
        }
      },
      "per_token_components": {
        "target_kv": {
          "unit": "bytes/token",
          "status": "unknown",
          "notes": "Hybrid KV runtime representation not measured; architectural FP16 full-attention payload is 65536 bytes/token, excluding recurrent state and workspace."
        }
      },
      "kv_token_multiplier": {
        "value": 1,
        "unit": "copies",
        "status": "assumed",
        "source": "run-plan.md",
        "observed_at": "2026-09-19T11:02:17.939934+00:00"
      },
      "metrics": {},
      "hard_failures": [],
      "artifact_bytes": {
        "weight": {
          "value": 17442399968,
          "unit": "bytes",
          "status": "confirmed",
          "source": "identity-summary.json",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        },
        "projector": {
          "value": 927607008,
          "unit": "bytes",
          "status": "confirmed",
          "source": "identity-summary.json",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      },
      "speculation": {
        "enabled": false,
        "draft_resident_bytes": {
          "value": 0,
          "unit": "bytes",
          "status": "assumed",
          "source": "Disabled campaign arm, not a claim about packaged MTP tensor allocation.",
          "observed_at": "2026-09-19T11:02:17.939934+00:00"
        }
      }
    }
  ],
  "disk_policy": {
    "free_bytes": {
      "value": 26295853056,
      "unit": "bytes",
      "status": "measured",
      "source": "baseline/native-observations.json",
      "observed_at": "2026-09-19T11:02:17.939934+00:00"
    },
    "free_reserve_bytes": {
      "value": 10737418240,
      "unit": "bytes",
      "status": "assumed",
      "source": "run-plan.md",
      "observed_at": "2026-09-19T11:02:17.939934+00:00"
    },
    "runtime_evidence_allowance_bytes": {
      "value": 1073741824,
      "unit": "bytes",
      "status": "assumed",
      "source": "run-plan.md",
      "observed_at": "2026-09-19T11:02:17.939934+00:00"
    },
    "temporary_download_bytes": {
      "unit": "bytes",
      "status": "unknown",
      "notes": "No managed native downloader exists; resume/rename/copy strategy unknown. Weight-only screen is optimistic."
    }
  },
  "additional_thresholds": {
    "reasoning_output_token_reduction_min": 0.2,
    "voice_p95_degradation_max": 0.2,
    "swap_growth_bytes_max": 0,
    "critical_gate_pass_rate": 1,
    "vision_relative_loss_max": 0.05,
    "efficiency_combination": "token reduction AND (E2E gain OR tasks/hour); calculator individual axes are not this final OR decision"
  }
}