{
  "schema": "anvil-serving.benchmark-run-plan/v1",
  "campaign_id": "2026-09-03-qwen38-ninfer-nvfp4-rtx5090",
  "declared_at": "2026-09-03T12:04:28Z",
  "arms": [
    {
      "order": 1,
      "id": "nospec",
      "recipe": "neroued/Qwen3.8-27B-NInfer-NVFP4-NoSpec-252K",
      "served_model": "qwen38-27b-ninfer-nvfp4-nospec-252k",
      "speculation": "disabled"
    },
    {
      "order": 2,
      "id": "mtp3",
      "recipe": "neroued/Qwen3.8-27B-NInfer-NVFP4-MTP3-252K",
      "served_model": "qwen38-27b-ninfer-nvfp4-mtp3-252k",
      "speculation": "MTP3 with optimized proposal head"
    }
  ],
  "held_constant": {
    "artifact_revision": "204e3d92c30d9d05f3300d2f52e443ad1edf6ddf",
    "runtime_revision": "e3aeaf8c0b6f83ae8f051780f0ad0d995d5a7bef",
    "base_image_digest": "sha256:b9f64abf7226fdb3463ca202bc99878ec847171e6c5f77bd34c8d1403fbf1eca",
    "context_tokens": 252928,
    "output_reserve_tokens": 8192,
    "kv_dtype": "int8",
    "concurrency": 1,
    "thinking": "disabled",
    "prefix_reuse": "enabled by engine default; performance requests must use unique prefixes and retain cache_n"
  },
  "gates": [
    "paper feasibility classification is benchmark-survivor",
    "complete pinned artifact cache",
    "managed recipe preview identifies exact artifact/runtime/GPU",
    "native health and exact /v1/models identity",
    "smoke before tools or context",
    "100 percent deterministic tool assertions",
    "244480-token retrieval with 8192-token output reserve",
    "MTP3 warm E2E improvement at least 10 percent versus matched no-spec",
    "no OOM, CUDA error, crash, restart, or unexplained request loss",
    "exact incumbent restoration and baseline route check"
  ],
  "mutation": {
    "current_candidate_lane_will_be_temporarily_unloaded": true,
    "route_alias_or_client_catalog_change": false,
    "candidate_port": 8081,
    "restore_incumbent_after_run": true,
    "promotion_authorized": false
  }
}