{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-02-glm53-sglang-sm120-qualification",
  "complete": true,
  "identity": {
    "model": "ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO@c3cbb9891b67c741bcbf6b176dd7af9265b069db",
    "image": "ghcr.io/ormandj/sglang-glm53-flash-sm120@sha256:0c0637959c3931829f05154087bbefd2c50003fb9b2010200ce0ec82f4d71a53",
    "served_model": "glm53-flash-ormandj-sglang-sm120-tp2-240k-c1-adaptive-mtp"
  },
  "workload": {
    "hardware": "2x NVIDIA RTX PRO 6000 Blackwell Max-Q, exclusive TP=2, PCIe without NVLink, WSL2",
    "selected_context_tokens": 245760,
    "selected_concurrency": 1,
    "thinking_controls": ["disabled", "enabled"]
  },
  "measurement": {
    "path": "online-direct",
    "warm_cold_state": "cold managed starts plus warm functional, capacity, quality, multimodal, and endurance stages",
    "sample_count": 160,
    "statistics": [
      "240K profile capacity r3 at nominal 4K, 120K, and 230K",
      "60-request 4K endurance",
      "15 deterministic coding-agent attempts",
      "12 deterministic image attempts"
    ]
  },
  "gates": [
    {"name": "full functional preflight", "status": "passed", "detail": "all 10 checks; 20/20 tools; 194227-token long tool prompt"},
    {"name": "thinking control", "status": "passed", "detail": "disabled forbids reasoning leakage; enabled requires dedicated reasoning"},
    {"name": "coding-agent quality", "status": "passed", "detail": "15/15 deterministic attempts at 2048 visible tokens"},
    {"name": "multimodal", "status": "passed", "detail": "12/12 deterministic image/OCR attempts"},
    {"name": "endurance", "status": "passed", "detail": "60/60 requests"},
    {"name": "post-workload reserve", "status": "passed", "detail": "3487 MiB free on each card; threshold 3072 MiB"},
    {"name": "exact restoration", "status": "passed", "detail": "original container id, image, model, port, mode owner, router, and shared-memory state restored"}
  ],
  "headline_results": [
    "131K adaptive MTP improved decode p50 by 39.7 percent at 4K and 29.2 percent at 120K versus matched no speculation",
    "240K selected profile decode p50 was 108.57, 93.35, and 95.00 tok/s at nominal 4K, 120K, and 230K",
    "240K selected profile effective prefill p50 was 16545, 5763, and 5608 tok/s at those points",
    "499712 C4 short reached 147.54 aggregate output tok/s but the profile and its C1 reductions failed reserve policy",
    "393216 C1 passed every functional and workload gate but fell to 2101 MiB free per card after workload"
  ],
  "failures": [
    "NCCL cuMem mapping failed with CUDA 999 until NCCL_CUMEM_ENABLE=0",
    "expandable_segments=True failed first distributed CUDA allocations until disabled",
    "Torch symmetric-memory logits gatherer raised SIGFPE until a source-hash-gated ordinary NCCL fallback patch",
    "checkpoint chat template leaked reasoning until a source-hash-gated enable_thinking template patch",
    "499712, 393216, and intermediate C1 profiles failed the 3 GiB post-workload reserve policy",
    "sparse-MLA CPB calibration rejected implausible fits and used its C++ heuristic"
  ],
  "limitations": [
    "Qualification is local to dual RTX PRO 6000 Blackwell Max-Q under WSL2 and PCIe without NVLink.",
    "Only C1 is qualified at the selected 245760-token profile; the high-concurrency external result is not reproduced.",
    "The 230K nominal harness point measured 189627 prompt tokens.",
    "Video was not qualified.",
    "No sparse-MLA calibration tune was adopted without an exact default-versus-tuned end-to-end A/B."
  ],
  "decision": {
    "evidence_labels": ["local-functional", "local-capacity", "local-performance", "local-quality", "local-multimodal", "local-endurance", "exact-restoration", "external-prior"],
    "decision_labels": ["verified", "challenger", "no-promotion"],
    "promotion_authorized": false,
    "selected_recipe": "configs/glm53-flash-ormandj-sglang-sm120-tp2-240k-c1-adaptive-mtp-recipe.toml",
    "next_gate": "human decision before any promotion; tune the exact sparse-MLA CPB path only with matched end-to-end A/B evidence"
  },
  "evidence": ["safe240k-preflight-all-low.json", "safe240k-preflight-thinking-enabled.json", "safe240k-quality-coding-agent-v2-disabled-r3-visible2048.json", "safe240k-multimodal-image-c1.json", "safe240k-endurance-c1-4k-r60.json", "post-workload-state.json", "restoration.json"]
}