{
  "schema": "anvil-serving.public-benchmark-summary/v1",
  "captured_date": "2026-08-26",
  "decision": "human-authorized-reference-promotion",
  "identity": {
    "checkpoint": "RadixArk/Qwen3.8-Flash-Next-NVFP4",
    "revision": "7b719225242aacd3dbd3f9407468c2ee9a9d2594",
    "served_model": "qwen38-flash-next-radixark-nvfp4-tp2-262k",
    "engine": "SGLang",
    "engine_revision": "d91c3682b0b429e4c70df63cd57f819588ce29b0",
    "image_digest": "sha256:59f06adce6f91401adf443bd168d45fdb2044d77671fd591c7c57a29d851cbae"
  },
  "recipe": {
    "hardware": "2x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition",
    "topology": "exclusive TP=2 over PCIe without NVLink under WSL2",
    "context_tokens": 262144,
    "client_prompt_budget_tokens": 253952,
    "client_output_reserve_tokens": 8192,
    "configured_concurrency": 1,
    "kv_pool_tokens": 416064,
    "quantization": "NVIDIA ModelOpt NVFP4 W4A4 routed experts; BF16 attention, GDN, QSA, shared experts, vision, and MTP; FP8 PLE tables dequantized to BF16 at load",
    "speculative_decoding": false,
    "compatibility_lane": "portable QSA sparse decode plus NCCL logits fallback"
  },
  "functional": {
    "direct_needle_prompt_tokens": 253325,
    "direct_needle_seconds": 17.606,
    "direct_long_tool_prompt_tokens": 105756,
    "direct_long_tool_seconds": 4.664,
    "routed_needle_prompt_tokens": 253325,
    "routed_needle_seconds": 28.38,
    "routed_tools_passed": 20,
    "routed_tools_attempted": 20,
    "streaming_tool_passed": true,
    "tool_result_continuation_passed": true,
    "responses_api_passed": true,
    "responses_exact_output": "READY",
    "client_catalog": {
      "alias": "llm.primary",
      "context_tokens": 262144,
      "max_output_tokens": 8192,
      "openclaw": "matched",
      "hermes": "matched",
      "pi": "matched"
    },
    "real_clients": {
      "openclaw": "passed",
      "hermes": "passed",
      "pi": "passed"
    }
  },
  "quality": {
    "thinking_mode": "disabled",
    "repetitions": 3,
    "intelligence_attempts": 6,
    "intelligence_passed": 6,
    "session_attempts": 3,
    "session_passed": 3,
    "tool_attempts": 3,
    "tool_passed": 3,
    "failures": 0
  },
  "capacity": {
    "four_k_c1": {
      "completed": 5,
      "failed": 0,
      "ttft_p50_ms": 214.612,
      "e2e_p50_ms": 2627.492,
      "decode_tok_s_p50": 12.801
    },
    "four_k_c2_queue_diagnostic": {
      "completed": 4,
      "failed": 0,
      "ttft_p50_ms": 2821.929,
      "e2e_p50_ms": 5290.454,
      "decode_tok_s_p50": 12.445,
      "interpretation": "requests completed behind the configured one-running-request scheduler; c2 is not the promoted admission contract"
    },
    "one_hundred_twenty_eight_k_c1": {
      "completed": 1,
      "failed": 0,
      "prompt_tokens": 125442,
      "ttft_ms": 11961.075,
      "e2e_ms": 13794.578,
      "decode_tok_s": 12.544
    },
    "native_boundary": {
      "largest_accepted_prompt_tokens_with_one_output": 262137,
      "accepted_visible_sequence_tokens": 262142,
      "first_rejected_prompt_tokens": 262138
    }
  },
  "retained_failures_and_fixes": [
    "SM120 WSL2 symmetric-memory logits initialization failed before the exact NCCL fallback was selected.",
    "The pinned sparse-attention FA4 CuTe path failed MLIR compilation on SM120; the qualified lane uses SGLang's device-agnostic QSA sparse decoder.",
    "The first long-tool generator undershot 100000 measured prompt tokens; a calibrated 125000-target rerun passed at 105756 measured tokens.",
    "Router admission originally rejected an exact upstream context contract and the Responses adapter rejected a bounded thinking control; both product defects were fixed and regression-tested.",
    "Hermes credential drift and a stale 393216-token model override, plus Pi's missing Anvil provider/environment-key reference, were repaired; all three clients now match the 262144/8192 router contract and passed without fallback."
  ],
  "publication": {
    "raw_operator_artifacts": "retained outside the public repository",
    "sanitization": "No credentials, addresses, GPU UUIDs, machine-local paths, or raw client configuration are included."
  }
}