{
  "schema": "anvil-serving.chat-template-ab-summary/v1",
  "date": "2026-08-20",
  "model": {
    "repository": "RadixArk/Qwen3.8-27B-NVFP4",
    "revision": "554ebba9b5f1b79dc11246341960360e6ef05ef4",
    "quantization": "ModelOpt NVFP4 W4A4 with FP8 projections and FP8 E4M3 KV",
    "context_tokens": 131072,
    "concurrency": 1
  },
  "runtime": {
    "engine": "SGLang",
    "image": "lmsysorg/sglang:qwen38-27b@sha256:506525a5907ea22c9d445afb7c03603959b912de034d86915cf17da814f1a124",
    "measured_gpu": "NVIDIA GeForce RTX 5090",
    "topology": "isolated direct loopback qualification endpoint",
    "pro_relationship": "unrelated"
  },
  "template": {
    "repository": "peculiar-ragdoll/Qwen-Sharp-Chat-Templates",
    "revision": "3dc34df52c63dd22ada21f96435e069deaa8d7da",
    "version": "Sharp v22.1",
    "application": "SGLang --chat-template from the pinned named-cache snapshot"
  },
  "thinking_enabled_mmlu_pro": {
    "shape": "10 items x 3 repetitions; 256 visible-answer tokens plus 1792 reasoning-headroom tokens",
    "stock": {
      "passed_attempts": 24,
      "attempts": 30,
      "completion_tokens": 24138,
      "reasoning_characters": 69798,
      "visible_characters": 8868,
      "mean_latency_ms": 10586.3
    },
    "sharp": {
      "passed_attempts": 24,
      "attempts": 30,
      "completion_tokens": 26736,
      "reasoning_characters": 66570,
      "visible_characters": 17424,
      "mean_latency_ms": 11716.6
    },
    "sharp_delta_percent": {
      "completion_tokens": 10.8,
      "reasoning_characters": -4.6,
      "visible_characters": 96.5,
      "mean_latency_ms": 10.7
    },
    "shared_failures": [
      "mmlu-pro-30-computer-science: reasoning_budget_exhausted in 3/3 attempts",
      "mmlu-pro-40-engineering: reasoning_budget_exhausted in 3/3 attempts"
    ]
  },
  "thinking_disabled_behavior": {
    "shape": "6 items x 3 repetitions; 256 visible-answer tokens; zero reasoning headroom",
    "stock": {
      "passed_attempts": 18,
      "attempts": 18,
      "completion_tokens": 1596,
      "visible_characters": 7017,
      "mean_latency_ms": 1267.6
    },
    "sharp": {
      "passed_attempts": 15,
      "attempts": 18,
      "completion_tokens": 1515,
      "visible_characters": 6723,
      "mean_latency_ms": 1204.7
    },
    "sharp_delta_percent": {
      "completion_tokens": -5.1,
      "visible_characters": -4.2,
      "mean_latency_ms": -5.0
    },
    "sharp_failure": "The ambiguous-request case requested four specific details but did not contain the diagnostic's required literal question mark."
  },
  "functional_gate": {
    "sharp_thinking_disabled": "passed",
    "checks": [
      "smoke",
      "json",
      "tools 20/20",
      "streaming tools",
      "tool-result continuation",
      "Responses API subset"
    ],
    "reasoning_leak": false
  },
  "artifact_caveats": [
    "Both quality artifacts retain inspector validation warnings for unrelated not-run built-in suites and absent aggregate chat timing fields.",
    "Thinking control was requested through chat_template_kwargs but is labeled requested_unverified in the quality artifacts.",
    "This bounded deterministic diagnostic is not the publisher's Claw-Eval, SWE-bench-Live, or full MMLU-Pro campaign."
  ],
  "decision": {
    "evidence": ["functional", "quality"],
    "label": "rejected",
    "promotion_authorized": false,
    "recommendation": "Do not adopt Sharp v22.1 for this pinned Qwen3.8 RTX 5090 profile from this evidence."
  }
}