{
  "schema": "anvil-serving.benchmark-run-plan/v1",
  "campaign_id": "2026-09-04-qwen38-27b-pro6000-possibility",
  "goal": "find the highest useful Qwen3.8 27B configuration on one or two local RTX PRO 6000 GPUs without treating any external headline as local proof",
  "promotion_authorized": false,
  "execution_order": [
    {"order": 1, "stage": "baseline", "arms": ["Inferact target-only", "Inferact DFlash2 K8"], "status": "complete"},
    {"order": 2, "stage": "SGLang optimization", "arms": ["K4", "K12", "K16", "chunk1K", "chunk8K", "torch.compile", "Mamba lazy96", "Mamba40"], "status": "complete", "selected": "K12 chunk1K extra_buffer Mamba96"},
    {"order": 3, "stage": "matched topology", "arms": ["TP1", "TP2", "two independent TP1 replicas"], "status": "complete", "selected": "DP2 for aggregate throughput", "rejected": "TP2"},
    {"order": 4, "stage": "target checkpoint", "arms": ["RadixArk target-only", "RadixArk DFlash2 K8", "RadixArk DFlash2 K12 chunk1K"], "status": "complete", "selected": "K8 as lower-TTFT tradeoff"},
    {"order": 5, "stage": "alternate runtime", "arms": ["kelnei vLLM 0.27.1 MTP2", "matched no-spec"], "status": "complete", "selected": "MTP2 within vLLM; SGLang remains faster for sustained output"},
    {"order": 6, "stage": "restore", "status": "complete", "evidence": "restoration.json"}
  ],
  "headline_comparison_rule": "Compare only the matched 100-request unique-canary 4K/C8-per-replica sustained-output artifacts. Do not compare their aggregate throughput with natural-completion screens or external workloads.",
  "advance_rules": [
    "all requests complete without restart or unexplained loss",
    "every unique request returns its own canary and no other canary",
    "strict structured JSON, tools, and Responses pass",
    "performance claim includes mean, p50, p95, and p99 for TTFT, effective prefill, decode, TPOT/ITL, and E2E",
    "topology winner must account for both throughput and communication overhead"
  ],
  "stop_rules": [
    "cross-request marker or tool contamination",
    "repeatable structured-output corruption",
    "unexplained request loss, OOM, restart, CUDA error, or parser corruption",
    "starting service cannot be restored through the managed path"
  ],
  "post_campaign_open_gates": [
    "broad repeated quality, agentic, SWE, and endurance",
    "multimodal and routed/client acceptance",
    "qualified load balancing and failure behavior for DP2",
    "power, clock, temperature, and energy-per-token capture",
    "native-Linux TP2 control",
    "DFlash2 license and deployment-use review"
  ]
}
