{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-03-qwen38-27b-rtx5090-quant-bakeoff",
  "complete": true,
  "identity": {
    "hardware": "single NVIDIA GeForce RTX 5090, 32607 MiB",
    "topology": "isolated direct managed recipe lane",
    "incumbent": "unsloth/Qwen3.8-27B-UD-Q4_K_XL-Native-MTP3-262K"
  },
  "workload": {
    "ranking_priority": ["ttft", "decode_tokens_per_second", "usable_context_and_kv", "concurrency"],
    "matched_prompt_tokens": 3613,
    "warm_repetitions_per_arm": 5,
    "long_context_actual_prompt_tokens": {"full_context_arms": 244002, "bounded_64k_arms": [53456, 53706]}
  },
  "measurement": {
    "path": "online-direct",
    "warm_cold_state": "cold labeled runs followed by warm repetitions",
    "candidate_arms": 14,
    "statistics": ["median", "p95", "aggregate throughput"]
  },
  "gates": [
    {"name": "idle GPU baseline before reserve decision", "status": "passed", "evidence": "gpu-idle-baseline.json"},
    {"name": "candidate preflight", "status": "passed for every arm that loaded"},
    {"name": "long-context exact retrieval", "status": "passed for every timed arm"},
    {"name": "Gittensor bounded quality and tools", "status": "passed", "evidence": "gittensor-nospec-quality.json"},
    {"name": "Unsloth Dynamic V3 NVFP4 preflight", "status": "passed on both arms, including tools 20/20", "evidence": "unsloth-nvfp4-mtp3-preflight.json"},
    {"name": "CometKim bounded quality and tools", "status": "failed", "detail": "tool schema 0/3; required zip argument missing", "evidence": "cometkim-mtp3-quality.json"},
    {"name": "exact incumbent restoration", "status": "passed", "evidence": "restoration-preflight.json"}
  ],
  "headline_results": [
    {"candidate": "Gittensor RTX5090 NVFP4 target-only", "context_tokens": 262144, "ttft_p50_ms": 50.9, "decode_p50_tok_s": 79.5, "c2_aggregate_tok_s": 128.7, "decision": "TTFT winner and preferred measured challenger; no promotion"},
    {"candidate": "CometKim NInfer NVFP4Full MTP3", "context_tokens": 262144, "ttft_p50_ms": 326.3, "decode_p50_tok_s": 228.0, "decision": "decode winner but rejected as a general-purpose winner after tool-schema failure"},
    {"candidate": "cdiamond iMatrix NVFP4 GGUF MTP8", "context_tokens": 262144, "ttft_p50_ms": 223.1, "decode_p50_tok_s": 96.0, "decision": "best full-context balanced fallback in this intake matrix"},
    {"candidate": "Unsloth Dynamic V3 NVFP4 MTP3", "context_tokens": 65536, "ttft_p50_ms": 388.7, "decode_p50_tok_s": 137.7, "decision": "strongest clean 64K speculative arm; tools 20/20 and 3198 MiB free"},
    {"candidate": "RedHatAI NVFP4 MTP4", "context_tokens": 65536, "ttft_p50_ms": 390.6, "decode_p50_tok_s": 118.9, "decision": "strong 64K long-context decode result"},
    {"candidate": "incumbent Unsloth GGUF MTP3", "context_tokens": 262144, "ttft_p50_ms": 232.6, "decode_p50_tok_s": 88.6, "decision": "restored incumbent with broader prior capability evidence"}
  ],
  "failures": [
    {"candidate": "Gittensor target plus DSpark", "class": "compatibility", "detail": "draft CUDA-graph capture shape mismatch; not OOM", "evidence": "gittensor-dspark-load-failure.json"},
    {"candidate": "CometKim MTP3", "class": "tool-schema", "detail": "0/3 strict tool calls supplied the required zip string", "evidence": "cometkim-mtp3-quality.json"},
    {"candidate": "QUASAR MTP2", "class": "workload-regression", "detail": "long-context decode fell from 66.6 tok/s without speculation to 37.5 tok/s with MTP2"}
  ],
  "limitations": [
    "The zero added reserve was a campaign-only exception made after a process-free idle baseline; it is not a new default.",
    "Gittensor's SGLang FP8 KV path reported absent calibrated KV scales and defaulted them to 1.0, so broad quality equivalence is unproven.",
    "Only Gittensor and CometKim received the campaign's repeated bounded quality suite; earlier GGUF and NInfer qualifications remain the broader evidence sources.",
    "The vLLM challengers were intentionally tested at publisher- or feasibility-bounded 64K profiles rather than unsafe 262K extrapolations.",
    "No route, client catalog, or promoted assignment changed."
  ],
  "decision": {
    "evidence_labels": ["functional", "capacity", "matched-performance", "bounded-quality", "compatibility-failure", "exact-restoration"],
    "decision_labels": ["challenger", "no-promotion"],
    "promotion_authorized": false,
    "winner": "gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 target-only on SGLang",
    "reason": "It won the operator-declared primary metric, warm median TTFT, while passing preflight, repeated tools/quality, 244002-token actual-prompt retrieval, and C2.",
    "next_gate": "Calibrate or validate FP8 KV quality, run broader agentic/SWE and routed-client gates, and obtain a compatible speculative draft before considering promotion."
  },
  "evidence": [
    "README.md",
    "publication-summary.md",
    "artifact-manifest.json",
    "gittensor-nospec-timing-4k-c1-warm.json",
    "gittensor-nospec-capacity-252928-c1.json",
    "gittensor-nospec-quality.json",
    "cometkim-mtp3-quality.json",
    "unsloth-nvfp4-mtp3-timing-4k-c1-warm.json",
    "unsloth-nvfp4-mtp3-runtime-metrics.json",
    "restoration-preflight.json"
  ]
}