{
  "schema": "anvil-serving.triton-bf16-result-summary/v1",
  "candidate": "thinkingcap-awq-5090",
  "profile": "16K C1 graph-enabled V2 MTP3 with target TRITON_ATTN and bfloat16 KV",
  "evidence_root": "native/candidate/triton-bf16",
  "recipe_semantic_digest": "fdcebf9caf431c10edd5635c74b1fe820fa474e78e76210527300721c04ce6d0",
  "startup": {
    "model_memory_gib": 22.14,
    "available_kv_cache_gib": 4.73,
    "gpu_kv_cache_tokens": 45180,
    "target_attention_backend": "TRITON_ATTN",
    "target_kv_cache_dtype": "bfloat16",
    "draft_kv_cache_dtype": "bfloat16 via standard V2 MTP inheritance",
    "cuda_graph_pool_memory_gib_actual": 0.1,
    "cuda_graph_pool_memory_gib_estimated": 0.14
  },
  "functional_gates": {
    "functional_observations_passed": 27,
    "reasoning_observations_passed": 8,
    "vision_attempts_passed": 18,
    "vision_attempts": 18
  },
  "capacity": {
    "workload": "nominal 4096 context, 32 exact code words, request canary, 45 completion tokens, C1, unique cache",
    "attempted_requests": 100,
    "strict_passes": 100,
    "strict_failures": 0,
    "aggregate_throughput_eligible": true,
    "ttft_p50_ms": 453.6534999788273,
    "decode_tok_s_p50": 158.49590269115166,
    "throughput_tok_s": 61.419572061303256
  },
  "quality_diagnostic": {
    "suite": "hf-mmlu-pro-10-repeated",
    "passed_attempts": 30,
    "attempts": 30,
    "finish_reason_counts": {
      "stop": 30
    },
    "scope": "small repeated diagnostic; not a full MMLU-Pro score or promotion comparison"
  },
  "comparison_boundary": {
    "publication_ready_for_chart_and_docs": true,
    "promotion_authorized": false,
    "limitations": [
      "Measured at 16K/C1 only; no long-context, concurrency, or FP8 claim follows."
    ]
  },
  "independent_source_review": {
    "path": "fp8-v2-review.json",
    "sha256": "857b05104429a6ee7ad2246d4910f3bf64f156d96f47cd4b9929143dbbaf19ba"
  },
  "redaction_ledger": "native/candidate/triton-bf16/redactions.json"
}