{
  "schema": "anvil-serving.mtp3-result-summary/v1",
  "candidate": "thinkingcap-awq-5090",
  "profile": "16K graph-enabled V2 MTP3 candidate",
  "evidence_root": "native/candidate/mtp3",
  "startup": {
    "model_memory_gib": 22.14,
    "available_kv_cache_gib": 4.71,
    "gpu_kv_cache_tokens": 44683
  },
  "functional_gates": {
    "functional_passed": 15,
    "reasoning_passed": 8,
    "vision_passed": 18,
    "vision_attempts": 18
  },
  "quality_diagnostic": {
    "suite": "hf-mmlu-pro-10-repeated",
    "passed_attempts": 27,
    "attempts": 30,
    "failed_eval_id": "mmlu-pro-30-computer-science",
    "failed_attempts": 3,
    "failure_class": "deterministic_check_failed",
    "finish_reason_counts": {
      "stop": 30
    },
    "qualification_boundary": "small repeated diagnostic only; not a full MMLU-Pro score or promotion comparison"
  },
  "capacity": {
    "workload": "strict 32-word, 4096-context, C1, unique cache, request canaries",
    "attempted_requests": 100,
    "strict_passes": 99,
    "strict_failures": 1,
    "failed_request_index": 89,
    "failure": "emitted 34 code words where 32 were required",
    "aggregate_throughput_eligible": false,
    "classification": "not a clean finalist; do not graph or claim a clean speedup"
  },
  "backend_interpretation": {
    "nested_drafter_triton_requested": true,
    "effective_drafter_triton_proven": false,
    "finding": "The pinned standard V2 MTP path inherits the target FLASH attention path; the nested speculative attention_backend request does not prove a Triton drafter.",
    "independent_review": {
      "path": "fp8-v2-review.json",
      "sha256": "857b05104429a6ee7ad2246d4910f3bf64f156d96f47cd4b9929143dbbaf19ba"
    }
  },
  "comparison_boundary": {
    "clean_matched_series": [
      "eager",
      "graphs"
    ],
    "mtp3_included_as_clean_series": false,
    "promoted": false
  },
  "redaction_ledger": "native/candidate/mtp3/redactions.json"
}