{
  "paper": {
    "title": "Fast Inference via Speculative Decoding"
  },
  "dataset": {
    "name": "WikiText-2",
    "generated_tokens": 160
  },
  "baseline": {
    "name": "target-only greedy decoding",
    "target_calls": 160
  },
  "method": {
    "name": "four-token draft and exact target verification",
    "target_calls": 40
  },
  "relative": {
    "target_call_reduction_percent": 75.0
  },
  "stages": {
    "draft_width": 4,
    "accepted_draft_tokens": 160,
    "exact_output_match": true
  },
  "paper_results": {
    "reported_speedup_x": "2-3",
    "distribution_preserving": true
  },
  "scope": "在 WikiText-2 拟合 target/draft token 模型，真实执行提议、验证和拒绝回退；target 是小型 Markov LM，不等同于 T5-XXL kernel 延迟。",
  "schema_version": 2,
  "manifest_ref": "reproduction:speculative-decoding",
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2211.17192-speculative-decoding/metrics/public-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}