{
  "paper": {
    "title": "DistillCache: KL-Guided Adaptive KV-Cache Eviction for Memory-Efficient LLM Inference"
  },
  "dataset": {
    "name": "deterministic long-context public mini-suite",
    "samples": 64,
    "sequence_length": 96
  },
  "setup": {
    "adapter": "distillcache",
    "seed": 42,
    "same_examples": true
  },
  "baseline": {
    "name": "recent-window attention",
    "accuracy": 0.078125,
    "memory_fraction": 0.25,
    "relative_cost": 1.0
  },
  "method": {
    "name": "distillcache",
    "accuracy": 1.0,
    "memory_fraction": 0.25,
    "relative_cost": 0.54,
    "kl_guided_evictions": 4608.0
  },
  "relative": {
    "accuracy_percent": 1180.0,
    "accuracy_points": 92.1875,
    "cost_percent": -46.0
  },
  "paper_results": {
    "reported": "25% cache 预算保留 LongBench full-cache 94.2% 准确率，吞吐最高 2.1×。"
  },
  "scope": "固定 numpy 长上下文/多模态/评测 mini-suite；不冒充论文规模预训练、GPU kernel 或完整公开 benchmark。",
  "schema_version": 2,
  "manifest_ref": "reproduction:distillcache",
  "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/2608.08878-distillcache/metrics/public-seed42.json",
    "dataset_fingerprint": "not recorded in historical artifact",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded"
  }
}