{
  "schema_version": 2,
  "manifest_ref": "reproduction:kvmem",
  "method": "kvmem-task-reference",
  "dataset": {
    "name": "WikiText-2 raw test",
    "revision": "b08601e04326c79dfdd32d625aee71d232d685c3",
    "text_sha256": "aca2f46735043bcfd0a44eca981d04627b9cdf74c4c9a04bf0856d04066f58fc"
  },
  "seed": 42,
  "checkpoint": {
    "model_id": "Qwen/Qwen3-4B-Instruct-2507",
    "revision": "cdbee75f17c01a7cc42f958dc650907174af0554"
  },
  "setup": {
    "examples": 6,
    "context_tokens": 2048,
    "target_tokens": 32,
    "workspace_tokens": 512,
    "accelerator": "NVIDIA A100"
  },
  "baseline": {
    "nll": 1.9173577563045472,
    "next_token_accuracy": 0.5729166666666666,
    "total_seconds": 1.3771060903867085,
    "peak_allocated_bytes": 10422372352
  },
  "recent_only": {
    "nll": 4.239119310318722,
    "next_token_accuracy": 0.390625,
    "total_seconds": 1.7594866041714947
  },
  "method_metrics": {
    "nll": 2.158501695045685,
    "next_token_accuracy": 0.5677083333333334,
    "total_seconds": 77.66178662857662,
    "peak_allocated_bytes": 10422375168
  },
  "evaluation_protocol": {
    "tier": "l2_execution_evidence",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "reference Python attention replacement; full backing cache retained; no speed or memory-saving claim"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2609.04852-kvmem/metrics/wikitext2-task-a100-seed42.json",
    "dataset_fingerprint": "sha256:aca2f46735043bcfd0a44eca981d04627b9cdf74c4c9a04bf0856d04066f58fc"
  }
}