{
  "schema_version": 1,
  "adapter_key": "kvmem",
  "validated_at": "2026-09-10",
  "correction": "Supersedes the single-text mechanism receipt with task-level next-token evaluation on six non-overlapping public WikiText-2 contexts. The reference implementation retains the full backing cache and makes no speed or memory-saving claim.",
  "accelerator": {"vendor": "NVIDIA", "model": "A100"},
  "command": [
    "python", "-m", "auto_research.reproductions.kv_task_evaluation",
    "--dataset-file", "wikitext-2-raw-v1-test.txt", "--context-tokens", "2048",
    "--examples", "6", "--seed", "42", "--output", "kv-task-evaluation.json"
  ],
  "dataset": {
    "name": "WikiText-2 raw test",
    "revision": "b08601e04326c79dfdd32d625aee71d232d685c3",
    "text_sha256": "aca2f46735043bcfd0a44eca981d04627b9cdf74c4c9a04bf0856d04066f58fc",
    "examples": 6
  },
  "checkpoint": {
    "model_id": "Qwen/Qwen3-4B-Instruct-2507",
    "revision": "cdbee75f17c01a7cc42f958dc650907174af0554"
  },
  "result": "passed",
  "metrics": {
    "seed": 42,
    "context_tokens": 2048,
    "target_tokens_per_example": 32,
    "workspace_tokens": 512,
    "full_next_token_accuracy": 0.5729166666666666,
    "recent_only_next_token_accuracy": 0.390625,
    "method_next_token_accuracy": 0.5677083333333334,
    "full_total_seconds": 1.3771060903867085,
    "method_total_seconds": 77.66178662857662,
    "full_peak_allocated_bytes": 10422372352,
    "method_peak_allocated_bytes": 10422375168
  },
  "provenance": {
    "commit": "f1a35df149bfd6dec7344d19b35de6540c70fc81",
    "artifact_path": "docs/gpu-validations/kvmem-a100-20260907.json",
    "raw_predictions_committed": false,
    "checkpoint_committed": false
  }
}