{
  "schema_version": 2,
  "status": "passed",
  "diagnostic_only": true,
  "paper_result_reproduced": false,
  "manifest_ref": "reproduction:kuafu",
  "dataset": {
    "name": "MRQA 2019 SQuAD train/dev",
    "train_sha256": "096378373b48ee7e0a67ce23ef67217b12fc11ed12c34bc94a235a8ccb299024",
    "dev_sha256": "b7f3214d3dfe2a9a368ccd8383c162787d08b978e5affd110bd0d9d58f53b0ed",
    "train_questions": 32,
    "dev_questions": 32,
    "sampling": "first question from each of 32 distinct context records per split"
  },
  "checkpoint": {
    "model_id": "Qwen/Qwen3-4B-Instruct-2507",
    "revision": "cdbee75f17c01a7cc42f958dc650907174af0554",
    "weights_committed": false
  },
  "budget": {
    "max_items": 4,
    "max_item_tokens": 24,
    "cached_tokens_per_item": 2,
    "compressed_supervised_stages": 3,
    "independent_control_qa_updates": 64,
    "independent_control_input": "raw_head_equal_cache_tokens",
    "fairness_caveat": "Q&A update counts are equal across the two Q&A stages, but total FLOPs are not: compression also trains an encoder, projector, and reconstruction stage."
  },
  "seeds": [
    {
      "seed": 42,
      "compressed_dev_em": 0.0,
      "compressed_dev_f1": 0.08904671717171717,
      "independent_control_dev_em": 0.0,
      "independent_control_dev_f1": 0.11462599900099898
    },
    {
      "seed": 43,
      "compressed_dev_em": 0.0,
      "compressed_dev_f1": 0.11303488178488177,
      "independent_control_dev_em": 0.0,
      "independent_control_dev_f1": 0.12999058233433233
    },
    {
      "seed": 44,
      "compressed_dev_em": 0.0,
      "compressed_dev_f1": 0.08341529793735676,
      "independent_control_dev_em": 0.0,
      "independent_control_dev_f1": 0.12999058233433233
    }
  ],
  "aggregate_metrics": {
    "compressed_dev_f1_mean": 0.09516563229798523,
    "independent_control_dev_f1_mean": 0.12486905455655455,
    "compressed_minus_control_dev_f1_mean": -0.029703422258569312,
    "compressed_winning_seeds": 0
  },
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "claim_policy": "Three-seed public small-sample diagnostic; no production, statistical significance, or paper-result claim."
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2609.31045-kuafu/metrics/mrqa-independent-control-seeds42-44.json",
    "dataset_fingerprint": "sha256:096378373b48ee7e0a67ce23ef67217b12fc11ed12c34bc94a235a8ccb299024",
    "raw_predictions_committed": false
  }
}