{
  "schema_version": 2,
  "method": "baff",
  "dataset": "controlled RTB simulation",
  "seeds": [
    42,
    43,
    44
  ],
  "runs": [
    {
      "paper": {
        "arxiv_id": "2609.08725",
        "title": "BAFF: Bid-Aware Filter Family for Mitigating Training Data Interference in RTB A/B Tests",
        "url": "https://arxiv.org/abs/2609.08725",
        "organization": "Dable"
      },
      "dataset": {
        "name": "controlled RTB simulation",
        "requests": 2400,
        "candidates": 30
      },
      "setup": {
        "adapter": "baff",
        "seed": 42,
        "same_split_and_candidates": true
      },
      "baseline": {
        "name": "naive pooled training",
        "parameter_deviation": 0.0554115075749221,
        "retained_control_fraction": 1.0
      },
      "method": {
        "name": "BAFF F(0.5,0.5)",
        "parameter_deviation": 0.05586013600227267,
        "retained_control_fraction": 0.9791666666666666
      },
      "relative": {
        "parameter_deviation_reduction_percent": -0.8096304305454585
      },
      "stages": {
        "d_ad_mean": 0.06244444444444445,
        "d_bp_mean": 0.023356297046995933,
        "filter_exactly_matches_equation": true
      },
      "paper_results": {
        "cpc_te_gap_naive": 9.85,
        "cpc_te_gap_baff": 0.44,
        "ctr_te_gap_naive_pp": 1.94,
        "ctr_te_gap_baff_pp": 0.4
      },
      "scope": "受控 RTB 仿真执行论文 F(k,l) 判据与 ridge 偏差诊断；不是论文线上 KPI，也不替代私有 DSP 日志。",
      "manifest_ref": "reproduction:baff"
    },
    {
      "paper": {
        "arxiv_id": "2609.08725",
        "title": "BAFF: Bid-Aware Filter Family for Mitigating Training Data Interference in RTB A/B Tests",
        "url": "https://arxiv.org/abs/2609.08725",
        "organization": "Dable"
      },
      "dataset": {
        "name": "controlled RTB simulation",
        "requests": 2400,
        "candidates": 30
      },
      "setup": {
        "adapter": "baff",
        "seed": 43,
        "same_split_and_candidates": true
      },
      "baseline": {
        "name": "naive pooled training",
        "parameter_deviation": 0.04554116250067161,
        "retained_control_fraction": 1.0
      },
      "method": {
        "name": "BAFF F(0.5,0.5)",
        "parameter_deviation": 0.04755823018759019,
        "retained_control_fraction": 0.9591666666666666
      },
      "relative": {
        "parameter_deviation_reduction_percent": -4.429108911940567
      },
      "stages": {
        "d_ad_mean": 0.09970833333333334,
        "d_bp_mean": 0.04248299025237329,
        "filter_exactly_matches_equation": true
      },
      "paper_results": {
        "cpc_te_gap_naive": 9.85,
        "cpc_te_gap_baff": 0.44,
        "ctr_te_gap_naive_pp": 1.94,
        "ctr_te_gap_baff_pp": 0.4
      },
      "scope": "受控 RTB 仿真执行论文 F(k,l) 判据与 ridge 偏差诊断；不是论文线上 KPI，也不替代私有 DSP 日志。",
      "manifest_ref": "reproduction:baff"
    },
    {
      "paper": {
        "arxiv_id": "2609.08725",
        "title": "BAFF: Bid-Aware Filter Family for Mitigating Training Data Interference in RTB A/B Tests",
        "url": "https://arxiv.org/abs/2609.08725",
        "organization": "Dable"
      },
      "dataset": {
        "name": "controlled RTB simulation",
        "requests": 2400,
        "candidates": 30
      },
      "setup": {
        "adapter": "baff",
        "seed": 44,
        "same_split_and_candidates": true
      },
      "baseline": {
        "name": "naive pooled training",
        "parameter_deviation": 0.048334301663495546,
        "retained_control_fraction": 1.0
      },
      "method": {
        "name": "BAFF F(0.5,0.5)",
        "parameter_deviation": 0.048272486315664234,
        "retained_control_fraction": 0.9991666666666666
      },
      "relative": {
        "parameter_deviation_reduction_percent": 0.12789126087239647
      },
      "stages": {
        "d_ad_mean": 0.027638888888888893,
        "d_bp_mean": 0.019645088153180917,
        "filter_exactly_matches_equation": true
      },
      "paper_results": {
        "cpc_te_gap_naive": 9.85,
        "cpc_te_gap_baff": 0.44,
        "ctr_te_gap_naive_pp": 1.94,
        "ctr_te_gap_baff_pp": 0.4
      },
      "scope": "受控 RTB 仿真执行论文 F(k,l) 判据与 ridge 偏差诊断；不是论文线上 KPI，也不替代私有 DSP 日志。",
      "manifest_ref": "reproduction:baff"
    }
  ],
  "aggregate_metrics": {
    "parameter_deviation_mean": 0.05056361750184237,
    "parameter_deviation_std": 0.004600801193615434,
    "retained_control_fraction_mean": 0.9791666666666666,
    "retained_control_fraction_std": 0.020000000000000018
  },
  "manifest_ref": "reproduction:baff",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "claim_policy": "public CPU mechanism comparison; not a production KPI reproduction"
  },
  "provenance": {
    "commit": "1b978289e839bfc12398533268c21e68f0f7eef6",
    "command": "PYTHONPATH=src python scripts/generate_sep_12_2026_artifacts.py",
    "dataset_fingerprint": "fed6f202d032852dffd662670266666bf7ad6145daf4f66f10cb6b6d50d03d7d",
    "artifact_path": "docs/reproductions/2609.08725-baff/metrics/controlled-rtb-seeds42-44.json"
  }
}