{
  "pdf_sha256": "2cfb6d1b55af14955b4634f3033bbf4f11740e9c7c1df770e6b748d370b67da3",
  "n_pages": 25,
  "benchmarks": [
    {
      "name": "DistilBERT",
      "d": 14,
      "m": 1440,
      "kind": "deep learning (text)"
    },
    {
      "name": "Estate",
      "d": 15,
      "m": 1722,
      "kind": "tabular"
    },
    {
      "name": "ViT16",
      "d": 16,
      "m": 1705,
      "kind": "deep learning (vision)"
    },
    {
      "name": "Cancer",
      "d": 30,
      "m": 2281,
      "kind": "tabular"
    },
    {
      "name": "IL60",
      "d": 60,
      "m": 5521,
      "kind": "tabular"
    },
    {
      "name": "CG60",
      "d": 60,
      "m": 5521,
      "kind": "tabular"
    },
    {
      "name": "NHANES",
      "d": 79,
      "m": 5864,
      "kind": "tabular"
    },
    {
      "name": "Crime",
      "d": 101,
      "m": 11126,
      "kind": "tabular"
    }
  ],
  "claim4": {
    "assertion_n_estimators": 8,
    "actual_n_estimator_rows": 11,
    "estimator_rows": [
      "MSR",
      "SV ARM",
      "PermutationSampling",
      "LeverageSHAP",
      "PolySHAP-3",
      "RegressionMSR",
      "Proxy",
      "FFD-RD",
      "FFD-RD-Corrected",
      "FourierSHAP",
      "OddSHAP"
    ],
    "n_estimators_matches": false,
    "assertion_n_benchmarks": 8,
    "actual_n_benchmarks": 8,
    "n_benchmarks_matches": true,
    "assertion_oddshap_rank": 1.5,
    "actual_oddshap_rank": 1.5,
    "oddshap_rank_matches": true,
    "oddshap_is_lowest_rank": true,
    "assertion_regressionmsr_rank": 2.25,
    "actual_regressionmsr_rank": 2.62,
    "regressionmsr_rank_matches": false,
    "second_best_estimator": "RegressionMSR",
    "all_ranks": {
      "MSR": 7.75,
      "SV ARM": 6.75,
      "PermutationSampling": 5.62,
      "LeverageSHAP": 3.25,
      "PolySHAP-3": 3.25,
      "RegressionMSR": 2.62,
      "Proxy": 5.0,
      "FFD-RD": 5.75,
      "FFD-RD-Corrected": 8.5,
      "FourierSHAP": 9.12,
      "OddSHAP": 1.5
    },
    "verdict": "FALSIFIED: OddSHAP's rank (1.50) and its first-place finish are correct, and there are indeed 8 benchmarks, but the table lists 11 estimator rows rather than 8, and RegressionMSR's average rank is 2.62, not 2.25."
  },
  "claim5": {
    "rows": [
      {
        "benchmark": "DistilBERT",
        "d": 14,
        "kind": "deep learning (text)",
        "leverageshap_mse": 7.7e-05,
        "oddshap_mse": 4.6e-05,
        "ratio": 1.673913043478261
      },
      {
        "benchmark": "Estate",
        "d": 15,
        "kind": "tabular",
        "leverageshap_mse": 0.00034,
        "oddshap_mse": 5.1e-07,
        "ratio": 666.6666666666667
      },
      {
        "benchmark": "ViT16",
        "d": 16,
        "kind": "deep learning (vision)",
        "leverageshap_mse": 3.5e-05,
        "oddshap_mse": 1.3e-05,
        "ratio": 2.692307692307692
      },
      {
        "benchmark": "Cancer",
        "d": 30,
        "kind": "tabular",
        "leverageshap_mse": 3.2e-05,
        "oddshap_mse": 4.2e-06,
        "ratio": 7.6190476190476195
      },
      {
        "benchmark": "IL60",
        "d": 60,
        "kind": "tabular",
        "leverageshap_mse": 2.6e-05,
        "oddshap_mse": 1.6e-06,
        "ratio": 16.25
      },
      {
        "benchmark": "CG60",
        "d": 60,
        "kind": "tabular",
        "leverageshap_mse": 2.5e-05,
        "oddshap_mse": 6.2e-06,
        "ratio": 4.032258064516129
      },
      {
        "benchmark": "NHANES",
        "d": 79,
        "kind": "tabular",
        "leverageshap_mse": 0.00062,
        "oddshap_mse": 5.8e-05,
        "ratio": 10.689655172413794
      },
      {
        "benchmark": "Crime",
        "d": 101,
        "kind": "tabular",
        "leverageshap_mse": 0.75,
        "oddshap_mse": 0.13,
        "ratio": 5.769230769230769
      }
    ],
    "claimed_interval": [
      6.0,
      62.0
    ],
    "tabular_only": {
      "n": 6,
      "min_ratio": 4.032258064516129,
      "max_ratio": 666.6666666666667,
      "ratios": {
        "Estate": 666.6666666666667,
        "Cancer": 7.6190476190476195,
        "IL60": 16.25,
        "CG60": 4.032258064516129,
        "NHANES": 10.689655172413794,
        "Crime": 5.769230769230769
      },
      "n_inside_interval": 3,
      "n_below_6": 2,
      "n_above_62": 1
    },
    "all_benchmarks": {
      "n": 8,
      "min_ratio": 1.673913043478261,
      "max_ratio": 666.6666666666667
    },
    "verdict": "FALSIFIED: on the paper's own tabular benchmarks the LeverageSHAP/OddSHAP MSE ratio spans 4.03x to 666.7x, which falls outside the claimed 6-62x band at BOTH ends."
  },
  "claim6": {
    "complete_rows_used": [
      "MSR",
      "SV ARM",
      "PermutationSampling",
      "LeverageSHAP",
      "RegressionMSR",
      "Proxy",
      "FourierSHAP",
      "OddSHAP"
    ],
    "partial_rows_excluded": [
      "PolySHAP-3",
      "FFD-RD",
      "FFD-RD-Corrected"
    ],
    "exclusion_reason": "these rows have blank cells in Table 1 and the linearised PDF text does not determine which benchmarks their numbers belong to",
    "rows": [
      {
        "benchmark": "DistilBERT",
        "d": 14,
        "best_baseline": "RegressionMSR",
        "best_baseline_mse": 3.1e-05,
        "oddshap_mse": 4.6e-05,
        "advantage": 0.6739130434782609
      },
      {
        "benchmark": "Estate",
        "d": 15,
        "best_baseline": "RegressionMSR",
        "best_baseline_mse": 1.3e-05,
        "oddshap_mse": 5.1e-07,
        "advantage": 25.49019607843137
      },
      {
        "benchmark": "ViT16",
        "d": 16,
        "best_baseline": "RegressionMSR",
        "best_baseline_mse": 1e-05,
        "oddshap_mse": 1.3e-05,
        "advantage": 0.7692307692307694
      },
      {
        "benchmark": "Cancer",
        "d": 30,
        "best_baseline": "LeverageSHAP",
        "best_baseline_mse": 3.2e-05,
        "oddshap_mse": 4.2e-06,
        "advantage": 7.6190476190476195
      },
      {
        "benchmark": "IL60",
        "d": 60,
        "best_baseline": "RegressionMSR",
        "best_baseline_mse": 9.7e-06,
        "oddshap_mse": 1.6e-06,
        "advantage": 6.062500000000001
      },
      {
        "benchmark": "CG60",
        "d": 60,
        "best_baseline": "LeverageSHAP",
        "best_baseline_mse": 2.5e-05,
        "oddshap_mse": 6.2e-06,
        "advantage": 4.032258064516129
      },
      {
        "benchmark": "NHANES",
        "d": 79,
        "best_baseline": "RegressionMSR",
        "best_baseline_mse": 0.00024,
        "oddshap_mse": 5.8e-05,
        "advantage": 4.137931034482759
      },
      {
        "benchmark": "Crime",
        "d": 101,
        "best_baseline": "RegressionMSR",
        "best_baseline_mse": 0.56,
        "oddshap_mse": 0.13,
        "advantage": 4.307692307692308
      }
    ],
    "max_advantage_benchmark": "Estate",
    "max_advantage_d": 15,
    "max_advantage": 25.49019607843137,
    "low_dim_d_lt_30": {
      "n": 3,
      "max_advantage": 25.49019607843137,
      "median_advantage": 0.7692307692307694
    },
    "high_dim_d_ge_30": {
      "n": 5,
      "max_advantage": 7.6190476190476195,
      "median_advantage": 4.307692307692308
    },
    "n_benchmarks_where_oddshap_loses": 2,
    "n_wins_d_lt_30": 1,
    "n_wins_d_ge_30": 5,
    "reading_maximum": {
      "argmax_d": 15,
      "argmax_advantage": 25.49019607843137,
      "supports_claim": false
    },
    "reading_median": {
      "median_lt_30": 0.7692307692307694,
      "median_ge_30": 4.307692307692308,
      "supports_claim": true
    },
    "reading_consistency": {
      "wins_lt_30": "1/3",
      "wins_ge_30": "5/5",
      "supports_claim": true
    },
    "verdict": "MIXED, and we decline to call it falsified. Table 1 at m~100d answers the claim differently depending on what 'largest advantage' means. Under a single-maximum reading the claim FAILS: the biggest advantage is 25.5x at d=15 (Estate), below 30. Under a median or consistency reading the claim HOLDS: the median advantage is 4.31x at d>=30 versus 0.77x below 30, and OddSHAP beats every complete-row baseline on 5/5 of the d>=30 benchmarks but only 1/3 of the smaller ones. The d<30 group is bimodal (one 25.5x win, two losses) while the d>=30 group is uniformly 4-8x, so the maximum is an outlier and the typical behaviour supports the paper."
  }
}