{
  "outcome_schema": "nmd-vcell-strict-outcome-collection/1.0",
  "schema_version": "1.1",
  "benchmark_release": "repeated-fold-v2.2",
  "record_count": 16,
  "records": [
    {
      "outcome_id": "STRICT-01",
      "model": "ridge_v1",
      "split": "val",
      "endpoint": "rmse_gain_vs_zero",
      "comparator": "zero delta",
      "n_genes": 20,
      "model_mean_rmse": 0.0875036571174859,
      "baseline_mean_rmse": 0.0930628716945647,
      "favorable_difference": 0.005559214577078808,
      "ci95": [
        -0.0040110260061919,
        0.0158380860695615
      ],
      "multiplicity_adjusted_p": null,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-02",
      "model": "ridge_v1",
      "split": "val",
      "endpoint": "rmse_gain_vs_train_mean",
      "comparator": "train mean",
      "n_genes": 20,
      "model_mean_rmse": 0.0875036571174859,
      "baseline_mean_rmse": 0.0951926868408918,
      "favorable_difference": 0.007689029723405905,
      "ci95": [
        0.0004062849748879,
        0.0148798867687583
      ],
      "multiplicity_adjusted_p": null,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-03",
      "model": "ridge_v1",
      "split": "test",
      "endpoint": "rmse_gain_vs_zero",
      "comparator": "zero delta",
      "n_genes": 21,
      "model_mean_rmse": 0.0877140478364058,
      "baseline_mean_rmse": 0.0937586594372987,
      "favorable_difference": 0.00604461160089291,
      "ci95": [
        -0.0010704531733478,
        0.0137440673563451
      ],
      "multiplicity_adjusted_p": 0.6415871682566349,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-04",
      "model": "ridge_v1",
      "split": "test",
      "endpoint": "rmse_gain_vs_train_mean",
      "comparator": "train mean",
      "n_genes": 21,
      "model_mean_rmse": 0.0877140478364058,
      "baseline_mean_rmse": 0.0940848111751533,
      "favorable_difference": 0.006370763338747504,
      "ci95": [
        -0.0013588148612706,
        0.0139120890936326
      ],
      "multiplicity_adjusted_p": 0.6415871682566349,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-05",
      "model": "mlp_v1",
      "split": "val",
      "endpoint": "rmse_gain_vs_zero",
      "comparator": "zero delta",
      "n_genes": 20,
      "model_mean_rmse": 0.0864128831773995,
      "baseline_mean_rmse": 0.0930628716945647,
      "favorable_difference": 0.006649988517165209,
      "ci95": [
        -0.0027104047685861,
        0.0167487915791571
      ],
      "multiplicity_adjusted_p": null,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-06",
      "model": "mlp_v1",
      "split": "val",
      "endpoint": "rmse_gain_vs_train_mean",
      "comparator": "train mean",
      "n_genes": 20,
      "model_mean_rmse": 0.0864128831773995,
      "baseline_mean_rmse": 0.0951926868408918,
      "favorable_difference": 0.008779803663492305,
      "ci95": [
        0.0008339052274823,
        0.0160476823849603
      ],
      "multiplicity_adjusted_p": null,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-07",
      "model": "mlp_v1",
      "split": "test",
      "endpoint": "rmse_gain_vs_zero",
      "comparator": "zero delta",
      "n_genes": 21,
      "model_mean_rmse": 0.0910465029023942,
      "baseline_mean_rmse": 0.0937586594372987,
      "favorable_difference": 0.002712156534904503,
      "ci95": [
        -0.00418536863955,
        0.009659337689213
      ],
      "multiplicity_adjusted_p": 1,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-08",
      "model": "mlp_v1",
      "split": "test",
      "endpoint": "rmse_gain_vs_train_mean",
      "comparator": "train mean",
      "n_genes": 21,
      "model_mean_rmse": 0.0910465029023942,
      "baseline_mean_rmse": 0.0940848111751533,
      "favorable_difference": 0.0030383082727590965,
      "ci95": [
        -0.0043203045569714,
        0.0098500274033063
      ],
      "multiplicity_adjusted_p": 1,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-09",
      "model": "ridge_v2",
      "split": "val",
      "endpoint": "rmse_gain_vs_zero",
      "comparator": "zero delta",
      "n_genes": 20,
      "model_mean_rmse": 0.0905455579981207,
      "baseline_mean_rmse": 0.0930628716945647,
      "favorable_difference": 0.002517313696444007,
      "ci95": [
        -0.0116165499202906,
        0.0141424705414101
      ],
      "multiplicity_adjusted_p": null,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-10",
      "model": "ridge_v2",
      "split": "val",
      "endpoint": "rmse_gain_vs_train_mean",
      "comparator": "train mean",
      "n_genes": 20,
      "model_mean_rmse": 0.0905455579981207,
      "baseline_mean_rmse": 0.0951926868408918,
      "favorable_difference": 0.004647128842771103,
      "ci95": [
        -0.0044592514727264,
        0.0129287437256425
      ],
      "multiplicity_adjusted_p": null,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-11",
      "model": "ridge_v2",
      "split": "test",
      "endpoint": "rmse_gain_vs_zero",
      "comparator": "zero delta",
      "n_genes": 21,
      "model_mean_rmse": 0.0869511488292898,
      "baseline_mean_rmse": 0.0937586594372987,
      "favorable_difference": 0.00680751060800891,
      "ci95": [
        0.0001784727975193,
        0.0141131128805379
      ],
      "multiplicity_adjusted_p": 0.4895902081958361,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-12",
      "model": "ridge_v2",
      "split": "test",
      "endpoint": "rmse_gain_vs_train_mean",
      "comparator": "train mean",
      "n_genes": 21,
      "model_mean_rmse": 0.0869511488292898,
      "baseline_mean_rmse": 0.0940848111751533,
      "favorable_difference": 0.0071336623458635035,
      "ci95": [
        -0.0007120130184505,
        0.0144327056594193
      ],
      "multiplicity_adjusted_p": 0.4895902081958361,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-13",
      "model": "mlp_v2",
      "split": "val",
      "endpoint": "rmse_gain_vs_zero",
      "comparator": "zero delta",
      "n_genes": 20,
      "model_mean_rmse": 0.0894924679771065,
      "baseline_mean_rmse": 0.0930628716945647,
      "favorable_difference": 0.0035704037174582037,
      "ci95": [
        -0.0069059940893202,
        0.0136394311720505
      ],
      "multiplicity_adjusted_p": null,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-14",
      "model": "mlp_v2",
      "split": "val",
      "endpoint": "rmse_gain_vs_train_mean",
      "comparator": "train mean",
      "n_genes": 20,
      "model_mean_rmse": 0.0894924679771065,
      "baseline_mean_rmse": 0.0951926868408918,
      "favorable_difference": 0.0057002188637853,
      "ci95": [
        -0.0018641833635047,
        0.0127864579204469
      ],
      "multiplicity_adjusted_p": null,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-15",
      "model": "mlp_v2",
      "split": "test",
      "endpoint": "rmse_gain_vs_zero",
      "comparator": "zero delta",
      "n_genes": 21,
      "model_mean_rmse": 0.0911172983192262,
      "baseline_mean_rmse": 0.0937586594372987,
      "favorable_difference": 0.002641361118072505,
      "ci95": [
        -0.0045699899838794,
        0.0101698261796541
      ],
      "multiplicity_adjusted_p": 1,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    },
    {
      "outcome_id": "STRICT-16",
      "model": "mlp_v2",
      "split": "test",
      "endpoint": "rmse_gain_vs_train_mean",
      "comparator": "train mean",
      "n_genes": 21,
      "model_mean_rmse": 0.0911172983192262,
      "baseline_mean_rmse": 0.0940848111751533,
      "favorable_difference": 0.0029675128559270986,
      "ci95": [
        -0.0041602398801062,
        0.0097235150723939
      ],
      "multiplicity_adjusted_p": 1,
      "paired_rmse_support": "not_supported",
      "raw_direction_support": "not_supported",
      "strict_gate": "fail",
      "legacy_rmse_only_gate": "pass"
    }
  ]
}