{
  "rows": 20,
  "results": {
    "arithmetic": {
      "rows": 16,
      "positives": 8,
      "methods": {
        "goodfire_frozen": {
          "original_auroc": 0.703125,
          "reformatted_auroc": 0.8125,
          "threshold_decision_flip_rate": 0.0
        },
        "difference_means": {
          "original_auroc": 0.484375,
          "reformatted_auroc": 0.515625,
          "threshold_decision_flip_rate": 0.0625
        },
        "covariance_lda": {
          "original_auroc": 0.953125,
          "reformatted_auroc": 0.90625,
          "threshold_decision_flip_rate": 0.0
        },
        "linear_layer12": {
          "original_auroc": 0.890625,
          "reformatted_auroc": 0.5625,
          "threshold_decision_flip_rate": 0.375
        },
        "linear_layer24": {
          "original_auroc": 0.796875,
          "reformatted_auroc": 0.953125,
          "threshold_decision_flip_rate": 0.1875
        },
        "linear_layer32": {
          "original_auroc": 0.875,
          "reformatted_auroc": 0.84375,
          "threshold_decision_flip_rate": 0.625
        },
        "sparse_l1": {
          "original_auroc": 0.78125,
          "reformatted_auroc": 0.84375,
          "threshold_decision_flip_rate": 0.125
        },
        "pca_linear": {
          "original_auroc": 0.78125,
          "reformatted_auroc": 0.90625,
          "threshold_decision_flip_rate": 0.5625
        },
        "rbf_kernel": {
          "original_auroc": 0.640625,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.0
        },
        "small_mlp": {
          "original_auroc": 0.75,
          "reformatted_auroc": 0.90625,
          "threshold_decision_flip_rate": 0.5625
        },
        "last_token": {
          "original_auroc": 0.546875,
          "reformatted_auroc": 0.78125,
          "threshold_decision_flip_rate": 0.1875
        },
        "multilayer": {
          "original_auroc": 0.84375,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.375
        },
        "pre_action": {
          "original_auroc": 0.625,
          "reformatted_auroc": 0.625,
          "threshold_decision_flip_rate": 0.0
        },
        "temporal_delta": {
          "original_auroc": 0.65625,
          "reformatted_auroc": 0.890625,
          "threshold_decision_flip_rate": 0.4375
        },
        "token_dispersion": {
          "original_auroc": 0.875,
          "reformatted_auroc": 0.875,
          "threshold_decision_flip_rate": 0.125
        },
        "cross_layer_increment": {
          "original_auroc": 0.828125,
          "reformatted_auroc": 0.984375,
          "threshold_decision_flip_rate": 0.375
        },
        "adapter_gradient": {
          "original_auroc": 0.53125,
          "reformatted_auroc": 0.34375,
          "threshold_decision_flip_rate": 0.1875
        },
        "gradient_centroids": {
          "original_auroc": 0.8125,
          "reformatted_auroc": 0.1875,
          "threshold_decision_flip_rate": 0.0
        },
        "likelihood_statistics": {
          "original_auroc": 0.53125,
          "reformatted_auroc": 0.40625,
          "threshold_decision_flip_rate": 0.0
        },
        "action_residualized": {
          "original_auroc": 0.9375,
          "reformatted_auroc": 0.609375,
          "threshold_decision_flip_rate": 0.0
        },
        "paired_contrast_pca": {
          "original_auroc": 0.890625,
          "reformatted_auroc": 0.9375,
          "threshold_decision_flip_rate": 0.0
        },
        "class_density": {
          "original_auroc": 0.78125,
          "reformatted_auroc": 0.984375,
          "threshold_decision_flip_rate": 0.5
        },
        "response_text": {
          "original_auroc": 0.625,
          "reformatted_auroc": 0.6875,
          "threshold_decision_flip_rate": 0.0
        },
        "context_and_text": {
          "original_auroc": 0.859375,
          "reformatted_auroc": 0.890625,
          "threshold_decision_flip_rate": 0.0625
        },
        "nuisance_only": {
          "original_auroc": 0.9609375,
          "reformatted_auroc": 0.9609375,
          "threshold_decision_flip_rate": 0.0
        },
        "permuted_label_control": {
          "original_auroc": 0.390625,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "random_direction": {
          "original_auroc": 0.5,
          "reformatted_auroc": 0.96875,
          "threshold_decision_flip_rate": 0.0
        },
        "mean_ensemble": {
          "original_auroc": 0.828125,
          "reformatted_auroc": 0.953125,
          "threshold_decision_flip_rate": 0.5625
        },
        "pca_consensus": {
          "original_auroc": 0.796875,
          "reformatted_auroc": 0.9375,
          "threshold_decision_flip_rate": 0.5625
        }
      }
    },
    "inventory": {
      "rows": 4,
      "positives": 2,
      "methods": {
        "goodfire_frozen": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "difference_means": {
          "original_auroc": 1.0,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.0
        },
        "covariance_lda": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.0
        },
        "linear_layer12": {
          "original_auroc": 0.75,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.0
        },
        "linear_layer24": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "linear_layer32": {
          "original_auroc": 0.75,
          "reformatted_auroc": 0.5,
          "threshold_decision_flip_rate": 1.0
        },
        "sparse_l1": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "pca_linear": {
          "original_auroc": 0.0,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.25
        },
        "rbf_kernel": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.5,
          "threshold_decision_flip_rate": 0.0
        },
        "small_mlp": {
          "original_auroc": 0.0,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.25
        },
        "last_token": {
          "original_auroc": 0.5,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.0
        },
        "multilayer": {
          "original_auroc": 0.5,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.0
        },
        "pre_action": {
          "original_auroc": 0.5,
          "reformatted_auroc": 0.5,
          "threshold_decision_flip_rate": 0.0
        },
        "temporal_delta": {
          "original_auroc": 0.5,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.25
        },
        "token_dispersion": {
          "original_auroc": 1.0,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.25
        },
        "cross_layer_increment": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.5,
          "threshold_decision_flip_rate": 0.0
        },
        "adapter_gradient": {
          "original_auroc": 0.25,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "gradient_centroids": {
          "original_auroc": 0.25,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "likelihood_statistics": {
          "original_auroc": 0.75,
          "reformatted_auroc": 0.75,
          "threshold_decision_flip_rate": 0.0
        },
        "action_residualized": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.0
        },
        "paired_contrast_pca": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "class_density": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "response_text": {
          "original_auroc": 1.0,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.0
        },
        "context_and_text": {
          "original_auroc": 1.0,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.25
        },
        "nuisance_only": {
          "original_auroc": 1.0,
          "reformatted_auroc": 1.0,
          "threshold_decision_flip_rate": 0.0
        },
        "permuted_label_control": {
          "original_auroc": 0.25,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.25
        },
        "random_direction": {
          "original_auroc": 0.5,
          "reformatted_auroc": 0.5,
          "threshold_decision_flip_rate": 0.0
        },
        "mean_ensemble": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.0,
          "threshold_decision_flip_rate": 0.0
        },
        "pca_consensus": {
          "original_auroc": 0.0,
          "reformatted_auroc": 0.25,
          "threshold_decision_flip_rate": 0.25
        }
      }
    }
  },
  "semantics_verified": true,
  "warning": "Selected balanced subset of spent stress rows; paired sensitivity only. No refit or threshold adjustment."
}