{
  "rows": [
    {
      "metric": "l2",
      "transformation": "identity_replay",
      "adjusted_rand": 1.0,
      "adjusted_rand_core_atoms_only": 1.0,
      "n_modules_original": 3,
      "n_modules_transformed": 3,
      "exact_set_recovery": 1.0,
      "mean_best_jaccard": 1.0,
      "median_best_jaccard": 1.0,
      "mean_best_jaccard_modules_le8": 1.0,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [
        2,
        2,
        2
      ],
      "per_module_best_jaccard": [
        1.0,
        1.0,
        1.0
      ],
      "note": "same data and rule, fresh run; upper bound on agreement achievable under any transformation"
    },
    {
      "metric": "l2",
      "transformation": "positive_powers_of_two_scaling",
      "gram_max_relative_change": 0.0,
      "adjusted_rand": 1.0,
      "adjusted_rand_core_atoms_only": 1.0,
      "n_modules_original": 3,
      "n_modules_transformed": 3,
      "exact_set_recovery": 1.0,
      "mean_best_jaccard": 1.0,
      "median_best_jaccard": 1.0,
      "mean_best_jaccard_modules_le8": 1.0,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [
        2,
        2,
        2
      ],
      "per_module_best_jaccard": [
        1.0,
        1.0,
        1.0
      ],
      "note": "c_j -> a_j c_j, v_j -> v_j/a_j applied to the cached coefficients, output weights and (Fisher) score projections; z_j preserved"
    },
    {
      "metric": "l2",
      "transformation": "positive_half_split_32_atoms",
      "cluster_count_mode": "free_C",
      "adjusted_rand": -0.0036128001617856094,
      "adjusted_rand_unsplit_atoms_only": -0.003494733857488641,
      "adjusted_rand_core_atoms_only": 0.0,
      "n_modules_original": 3,
      "n_modules_transformed": 43,
      "exact_set_recovery": 0.0,
      "mean_best_jaccard": 0.0,
      "median_best_jaccard": 0.0,
      "mean_best_jaccard_modules_le8": 0.0,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [
        2,
        2,
        2
      ],
      "per_module_best_jaccard": [
        0.0,
        0.0,
        0.0
      ],
      "reference_C_original": 3,
      "reference_C_transformed": 31,
      "split_copies_co_assigned": 0.96875,
      "split_copies_co_assigned_core_only": null,
      "n_split_atoms_in_a_core_module": 0,
      "split_atoms_in_a_core_module_originally": 0.0,
      "affinity_max_relative_change_on_unsplit_atoms": {
        "fcm": 0.0,
        "cosine": 0.0
      },
      "collinear_collapse": null,
      "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction."
    },
    {
      "metric": "l2",
      "transformation": "positive_half_split_32_atoms",
      "cluster_count_mode": "matched_C",
      "adjusted_rand": -0.001934228314481383,
      "adjusted_rand_unsplit_atoms_only": 0.0,
      "adjusted_rand_core_atoms_only": 0.0,
      "n_modules_original": 3,
      "n_modules_transformed": 3,
      "exact_set_recovery": 0.0,
      "mean_best_jaccard": 0.0,
      "median_best_jaccard": 0.0,
      "mean_best_jaccard_modules_le8": 0.0,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [
        2,
        2,
        2
      ],
      "per_module_best_jaccard": [
        0.0,
        0.0,
        0.0
      ],
      "reference_C_original": 3,
      "reference_C_transformed": 3,
      "split_copies_co_assigned": 0.09375,
      "split_copies_co_assigned_core_only": null,
      "n_split_atoms_in_a_core_module": 0,
      "split_atoms_in_a_core_module_originally": 0.0,
      "affinity_max_relative_change_on_unsplit_atoms": {
        "fcm": 0.0,
        "cosine": 0.0
      },
      "collinear_collapse": null,
      "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction."
    },
    {
      "metric": "l2",
      "transformation": "positive_half_split_32_atoms",
      "cluster_count_mode": "matched_C_collinear_collapsed",
      "adjusted_rand": 1.0,
      "adjusted_rand_unsplit_atoms_only": 1.0,
      "adjusted_rand_core_atoms_only": 1.0,
      "n_modules_original": 3,
      "n_modules_transformed": 3,
      "exact_set_recovery": 1.0,
      "mean_best_jaccard": 1.0,
      "median_best_jaccard": 1.0,
      "mean_best_jaccard_modules_le8": 1.0,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [
        2,
        2,
        2
      ],
      "per_module_best_jaccard": [
        1.0,
        1.0,
        1.0
      ],
      "reference_C_original": 3,
      "reference_C_transformed": 3,
      "split_copies_co_assigned": 0.0,
      "split_copies_co_assigned_core_only": null,
      "n_split_atoms_in_a_core_module": 0,
      "split_atoms_in_a_core_module_originally": 0.0,
      "affinity_max_relative_change_on_unsplit_atoms": {
        "fcm": 0.0,
        "cosine": 0.0
      },
      "collinear_collapse": {
        "atoms": 3104,
        "merged_groups": 3072,
        "merged_gram_max_relative_difference_from_original": 0.0,
        "collinear_groups_in_original_layer": 3072,
        "original_width": 3072
      },
      "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction."
    },
    {
      "metric": "l2",
      "transformation": "baseline_raw_covariance_affinity_under_scaling",
      "adjusted_rand": 0.8118744351996922,
      "n_modules_original": 3,
      "n_modules_transformed": 3,
      "note": "naive co-activation clustering on W=|E[c_i c_j]| (no mass normalization); the same function-preserving scaling changes this affinity"
    },
    {
      "metric": "fisher",
      "transformation": "identity_replay",
      "adjusted_rand": 1.0,
      "adjusted_rand_core_atoms_only": null,
      "n_modules_original": 0,
      "n_modules_transformed": 0,
      "exact_set_recovery": 0.0,
      "mean_best_jaccard": null,
      "median_best_jaccard": null,
      "mean_best_jaccard_modules_le8": null,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [],
      "per_module_best_jaccard": [],
      "note": "same data and rule, fresh run; upper bound on agreement achievable under any transformation"
    },
    {
      "metric": "fisher",
      "transformation": "positive_powers_of_two_scaling",
      "gram_max_relative_change": 0.0,
      "adjusted_rand": 1.0,
      "adjusted_rand_core_atoms_only": null,
      "n_modules_original": 0,
      "n_modules_transformed": 0,
      "exact_set_recovery": 0.0,
      "mean_best_jaccard": null,
      "median_best_jaccard": null,
      "mean_best_jaccard_modules_le8": null,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [],
      "per_module_best_jaccard": [],
      "note": "c_j -> a_j c_j, v_j -> v_j/a_j applied to the cached coefficients, output weights and (Fisher) score projections; z_j preserved"
    },
    {
      "metric": "fisher",
      "transformation": "positive_half_split_32_atoms",
      "cluster_count_mode": "free_C",
      "adjusted_rand": 0.0,
      "adjusted_rand_unsplit_atoms_only": 0.0,
      "adjusted_rand_core_atoms_only": null,
      "n_modules_original": 0,
      "n_modules_transformed": 63,
      "exact_set_recovery": 0.0,
      "mean_best_jaccard": null,
      "median_best_jaccard": null,
      "mean_best_jaccard_modules_le8": null,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [],
      "per_module_best_jaccard": [],
      "reference_C_original": 2,
      "reference_C_transformed": 30,
      "split_copies_co_assigned": 0.96875,
      "split_copies_co_assigned_core_only": null,
      "n_split_atoms_in_a_core_module": 0,
      "split_atoms_in_a_core_module_originally": 0.0,
      "affinity_max_relative_change_on_unsplit_atoms": {
        "fcm": 0.0,
        "cosine": 0.0
      },
      "collinear_collapse": null,
      "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction."
    },
    {
      "metric": "fisher",
      "transformation": "positive_half_split_32_atoms",
      "cluster_count_mode": "matched_C",
      "adjusted_rand": 0.0,
      "adjusted_rand_unsplit_atoms_only": 1.0,
      "adjusted_rand_core_atoms_only": null,
      "n_modules_original": 0,
      "n_modules_transformed": 1,
      "exact_set_recovery": 0.0,
      "mean_best_jaccard": null,
      "median_best_jaccard": null,
      "mean_best_jaccard_modules_le8": null,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [],
      "per_module_best_jaccard": [],
      "reference_C_original": 2,
      "reference_C_transformed": 2,
      "split_copies_co_assigned": 0.03125,
      "split_copies_co_assigned_core_only": null,
      "n_split_atoms_in_a_core_module": 0,
      "split_atoms_in_a_core_module_originally": 0.0,
      "affinity_max_relative_change_on_unsplit_atoms": {
        "fcm": 0.0,
        "cosine": 0.0
      },
      "collinear_collapse": null,
      "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction."
    },
    {
      "metric": "fisher",
      "transformation": "positive_half_split_32_atoms",
      "cluster_count_mode": "matched_C_collinear_collapsed",
      "adjusted_rand": 1.0,
      "adjusted_rand_unsplit_atoms_only": 1.0,
      "adjusted_rand_core_atoms_only": null,
      "n_modules_original": 0,
      "n_modules_transformed": 0,
      "exact_set_recovery": 0.0,
      "mean_best_jaccard": null,
      "median_best_jaccard": null,
      "mean_best_jaccard_modules_le8": null,
      "mean_best_jaccard_modules_gt8": null,
      "module_sizes": [],
      "per_module_best_jaccard": [],
      "reference_C_original": 2,
      "reference_C_transformed": 2,
      "split_copies_co_assigned": 0.0,
      "split_copies_co_assigned_core_only": null,
      "n_split_atoms_in_a_core_module": 0,
      "split_atoms_in_a_core_module_originally": 0.0,
      "affinity_max_relative_change_on_unsplit_atoms": {
        "fcm": 0.0,
        "cosine": 0.0
      },
      "collinear_collapse": {
        "atoms": 3104,
        "merged_groups": 3072,
        "merged_gram_max_relative_difference_from_original": 0.0,
        "collinear_groups_in_original_layer": 3072,
        "original_width": 3072
      },
      "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction."
    },
    {
      "metric": "fisher",
      "transformation": "baseline_raw_covariance_affinity_under_scaling",
      "adjusted_rand": 0.8118744351996922,
      "n_modules_original": 3,
      "n_modules_transformed": 3,
      "note": "naive co-activation clustering on W=|E[c_i c_j]| (no mass normalization); the same function-preserving scaling changes this affinity"
    }
  ],
  "settings": {
    "root": "results/comparison/random_gpt2",
    "out": "results/modules/random_gpt2",
    "metrics": [
      "l2",
      "fisher"
    ],
    "bootstraps": 12,
    "split_atoms": 32,
    "seed": 17,
    "affinity": "square",
    "knn": 0,
    "normalization": "fcm",
    "cluster_rule": "reference",
    "clusters": null,
    "min_clusters": 2,
    "max_clusters": 32,
    "threshold": 0.8,
    "min_size": 2,
    "max_module_fraction": 0.5,
    "restarts": 5
  },
  "seconds": 29.21317982673645,
  "scope": "partition identity only; semantic and causal changes under reparameterization are not remeasured here",
  "ari_convention": "the rest module is treated as one additional label, so ARI scores the whole assignment including which neurons are declared unstructured"
}