{ "experiment": "P5 eigen-role probe at PRE-POOL token level (R_role from COMPOSITION operator)", "config": { "n_pairs": 2500, "n_matched_pairs": 2500, "n_span_tokens": 131416, "n_pi_planes": 32, "n_id_planes": 32, "seed": 0, "d": 1024, "role_source": "tae_comp_W_order.pt" }, "R_role_meta": { "orth_dev": 0.0, "source": "tae_comp_W_order.pt S_matrix (z_AB->z_BA), polar-orthogonal part", "singular_value_spread": [ 1.9996, 0.0 ] }, "role_plane_angles": { "pi_subspace_mean_rad": 3.1039, "pi_subspace_min_rad": 3.066, "near_identity_mean_rad": 0.00337, "rand_planes_mean_rad": 0.9819, "max_plane_rad": 3.1416, "median_plane_rad": 0.1141 }, "prediction": "pi-subspace does NOT separate spans better than matched random subspace (R_role = generic content-mixing rotation)", "falsifier": "pi-subspace cleanly separates spans (>0.90 acc/auc) AND beats random subspace", "span_separation": { "raw_full_space": { "acc": 0.9786, "auc": 0.9968, "n_train": 105234, "n_test": 26182 }, "pi_eigen_subspace": { "acc": 0.9284, "auc": 0.9704, "n_train": 105234, "n_test": 26182 }, "near_identity_subspace": { "acc": 0.6298, "auc": 0.6517, "n_train": 105234, "n_test": 26182 }, "random_orthonormal_subspace": { "acc": 0.846, "auc": 0.9007, "n_train": 105234, "n_test": 26182 }, "random_eigenplane_subspace": { "acc": 0.8773, "auc": 0.9362, "n_train": 105234, "n_test": 26182 }, "pi_minus_identity_acc": 0.2986, "pi_minus_random_acc": 0.0824, "pi_minus_random_auc": 0.0697, "subspace_mean_coord_norm": { "pi": 0.1727, "near_identity": 0.237, "random_orthonormal": 0.2088, "random_eigenplane": 0.2006 } }, "confounds": { "pi_after_R_pos_inv": { "acc": 0.6949, "auc": 0.7558, "n_train": 105234, "n_test": 26182 }, "R_pos_inv_acc_drop": 0.2335, "pi_after_R_role_inv": { "acc": 0.9284, "auc": 0.9704, "n_train": 105234, "n_test": 26182 }, "R_role_inv_acc_drop": 0.0, "position_stratified": { "n_tokens": 29302, "n_positions_used": 44, "pi": { "acc": 0.8357, "auc": 0.903, "n_train": 23471, "n_test": 5831 }, "random_orthonormal": { "acc": 0.7652, "auc": 0.8278, "n_train": 23471, "n_test": 5831 }, "raw_full": { "acc": 0.947, "auc": 0.9854, "n_train": 23471, "n_test": 5831 }, "pi_minus_random_acc": 0.0705 } }, "verdict": { "falsified": true, "summary": "FALSIFIED: pi-subspace cleanly separates spans AND beats random -> role reading RESURRECTED at token level (comp-operator R_role)", "key_numbers": { "pi_acc": 0.9284, "pi_auc": 0.9704, "random_acc": 0.846, "random_auc": 0.9007, "near_id_acc": 0.6298, "raw_acc": 0.9786, "pi_minus_random_acc": 0.0824 } }, "examples": [ { "A": "The fate of the ACA repeal bill hangs in the balance, and the impact on women's health is a pressing concern.", "B": "The American people deserve a healthcare system that prioritizes prevention, equality, and access to essential services," }, { "A": "The American people deserve a healthcare system that prioritizes prevention, equality, and access to essential services,", "B": "The outcome of this vote will have far-reaching consequences for the health and well-being of millions of Americans, par" }, { "A": "This news comes as a shock to many Americans, who are already feeling the pinch of rising grocery bills.", "B": "The USDA attributes the price increase to higher costs for corn, soybeans, and wheat, as well as increased demand for me" }, { "A": "According to a recent survey, American beef consumption is shifting towards more grass-fed and leaner options.", "B": "This trend is driven by consumer demand for healthier, more sustainable meat options." } ] }