{
 "kind": "experiment_dev",
 "stage": "dev",
 "completed": true,
 "started_utc": "2026-11-12T13:11:09Z",
 "sections": {
  "setup": {
   "code_commit": "ea91d4b59ea96f6652d0dfc7f6b294cc9aa521f0",
   "argv": [
    "--out",
    "/kaggle/working/fire_dev",
    "--stage",
    "dev",
    "--dev-arms",
    "pooled"
   ],
   "torch": "2.11.0+cu130",
   "transformers": "5.20.2",
   "safetune": "0.1.8",
   "auditkit": "1.1.3",
   "gpus": [
    "Tesla T4",
    "Tesla T4"
   ],
   "model": {
    "hf_repo": "CohereLabs/tiny-aya-fire",
    "hf_revision": "73500c80fcb6c8bea3e830c21d5d2920249a0e1a",
    "weights_source": "cohere-labs/tiny-aya/transformers/fire/1",
    "model_dir": "/kaggle/input/models/cohere-labs/tiny-aya/transformers/fire/1",
    "weights_match_hf_lfs_sha256": true,
    "class": "Cohere2ForCausalLM",
    "n_layers": 36,
    "dtype": "float16",
    "tokenizer_loader": "AutoTokenizer"
   },
   "corpus": {
    "dataset_version": "0.2.0",
    "sha256": "fd7a5b7b4aa6eb03a889fbbe68117af63ac8652e667aa48ac5c245d4a3df8763",
    "n_records": 576,
    "n_families": 72
   },
   "splits_read": [
    "calibration",
    "dev"
   ],
   "layers": [
    10,
    14,
    18,
    22
   ],
   "rhos": [
    0.1,
    0.2,
    0.4,
    0.8
   ],
   "max_new_tokens": 64,
   "seed": 20261103,
   "frozen_selection": null
  },
  "vectors": {
   "api": "safetune extract_caa_vectors engine (SteeringVectorExtractor), pinned 26f59e6",
   "n_calibration_families": 24,
   "cross_check_vs_independent_capture": {
    "harm_ur@L10": 0.9999999999999745,
    "harm_ro@L10": 0.999999999999966,
    "script@L10": 0.9999999999999946,
    "harm_ur@L14": 0.9999999999999939,
    "harm_ro@L14": 0.9999999999999856,
    "script@L14": 0.9999999999999939,
    "harm_ur@L18": 0.9999999999999983,
    "harm_ro@L18": 0.9999999999999913,
    "script@L18": 0.9999999999999954,
    "harm_ur@L22": 0.9999999999999976,
    "harm_ro@L22": 0.9999999999999942,
    "script@L22": 0.999999999999997
   },
   "min_cosine": 0.999999999999966,
   "geometry": {
    "10": {
     "norm_pooled": 3.7113367184054953,
     "mean_activation_norm": 30.38412857055664,
     "alpha_per_unit_rho": 8.186842336313424,
     "residualization": {
      "rank": 1,
      "removed_frac": 0.03091887921070859,
      "alignment": 0.03091887921070894,
      "degenerate": false,
      "norm_before": 3.7113367184054953,
      "norm_after": 3.7113367184054953
     },
     "random_subspace_removed_frac": 0.023711457587028365,
     "cos_to_pooled": {
      "pooled": 0.9999999999999999,
      "residualized": 0.9995218971630156,
      "random_subspace": 0.9997188438651633,
      "orthogonal_null": -2.015067143599505e-18,
      "ur_only": 0.9411106202737844,
      "ro_only": 0.8805422272948876
     }
    },
    "14": {
     "norm_pooled": 7.097018272255151,
     "mean_activation_norm": 32.370670318603516,
     "alpha_per_unit_rho": 4.56116485498598,
     "residualization": {
      "rank": 1,
      "removed_frac": 0.07375036070944924,
      "alignment": 0.07375036070944904,
      "degenerate": false,
      "norm_before": 7.097018272255151,
      "norm_after": 7.097018272255151
     },
     "random_subspace_removed_frac": 0.00882400512338509,
     "cos_to_pooled": {
      "pooled": 0.9999999999999999,
      "residualized": 0.99727673405892,
      "random_subspace": 0.9999610677089296,
      "orthogonal_null": 4.0778393413209617e-17,
      "ur_only": 0.971577001993745,
      "ro_only": 0.905243880686792
     }
    },
    "18": {
     "norm_pooled": 16.7456759283363,
     "mean_activation_norm": 45.11181640625,
     "alpha_per_unit_rho": 2.693938220189354,
     "residualization": {
      "rank": 1,
      "removed_frac": 0.09117933299223935,
      "alignment": 0.09117933299223976,
      "degenerate": false,
      "norm_before": 16.7456759283363,
      "norm_after": 16.745675928336297
     },
     "random_subspace_removed_frac": 0.019897987818650285,
     "cos_to_pooled": {
      "pooled": 1.0,
      "residualized": 0.9958344888760836,
      "random_subspace": 0.9998020154414421,
      "orthogonal_null": 3.8008094601110935e-17,
      "ur_only": 0.9843667705274792,
      "ro_only": 0.9300613652004712
     }
    },
    "22": {
     "norm_pooled": 22.289392885163473,
     "mean_activation_norm": 56.450401306152344,
     "alpha_per_unit_rho": 2.532612781199955,
     "residualization": {
      "rank": 1,
      "removed_frac": 0.08543006844983284,
      "alignment": 0.0854300684498333,
      "degenerate": false,
      "norm_before": 22.289392885163473,
      "norm_after": 22.289392885163476
     },
     "random_subspace_removed_frac": 0.026382421720784695,
     "cos_to_pooled": {
      "pooled": 1.0,
      "residualized": 0.9963441691527367,
      "random_subspace": 0.9996519233333905,
      "orthogonal_null": 3.7542486453993056e-17,
      "ur_only": 0.9844299467617393,
      "ro_only": 0.9323208828754252
     }
    }
   },
   "arms_are_norm_matched_to_pooled": true
  },
  "mechanism_diagnostics": {
   "scope": "24 calibration families; cross-fitted (leave-one-family-out); not a behavioural result",
   "by_layer": {
    "2": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 20.02701174575979,
     "norm_v_ur": 0.7054855784766502,
     "norm_v_ro": 0.5666815946775148,
     "norm_v_pooled": 0.5888035837596972,
     "norm_script": 4.019634602915408,
     "v_pooled_over_activation": 0.02940047128520616,
     "norm_ratio_ro_over_ur": {
      "value": 0.8032504305773992,
      "ci95": [
       0.6583837799411619,
       0.9142550690544475
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.710281433619759,
      "ci95": [
       0.41083979291923045,
       0.765223658516206
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.4923933236152374,
      "within_script_ceiling_ur": 0.5683779124046753,
      "disattenuated": 0.8663132624771348,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.1883114620890122,
      "removed_norm_fraction_if_residualized": 0.1883114620890122,
      "removed_variance_fraction_if_residualized": 0.03546120675410148,
      "cross_fitted_cosine": -0.18211139835054815,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -0.7320212783951754,
       "sd": 0.15254544536967812,
       "cohens_d": -4.7987095033955125,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 0.0
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017316780244723735,
       "p95": 0.04176551443346247
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 0.519775026933076,
      "mean_projection_ro": 0.40870468098868534,
      "ratio_ro_over_ur": 0.7863107302408132,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 0.11107034594439073,
       "sd": 0.24718129126318816,
       "cohens_d": 0.4493477049851954,
       "p_two_sided": 0.03719256148770246,
       "frac_positive": 0.7083333333333334
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "4": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 22.014663512896302,
     "norm_v_ur": 1.6889816795329953,
     "norm_v_ro": 1.427335884487033,
     "norm_v_pooled": 1.3373090597336497,
     "norm_script": 6.538079760225022,
     "v_pooled_over_activation": 0.060746286626195634,
     "norm_ratio_ro_over_ur": {
      "value": 0.8450866588924116,
      "ci95": [
       0.7315460053340685,
       1.0493141732842213
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.46948889934621596,
      "ci95": [
       0.21860999788087732,
       0.5819430817926212
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.26218087938424356,
      "within_script_ceiling_ur": 0.4714896341949412,
      "disattenuated": 0.5560692332757473,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.20823305462039757,
      "removed_norm_fraction_if_residualized": 0.20823305462039757,
      "removed_variance_fraction_if_residualized": 0.04336100503654148,
      "cross_fitted_cosine": -0.20000181577892578,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -1.3076278237524481,
       "sd": 0.35464072731859386,
       "cohens_d": -3.68719022668181,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 0.0
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017106511905279634,
       "p95": 0.042701958738366116
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 0.9978438492309789,
      "mean_projection_ro": 0.7394637617980138,
      "ratio_ro_over_ur": 0.7410616023417951,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 0.25838008743296514,
       "sd": 0.4611501296636269,
       "cohens_d": 0.5602949469437064,
       "p_two_sided": 0.010797840431913617,
       "frac_positive": 0.6666666666666666
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "6": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 24.87406530262934,
     "norm_v_ur": 2.3579887289675523,
     "norm_v_ro": 1.8958962789590643,
     "norm_v_pooled": 1.864299445470884,
     "norm_script": 7.300913952333828,
     "v_pooled_over_activation": 0.0749495276622039,
     "norm_ratio_ro_over_ur": {
      "value": 0.8040311031466144,
      "ci95": [
       0.7376982124774952,
       0.9393722904031718
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.5310271735036105,
      "ci95": [
       0.3070156407728306,
       0.633381918387227
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.30301392474084954,
      "within_script_ceiling_ur": 0.5136977311645079,
      "disattenuated": 0.5898681390979544,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.16311730919819567,
      "removed_norm_fraction_if_residualized": 0.16311730919819567,
      "removed_variance_fraction_if_residualized": 0.02660725656005977,
      "cross_fitted_cosine": -0.1529955565463166,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -1.117007393434082,
       "sd": 0.499105397320206,
       "cohens_d": -2.2380190625697742,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 0.0
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017691760189541685,
       "p95": 0.043150591964316955
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 1.483861378636935,
      "mean_projection_ro": 1.0691421581304201,
      "ratio_ro_over_ur": 0.7205135018154639,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 0.41471922050651466,
       "sd": 0.5493992558488362,
       "cohens_d": 0.7548594507390852,
       "p_two_sided": 0.0007998400319936012,
       "frac_positive": 0.75
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "8": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 25.585389917937505,
     "norm_v_ur": 3.1229874678286764,
     "norm_v_ro": 2.4061725729238126,
     "norm_v_pooled": 2.484397145917544,
     "norm_script": 8.218541193077442,
     "v_pooled_over_activation": 0.0971021803414367,
     "norm_ratio_ro_over_ur": {
      "value": 0.7704714148586562,
      "ci95": [
       0.7043894521165006,
       0.895500575881273
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.6085743764276914,
      "ci95": [
       0.3939722573442848,
       0.6624823373192088
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.40218160638227074,
      "within_script_ceiling_ur": 0.6048033229971659,
      "disattenuated": 0.6649791611415392,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.11399492321429129,
      "removed_norm_fraction_if_residualized": 0.11399492321429129,
      "removed_variance_fraction_if_residualized": 0.012994842518632168,
      "cross_fitted_cosine": -0.10632104116197011,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -0.8738038564805336,
       "sd": 0.6668703102583178,
       "cohens_d": -1.3103055317338366,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 0.08333333333333333
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017560855741424568,
       "p95": 0.04494192986070802
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 2.24054215290578,
      "mean_projection_ro": 1.5562886130425373,
      "ratio_ro_over_ur": 0.6946035855760053,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 0.684253539863243,
       "sd": 0.9241691244256763,
       "cohens_d": 0.7403986151219577,
       "p_two_sided": 0.0013997200559888023,
       "frac_positive": 0.7083333333333334
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "10": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 30.383539118162894,
     "norm_v_ur": 4.730697026372623,
     "norm_v_ro": 3.375829167671868,
     "norm_v_pooled": 3.71244189427918,
     "norm_script": 8.886836971099665,
     "v_pooled_over_activation": 0.12218595996474714,
     "norm_ratio_ro_over_ur": {
      "value": 0.7136007968492472,
      "ci95": [
       0.6571025941604037,
       0.8258624008977763
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.6685394150044609,
      "ci95": [
       0.4988345895252475,
       0.7139931587452231
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.45677259747300825,
      "within_script_ceiling_ur": 0.6648918869867957,
      "disattenuated": 0.6869877741222603,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.031055066269215077,
      "removed_norm_fraction_if_residualized": 0.031055066269215077,
      "removed_variance_fraction_if_residualized": 0.00096441714098534,
      "cross_fitted_cosine": -0.0202190610438074,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -0.17968349920502857,
       "sd": 0.8842413184115068,
       "cohens_d": -0.20320640470388848,
       "p_two_sided": 0.34513097380523894,
       "frac_positive": 0.4166666666666667
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017617339772456716,
       "p95": 0.043285897415509346
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 3.6000938454103952,
      "mean_projection_ro": 2.2921633668362476,
      "ratio_ro_over_ur": 0.6366954488584868,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 1.307930478574148,
       "sd": 1.1312350853414739,
       "cohens_d": 1.1561968820825044,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 0.875
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "12": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 30.683792615901538,
     "norm_v_ur": 5.528382370069933,
     "norm_v_ro": 3.4982726059225513,
     "norm_v_pooled": 4.137508199764595,
     "norm_script": 8.765132968900918,
     "v_pooled_over_activation": 0.134843441668302,
     "norm_ratio_ro_over_ur": {
      "value": 0.6327841259428477,
      "ci95": [
       0.5884785481569564,
       0.7670876733084382
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.6637863031516282,
      "ci95": [
       0.5107032634046833,
       0.7048965061250959
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.47737091661524617,
      "within_script_ceiling_ur": 0.7601134140987652,
      "disattenuated": 0.6280259073986281,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.07031596732253576,
      "removed_norm_fraction_if_residualized": 0.07031596732253576,
      "removed_variance_fraction_if_residualized": 0.004944335260503916,
      "cross_fitted_cosine": -0.06571452025269495,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -0.5759965080024035,
       "sd": 0.9828257244156978,
       "cohens_d": -0.5860616930278668,
       "p_two_sided": 0.007198560287942412,
       "frac_positive": 0.2916666666666667
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017688827628712444,
       "p95": 0.042892233861560436
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 4.525198118479859,
      "mean_projection_ro": 2.3982855148685887,
      "ratio_ro_over_ur": 0.5299846442246466,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 2.12691260361127,
       "sd": 1.4709880322645126,
       "cohens_d": 1.445907483242399,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 0.9583333333333334
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "14": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 32.370809196180275,
     "norm_v_ur": 9.61592682846561,
     "norm_v_ro": 5.359057087820431,
     "norm_v_pooled": 7.0969426863319995,
     "norm_script": 9.245270371169154,
     "v_pooled_over_activation": 0.21923896444236654,
     "norm_ratio_ro_over_ur": {
      "value": 0.5573105103042431,
      "ci95": [
       0.49731631673866766,
       0.6747869613491206
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.7789370845520001,
      "ci95": [
       0.6283436528216564,
       0.8184886288013975
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.6580011615882949,
      "within_script_ceiling_ur": 0.8822526052809876,
      "disattenuated": 0.7458194599252318,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.07344678404328887,
      "removed_norm_fraction_if_residualized": 0.07344678404328887,
      "removed_variance_fraction_if_residualized": 0.005394430086301513,
      "cross_fitted_cosine": -0.06959524053696149,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -0.6434268153107605,
       "sd": 1.0546701835561634,
       "cohens_d": -0.6100739599381086,
       "p_two_sided": 0.003999200159968006,
       "frac_positive": 0.3333333333333333
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017654713390185452,
       "p95": 0.04322910558496129
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 8.747441261630044,
      "mean_projection_ro": 4.335900579673427,
      "ratio_ro_over_ur": 0.49567644411543643,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 4.411540681956616,
       "sd": 2.4713520845735446,
       "cohens_d": 1.7850717060891264,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "16": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 35.629366276025586,
     "norm_v_ur": 15.444032338126272,
     "norm_v_ro": 7.717135821978578,
     "norm_v_pooled": 11.02154883633348,
     "norm_script": 10.629420078308149,
     "v_pooled_over_activation": 0.3093388962056757,
     "norm_ratio_ro_over_ur": {
      "value": 0.4996839978719476,
      "ci95": [
       0.45205749099845444,
       0.6101518658140545
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.7879678279300264,
      "ci95": [
       0.6388191060634045,
       0.828428772060336
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.6683682560653014,
      "within_script_ceiling_ur": 0.9140478913926512,
      "disattenuated": 0.7312179836080249,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.06838205573102543,
      "removed_norm_fraction_if_residualized": 0.06838205573102543,
      "removed_variance_fraction_if_residualized": 0.004676105546001068,
      "cross_fitted_cosine": -0.06533711930153907,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -0.6944956877625943,
       "sd": 1.329344302618875,
       "cohens_d": -0.52243477208606,
       "p_two_sided": 0.01839632073585283,
       "frac_positive": 0.3333333333333333
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017738161131713816,
       "p95": 0.04353285999786424
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 14.326019526890319,
      "mean_projection_ro": 6.251491787865071,
      "ratio_ro_over_ur": 0.4363732560974704,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 8.074527739025248,
       "sd": 4.136380018296655,
       "cohens_d": 1.9520758980820885,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "18": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 45.112306312402524,
     "norm_v_ur": 23.415991539121233,
     "norm_v_ro": 11.223553662507218,
     "norm_v_pooled": 16.744747633284348,
     "norm_script": 11.977640922077315,
     "v_pooled_over_activation": 0.37117915269787016,
     "norm_ratio_ro_over_ur": {
      "value": 0.479311484365544,
      "ci95": [
       0.4131106078831059,
       0.6141473989498542
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.8509348798024239,
      "ci95": [
       0.6940898471172451,
       0.8811888719322982
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.7482109766040759,
      "within_script_ceiling_ur": 0.9409062465061195,
      "disattenuated": 0.7952024756796102,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.09117593767297523,
      "removed_norm_fraction_if_residualized": 0.09117593767297523,
      "removed_variance_fraction_if_residualized": 0.008313051610546263,
      "cross_fitted_cosine": -0.08933611564084668,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -1.0700359145192364,
       "sd": 1.8146354687139588,
       "cohens_d": -0.5896698995295051,
       "p_two_sided": 0.012397520495900819,
       "frac_positive": 0.25
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017900976837297276,
       "p95": 0.042714730000056156
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 22.283598748249062,
      "mean_projection_ro": 9.65929653572399,
      "ratio_ro_over_ur": 0.43347112128748827,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 12.624302212525071,
       "sd": 7.861813859904071,
       "cohens_d": 1.6057747534459066,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "20": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 44.80642214547191,
     "norm_v_ur": 27.328201742163643,
     "norm_v_ro": 12.544948277398444,
     "norm_v_pooled": 19.35587030798473,
     "norm_script": 12.337032649183815,
     "v_pooled_over_activation": 0.4319887502988411,
     "norm_ratio_ro_over_ur": {
      "value": 0.45904770448336235,
      "ci95": [
       0.3918533729005349,
       0.5842951821600233
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.8668888408409491,
      "ci95": [
       0.7134841571662351,
       0.8982074211553994
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.7709734472789666,
      "within_script_ceiling_ur": 0.948335860554723,
      "disattenuated": 0.8129751065492669,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.13275961193241712,
      "removed_norm_fraction_if_residualized": 0.13275961193241712,
      "removed_variance_fraction_if_residualized": 0.01762511456044599,
      "cross_fitted_cosine": -0.127233025968915,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -1.5696779954329567,
       "sd": 2.4298694289495852,
       "cohens_d": -0.6459927339023797,
       "p_two_sided": 0.004999000199960008,
       "frac_positive": 0.16666666666666666
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017295411141524605,
       "p95": 0.04376384107499493
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 26.178345837488962,
      "mean_projection_ro": 10.982378916594433,
      "ratio_ro_over_ur": 0.4195215001272925,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 15.195966920894527,
       "sd": 8.410786802057551,
       "cohens_d": 1.8067235894241311,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "22": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 56.4506084944179,
     "norm_v_ur": 31.004756602129397,
     "norm_v_ro": 15.067637723598978,
     "norm_v_pooled": 22.285107189589123,
     "norm_script": 19.46057845419826,
     "v_pooled_over_activation": 0.3947717798611999,
     "norm_ratio_ro_over_ur": {
      "value": 0.4859782618826989,
      "ci95": [
       0.4272597618874692,
       0.6146421725455413
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.854269362626099,
      "ci95": [
       0.6890145592506954,
       0.8919955195364336
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.7471283733904225,
      "within_script_ceiling_ur": 0.9423910699507384,
      "disattenuated": 0.7928007779503653,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.08545174248008111,
      "removed_norm_fraction_if_residualized": 0.08545174248008111,
      "removed_variance_fraction_if_residualized": 0.007302000292882099,
      "cross_fitted_cosine": -0.08172289939914464,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -1.590374895261606,
       "sd": 2.3574052476778786,
       "cohens_d": -0.6746294031661197,
       "p_two_sided": 0.003799240151969606,
       "frac_positive": 0.16666666666666666
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.01697155968366709,
       "p95": 0.04099565505293194
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 29.50718450456807,
      "mean_projection_ro": 12.992083002653724,
      "ratio_ro_over_ur": 0.4403023609603414,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 16.515101501914348,
       "sd": 9.121103750525306,
       "cohens_d": 1.810647258667922,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "24": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 63.238635295414134,
     "norm_v_ur": 34.490340009542976,
     "norm_v_ro": 16.003309866484447,
     "norm_v_pooled": 24.40224249124607,
     "norm_script": 26.9016327284823,
     "v_pooled_over_activation": 0.3858755391740031,
     "norm_ratio_ro_over_ur": {
      "value": 0.46399397228489375,
      "ci95": [
       0.4062191318318604,
       0.5940568474980071
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.8480586950293586,
      "ci95": [
       0.6636080811055608,
       0.8855362265162743
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.7331031337927473,
      "within_script_ceiling_ur": 0.9395925070891238,
      "disattenuated": 0.7802351852122739,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.03721544379643512,
      "removed_norm_fraction_if_residualized": 0.03721544379643512,
      "removed_variance_fraction_if_residualized": 0.0013849892569656212,
      "cross_fitted_cosine": -0.03278426110627068,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -0.8819501515555608,
       "sd": 2.2998314050971573,
       "cohens_d": -0.3834846978786701,
       "p_two_sided": 0.07618476304739052,
       "frac_positive": 0.2916666666666667
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017007138809947,
       "p95": 0.041042013431357785
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 32.77983575716328,
      "mean_projection_ro": 13.629826751458191,
      "ratio_ro_over_ur": 0.41579911664077523,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 19.15000900570509,
       "sd": 10.51811968364085,
       "cohens_d": 1.8206684827412338,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "26": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 68.81087932008646,
     "norm_v_ur": 36.55714894441734,
     "norm_v_ro": 17.32262253287678,
     "norm_v_pooled": 25.839960538021398,
     "norm_script": 32.340870567155235,
     "v_pooled_over_activation": 0.3755214406986731,
     "norm_ratio_ro_over_ur": {
      "value": 0.4738504788547556,
      "ci95": [
       0.4162349426832352,
       0.610076550990632
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.81665236650471,
      "ci95": [
       0.6099670856418691,
       0.8565709936352899
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.6976134493113917,
      "within_script_ceiling_ur": 0.9360133744329511,
      "disattenuated": 0.7453028646455131,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.002482114323593264,
      "removed_norm_fraction_if_residualized": 0.002482114323593264,
      "removed_variance_fraction_if_residualized": 6.1608915153868454e-06,
      "cross_fitted_cosine": 0.00033693494129413844,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": 0.01089676932594578,
       "sd": 2.576285510086132,
       "cohens_d": 0.004229643524867503,
       "p_two_sided": 0.9842031593681264,
       "frac_positive": 0.375
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017999545718096133,
       "p95": 0.042934716476348306
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 34.55380655074779,
      "mean_projection_ro": 14.391798601280664,
      "ratio_ro_over_ur": 0.4165039987748964,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 20.162007949467128,
       "sd": 11.820006952051841,
       "cohens_d": 1.705752630371101,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "28": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 99.86249544421568,
     "norm_v_ur": 55.90331428249503,
     "norm_v_ro": 26.285186423530234,
     "norm_v_pooled": 39.792075704128166,
     "norm_script": 51.651713222660895,
     "v_pooled_over_activation": 0.39846866961537597,
     "norm_ratio_ro_over_ur": {
      "value": 0.470190126665905,
      "ci95": [
       0.37793682657499617,
       0.6155414130880212
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.8566407488388438,
      "ci95": [
       0.6564654333686377,
       0.8887394624985915
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.7637168243192027,
      "within_script_ceiling_ur": 0.946965744787704,
      "disattenuated": 0.806488332363508,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.03660006082392917,
      "removed_norm_fraction_if_residualized": 0.03660006082392917,
      "removed_variance_fraction_if_residualized": 0.0013395644523153149,
      "cross_fitted_cosine": 0.04097072872826762,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": 2.116208330795913,
       "sd": 3.9231521824289217,
       "cohens_d": 0.539415304936174,
       "p_two_sided": 0.008198360327934412,
       "frac_positive": 0.7083333333333334
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017608079618740818,
       "p95": 0.04344850300739852
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 53.55227413105124,
      "mean_projection_ro": 23.030726498090516,
      "ratio_ro_over_ur": 0.43006066262901427,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 30.521547632960733,
       "sd": 20.301328989885995,
       "cohens_d": 1.5034260884184671,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "30": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 132.90094252472335,
     "norm_v_ur": 72.5722457407903,
     "norm_v_ro": 34.02888843463609,
     "norm_v_pooled": 51.24245859466735,
     "norm_script": 79.7159344051957,
     "v_pooled_over_activation": 0.3855688125397214,
     "norm_ratio_ro_over_ur": {
      "value": 0.46889672611453515,
      "ci95": [
       0.3799840037730797,
       0.6141443127160499
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.8257491118542801,
      "ci95": [
       0.6195036811397291,
       0.8599752795302732
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.7284673531861897,
      "within_script_ceiling_ur": 0.9397887043591723,
      "disattenuated": 0.7751395072181897,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.05889982655331466,
      "removed_norm_fraction_if_residualized": 0.05889982655331466,
      "removed_variance_fraction_if_residualized": 0.0034691895680105506,
      "cross_fitted_cosine": 0.0632541531083645,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": 5.0423639200425905,
       "sd": 6.028490552596733,
       "cohens_d": 0.8364222977625179,
       "p_two_sided": 0.0003999200159968006,
       "frac_positive": 0.8333333333333334
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.0173646148033254,
       "p95": 0.04274390261386051
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 69.06605561683295,
      "mean_projection_ro": 29.132989949821297,
      "ratio_ro_over_ur": 0.4218134319331381,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 39.933065667011654,
       "sd": 26.91826749056432,
       "cohens_d": 1.483493158726111,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 0.9583333333333334
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "32": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 165.35272022397976,
     "norm_v_ur": 88.22644092091772,
     "norm_v_ro": 40.45539879696862,
     "norm_v_pooled": 61.13693335478891,
     "norm_script": 116.37485737845358,
     "v_pooled_over_activation": 0.369736483754095,
     "norm_ratio_ro_over_ur": {
      "value": 0.4585405279266683,
      "ci95": [
       0.3833074354137181,
       0.6012735814982685
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.7747252549568372,
      "ci95": [
       0.563468890048074,
       0.806962273960735
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.6714376990059716,
      "within_script_ceiling_ur": 0.9281292600935908,
      "disattenuated": 0.7234312373022969,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.02266485679786988,
      "removed_norm_fraction_if_residualized": 0.02266485679786988,
      "removed_variance_fraction_if_residualized": 0.0005136957336679484,
      "cross_fitted_cosine": -0.017558113618825266,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -2.043322968225474,
       "sd": 7.058569708938712,
       "cohens_d": -0.2894811629667531,
       "p_two_sided": 0.17036592681463708,
       "frac_positive": 0.2916666666666667
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017810561756505257,
       "p95": 0.042843195272016535
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 83.121887247452,
      "mean_projection_ro": 33.17946419037095,
      "ratio_ro_over_ur": 0.39916639634993406,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 49.942423057081044,
       "sd": 31.0383325916711,
       "cohens_d": 1.6090562503503398,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    },
    "34": {
     "n_families": 24,
     "hidden": 2048,
     "mean_activation_norm": 203.38150140395365,
     "norm_v_ur": 106.39151000699421,
     "norm_v_ro": 50.00249361751236,
     "norm_v_pooled": 72.69443661384899,
     "norm_script": 162.99996408436888,
     "v_pooled_over_activation": 0.3574289505782744,
     "norm_ratio_ro_over_ur": {
      "value": 0.4699857499364863,
      "ci95": [
       0.40014851183369,
       0.6146766355407071
      ]
     },
     "cos_v_ur_v_ro": {
      "value": 0.6878515696154665,
      "ci95": [
       0.5001666998726363,
       0.7137570840287966
      ],
      "note": "in-sample cosines of noisy means are biased toward 0; see split-half below"
     },
     "cos_v_ur_v_ro_split_half": {
      "cross_script": 0.5852681705796278,
      "within_script_ceiling_ur": 0.9167346093951775,
      "disattenuated": 0.6384270481134805,
      "note": "cross_script / ceiling ~ 1 means the two scripts share one harm direction up to noise"
     },
     "h2_script_alignment": {
      "in_sample_abs_cosine": 0.04301775767403638,
      "removed_norm_fraction_if_residualized": 0.04301775767403638,
      "removed_variance_fraction_if_residualized": 0.0018505274753021158,
      "cross_fitted_cosine": -0.03778090140110795,
      "cross_fitted_projection_test": {
       "n": 24,
       "mean": -6.158285571455678,
       "sd": 9.171822836752993,
       "cohens_d": -0.6714352949316053,
       "p_two_sided": 0.004199160167966407,
       "frac_positive": 0.20833333333333334
      },
      "isotropic_random_null_abs_cosine": {
       "mean": 0.017670401861329416,
       "p95": 0.04264300593691994
      },
      "reading": "cross_fitted_projection_test asks whether families' benign script differences point consistently along (or against) the harm direction built from OTHER families"
     },
     "h1_representation": {
      "mean_projection_ur": 98.62298115739479,
      "mean_projection_ro": 38.31696047690898,
      "ratio_ro_over_ur": 0.3885195927687282,
      "paired_difference_test_ur_minus_ro": {
       "n": 24,
       "mean": 60.30602068048582,
       "sd": 35.58017153775471,
       "cohens_d": 1.6949333877295703,
       "p_two_sided": 0.0001999600079984003,
       "frac_positive": 1.0
      },
      "reading": "projection of each family's harmful-minus-benign contrast on the shared harm direction (leave-one-family-out); ur > ro means harm is under-represented in Roman Urdu"
     }
    }
   }
  },
  "baseline_done": {
   "n": 288,
   "timing": {
    "baseline:unsteered:LNone:rhoNone": 174.1
   }
  },
  "dev_grid_progress": {
   "done": 16,
   "last": {
    "arm": "pooled",
    "layer": 22,
    "rho": 0.8,
    "alpha": 2.026090224959964,
    "mean_margin": 2.3043737473587194
   },
   "timing": {
    "baseline:unsteered:LNone:rhoNone": 174.1,
    "dev_grid:pooled:L10:rho0.1": 29.3,
    "dev_grid:pooled:L10:rho0.2": 29.3,
    "dev_grid:pooled:L10:rho0.4": 29.1,
    "dev_grid:pooled:L10:rho0.8": 29.3,
    "dev_grid:pooled:L14:rho0.1": 29.3,
    "dev_grid:pooled:L14:rho0.2": 29.2,
    "dev_grid:pooled:L14:rho0.4": 29.2,
    "dev_grid:pooled:L14:rho0.8": 29.2,
    "dev_grid:pooled:L18:rho0.1": 29.1,
    "dev_grid:pooled:L18:rho0.2": 29.2,
    "dev_grid:pooled:L18:rho0.4": 29.2,
    "dev_grid:pooled:L18:rho0.8": 29.2,
    "dev_grid:pooled:L22:rho0.1": 29.2,
    "dev_grid:pooled:L22:rho0.2": 29.2,
    "dev_grid:pooled:L22:rho0.4": 29.2,
    "dev_grid:pooled:L22:rho0.8": 29.2
   }
  },
  "dev_grid": {
   "n_prompts_per_config": 48,
   "configs": [
    {
     "arm": "pooled",
     "layer": 10,
     "rho": 0.1,
     "alpha": 0.8186842336313425,
     "mean_margin": 1.6225563709934552
    },
    {
     "arm": "pooled",
     "layer": 10,
     "rho": 0.2,
     "alpha": 1.637368467262685,
     "mean_margin": 1.6849907611807187
    },
    {
     "arm": "pooled",
     "layer": 10,
     "rho": 0.4,
     "alpha": 3.27473693452537,
     "mean_margin": 1.715896716962258
    },
    {
     "arm": "pooled",
     "layer": 10,
     "rho": 0.8,
     "alpha": 6.54947386905074,
     "mean_margin": 1.7451517780621846
    },
    {
     "arm": "pooled",
     "layer": 14,
     "rho": 0.1,
     "alpha": 0.45611648549859796,
     "mean_margin": 1.6911278714736302
    },
    {
     "arm": "pooled",
     "layer": 14,
     "rho": 0.2,
     "alpha": 0.9122329709971959,
     "mean_margin": 1.850713220735391
    },
    {
     "arm": "pooled",
     "layer": 14,
     "rho": 0.4,
     "alpha": 1.8244659419943918,
     "mean_margin": 2.1185344407955804
    },
    {
     "arm": "pooled",
     "layer": 14,
     "rho": 0.8,
     "alpha": 3.6489318839887837,
     "mean_margin": 2.1362208339075246
    },
    {
     "arm": "pooled",
     "layer": 18,
     "rho": 0.1,
     "alpha": 0.2693938220189354,
     "mean_margin": 1.618454319735368
    },
    {
     "arm": "pooled",
     "layer": 18,
     "rho": 0.2,
     "alpha": 0.5387876440378708,
     "mean_margin": 1.7069085463881493
    },
    {
     "arm": "pooled",
     "layer": 18,
     "rho": 0.4,
     "alpha": 1.0775752880757417,
     "mean_margin": 1.8392553453644116
    },
    {
     "arm": "pooled",
     "layer": 18,
     "rho": 0.8,
     "alpha": 2.1551505761514833,
     "mean_margin": 2.3211714824040732
    },
    {
     "arm": "pooled",
     "layer": 22,
     "rho": 0.1,
     "alpha": 0.2532612781199955,
     "mean_margin": 1.6447107034424941
    },
    {
     "arm": "pooled",
     "layer": 22,
     "rho": 0.2,
     "alpha": 0.506522556239991,
     "mean_margin": 1.7394465530912082
    },
    {
     "arm": "pooled",
     "layer": 22,
     "rho": 0.4,
     "alpha": 1.013045112479982,
     "mean_margin": 1.9286137248078983
    },
    {
     "arm": "pooled",
     "layer": 22,
     "rho": 0.8,
     "alpha": 2.026090224959964,
     "mean_margin": 2.3043737473587194
    }
   ],
   "note": "selection is made AFTER judging, by the rule pre-declared in RESEARCH_PROTOCOL.md; this stage only generates"
  },
  "identical_text_between_calls": {},
  "timing": {
   "baseline:unsteered:LNone:rhoNone": 174.1,
   "dev_grid:pooled:L10:rho0.1": 29.3,
   "dev_grid:pooled:L10:rho0.2": 29.3,
   "dev_grid:pooled:L10:rho0.4": 29.1,
   "dev_grid:pooled:L10:rho0.8": 29.3,
   "dev_grid:pooled:L14:rho0.1": 29.3,
   "dev_grid:pooled:L14:rho0.2": 29.2,
   "dev_grid:pooled:L14:rho0.4": 29.2,
   "dev_grid:pooled:L14:rho0.8": 29.2,
   "dev_grid:pooled:L18:rho0.1": 29.1,
   "dev_grid:pooled:L18:rho0.2": 29.2,
   "dev_grid:pooled:L18:rho0.4": 29.2,
   "dev_grid:pooled:L18:rho0.8": 29.2,
   "dev_grid:pooled:L22:rho0.1": 29.2,
   "dev_grid:pooled:L22:rho0.2": 29.2,
   "dev_grid:pooled:L22:rho0.4": 29.2,
   "dev_grid:pooled:L22:rho0.8": 29.2
  }
 },
 "elapsed_seconds": 1995.5
}