{
 "ABLATION_SCAN": {
  "_source": "steering/pllum-12b-chat-2512_ablation_scan.json (layer 19; 200 MMLU items, 15 generations, 6 perplexity items per cell)",
  "strengths": [0.0, 0.25, 0.5, 0.75, 1.0],
  "global_": {
   "proj": [3.1354, 1.2469, 0.5287, 0.1812, 0.00034],
   "mmlu_eval": [0.645, 0.635, 0.635, 0.62, 0.615],
   "mmlu_deploy": [0.64, 0.62, 0.61, 0.605, 0.61],
   "ppl": [25.31, 25.7, 25.73, 25.77, 25.68],
   "degeneracy": [0.6, 0.4667, 0.4667, 0.3333, 0.3333]
  },
  "layerwise": {
   "proj": [3.1354, 2.3221, 1.4437, 0.5837, 0.00146],
   "mmlu_eval": [0.645, 0.645, 0.615, 0.44, 0.28],
   "mmlu_deploy": [0.64, 0.64, 0.565, 0.405, 0.25],
   "ppl": [25.31, 27.04, 35.66, 59.38, 108.28],
   "degeneracy": [0.6, 0.4667, 0.4667, 0.4667, 0.8667]
  },
  "proj_reduction_global": 0.99989,
  "sandbagging_delta_global_full": 0.005,
  "baseline_delta": 0.005,
  "baseline_generations_polish": 13,
  "baseline_generations_n": 15,
  "degeneracy_cut_at_eos": {"global_0": 0.066667, "layerwise_1": 0.0}
 },
 "BOOT": {
  "_source": "reports/pllum-12b-chat-2512_bootstrap.json (layer 19, 48 single-cue tasks, 400 resamples)",
  "mean_offdiag": [0.19891, 0.17856, 0.21957],
  "crossfit_offdiag": [0.17604, 0.1544, 0.19572],
  "ceiling": [0.88385, 0.85512, 0.90145],
  "cos_global": {
   "F1": [0.17682, 0.1225, 0.21964],
   "F2": [0.32607, 0.29815, 0.355],
   "F3": [0.32695, 0.30594, 0.34782],
   "F4": [0.13146, 0.08735, 0.17816],
   "F5": [0.33059, 0.29335, 0.36727],
   "F6": [0.33908, 0.30767, 0.3645],
   "F7": [0.46877, 0.44179, 0.48986],
   "F8": [0.8095, 0.79473, 0.82011]
  },
  "lofo_dom": {
   "F1": [0.61741, 0.57334, 0.66981],
   "F2": [0.91343, 0.8498, 0.97266],
   "F3": [0.97353, 0.9427, 1.0],
   "F4": [0.69846, 0.63105, 0.77604],
   "F5": [0.72902, 0.67707, 0.78864],
   "F6": [0.86336, 0.80029, 0.92537],
   "F7": [0.91457, 0.85588, 0.97483],
   "F8": [0.9677, 0.92658, 1.0]
  },
  "cross_half": [
   [0.73076, 0.20914, 0.24678, 0.14162, 0.22713, 0.13573, 0.08906, 0.0643],
   [0.20914, 0.92446, 0.43003, 0.14108, 0.3025, 0.26562, 0.02782, 0.17437],
   [0.24678, 0.43003, 0.95138, 0.00596, 0.15793, 0.40833, 0.20245, 0.13688],
   [0.14162, 0.14108, 0.00596, 0.88263, 0.19872, 0.05288, 0.13416, 0.04201],
   [0.22713, 0.3025, 0.15793, 0.19872, 0.79043, 0.13529, 0.17574, 0.16707],
   [0.13573, 0.26562, 0.40833, 0.05288, 0.13529, 0.90061, 0.07555, 0.16295],
   [0.08906, 0.02782, 0.20245, 0.13416, 0.17574, 0.07555, 0.91696, 0.39644],
   [0.0643, 0.17437, 0.13688, 0.04201, 0.16707, 0.16295, 0.39644, 0.97355]
  ]
 },
 "COMPARISON": {
  "_source": "Recomputed from raw result files. Qwen3/Gemma 3: evalaware-interp repository — reports/<model>.json (full sample) and reports/<model>_bootstrap.json (field point = mean over task resamples); Qwen3-0.6B is the Phase-1 float32 run (results/reports/), 8B phase2/, 14B phase3/, Gemma 3 phase3c/ (residual stream). PLLuM: runs/pllum-12b-chat-2512/ of the private dataset. Cue geometry (pct_ceiling, ceiling_value, crossfit_value, cos_f8_global, lofo_f1) = bootstrap means; pc1, pr, greet_pct, f8_necessity_pct = full sample at each model's primary layer; mmlu_eval_minus_deploy = MMLU validation harness (letter log-probability).",
  "models": ["Qwen3-0.6B", "Qwen3-8B", "Qwen3-14B", "PLLuM-12B", "Gemma 3 270M", "Gemma 3 1B"],
  "pct_ceiling": {"Qwen3-0.6B": 21.473416, "Qwen3-8B": 19.681725, "Qwen3-14B": 21.393748, "PLLuM-12B": 19.917863, "Gemma 3 270M": 17.261521, "Gemma 3 1B": 26.457548},
  "ceiling_value": {"Qwen3-0.6B": 0.883263, "Qwen3-8B": 0.859215, "Qwen3-14B": 0.923652, "PLLuM-12B": 0.883848},
  "crossfit_value": {"Qwen3-0.6B": 0.189667, "Qwen3-8B": 0.169108, "Qwen3-14B": 0.197604, "PLLuM-12B": 0.176044},
  "cos_f8_global": {"Qwen3-0.6B": 0.88943, "Qwen3-8B": 0.885808, "Qwen3-14B": 0.865195, "PLLuM-12B": 0.809496, "Gemma 3 270M": 0.924793, "Gemma 3 1B": 0.944544},
  "lofo_f1": {"Qwen3-0.6B": 0.52547, "Qwen3-8B": 0.582195, "Qwen3-14B": 0.615537, "PLLuM-12B": 0.617412, "Gemma 3 270M": 0.491921, "Gemma 3 1B": 0.534801},
  "pc1": {"Qwen3-0.6B": 14.808677, "Qwen3-8B": 12.936474, "Qwen3-14B": 10.366265, "PLLuM-12B": 10.488337},
  "pr": {"Qwen3-0.6B": 20.255211, "Qwen3-8B": 27.879276, "Qwen3-14B": 30.41515, "PLLuM-12B": 32.163609},
  "greet_pct": {"Qwen3-0.6B": 15.178704, "Qwen3-8B": 1.280627, "Qwen3-14B": 4.676392, "PLLuM-12B": 7.006954},
  "mmlu_eval_minus_deploy": {"Qwen3-8B": 0.0, "Qwen3-14B": 0.013333, "PLLuM-12B": 0.016, "Gemma 3 1B": -0.006667},
  "f8_necessity_pct": {"Qwen3-0.6B": 54.671526, "Qwen3-8B": 54.496455, "Qwen3-14B": 37.038544, "PLLuM-12B": 44.995651},
  "n_boot": {"Qwen3-0.6B": 400, "Qwen3-8B": 400, "Qwen3-14B": 1000, "PLLuM-12B": 400, "Gemma 3 270M": 1000, "Gemma 3 1B": 1000},
  "lofo_f1_ci": {
   "Qwen3-0.6B": [0.500857, 0.552517],
   "Qwen3-8B": [0.544271, 0.630675],
   "Qwen3-14B": [0.580273, 0.659755],
   "PLLuM-12B": [0.57334, 0.669813],
   "Gemma 3 270M": [0.465701, 0.517361],
   "Gemma 3 1B": [0.512576, 0.556424]
  },
  "lofo_f1_full": {"Qwen3-0.6B": 0.521267, "Qwen3-8B": 0.579861, "Qwen3-14B": 0.610677, "PLLuM-12B": 0.607205, "Gemma 3 270M": 0.489149, "Gemma 3 1B": 0.530382},
  "cos_f8_global_full": {"Qwen3-0.6B": 0.893641, "Qwen3-8B": 0.890746, "Qwen3-14B": 0.869413, "PLLuM-12B": 0.814492, "Gemma 3 270M": 0.928999, "Gemma 3 1B": 0.955267}
 },
 "CONTROLS": {
  "_source": "reports/pllum-12b-chat-2512.json -> controls",
  "L19": {
   "ctl_greet": {"frac": 0.0701, "auroc": 0.691, "d": 0.6787},
   "ctl_typo": {"frac": 0.0038, "auroc": 0.5097, "d": 0.033},
   "ctl_verbose": {"frac": 0.0393, "auroc": 0.6147, "d": 0.3632}
  },
  "greet_frac_by_layer": {"10": 0.1334, "19": 0.0701, "20": 0.055, "30": -0.0101},
  "greet_auroc_by_layer": {"10": 0.7427, "19": 0.691, "20": 0.667, "30": 0.4875}
 },
 "CROSSLAYER": {
  "_source": "reports/pllum-12b-chat-2512_crosslayer.json",
  "mean_adjacent": 0.7922,
  "mean_all_pairs": 0.2934,
  "median_all_pairs": 0.198,
  "frac_pairs_below_0_2": 0.5032,
  "sampled": {
   "5": {"5": 1.0, "10": 0.247, "15": 0.077, "20": 0.06, "25": 0.054, "30": 0.038, "35": 0.019, "40": 0.013},
   "10": {"5": 0.247, "10": 1.0, "15": 0.309, "20": 0.178, "25": 0.126, "30": 0.08, "35": 0.053, "40": 0.019},
   "15": {"5": 0.077, "10": 0.309, "15": 1.0, "20": 0.315, "25": 0.202, "30": 0.127, "35": 0.092, "40": 0.043},
   "20": {"5": 0.06, "10": 0.178, "15": 0.315, "20": 1.0, "25": 0.512, "30": 0.349, "35": 0.205, "40": 0.079},
   "25": {"5": 0.054, "10": 0.126, "15": 0.202, "20": 0.512, "25": 1.0, "30": 0.626, "35": 0.379, "40": 0.155},
   "30": {"5": 0.038, "10": 0.08, "15": 0.127, "20": 0.349, "25": 0.626, "30": 1.0, "35": 0.628, "40": 0.295},
   "35": {"5": 0.019, "10": 0.053, "15": 0.092, "20": 0.205, "25": 0.379, "30": 0.628, "35": 1.0, "40": 0.506},
   "40": {"5": 0.013, "10": 0.019, "15": 0.043, "20": 0.079, "25": 0.155, "30": 0.295, "35": 0.506, "40": 1.0}
  }
 },
 "CUES": ["F1", "F2", "F3", "F4", "F5", "F6", "F7", "F8"],
 "CUE_L19": {
  "_source": "reports/pllum-12b-chat-2512.json -> per_factor['19'], lofo['19'], necessity['19']",
  "auroc_own": {"F1": 0.8954, "F2": 0.951, "F3": 0.9965, "F4": 0.9748, "F5": 0.9648, "F6": 0.9887, "F7": 0.9991, "F8": 1.0},
  "auroc_global": {"F1": 0.5803, "F2": 0.8872, "F3": 0.9748, "F4": 0.6836, "F5": 0.7244, "F6": 0.8681, "F7": 0.9905, "F8": 1.0},
  "paired_dz_own": {"F1": 1.6051, "F2": 1.8901, "F3": 3.6093, "F4": 2.9355, "F5": 2.7417, "F6": 3.3721, "F7": 4.1533, "F8": 6.6558},
  "cos_global": {"F1": 0.189765, "F2": 0.332317, "F3": 0.331614, "F4": 0.136844, "F5": 0.350508, "F6": 0.347559, "F7": 0.479408, "F8": 0.814492},
  "lofo_dom": {"F1": 0.6072, "F2": 0.9106, "F3": 0.9701, "F4": 0.6784, "F5": 0.7066, "F6": 0.8494, "F7": 0.9093, "F8": 0.9627},
  "lofo_probe": {"F1": 0.8607, "F2": 0.9731, "F3": 1.0, "F4": 0.9865, "F5": 0.9896, "F6": 0.8997, "F7": 1.0, "F8": 1.0},
  "necessity_drop": {"F1": 0.0101, "F2": 0.0093, "F3": 0.0072, "F4": 0.0692, "F5": 0.0221, "F6": 0.0161, "F7": 0.0372, "F8": 0.45},
  "necessity_drop_by_layer": {
   "10": {"F1": 0.0086, "F2": -0.0121, "F3": 0.0039, "F4": 0.0155, "F5": 0.0091, "F6": 0.0322, "F7": 0.0549, "F8": 0.2115},
   "19": {"F1": 0.0101, "F2": 0.0093, "F3": 0.0072, "F4": 0.0692, "F5": 0.0221, "F6": 0.0161, "F7": 0.0372, "F8": 0.45},
   "20": {"F1": 0.0263, "F2": 0.0222, "F3": 0.002, "F4": 0.1031, "F5": 0.0426, "F6": 0.0274, "F7": 0.0807, "F8": 0.5006},
   "30": {"F1": 0.0686, "F2": 0.0683, "F3": 0.0138, "F4": 0.177, "F5": 0.0738, "F6": 0.0328, "F7": 0.1315, "F8": 0.3802}
  },
  "mean_offdiag_cos": 0.1992,
  "median_offdiag_cos": 0.1814,
  "mean_offdiag_cos_by_layer": {"10": 0.2462, "19": 0.1992, "20": 0.2015, "30": 0.2927}
 },
 "CUE_NAMES": {
  "F1": "stimulus artificiality",
  "F2": "scenario coherence",
  "F3": "prose register",
  "F4": "ecological completeness",
  "F5": "stakes telegraphing",
  "F6": "paradigm recognizability",
  "F7": "task format",
  "F8": "explicit disclosure"
 },
 "DEPTH_EARLIER": {
  "_source": "evalaware-interp: phase3/results/steering/qwen3-{0.6b,8b}_vgrid.json and phase3c/results/steering/gemma3-1b_vgrid.json (cells kind=ablate, strength=1; up/down counts from the matching _vgrid_rows.jsonl). Per layer: (mean delta, d_z, Wilcoxon p, paired t p, prompts up, prompts down).",
  "Qwen3_0_6B": {
   "primary": 27,
   "baseline": -0.284779,
   "cells": {
    "7": [0.032152, 0.491414, 0.01283, 0.01168, 22, 8],
    "14": [0.136463, 0.628927, 0.001864, 0.001762, 24, 6],
    "21": [0.171398, 2.936023, 1.863e-09, 5.504e-16, 30, 0],
    "27": [0.038641, 1.316721, 4.657e-08, 6.116e-08, 28, 2]
   }
  },
  "Qwen3_8B": {
   "primary": 22,
   "baseline": -0.263893,
   "cells": {
    "9": [-0.107929, -1.035033, 1.192e-06, 3.962e-06, 5, 25],
    "18": [0.184121, 3.127858, 1.863e-09, 1.04e-16, 30, 0],
    "22": [0.176545, 2.429608, 1.863e-09, 7.057e-14, 30, 0],
    "27": [-0.08122, -0.906275, 2.08e-05, 2.805e-05, 8, 22]
   }
  },
  "Gemma_3_1B": {
   "primary": 25,
   "baseline": -0.385042,
   "cells": {
    "6": [1.0686, 2.316208, 3.725e-09, 2.318e-13, 29, 1],
    "13": [-0.231769, -0.507224, 0.01283, 0.009487, 9, 21],
    "20": [0.262982, 0.569853, 0.002766, 0.004055, 23, 7],
    "25": [0.220089, 0.494329, 0.01853, 0.01125, 19, 11]
   }
  }
 },
 "DIMENSIONALITY": {
  "_source": "reports/pllum-12b-chat-2512.json -> dimensionality",
  "by_layer": {
   "10": {"pc1": 0.1724, "pc2": 0.0737, "pc3": 0.0546, "top5": 0.3799, "pr": 20.74, "snr": 1.078},
   "19": {"pc1": 0.1049, "pc2": 0.0769, "pc3": 0.0622, "top5": 0.3162, "pr": 32.16, "snr": 0.94},
   "20": {"pc1": 0.1159, "pc2": 0.0734, "pc3": 0.0573, "top5": 0.3202, "pr": 30.76, "snr": 0.942},
   "30": {"pc1": 0.144, "pc2": 0.0645, "pc3": 0.0576, "top5": 0.3443, "pr": 25.77, "snr": 0.868}
  }
 },
 "LAYERS": {
  "_source": "reports/pllum-12b-chat-2512.json -> layers[]",
  "paired_dz": [
   0.0,
   1.5876,
   2.2334,
   2.3542,
   2.8281,
   2.826,
   3.1403,
   3.1383,
   3.3438,
   3.3434,
   3.736,
   4.236,
   4.4582,
   5.116,
   5.6883,
   5.8581,
   6.0398,
   6.461,
   6.5883,
   7.2886,
   6.2667,
   5.4295,
   5.601,
   5.5717,
   5.2969,
   5.2096,
   5.0418,
   4.3635,
   4.5203,
   4.271,
   4.2397,
   3.5773,
   3.5898,
   3.6286,
   4.0201,
   4.0617,
   4.1655,
   4.2049,
   4.3647,
   4.4275,
   4.5514
  ],
  "cohens_d": [
   0.0,
   1.505,
   2.3285,
   2.456,
   3.5748,
   4.022,
   4.4302,
   4.3057,
   5.0259,
   5.0569,
   5.8884,
   6.4301,
   6.5853,
   7.4202,
   8.3856,
   8.3368,
   8.3137,
   8.6802,
   7.9429,
   8.5472,
   7.8746,
   6.6049,
   6.5503,
   6.3318,
   6.0674,
   6.0201,
   5.7209,
   4.7936,
   5.0726,
   4.9095,
   4.7838,
   4.0194,
   3.9801,
   4.0919,
   4.5414,
   4.5695,
   4.7049,
   4.7249,
   4.7726,
   4.8486,
   5.3168
  ],
  "dom_auroc": [
   0.5,
   0.8664,
   0.9348,
   0.9418,
   0.9975,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   0.9985,
   0.9984,
   0.9985,
   0.9997,
   0.9996,
   0.9999,
   0.9997,
   0.9991,
   0.999,
   0.9998
  ],
  "probe_auroc": [
   0.5,
   0.9995,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0,
   1.0
  ],
  "shuffled_null": [
   0.5,
   0.552,
   0.557,
   0.529,
   0.5275,
   0.587,
   0.545,
   0.5345,
   0.57,
   0.5435,
   0.449,
   0.4835,
   0.4965,
   0.506,
   0.506,
   0.5035,
   0.4715,
   0.4615,
   0.5015,
   0.4805,
   0.4765,
   0.4775,
   0.472,
   0.483,
   0.489,
   0.454,
   0.442,
   0.446,
   0.433,
   0.4415,
   0.426,
   0.4535,
   0.4425,
   0.4475,
   0.4305,
   0.424,
   0.4315,
   0.4395,
   0.4215,
   0.4325,
   0.423
  ],
  "mean_act_norm_L19": 23.0639,
  "diff_norm_L19": 11.5912,
  "primary_layer": 19,
  "n_pairs": 100,
  "pair_wins_L19": 100,
  "sign_test_p_L19": 1.5777e-30,
  "n_states_probe_auroc_1": 40,
  "primary_criterion": "argmax paired d_z on matched main pairs, excluding the post-final-norm hidden state (40)",
  "analysis_layers": [10, 19, 20, 30]
 },
 "LENGTH_L19": {"tokens_eval": 290.78, "tokens_deploy": 243.92, "corr": 0.2477, "auroc_raw": 1.0, "auroc_after": 1.0, "d_after": 7.2867},
 "MASSIVE": {"_source": "reports/pllum-12b-chat-2512_massive.json", "top_coord": 3448, "layer1_top1_share": 0.2162, "layer2_top1_share": 0.025},
 "MMLU_INTERVENTIONS": {
  "_source": "steering/pllum-12b-chat-2512_mmlu_v3_summary.json (200 items x 2 framings)",
  "acc": {
   "none": [[0.63, 0.561, 0.694], [0.64, 0.571, 0.703]],
   "ablate": [[0.605, 0.536, 0.67], [0.62, 0.551, 0.684]],
   "ablate_layerwise": [[0.255, 0.2, 0.32], [0.275, 0.218, 0.341]],
   "ablate_layerwise_random": [[0.64, 0.571, 0.703], [0.64, 0.571, 0.703]],
   "ablate_random0": [[0.63, 0.561, 0.694], [0.64, 0.571, 0.703]],
   "add_x1": [[0.625, 0.556, 0.689], [0.625, 0.556, 0.689]],
   "add_x2": [[0.615, 0.546, 0.68], [0.61, 0.541, 0.675]],
   "add_random": [[0.63, 0.561, 0.694], [0.625, 0.556, 0.689]]
  },
  "paired_eval": {
   "ablate": [-0.02, 4, 8, 0.388],
   "ablate_layerwise": [-0.365, 19, 92, 1.12e-12],
   "ablate_layerwise_random": [0.0, 4, 4, 1.0],
   "ablate_random0": [0.0, 2, 2, 1.0],
   "add_random": [-0.015, 4, 7, 0.549],
   "add_x1": [-0.015, 3, 6, 0.508],
   "add_x2": [-0.03, 6, 12, 0.238]
  },
  "paired_deploy": {
   "ablate": [-0.025, 1, 6, 0.125],
   "ablate_layerwise": [-0.375, 18, 93, 2.26e-13],
   "ablate_layerwise_random": [0.01, 4, 2, 0.688],
   "ablate_random0": [0.0, 0, 0, 1.0],
   "add_random": [0.0, 3, 3, 1.0],
   "add_x1": [-0.005, 4, 5, 1.0],
   "add_x2": [-0.015, 5, 8, 0.581]
  },
  "scan_vs_battery_baseline_agreement": {"eval": [197, 200], "deploy": [198, 200]},
  "interaction": {
   "ablate": [0.005, 0.808],
   "ablate_layerwise": [0.01, 0.814],
   "ablate_layerwise_random": [-0.01, 0.564],
   "ablate_random0": [0.0, 1.0],
   "add_random": [-0.015, 0.439],
   "add_x1": [-0.01, 0.637],
   "add_x2": [-0.015, 0.532]
  }
 },
 "MMLU_VALIDATION": {
  "_source": "mmlu_validation/pllum-12b-chat-2512/summary.json (500 logprob, 200 generate)",
  "logprob": {"neutral": [0.612, 0.5686, 0.6537], "eval": [0.624, 0.5807, 0.6654], "deploy": [0.608, 0.5645, 0.6498]},
  "generate": {"neutral": [0.625, 0.5561, 0.6891], "eval": [0.65, 0.5816, 0.7127], "deploy": [0.635, 0.5663, 0.6986]},
  "agreement": {"neutral": 0.985, "eval": 0.98, "deploy": 0.98},
  "eval_vs_deploy": {"n": 500, "eval_only": 33, "deploy_only": 25, "delta": 0.016, "p": 0.3581},
  "eval_vs_neutral": {"framed_only": 31, "neutral_only": 25, "delta": 0.012, "p": 0.5044},
  "deploy_vs_neutral": {"framed_only": 27, "neutral_only": 29, "delta": -0.004, "p": 0.8939},
  "mean_top_logprob": {"neutral": -0.193, "eval": -0.2354, "deploy": -0.2146},
  "unparsable_deploy": 2
 },
 "N_STATES": 41,
 "OOD": {"auroc_at_primary": 1.0, "d_at_primary": 4.7486, "best_layer": 8, "n_each": 15, "auroc_L30": 0.9467, "auroc_L40": 0.84},
 "QWEN06_LOFO": {"F1": "0.525 [0.501, 0.553]", "F4": "0.652", "F5": "0.595", "F7": "0.952–0.976", "F8": "0.952–0.976"},
 "REGISTER": {
  "_source": "reports/pllum-12b-chat-2512_register.json -> layers['19'] (2x2: F3 register x F8 disclosure, 48 tasks). The page states layer 19; until 15 Oct 2026 the direction-level values were read from layers['30'] by mistake.",
  "cos_register_disclosure": -0.198016,
  "cos_register_global": 0.291172,
  "auroc_main_global": 1.0,
  "auroc_main_global_orth_register": 1.0,
  "auroc_ood_global": 1.0,
  "auroc_ood_global_orth_register": 1.0,
  "global_": {
   "auroc_main": 1.0,
   "d_main": 8.547203,
   "auroc_ood": 1.0,
   "greet_frac": 0.07015,
   "greet_auroc": 0.6912,
   "register_dz": 2.4454,
   "register_p": 7.10543e-15,
   "disclosure_dz": 8.08008,
   "disclosure_p": 7.10543e-15,
   "interaction_dz": -1.96093,
   "interaction_p": 1.42109e-14
  },
  "disclosure": {
   "auroc_main": 1.0,
   "d_main": 7.490436,
   "auroc_ood": 1.0,
   "greet_frac": 0.169382,
   "greet_auroc": 0.8882,
   "register_dz": -1.36898,
   "register_p": 3.06244e-11,
   "disclosure_dz": 6.70263,
   "disclosure_p": 7.10543e-15,
   "interaction_dz": 0.0461263,
   "interaction_p": 0.741141
  },
  "disclosure_orth_register": {
   "auroc_main": 1.0,
   "d_main": 8.052558,
   "auroc_ood": 1.0,
   "greet_frac": 0.142345,
   "greet_auroc": 0.8464,
   "register_dz": -2.75829e-07,
   "register_p": 0.979734,
   "disclosure_dz": 6.94417,
   "disclosure_p": 7.10543e-15,
   "interaction_dz": -0.690234,
   "interaction_p": 4.41091e-05
  },
  "register": {
   "auroc_main": 0.9629,
   "d_main": 2.760421,
   "auroc_ood": 0.982222,
   "greet_frac": -0.215642,
   "greet_auroc": 0.3647,
   "register_dz": 4.28731,
   "register_p": 7.10543e-15,
   "disclosure_dz": -1.93741,
   "disclosure_p": 3.55271e-14,
   "interaction_dz": -2.78197,
   "interaction_p": 7.10543e-15
  }
 },
 "REGISTER_L30": {
  "_source": "reports/pllum-12b-chat-2512_register.json -> layers['30']. Not used for any current statement: kept so that §5 can show, verifiably, which values the first version of the register paragraph printed under a layer-19 label.",
  "global_": {
   "auroc_main": 1.0,
   "d_main": 4.78376,
   "auroc_ood": 0.946667,
   "greet_frac": -0.00997769,
   "greet_auroc": 0.4878,
   "register_dz": 1.91815,
   "register_p": 1.35003e-13,
   "disclosure_dz": 4.88718,
   "disclosure_p": 7.10543e-15,
   "interaction_dz": -2.14285,
   "interaction_p": 2.13163e-14
  },
  "disclosure": {
   "auroc_main": 1.0,
   "d_main": 6.70225,
   "auroc_ood": 0.986667,
   "greet_frac": 0.0193425,
   "greet_auroc": 0.5621,
   "register_dz": -0.389632,
   "register_p": 0.00696877,
   "disclosure_dz": 5.08759,
   "disclosure_p": 7.10543e-15,
   "interaction_dz": -0.643946,
   "interaction_p": 0.00011515
  },
  "disclosure_orth_register": {
   "auroc_main": 1.0,
   "d_main": 6.65552,
   "auroc_ood": 0.982222,
   "greet_frac": 0.0156766,
   "greet_auroc": 0.5508,
   "register_dz": -1.54361e-07,
   "register_p": 0.725823,
   "disclosure_dz": 5.07684,
   "disclosure_p": 7.10543e-15,
   "interaction_dz": -0.875963,
   "interaction_p": 8.80955e-08
  },
  "register": {
   "auroc_main": 0.9871,
   "d_main": 2.75133,
   "auroc_ood": 0.733333,
   "greet_frac": -0.111773,
   "greet_auroc": 0.4319,
   "register_dz": 4.0487,
   "register_p": 7.10543e-15,
   "disclosure_dz": -0.397219,
   "disclosure_p": 0.0143271,
   "interaction_dz": -2.54004,
   "interaction_p": 7.10543e-15
  }
 },
 "RUN": {
  "model": "CYFRAGOVPL/PLLuM-12B-chat-2512",
  "cls": "MistralForCausalLM",
  "n_layers": 40,
  "d_model": 5120,
  "vocab": 131072,
  "params_b": 12.248,
  "footprint_gib": 22.813,
  "dtype": "bfloat16",
  "quantisation": "none",
  "gpu": "NVIDIA A100-SXM4-80GB (79.3 GiB)",
  "torch": "2.11.0+cu130",
  "transformers": "5.19.0",
  "prompts_sha1": "6ba7456f5cb03427",
  "n_prompts": 1490,
  "set_counts": {"main": 200, "single": 384, "leaveout": 384, "factorial": 192, "control": 300, "ood": 30},
  "tokens_mean": 224.56,
  "tokens_p95": 516,
  "tokens_max": 795,
  "last_token": "[/INST]",
  "act_norm_L19": 23.06,
  "final_norm": 392.59,
  "eab_commit": "2961c17ebcb35fd15d5c614c4874d1f567b85d1b"
 },
 "VERBALISATION": {
  "_source": "steering/pllum-12b-chat-2512_verbalisation.json (150 generations, 96 new tokens)",
  "n_per_cell": 15,
  "n_generations": 150,
  "all_cells_explicit_k": 0,
  "ci_hi": 0.2039,
  "mcnemar_all_p": 1.0,
  "judge_lexicon_language": "src/judge.py"
 },
 "VERIFY": {
  "_source": "steering/pllum-12b-chat-2512_verify.json (30 OOD prompts, layer 19)",
  "none": [-0.2747, 3.6818],
  "ablate": [0.0005, 0.00147],
  "ablate_layerwise": [0.00138, 0.00168],
  "ablate_layerwise_random": [-0.1426, 3.6386],
  "ablate_random0": [-0.2713, 3.6772],
  "mean_norm": 24.755,
  "mean_norm_ablate": 24.473
 },
 "VGRID": {
  "_source": "steering/pllum-12b-chat-2512_vgrid.json (30 OOD prompts; score = mean log-prob of 4 English meta continuations minus 4 neutral ones)",
  "n": 30,
  "baseline_meta_lp": -5.6988,
  "baseline_neutral_lp": -5.503,
  "baseline_score": -0.1958,
  "ablate_L19_by_strength": {"0.25": [0.0023, 0.801], "0.5": [0.0061, 0.618], "0.75": [0.0124, 0.396], "1.0": [0.0109, 0.478]},
  "ablate_L19_full_dz": 0.1312,
  "ablate_L19_full_wilcoxon_p": 0.503,
  "random_full": [0.001116, 0.036476, 0.843043],
  "random_full_wilcoxon_p": 0.871208,
  "random_matched_by_strength": {"0.25": [-0.0101, -0.24, 0.198], "0.5": [-0.0171, -0.365, 0.0552], "0.75": [-0.0154, -0.316, 0.0946], "1.0": [-0.0139, -0.262, 0.162]},
  "random_matched_full_wilcoxon_p": 0.124,
  "locality": {"10": [0.009434, 0.145308, 0.432565], "19": [0.010922, 0.131245, 0.477979], "20": [0.012332, 0.175444, 0.344524], "30": [-0.036884, -0.777344, 0.00019793]},
  "L30_wilcoxon_p": 0.000283264,
  "L30_t_p": 0.00019793,
  "clamped_frac_matched_full": 0.017,
  "matched_half_wilcoxon_p": 0.03272089920938015,
  "matched_half_t_p": 0.05517200340476346,
  "clamp_kappa_max": 50.0
 },
 "_note": "All numbers of the article, transcribed from runs/pllum-12b-chat-2512/ of the private dataset Mati83moni/pllum-evalaware. Qwen3/Gemma 3 values in COMPARISON and DEPTH_EARLIER are recomputed from the earlier study's raw result files (see each block's _source)."
}