OUTPUT of run 3 of at most 60:
```
=== DIAGONAL-SCALE ANISOTROPY ===
Energy shares below concern parameter vectors, not activations.
L00 input_layernorm              top1=0.02906 top8=0.12802 largest_coordinates=[(299, 62.75), (418, 51.0), (76, 48.25)]
L00 post_attention_layernorm     top1=0.75709 top8=0.86450 largest_coordinates=[(400, 52.75), (24, 12.625), (639, 8.09375)]
L03 input_layernorm              top1=0.00686 top8=0.04535 largest_coordinates=[(585, 63.0), (115, 57.5), (514, 56.75)]
L03 post_attention_layernorm     top1=0.49057 top8=0.62920 largest_coordinates=[(400, 222.0), (163, 74.0), (120, 40.5)]
L04 input_layernorm              top1=0.01641 top8=0.10576 largest_coordinates=[(152, 163.0), (420, 149.0), (436, 147.0)]
L04 post_attention_layernorm     top1=0.44364 top8=0.90189 largest_coordinates=[(400, 222.0), (163, 210.0), (558, 54.0)]
L11 input_layernorm              top1=0.04677 top8=0.29843 largest_coordinates=[(54, 53.25), (4, 49.75), (293, 49.25)]
L11 post_attention_layernorm     top1=0.01475 top8=0.10477 largest_coordinates=[(517, 569.0), (360, 553.0), (460, 537.0)]

=== REAL DIFFERENCES VERSUS ALIGNMENT-DISRUPTED CONTROLS ===
Control summaries give [minimum, median, maximum] over 12 scrambles.
PR = spectral participation; smaller means stronger concentration.
Control comparisons are descriptive, not behavioral tests or p-values.

L04H1
  Actual raw OV:    top1=0.11811 top8=0.44578 PR=26.936
  Actual scaled OV: top1=0.64531 top8=0.95054 PR=2.125
  Scale-change residual / full difference norm: 0.0001292
  Control raw PR:    [31.40028914424846, 32.23855096449642, 33.41973838489204]
  Control scaled PR: [2.0007317306524297, 2.209492325672588, 2.332306092106607]
  Control scaled norm / actual: [0.8163644806553322, 0.8678860441817627, 0.8892529853756277]
  Control energy on observed leading axis / actual: [7.550987134196859e-08, 0.002507101650432073, 0.009157045336634148]
  Leading output-axis coordinates (index, squared loading): [(400, 0.5690451014602473), (163, 0.42180551481214934), (477, 0.002177254548642163)]
  Leading input-axis coordinates (index, squared loading): [(45, 0.08069606123817241), (156, 0.06486337879927362), (338, 0.04288983146751946)]

L03H3
  Actual raw OV:    top1=0.13440 top8=0.53480 PR=19.864
  Actual scaled OV: top1=0.64568 top8=0.84719 PR=2.351
  Scale-change residual / full difference norm: 0.000177579
  Control raw PR:    [20.997515457423717, 21.378699722161734, 22.072154705981685]
  Control scaled PR: [2.2818139081698354, 2.4368051835941014, 2.6554937448358316]
  Control scaled norm / actual: [0.8721084636924287, 0.9081716486822737, 0.9458660423971912]
  Control energy on observed leading axis / actual: [1.1445190799675947e-07, 0.004907290025939455, 0.022317580394001113]
  Leading output-axis coordinates (index, squared loading): [(400, 0.9745544682805871), (267, 0.002749711634643326), (608, 0.002217160900187809)]
  Leading input-axis coordinates (index, squared loading): [(424, 0.0313884832373823), (156, 0.026693773271802), (91, 0.024095312203943564)]

L00H1
  Actual raw OV:    top1=0.11368 top8=0.41926 PR=30.530
  Actual scaled OV: top1=0.81876 top8=0.92912 PR=1.486
  Scale-change residual / full difference norm: 0.000714655
  Control raw PR:    [30.800712195850934, 32.636061488658314, 35.472051148453474]
  Control scaled PR: [1.406860198998622, 1.4662868367445014, 1.5503510631405084]
  Control scaled norm / actual: [0.9095983251190511, 0.9512727575374198, 1.007153128344585]
  Control energy on observed leading axis / actual: [0.003160436838063406, 0.005715963837967568, 0.02478324356728079]
  Leading output-axis coordinates (index, squared loading): [(400, 0.985933175056207), (381, 0.0017288631713841124), (267, 0.0013265867299134954)]
  Leading input-axis coordinates (index, squared loading): [(418, 0.034825271373035926), (0, 0.022209230266740043), (507, 0.02214148450907083)]

L04H2
  Actual raw OV:    top1=0.15594 top8=0.54461 PR=19.011
  Actual scaled OV: top1=0.72667 top8=0.96329 PR=1.795
  Scale-change residual / full difference norm: 0.000177646
  Control raw PR:    [23.8877657338594, 24.23791423472676, 25.161138809022358]
  Control scaled PR: [1.8035042576682434, 2.052450421196518, 2.1925756244775005]
  Control scaled norm / actual: [0.7264550269137751, 0.784523281226999, 0.8478494272015409]
  Control energy on observed leading axis / actual: [0.00020915442657921, 0.0086128973928174, 0.035806778005535604]
  Leading output-axis coordinates (index, squared loading): [(400, 0.6366655598035074), (163, 0.3467922232506595), (477, 0.004159614282457591)]
  Leading input-axis coordinates (index, squared loading): [(497, 0.0496841457070909), (496, 0.04763429464129687), (45, 0.046113559074802746)]

L04H0
  Actual raw OV:    top1=0.15407 top8=0.48409 PR=21.417
  Actual scaled OV: top1=0.62179 top8=0.95405 PR=2.165
  Scale-change residual / full difference norm: 0.000195639
  Control raw PR:    [23.09494203245452, 24.55700513779017, 26.286226149502813]
  Control scaled PR: [2.1124542357131886, 2.389097216479799, 2.6673636982310147]
  Control scaled norm / actual: [0.8208750042940458, 0.8560169197465421, 0.9305258165630459]
  Control energy on observed leading axis / actual: [3.1145167512844713e-07, 0.0024425655968831382, 0.014052880761554991]
  Leading output-axis coordinates (index, squared loading): [(163, 0.5702942235946331), (400, 0.4117780862353051), (477, 0.0055854239028887975)]
  Leading input-axis coordinates (index, squared loading): [(376, 0.05880275562232384), (360, 0.05397091183747538), (582, 0.04656320983779083)]

L11H1
  Actual raw OV:    top1=0.09723 top8=0.39531 PR=34.506
  Actual scaled OV: top1=0.09612 top8=0.55462 PR=20.184
  Scale-change residual / full difference norm: 0.000169579
  Control raw PR:    [40.44977281058236, 53.031312663900366, 54.546073294989725]
  Control scaled PR: [17.674551135207754, 18.975413996223338, 20.02916304046258]
  Control scaled norm / actual: [0.9716019511195281, 0.9930008968673084, 1.0242843782089976]
  Control energy on observed leading axis / actual: [7.887640891307024e-05, 0.0036572662461458036, 0.029432936335864173]
  Leading output-axis coordinates (index, squared loading): [(163, 0.07237685481268219), (547, 0.06576598975630432), (149, 0.05398817367347797)]
  Leading input-axis coordinates (index, squared loading): [(417, 0.2530731848086276), (293, 0.1738153832985887), (26, 0.13251977460118614)]

=== COMPACT COMPARISON ===
 L H real_raw_PR real_scaled_PR control_raw_PR_median control_scaled_PR_median
 4 1      26.936          2.125                32.239                    2.209
 3 3      19.864          2.351                21.379                    2.437
 0 1      30.530          1.486                32.636                    1.466
 4 2      19.011          1.795                24.238                    2.052
 4 0      21.417          2.165                24.557                    2.389
11 1      34.506         20.184                53.031                   18.975

Interpretation limits:
- Low rank reproduced by controls can arise from surrounding weights/scales.
- Failure of these particular controls to reproduce a feature does not prove
  that the feature implements sycophancy, truthfulness, or a causal bottleneck.
- All four heads in a layer share the same V projection in this architecture;
  their observations must not be counted as independent evidence.
- Only weight matrices, their differences, and their products were analyzed.

```