OUTPUT of run 10 of at most 60:
```
=== ROTARY-FREQUENCY COEFFICIENT ANALYSIS ===
Frequency bands: ['period<=32', '32<period<=512', '512<period<=8192', 'period>8192']
Energies are sums of coefficient-matrix Frobenius norms, not scores or probabilities.
No positions or hidden-state vectors were evaluated.

=== LARGEST RELATIVE CHANGES ACROSS COSINE AND SINE COEFFICIENTS ===
layer head type | relative_change sine_change_share cosine_sum_ratio | top frequency indices; periods; energy shares
11 0 full_attention    | 0.029355 0.71941 1.60938 | [88, 60, 70]; [83787.62, 4080.19, 12006.87]; [0.25647, 0.05275, 0.03682]
11 1 full_attention    | 0.027430 0.66029 1.77208 | [88, 60, 70]; [83787.62, 4080.19, 12006.87]; [0.17502, 0.0396, 0.03706]
12 0 sliding_attention | 0.021802 0.66817 1.35719 | [111, 117, 72]; [18489.7, 28472.78, 1117.33]; [0.07331, 0.05243, 0.03818]
12 2 sliding_attention | 0.021751 0.67046 1.52224 | [111, 114, 72]; [18489.7, 22944.57, 1117.33]; [0.05701, 0.04315, 0.0375]
11 3 full_attention    | 0.021608 0.61165 1.56860 | [88, 60, 63]; [83787.62, 4080.19, 5640.33]; [0.14078, 0.03904, 0.02907]
17 1 full_attention    | 0.021080 0.67326 2.28605 | [87, 107, 102]; [75215.07, 651335.68, 379690.62]; [0.17928, 0.04376, 0.0426]
12 3 sliding_attention | 0.019655 0.65735 1.45245 | [111, 115, 118]; [18489.7, 24656.41, 30597.07]; [0.04196, 0.03746, 0.03505]
 5 3 full_attention    | 0.019417 0.60798 1.31771 | [112, 121, 122]; [1117325.91, 2951582.1, 3287985.31]; [0.05646, 0.05194, 0.04333]
17 0 full_attention    | 0.018100 0.71402 1.66524 | [87, 110, 100]; [75215.07, 900388.35, 305970.72]; [0.19946, 0.04899, 0.04359]
11 2 full_attention    | 0.017406 0.59394 1.31216 | [88, 108, 126]; [83787.62, 725570.92, 5063255.79]; [0.1495, 0.03618, 0.0292]
12 1 sliding_attention | 0.016833 0.67250 1.51748 | [115, 111, 117]; [24656.41, 18489.7, 28472.78]; [0.04724, 0.0416, 0.04158]
10 1 sliding_attention | 0.016265 0.81725 1.82820 | [75, 70, 113]; [1386.53, 967.56, 21351.57]; [0.09105, 0.04256, 0.03027]
 5 0 full_attention    | 0.016091 0.59353 1.33354 | [112, 114, 124]; [1117325.91, 1386531.91, 4080185.13]; [0.0577, 0.04423, 0.0429]
 5 2 full_attention    | 0.016079 0.59621 1.18965 | [112, 123, 127]; [1117325.91, 3662729.69, 5640334.6]; [0.06584, 0.04223, 0.04004]
17 2 full_attention    | 0.015415 0.68535 1.32725 | [87, 100, 110]; [75215.07, 305970.72, 900388.35]; [0.12889, 0.0602, 0.03848]
14 1 sliding_attention | 0.015037 0.76879 2.17156 | [122, 104, 107]; [40801.85, 11173.26, 13865.32]; [0.06274, 0.0424, 0.03423]
13 3 sliding_attention | 0.014898 0.64529 1.41928 | [71, 75, 113]; [1039.75, 1386.53, 21351.57]; [0.07351, 0.04731, 0.03646]
14 0 sliding_attention | 0.014605 0.77312 2.57787 | [79, 115, 104]; [1848.97, 24656.41, 11173.26]; [0.04065, 0.04063, 0.03528]
 5 1 full_attention    | 0.014405 0.60537 1.12691 | [112, 111, 121]; [1117325.91, 1003009.09, 2951582.1]; [0.08423, 0.04747, 0.04254]
14 3 sliding_attention | 0.014355 0.77821 2.63944 | [104, 107, 106]; [11173.26, 13865.32, 12902.68]; [0.04646, 0.04556, 0.04231]

=== PER-LAYER FREQUENCY-BAND CHANGES ===
Shares aggregate all heads; enrichment divides change share by baseline coefficient-energy share.
layer | change shares in four bands | enrichment in four bands
 0 | 0.0387 0.1212 0.3998 0.4403 | 1.6173 1.4802 0.9856 0.9012
 1 | 0.0869 0.2156 0.3976 0.2999 | 0.9185 0.9795 1.0453 0.9836
 2 | 0.0768 0.2823 0.4011 0.2398 | 1.1756 1.0857 0.9966 0.8811
 3 | 0.0536 0.1797 0.3956 0.3711 | 3.3574 2.4926 0.9607 0.7419
 4 | 0.0255 0.1308 0.6491 0.1945 | 1.4321 1.0807 0.9920 0.9407
 5 | 0.0019 0.0098 0.0386 0.9498 | 3.4793 1.9433 1.1659 0.9880
 6 | 0.0309 0.2206 0.4801 0.2684 | 2.2381 1.6380 1.0164 0.7078
 7 | 0.0241 0.1144 0.3206 0.5409 | 3.2891 2.1632 1.0937 0.8364
 8 | 0.0424 0.2700 0.5923 0.0953 | 1.6291 1.2785 0.9187 0.8070
 9 | 0.0313 0.1422 0.3657 0.4608 | 3.9222 2.3708 1.0769 0.7777
10 | 0.0296 0.1246 0.4094 0.4364 | 3.2132 2.2470 1.0880 0.7806
11 | 0.0449 0.0983 0.2627 0.5940 | 1.7219 1.4817 1.6227 0.7967
12 | 0.0183 0.0897 0.3690 0.5230 | 5.5765 2.0783 1.1082 0.8427
13 | 0.0918 0.3604 0.3901 0.1578 | 1.8572 1.7087 0.8339 0.5802
14 | 0.0171 0.1319 0.3666 0.4844 | 1.5447 1.4954 1.0358 0.8859
15 | 0.0504 0.1464 0.3974 0.4057 | 1.7737 1.8736 1.1290 0.7494
16 | 0.0876 0.2451 0.4186 0.2488 | 1.2344 1.2041 1.0422 0.7681
17 | 0.0129 0.0271 0.1032 0.8568 | 3.2363 2.5804 0.7567 1.0091

=== LARGEST ENRICHMENT OF RAPIDLY ROTATING COEFFICIENT CHANGES ===
layer=11 head=2 rapid_enrichment=47.38269 rapid_change_share=0.02582 rapid_base_share=0.00054
layer= 3 head=3 rapid_enrichment=26.09724 rapid_change_share=0.01945 rapid_base_share=0.00075
layer=10 head=2 rapid_enrichment=25.21026 rapid_change_share=0.01429 rapid_base_share=0.00057
layer=17 head=3 rapid_enrichment=21.55911 rapid_change_share=0.00944 rapid_base_share=0.00044
layer=17 head=2 rapid_enrichment=17.21124 rapid_change_share=0.00763 rapid_base_share=0.00044
layer=13 head=3 rapid_enrichment=10.66883 rapid_change_share=0.07341 rapid_base_share=0.00688
layer= 5 head=1 rapid_enrichment=9.93940 rapid_change_share=0.00102 rapid_base_share=0.00010
layer= 1 head=2 rapid_enrichment=9.19873 rapid_change_share=0.05964 rapid_base_share=0.00648
layer=17 head=0 rapid_enrichment=8.68322 rapid_change_share=0.00656 rapid_base_share=0.00076
layer=12 head=3 rapid_enrichment=8.04082 rapid_change_share=0.02044 rapid_base_share=0.00254
layer=12 head=1 rapid_enrichment=7.42610 rapid_change_share=0.01497 rapid_base_share=0.00202
layer= 9 head=3 rapid_enrichment=5.80383 rapid_change_share=0.02309 rapid_base_share=0.00398

=== LARGEST SIGNED SHIFTS IN SLOW-FREQUENCY BASELINE ENERGY SHARE ===
Positive means B has a larger share in periods >512; this does not mean more or less sycophancy.
layer=11 head=0 A=0.8162705 B=0.8160138 B_minus_A=-0.0002567
layer=11 head=1 A=0.7575919 B=0.7573616 B_minus_A=-0.0002303
layer=16 head=0 A=0.6804153 B=0.6806244 B_minus_A=+0.0002092
layer=13 head=1 A=0.5927340 B=0.5928897 B_minus_A=+0.0001557
layer= 8 head=0 A=0.7663068 B=0.7664356 B_minus_A=+0.0001288
layer=10 head=1 A=0.7867615 B=0.7868903 B_minus_A=+0.0001288
layer=17 head=1 A=0.9338621 B=0.9337351 B_minus_A=-0.0001270
layer= 8 head=1 A=0.7816954 B=0.7818140 B_minus_A=+0.0001187
layer= 6 head=2 A=0.6917203 B=0.6916351 B_minus_A=-0.0000852
layer=16 head=1 A=0.6066903 B=0.6067736 B_minus_A=+0.0000833
layer=16 head=2 A=0.6880202 B=0.6879369 B_minus_A=-0.0000833
layer=16 head=3 A=0.9106752 B=0.9105982 B_minus_A=-0.0000770

=== DOES ZERO-OFFSET SUMMATION HIDE FREQUENCY-SPECIFIC CHANGE? ===
Ratio = ||sum_j delta C_j||^2 / sum_j ||delta C_j||^2.
Below/above 1 indicates cancellation/reinforcement when cosine coefficients are summed.
layer= 1 head=1 ratio=1.084506
layer= 5 head=1 ratio=1.126906
layer= 1 head=3 ratio=1.134746
layer= 2 head=2 ratio=1.176980
layer= 5 head=2 ratio=1.189654
layer= 2 head=1 ratio=1.205085
layer= 0 head=1 ratio=1.213776
layer= 1 head=2 ratio=1.221996
Median ratio: 1.519863082659428
Maximum ratio: 2.6394378587396146

Interpretation limits:
- Uses the standard half-split Gemma rotary convention and frequencies specified by config.
- Frequency coefficient norms do not determine attention distance preferences on actual text.
- Relative-position effects also depend on content, query/key RMS denominators and softmax competition.
- This does not identify the location or representation of the user's guess.
- Only algebraic weight coefficients and architecture constants were analyzed.

```