=== ROTARY ARCHITECTURE FIELDS ===
{"head_dim": 256, "layer_types": ["sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention"], "rope_parameters": {"full_attention": {"rope_theta": 1000000.0, "rope_type": "default"}, "sliding_attention": {"rope_theta": 10000.0, "rope_type": "default"}}}

=== NEW HYPOTHESIS: POSITIONAL MATCHING CHANGES HIDDEN AT ZERO DISPLACEMENT ===
Using Gemma's half-split rotary pairing: channel i with channel i+head_dim/2.
For each rotary plane, define purely weight-derived coefficient matrices:
C_i = q_i outer k_i + q_partner outer k_partner
S_i = q_i outer k_partner - q_partner outer k_i
The sign convention for S depends on displacement orientation; its difference norms do not.
No positions, sin/cos evaluations, input vectors, or attention scores are used.
The coefficient bank is not an evaluated model or an input-distribution average.

=== PER-LAYER ROTARY-COEFFICIENT RELATIVE DIFFERENCES ===
Heads 0,1,2,3; raw and learned-diagonal-scaled coefficient banks.
L00 raw: 0.005957 0.007189 0.006249 0.006829  scaled: 0.009724 0.008415 0.008645 0.009117
L01 raw: 0.006280 0.007564 0.005387 0.007183  scaled: 0.009125 0.009737 0.010768 0.009621
L02 raw: 0.008370 0.008502 0.009324 0.009011  scaled: 0.009421 0.008591 0.010752 0.011764
L03 raw: 0.008776 0.011027 0.009886 0.007958  scaled: 0.008208 0.014585 0.010488 0.006682
L04 raw: 0.008728 0.008557 0.008399 0.009464  scaled: 0.010042 0.010673 0.009929 0.010543
L05 raw: 0.011430 0.010466 0.010657 0.011286  scaled: 0.015830 0.014031 0.015988 0.021143
L06 raw: 0.009237 0.009855 0.008657 0.008907  scaled: 0.012165 0.013160 0.013843 0.011447
L07 raw: 0.007826 0.008424 0.008186 0.008144  scaled: 0.008999 0.011310 0.009790 0.012291
L08 raw: 0.009206 0.008903 0.009555 0.009378  scaled: 0.010148 0.009800 0.011195 0.010450
L09 raw: 0.009779 0.009484 0.009270 0.008947  scaled: 0.011587 0.009677 0.011122 0.009480
L10 raw: 0.009902 0.009045 0.008853 0.009647  scaled: 0.013206 0.016471 0.009198 0.012304
L11 raw: 0.012788 0.012602 0.008528 0.010900  scaled: 0.028271 0.025918 0.016979 0.021837
L12 raw: 0.008432 0.008985 0.009176 0.008734  scaled: 0.022107 0.018186 0.023879 0.021264
L13 raw: 0.009517 0.010250 0.010473 0.008114  scaled: 0.012743 0.013668 0.014471 0.013402
L14 raw: 0.006164 0.006487 0.006389 0.006431  scaled: 0.013014 0.014988 0.013375 0.013230
L15 raw: 0.009579 0.010492 0.009717 0.009184  scaled: 0.014832 0.013756 0.014412 0.011955
L16 raw: 0.009343 0.009652 0.009410 0.008518  scaled: 0.011380 0.011896 0.011511 0.010833
L17 raw: 0.008544 0.009678 0.009363 0.008641  scaled: 0.015200 0.018420 0.013727 0.013518

=== TOP 20 SCALED ROTARY-COEFFICIENT DISCREPANCIES ===
Q_share allocates coefficient-difference energy, not behavioural influence.
Bands run from earlier to later rotary frequency indices.
layer head bank_rel zero_rel sine_share% Q_share% cosine_sum_ratio band_delta_energy%[4]
   11    0 0.028271 0.046603 68.55 23.35 1.6165 [10.96, 33.09, 44.5, 11.45]
   11    1 0.025918 0.038063 63.52 18.18 1.5327 [18.47, 30.83, 41.22, 9.48]
   12    2 0.023879 0.031121 67.15 74.64 1.5297 [3.87, 9.52, 28.68, 57.94]
   12    0 0.022107 0.030236 64.54 74.21 1.7164 [4.23, 10.54, 30.3, 54.93]
   11    3 0.021837 0.027371 59.43 34.74 1.5815 [5.63, 27.2, 39.7, 27.48]
   12    3 0.021264 0.025833 66.13 73.76 1.4010 [3.05, 8.72, 30.99, 57.24]
    5    3 0.021143 0.024607 60.95 89.26 1.2415 [1.63, 5.53, 16.92, 75.93]
   17    1 0.018420 0.022978 67.63 80.58 2.1284 [6.36, 11.96, 38.02, 43.67]
   12    1 0.018186 0.023847 66.66 67.20 1.7038 [2.61, 8.03, 32.19, 57.17]
   11    2 0.016979 0.018118 58.00 40.77 1.3575 [3.92, 10.4, 43.76, 41.92]
   10    1 0.016471 0.032763 81.66 56.56 1.7973 [9.13, 14.72, 35.6, 40.55]
    5    2 0.015988 0.018362 59.55 89.13 1.2662 [0.63, 3.87, 16.86, 78.64]
    5    0 0.015830 0.022973 58.44 84.43 1.3285 [0.72, 3.25, 17.66, 78.37]
   17    0 0.015200 0.015541 70.53 81.24 1.4819 [2.16, 9.02, 39.92, 48.9]
   14    1 0.014988 0.025865 76.61 69.54 2.1132 [5.08, 14.22, 30.05, 50.65]
   15    0 0.014832 0.022713 80.49 54.65 1.6871 [8.92, 14.49, 36.21, 40.38]
    3    1 0.014585 0.018886 62.39 63.49 1.3209 [12.74, 17.48, 33.53, 36.25]
   13    2 0.014471 0.023653 69.57 45.57 1.8083 [18.77, 35.15, 33.2, 12.88]
   15    2 0.014412 0.018812 81.25 51.52 1.6794 [9.21, 16.14, 34.0, 40.65]
    5    1 0.014031 0.014135 60.15 92.13 1.1853 [0.38, 2.01, 13.76, 83.85]

Top-10 overlap, rotary bank versus zero-displacement comparison: 8
New in rotary-bank top 10: [(11, 2), (17, 1)]
Only in zero-displacement top 10: [(10, 1), (14, 1)]

=== SIGNED CHANGE IN COEFFICIENT-ENERGY DISTRIBUTION OVER FREQUENCY BANDS ===
B minus A, percentage points; not an inference about attended distances.
L11H0: [-0.019549, -0.030671, 0.062395, -0.012175]
L11H1: [-0.01585, 0.005983, -0.00047, 0.010337]
L12H2: [-0.002743, -0.015455, 0.019961, -0.001763]
L12H0: [0.006481, -0.00403, -0.036748, 0.034298]
L11H3: [-0.001454, 0.017631, 0.014981, -0.031158]
L12H3: [0.001788, 0.002254, 0.02097, -0.025012]
L05H3: [-0.001304, -0.010513, -0.005898, 0.017714]
L17H1: [0.017708, 0.032193, -0.056049, 0.006148]
L12H1: [0.000411, 0.00312, 0.02533, -0.028861]
L11H2: [-0.000254, -0.001863, 0.002059, 5.8e-05]

Sine-difference energy share across all scaled heads [min,median,max]: [0.58   0.6857 0.9138]
No coefficient has been associated with the user's guess or the correct answer.
