OUTPUT of run 13 of at most 60:
```
=== SYMBOLIC MLP COEFFICIENTS ===
Coefficients: {2: 0.5, 3: np.float64(0.3989422804014327), 5: np.float64(-0.06681947717750458)}
Degree 2 is the exact even part of the specified numerator.
Degrees 3 and 5 are its leading odd Taylor terms, not the complete odd part.
No activation function or input vector is evaluated.

=== PER-UNIT RANKING STABILITY ACROSS DEGREES ===
layer | top10 overlap 2/3 2/5 | rank correlation 2/3 2/5 | top3 units by degree 2;3;5
 0 |  7  6 | 0.96137 0.81086 | [1130, 1405, 1856]; [1130, 1856, 1405]; [1130, 1856, 878]
 1 |  5  2 | 0.96677 0.84257 | [239, 835, 1355]; [239, 835, 1355]; [1725, 239, 167]
 2 |  6  3 | 0.96266 0.81979 | [1831, 154, 925]; [1831, 624, 6]; [1831, 624, 6]
 3 |  4  4 | 0.90806 0.76911 | [1414, 226, 986]; [82, 1671, 226]; [82, 1671, 303]
 4 |  7  3 | 0.93608 0.79985 | [1044, 1527, 1353]; [1044, 1527, 280]; [280, 1527, 1044]
 5 |  5  2 | 0.90558 0.74586 | [584, 1759, 919]; [584, 1579, 708]; [1579, 708, 584]
 6 |  5  4 | 0.90141 0.72208 | [1122, 1824, 314]; [1122, 1824, 314]; [1824, 1122, 314]
 7 |  8  5 | 0.87734 0.67907 | [1455, 1057, 1032]; [1455, 1300, 1032]; [1300, 1455, 656]
 8 |  5  4 | 0.92385 0.80999 | [188, 1558, 487]; [487, 522, 1394]; [1819, 522, 1394]
 9 |  5  3 | 0.90704 0.78825 | [448, 1782, 808]; [448, 1782, 1472]; [448, 1766, 1644]
10 |  7  6 | 0.95159 0.87721 | [1363, 1564, 1529]; [1363, 1564, 1529]; [1363, 1529, 1564]
11 |  7  6 | 0.95505 0.87517 | [1815, 716, 1576]; [1815, 1576, 716]; [1815, 1576, 716]
12 |  7  6 | 0.92149 0.80610 | [1528, 1520, 259]; [1528, 1520, 1584]; [1528, 1584, 1520]
13 |  8  7 | 0.91506 0.80237 | [16, 1575, 1517]; [16, 1575, 369]; [16, 369, 1517]
14 |  8  8 | 0.89940 0.76837 | [641, 685, 663]; [641, 1529, 685]; [641, 1529, 1801]
15 |  7  4 | 0.88254 0.73155 | [1369, 1226, 651]; [1369, 651, 1940]; [1369, 651, 1722]
16 |  8  5 | 0.85139 0.66750 | [893, 730, 1591]; [893, 730, 324]; [893, 730, 204]
17 |  8  5 | 0.83204 0.63666 | [1916, 1312, 1026]; [1916, 1312, 1026]; [1312, 1916, 60]

=== COMPLETE COEFFICIENT-TENSOR CHANGES, ALL UNITS ===
Do not compare absolute norms across degrees as behavioral importance.
layer degree | whole_delta_norm relative_delta coherence | top10_attribution negative_attribution
 2 2 |       1110.15 0.0191339 1.019204 | 0.072021 -0.000300
 2 3 |         23668 0.0230627 1.011657 | 0.370168 -0.000422
 2 5 |   1.39395e+07 0.0584427 1.000935 | 0.981190 -0.000020
 3 2 |       3060.04 0.0231088 1.028342 | 0.057108 -0.000003
 3 3 |       68710.3 0.0220670 1.006669 | 0.186033 +0.000000
 3 5 |     3.459e+07 0.0263037 1.012198 | 0.867208 +0.000000
10 2 |       476.895 0.0205425 1.146846 | 0.197047 -0.000211
10 3 |       5888.01 0.0202307 0.935775 | 0.785392 -0.000100
10 5 |    1.3339e+06 0.0225728 0.905135 | 0.995845 -0.000039
12 2 |       938.298 0.0149434 1.101663 | 0.039206 +0.000000
12 3 |       5511.32 0.0149124 1.018856 | 0.257964 -0.000006
12 5 |        279853 0.0142299 1.000822 | 0.835651 -0.000011
16 2 |        579.16 0.0178494 1.068137 | 0.048855 +0.000000
16 3 |       1888.89 0.0175032 1.005758 | 0.334902 +0.000000
16 5 |       39790.5 0.0148799 1.001467 | 0.968605 -0.000085

=== TOP FIVE CONTRIBUTORS TO EACH COMPLETE COEFFICIENT CHANGE ===
layer= 2 degree=2 units_and_signed_attributions=[(1831, 0.039599), (6, 0.00433), (154, 0.0042421), (918, 0.0039057), (798, 0.0037884)]
layer= 2 degree=3 units_and_signed_attributions=[(1831, 0.3272498), (6, 0.0083252), (624, 0.006858), (798, 0.0059358), (1790, 0.0052415)]
layer= 2 degree=5 units_and_signed_attributions=[(1831, 0.9734124), (624, 0.0029838), (6, 0.001558), (1790, 0.0009619), (798, 0.0005483)]
layer= 3 degree=2 units_and_signed_attributions=[(1414, 0.0125146), (226, 0.0066545), (1671, 0.0058986), (986, 0.0054151), (82, 0.0050591)]
layer= 3 degree=3 units_and_signed_attributions=[(82, 0.0565295), (1671, 0.0321259), (226, 0.0193306), (303, 0.0188772), (1635, 0.0125792)]
layer= 3 degree=5 units_and_signed_attributions=[(82, 0.5803139), (1671, 0.0870212), (303, 0.0728888), (1073, 0.0647292), (226, 0.0129347)]
layer=10 degree=2 units_and_signed_attributions=[(1363, 0.0780284), (1564, 0.0346243), (1529, 0.0296607), (1106, 0.0127541), (900, 0.0111535)]
layer=10 degree=3 units_and_signed_attributions=[(1363, 0.4376607), (1564, 0.1600499), (1529, 0.0874367), (1869, 0.03431), (900, 0.0258878)]
layer=10 degree=5 units_and_signed_attributions=[(1363, 0.7522091), (1564, 0.1219092), (1529, 0.0859005), (1869, 0.0241134), (900, 0.0059698)]
layer=12 degree=2 units_and_signed_attributions=[(1528, 0.0080724), (1520, 0.0040502), (259, 0.0037781), (22, 0.0037572), (2027, 0.0035525)]
layer=12 degree=3 units_and_signed_attributions=[(1528, 0.0972558), (1520, 0.027475), (1584, 0.0246763), (2027, 0.0225919), (6, 0.0184813)]
layer=12 degree=5 units_and_signed_attributions=[(1528, 0.5398828), (1584, 0.0678294), (1520, 0.0467219), (2027, 0.0442919), (138, 0.0324524)]
layer=16 degree=2 units_and_signed_attributions=[(893, 0.0156962), (730, 0.0071423), (1591, 0.0069032), (287, 0.004465), (324, 0.0032879)]
layer=16 degree=3 units_and_signed_attributions=[(893, 0.2290954), (730, 0.0511411), (324, 0.0111312), (1439, 0.0078142), (238, 0.0067812)]
layer=16 degree=5 units_and_signed_attributions=[(893, 0.8919524), (730, 0.0586196), (324, 0.0032986), (204, 0.0031872), (1900, 0.0026886)]

=== PREVIOUS CANDIDATES ACROSS ALL THREE DEGREES ===
layer unit | ranks at degrees 2/3/5 | signed attributions at degrees 2/3/5
 2 1831 | [1, 1, 1] | +0.0395990 +0.3272498 +0.9734124
 3 1414 | [1, 20, 322] | +0.0125146 +0.0053498 +0.0000498
10 1363 | [1, 1, 1] | +0.0780284 +0.4376607 +0.7522091
10 1564 | [2, 2, 2] | +0.0346243 +0.1600499 +0.1219092
10 1529 | [3, 3, 3] | +0.0296607 +0.0874367 +0.0859005
12 1528 | [1, 1, 1] | +0.0080724 +0.0972558 +0.5398828
12 1520 | [2, 2, 3] | +0.0040502 +0.0274750 +0.0467219
12  259 | [3, 12, 19] | +0.0037781 +0.0121461 +0.0061802
16  893 | [1, 1, 1] | +0.0156962 +0.2290954 +0.8919524

=== WHOLE-TENSOR TOP-TEN OVERLAPS ===
layer= 2 degree2_vs3=7/10 degree2_vs5=4/10
layer= 3 degree2_vs3=5/10 degree2_vs5=4/10
layer=10 degree2_vs3=7/10 degree2_vs5=7/10
layer=12 degree2_vs3=6/10 degree2_vs5=6/10
layer=16 degree2_vs3=8/10 degree2_vs5=5/10

Interpretation limits:
- These are symbolic coefficients of the unnormalized, diagonal-gain-weighted MLP numerator.
- Higher-degree coefficient magnitudes do not establish dominance on actual hidden states.
- A truncated Taylor description does not determine behavior in saturated gating regimes.
- RMS denominators, semantic grounding and output decoding remain absent.
- No inputs, activations, logits, model execution, or training were used.

```