=== GATE/UP SIGN GEOMETRY ===
No gate, up response, GELU, or hidden state is evaluated.
For weight vectors g,u, sym(g outer u) has nonzero eigenvalues
  (g.u +/- ||g|| ||u||)/2.
Unless g and u are collinear, this quadratic form is indefinite.
GELU's multiplicative sign factor is nonnegative; this is a symbolic
sign observation, not an activation calculation.
rho is a cosine between PARAMETER vectors, optionally diagonally scaled.
small/wrong-sign eigen-capacity ratio = (1-|rho|)/(1+|rho|).
This ratio is NOT a probability of either coefficient sign.

L geometry | median |rho|, p95, max | units |rho|>=.9/.99 | maximum A/B cosine change
 0 raw  | 0.045 0.176 0.837 | 0/0 | 0.00365
 0 diag | 0.054 0.207 0.846 | 0/0 | 0.00288
 1 raw  | 0.038 0.138 0.817 | 0/0 | 0.00244
 1 diag | 0.043 0.146 0.902 | 1/0 | 0.00320
 2 raw  | 0.049 0.196 0.785 | 0/0 | 0.00354
 2 diag | 0.053 0.191 0.870 | 0/0 | 0.00359
 3 raw  | 0.065 0.274 0.758 | 0/0 | 0.00304
 3 diag | 0.080 0.296 0.900 | 1/0 | 0.00769
 4 raw  | 0.083 0.380 0.808 | 0/0 | 0.00264
 4 diag | 0.119 0.352 0.675 | 0/0 | 0.00416
 5 raw  | 0.095 0.265 0.669 | 0/0 | 0.00443
 5 diag | 0.112 0.293 0.555 | 0/0 | 0.00431
 6 raw  | 0.092 0.296 0.656 | 0/0 | 0.00318
 6 diag | 0.105 0.304 0.628 | 0/0 | 0.00383
 7 raw  | 0.090 0.284 0.728 | 0/0 | 0.00731
 7 diag | 0.099 0.321 0.813 | 0/0 | 0.00718
 8 raw  | 0.109 0.300 0.768 | 0/0 | 0.00373
 8 diag | 0.115 0.332 0.717 | 0/0 | 0.00410
 9 raw  | 0.084 0.266 0.639 | 0/0 | 0.00246
 9 diag | 0.102 0.335 0.779 | 0/0 | 0.00441
10 raw  | 0.103 0.326 0.775 | 0/0 | 0.00391
10 diag | 0.136 0.428 0.837 | 0/0 | 0.00704
11 raw  | 0.110 0.338 0.837 | 0/0 | 0.00225
11 diag | 0.143 0.452 0.891 | 0/0 | 0.01039
12 raw  | 0.079 0.307 0.749 | 0/0 | 0.00247
12 diag | 0.129 0.420 0.849 | 0/0 | 0.00547
13 raw  | 0.045 0.262 0.735 | 0/0 | 0.00190
13 diag | 0.111 0.370 0.855 | 0/0 | 0.00528
14 raw  | 0.040 0.272 0.723 | 0/0 | 0.00192
14 diag | 0.114 0.360 0.782 | 0/0 | 0.00505
15 raw  | 0.036 0.243 0.700 | 0/0 | 0.00169
15 diag | 0.113 0.355 0.700 | 0/0 | 0.00619
16 raw  | 0.033 0.271 0.658 | 0/0 | 0.00194
16 diag | 0.114 0.364 0.786 | 0/0 | 0.00576
17 raw  | 0.042 0.340 0.816 | 0/0 | 0.00231
17 diag | 0.128 0.442 0.874 | 0/0 | 0.00742

=== PREVIOUSLY SELECTED UNITS ===
L unit | raw rho A/B | diagonal rho A/B | diagonal wrong-sign capacity ratio
13  163 | -0.0384/-0.0387 | +0.0555/+0.0550 | 0.8953
13 1839 | -0.0592/-0.0585 | -0.0040/-0.0045 | 0.9915
16   36 | -0.0375/-0.0376 | -0.0589/-0.0615 | 0.8864
16 1263 | +0.0789/+0.0784 | +0.0550/+0.0542 | 0.8965
16 1125 | -0.3621/-0.3616 | -0.1879/-0.1881 | 0.6834
17  956 | +0.0471/+0.0471 | -0.1026/-0.1020 | 0.8144
17 1149 | +0.0429/+0.0421 | +0.2005/+0.1981 | 0.6676
17 1940 | -0.0093/-0.0092 | +0.0594/+0.0607 | 0.8867
17  824 | +0.0769/+0.0775 | -0.0797/-0.0772 | 0.8545
17  657 | +0.0134/+0.0132 | +0.0938/+0.0935 | 0.8287
17  790 | +0.0895/+0.0893 | +0.1558/+0.1577 | 0.7290
17 1393 | +0.0361/+0.0365 | +0.1631/+0.1643 | 0.7186
17 1643 | -0.1517/-0.1516 | -0.1426/-0.1431 | 0.7501
17 1865 | -0.1169/-0.1170 | -0.4498/-0.4523 | 0.3783
17 1773 | +0.0256/+0.0253 | -0.0291/-0.0273 | 0.9452
17  152 | +0.2374/+0.2376 | +0.4725/+0.4734 | 0.3578
17  145 | +0.0154/+0.0154 | +0.2381/+0.2398 | 0.6143
17 1801 | -0.0332/-0.0329 | +0.1239/+0.1256 | 0.7782
17 1353 | -0.2142/-0.2144 | -0.1667/-0.1680 | 0.7133
17 1314 | +0.0364/+0.0364 | -0.0601/-0.0610 | 0.8858
17 1641 | -0.0199/-0.0199 | +0.0234/+0.0231 | 0.9545
17  724 | -0.0131/-0.0136 | -0.0423/-0.0423 | 0.9188
17 1077 | +0.0695/+0.0700 | +0.1636/+0.1628 | 0.7194
17 1027 | +0.1637/+0.1627 | +0.2667/+0.2662 | 0.5792
17 1516 | -0.0183/-0.0192 | -0.1312/-0.1333 | 0.7664

=== LARGEST CONDITIONAL CAPACITY INGREDIENTS, THREE PER LAYER ===
These use weight norms and task-output projection, not activity.
Unknown attention, gate magnitude, and norm denominators prevent a
behavioural importance ranking from these bounds.
L unit | capacity ingredient | projected output change | rho | wrong-sign ratio
 0 1352 | 16.424 | 0.18324 | -0.0970 | 0.8231
 0 1130 | 14.749 | 0.074053 | +0.8461 | 0.0834
 0  878 | 13.422 | 0.10185 | -0.3757 | 0.4538
 1  681 | 296.28 | 0.17433 | +0.1845 | 0.6884
 1  802 | 295.21 | 0.23596 | -0.0712 | 0.8671
 1 1555 | 291.41 | 0.23822 | +0.0676 | 0.8733
 2  624 | 338.92 | 0.18389 | -0.6563 | 0.2075
 2 1831 | 327.69 | 0.14629 | -0.8698 | 0.0696
 2   35 | 264.4 | 0.22064 | -0.0909 | 0.8333
 3  146 | 1183.7 | 0.41524 | -0.3217 | 0.5132
 3   82 | 1106.4 | 0.47398 | +0.6093 | 0.2428
 3  861 | 1097.4 | 0.46772 | -0.3912 | 0.4376
 4  563 | 226.72 | 0.23077 | +0.3972 | 0.4314
 4  937 | 195.07 | 0.35475 | +0.5258 | 0.3108
 4  658 | 170.36 | 0.32217 | +0.1817 | 0.6925
 5  226 | 152.8 | 0.25747 | +0.3006 | 0.5378
 5 1267 | 130.43 | 0.25576 | -0.0733 | 0.8634
 5  725 | 123.82 | 0.23844 | -0.1979 | 0.6696
 6 1824 | 151.75 | 0.15322 | -0.4908 | 0.3416
 6  690 | 81.227 | 0.27311 | -0.3348 | 0.4984
 6  314 | 73.394 | 0.20293 | +0.0075 | 0.9851
 7 1455 | 141.67 | 0.2026 | +0.3300 | 0.5038
 7 1300 | 96.392 | 0.19586 | +0.7919 | 0.1161
 7 1032 | 74.398 | 0.21438 | -0.8125 | 0.1034
 8  487 | 98.074 | 0.39324 | +0.3095 | 0.5273
 8 1379 | 56.354 | 0.2282 | +0.1775 | 0.6986
 8  229 | 50.625 | 0.25489 | -0.4197 | 0.4088
 9 1944 | 34.401 | 0.44071 | +0.2989 | 0.5398
 9  448 | 32.586 | 0.34143 | -0.0364 | 0.9297
 9  563 | 31.435 | 0.38758 | -0.3862 | 0.4428
10 1363 | 423.31 | 0.57536 | -0.6959 | 0.1793
10 1564 | 359.65 | 0.67788 | +0.6323 | 0.2252
10 1529 | 285.62 | 0.69684 | -0.7135 | 0.1672
11  716 | 590.38 | 1.8234 | +0.8775 | 0.0653
11 1685 | 510.19 | 2.2317 | -0.4277 | 0.4008
11 1388 | 497.48 | 1.8996 | +0.3674 | 0.4626
12 1528 | 464.85 | 3.4425 | -0.7582 | 0.1375
12  831 | 341.67 | 3.6422 | -0.4660 | 0.3642
12 1520 | 289.76 | 2.2969 | -0.2492 | 0.6011
13   16 | 349.79 | 1.9324 | -0.6191 | 0.2352
13  716 | 300.66 | 3.9661 | -0.4658 | 0.3644
13 1915 | 291.93 | 3.3045 | +0.3769 | 0.4526
14  641 | 627.52 | 3.6431 | -0.7782 | 0.1247
14  392 | 482.11 | 7.1402 | +0.1623 | 0.7207
14 1767 | 376.82 | 5.9636 | +0.0907 | 0.8336
15 1369 | 451.71 | 6.1368 | -0.5049 | 0.3290
15 1940 | 282.39 | 10.176 | -0.3713 | 0.4585
15 1337 | 266.08 | 10.078 | +0.1780 | 0.6978
16  893 | 627.46 | 8.2504 | +0.7862 | 0.1197
16  730 | 454.29 | 9.9133 | +0.5217 | 0.3143
16  931 | 371.4 | 17.403 | -0.0462 | 0.9118
17 1312 | 645.05 | 14.372 | +0.6500 | 0.2121
17 1916 | 629.2 | 9.3071 | +0.8740 | 0.0672
17  687 | 538.28 | 13.165 | +0.5867 | 0.2604

Exact sign-definiteness would require collinearity, not merely a large cosine.
No unit coefficients or their distributions have been estimated.
