OUTPUT of run 19 of at most 60:
```
=== SOURCE-UNIT LOCALIZATION OF MLP15 -> L16 UNIT1593 ===
Degree means the source numerator's formal coefficient degree.
Shared expansion constants cancel from the reported ratios.

=== DEGREE 2 ===
Target1593 direction-normalized rank: 4 of 2048
Target1593 coupling gain: 0.39204611097858383
Target1593 source/control: 14.294257626398242
Top receiving units: [(1682, 0.463142), (1801, 0.413629), (332, 0.395742), (1593, 0.392046), (888, 0.372927), (1955, 0.367964), (407, 0.338054), (375, 0.233802)]
Source/receiver/cross fractions of FULL connection difference: [0.9831 0.0023 0.0146]
Cosine(source-only difference, full difference): 0.9988907690095896
Source-unit reinforcement ratio: 1.0212765171436275
Negative source contributions: 564 ; summed signed share: -0.10202438819950885

Leading positive source-unit contributions:
unit source_share full_connection_rank source_read_coeff_A source_read_coeff_B
 651   +0.0187461                    1          +0.1939231          -0.3599358
1369   +0.0107866                    2           -2.537085           -2.222483
 440   +0.0107453                    3           +5.963878           +6.933504
 765   +0.0072590                    4          -0.5709437          +0.2970007
1360   +0.0067295                    5           +2.149811           +1.122754
1052   +0.0066563                    6           +1.307962          +0.4639941
 667   +0.0058536                    7           -2.187561           -3.000016
 343   +0.0055977                    8           +1.890165           +2.641172
 952   +0.0055676                    9            -3.32053           -2.144904
 335   +0.0051027                   10           +9.962394           +9.345648
1347   +0.0050768                   11           +3.403291           +4.637927
1744   +0.0049030                   12           +3.745679           +4.580612

Leading opposing source-unit contributions:
unit=1626 signed_share=-0.0011697
unit=1317 signed_share=-0.0011419
unit=1205 signed_share=-0.0011250
unit=1461 signed_share=-0.0010772
unit= 828 signed_share=-0.0010518

Partial coefficient-difference reconstruction:
 K top_signed_overlap_explained top_absolute_overlap_explained uniform_subset_expected_explained
  1                    0.0236895                      0.0236895                          0.0004985
  3                    0.0468827                      0.0468827                          0.0014953
  5                    0.0650161                      0.0650161                          0.0024922
 10                    0.0978287                      0.0978287                          0.0049841
 20                    0.1499357                      0.1499357                          0.0099672
 50                    0.2695953                      0.2695953                          0.0249105
100                    0.3971398                      0.3971398                          0.0497962
256                    0.5826619                      0.5826619                          0.1272798
512                    0.7087855                      0.7115065                          0.2539081
Source covariance/Gram accounting discrepancy: 4.458878278572005e-15
Full midpoint/Gram accounting discrepancy: 6.806108386720957e-15

=== DEGREE 3 ===
Target1593 direction-normalized rank: 3 of 2048
Target1593 coupling gain: 0.44262042868261614
Target1593 source/control: 15.996272649731468
Top receiving units: [(1682, 0.511791), (888, 0.461432), (1593, 0.44262), (1801, 0.435113), (407, 0.410965), (1955, 0.341717), (332, 0.316471), (2007, 0.210518)]
Source/receiver/cross fractions of FULL connection difference: [0.9709 0.0017 0.0274]
Cosine(source-only difference, full difference): 0.9992659678601477
Source-unit reinforcement ratio: 1.0040598665616194
Negative source contributions: 290 ; summed signed share: -0.007952775768184935

Leading positive source-unit contributions:
unit source_share full_connection_rank source_read_coeff_A source_read_coeff_B
 651   +0.1774839                    1          +0.1939231          -0.3599358
1369   +0.1667789                    2           -2.537085           -2.222483
 667   +0.0303262                    3           -2.187561           -3.000016
1267   +0.0209492                    4           -1.673749            -2.40524
 765   +0.0169505                    5          -0.5709437          +0.2970007
2035   +0.0153683                    6          -0.5629453           -1.216813
1746   +0.0127273                    7           +2.418035           +2.879887
 604   +0.0122291                    8          +0.1884581          -0.4275197
1497   +0.0079959                    9           +5.429302            +4.51064
 440   +0.0075464                   10           +5.963878           +6.933504
1353   +0.0072025                   13            -3.32434           -2.952624
1337   +0.0069642                   11           -2.243739           -1.528903

Leading opposing source-unit contributions:
unit=1676 signed_share=-0.0005915
unit=2031 signed_share=-0.0002116
unit= 545 signed_share=-0.0002090
unit= 164 signed_share=-0.0001879
unit= 544 signed_share=-0.0001740

Partial coefficient-difference reconstruction:
 K top_signed_overlap_explained top_absolute_overlap_explained uniform_subset_expected_explained
  1                    0.1775931                      0.1775931                          0.0004903
  3                    0.3754975                      0.3754975                          0.0014708
  5                    0.4150356                      0.4150356                          0.0024513
 10                    0.4677370                      0.4677370                          0.0049025
 20                    0.5252081                      0.5252081                          0.0098047
 50                    0.6081805                      0.6081805                          0.0245104
100                    0.6864761                      0.6864761                          0.0490160
256                    0.8003623                      0.8000228                          0.1254425
512                    0.8853078                      0.8853568                          0.2507585
Source covariance/Gram accounting discrepancy: 3.767690341713002e-15
Full midpoint/Gram accounting discrepancy: 3.2922113539272044e-15

=== DEGREE 5 ===
Target1593 direction-normalized rank: 5 of 2048
Target1593 coupling gain: 0.6169462023689705
Target1593 source/control: 20.96686220217719
Top receiving units: [(888, 0.919267), (1682, 0.828612), (407, 0.755466), (1801, 0.664041), (1593, 0.616946), (1073, 0.313773), (654, 0.313288), (1977, 0.312831)]
Source/receiver/cross fractions of FULL connection difference: [0.9444 0.0011 0.0546]
Cosine(source-only difference, full difference): 0.999868468241194
Source-unit reinforcement ratio: 0.9997883541398571
Negative source contributions: 265 ; summed signed share: -0.00025407290164997836

Leading positive source-unit contributions:
unit source_share full_connection_rank source_read_coeff_A source_read_coeff_B
1369   +0.5474345                    1           -2.537085           -2.222483
 651   +0.4064683                    2          +0.1939231          -0.3599358
2035   +0.0075479                    3          -0.5629453           -1.216813
1267   +0.0059335                    5           -1.673749            -2.40524
1746   +0.0057708                    4           +2.418035           +2.879887
 667   +0.0053242                    6           -2.187561           -3.000016
 765   +0.0024782                    7          -0.5709437          +0.2970007
1397   +0.0018797                    8          +0.2496807          +0.5314327
1353   +0.0016112                   10            -3.32434           -2.952624
 641   +0.0014651                    9           -1.169562          -0.7646071
 105   +0.0013380                   11           +1.813529           +1.607407
1940   +0.0011079                   14          +0.5621159          +0.3413301

Leading opposing source-unit contributions:
unit= 436 signed_share=-0.0000789
unit=1676 signed_share=-0.0000535
unit=2000 signed_share=-0.0000124
unit=1696 signed_share=-0.0000120
unit= 544 signed_share=-0.0000092

Partial coefficient-difference reconstruction:
 K top_signed_overlap_explained top_absolute_overlap_explained uniform_subset_expected_explained
  1                    0.5474657                      0.5474657                          0.0004882
  3                    0.9615220                      0.9615220                          0.0014645
  5                    0.9730451                      0.9730451                          0.0024409
 10                    0.9857865                      0.9857865                          0.0048818
 20                    0.9932770                      0.9932770                          0.0097636
 50                    0.9969633                      0.9969147                          0.0244090
100                    0.9985335                      0.9985191                          0.0488183
256                    0.9993300                      0.9993440                          0.1249768
512                    0.9996529                      0.9996758                          0.2499603
Source covariance/Gram accounting discrepancy: 5.943000272958222e-14
Full midpoint/Gram accounting discrepancy: 4.721981367830752e-14

=== CROSS-ORDER SOURCE-UNIT CONSISTENCY ===
Units in the top 20 at every examined degree: [604, 651, 667, 765, 1369]
unit= 440 ranks_degree2/3/5=[3, 10, 37] shares=[0.0107 0.0075 0.0001]
unit= 651 ranks_degree2/3/5=[1, 1, 2] shares=[0.0187 0.1775 0.4065]
unit= 667 ranks_degree2/3/5=[7, 3, 6] shares=[0.0059 0.0303 0.0053]
unit= 765 ranks_degree2/3/5=[4, 5, 7] shares=[0.0073 0.017  0.0025]
unit=1267 ranks_degree2/3/5=[21, 4, 4] shares=[0.0044 0.0209 0.0059]
unit=1360 ranks_degree2/3/5=[5, 24, 61] shares=[0.0067 0.004  0.    ]
unit=1369 ranks_degree2/3/5=[2, 2, 1] shares=[0.0108 0.1668 0.5474]
unit=1746 ranks_degree2/3/5=[50, 7, 5] shares=[0.003  0.0127 0.0058]
unit=2035 ranks_degree2/3/5=[115, 6, 3] shares=[0.0021 0.0154 0.0075]

Elapsed seconds: 1.78

Interpretation limits:
A coefficient reconstruction is not a model ablation or behavioral reconstruction.
Signed source read coefficients are conditional weight contractions,
not measured gate inputs, activation signs, or evidence of inhibition/excitation
of a semantic feature on real examples.
Receiver selectivity and sparse coefficient structure would still not
identify guesses, correct answers, or the sycophantic checkpoint.

```