OUTPUT of run 4 of at most 100: ``` Key-path matching: tr(QK_l @ OV_0)/(||QK_l||F ||OV_0||F) L1H0 [0.0018 0.0341 0.0502 0.0058 0.0292 0.0479] L1H1 [0.0004 0.0065 0.025 0.0141 0.0042 0.0212] L1H2 [-0.0018 0.0296 0.0418 0.0121 0.0312 0.0467] L1H3 [0.0012 0.0283 0.0453 0.0051 0.0242 0.0429] L1H4 [0.0007 0.0167 0.0297 0.001 0.0233 0.0368] L1H5 [-0.0012 0.0195 0.0396 0.0112 0.0194 0.0307] L2H0 [-0.0004 0.0385 0.0584 0.0038 0.0229 0.0667] L2H1 [0.0014 0.021 0.0334 0.0005 0.0129 0.0425] L2H2 [0.0025 0.0257 0.0391 0.0053 0.0136 0.0461] L2H3 [0.0031 0.0164 0.0226 0.0047 0.0122 0.0247] L2H4 [-0.0014 0.0299 0.0451 0.0088 0.022 0.0512] L2H5 [0.0008 0.0211 0.0362 0.0006 0.0135 0.0475] L3H0 [-0.0025 0.0081 0.0128 -0.0104 0.0049 0.013 ] L3H1 [ 0.0024 0.0098 0.0134 -0.0017 0.0037 0.0139] L3H2 [-0.0004 0.0069 0.0119 0.0018 0.0006 0.0036] L3H3 [-0.008 0.0019 0.0021 0.0626 0.0024 -0.0016] L3H4 [0.0052 0.0093 0.0092 0.0064 0.0104 0.0089] L3H5 [-0.0004 0.0024 0.0073 -0.0032 0. 0.0088] Largest matching compositions across all preceding heads: L1H0 [('L0H2', np.float64(0.0502)), ('L0H5', np.float64(0.0479)), ('L0H1', np.float64(0.0341)), ('L0H4', np.float64(0.0292))] L1H1 [('L0H2', np.float64(0.025)), ('L0H5', np.float64(0.0212)), ('L0H3', np.float64(0.0141)), ('L0H1', np.float64(0.0065))] L1H2 [('L0H5', np.float64(0.0467)), ('L0H2', np.float64(0.0418)), ('L0H4', np.float64(0.0312)), ('L0H1', np.float64(0.0296))] L1H3 [('L0H2', np.float64(0.0453)), ('L0H5', np.float64(0.0429)), ('L0H1', np.float64(0.0283)), ('L0H4', np.float64(0.0242))] L1H4 [('L0H5', np.float64(0.0368)), ('L0H2', np.float64(0.0297)), ('L0H4', np.float64(0.0233)), ('L0H1', np.float64(0.0167))] L1H5 [('L0H2', np.float64(0.0396)), ('L0H5', np.float64(0.0307)), ('L0H1', np.float64(0.0195)), ('L0H4', np.float64(0.0194))] L2H0 [('L0H5', np.float64(0.0667)), ('L0H2', np.float64(0.0584)), ('L0H1', np.float64(0.0385)), ('L0H4', np.float64(0.0229))] L2H1 [('L0H5', np.float64(0.0425)), ('L0H2', np.float64(0.0334)), ('L0H1', np.float64(0.021)), ('L0H4', np.float64(0.0129))] L2H2 [('L0H5', np.float64(0.0461)), ('L0H2', np.float64(0.0391)), ('L0H1', np.float64(0.0257)), ('L0H4', np.float64(0.0136))] L2H3 [('L0H5', np.float64(0.0247)), ('L0H2', np.float64(0.0226)), ('L0H1', np.float64(0.0164)), ('L0H4', np.float64(0.0122))] L2H4 [('L0H5', np.float64(0.0512)), ('L0H2', np.float64(0.0451)), ('L0H1', np.float64(0.0299)), ('L0H4', np.float64(0.022))] L2H5 [('L0H5', np.float64(0.0475)), ('L0H2', np.float64(0.0362)), ('L0H1', np.float64(0.0211)), ('L0H4', np.float64(0.0135))] L3H0 [('L0H5', np.float64(0.013)), ('L0H2', np.float64(0.0128)), ('L0H1', np.float64(0.0081)), ('L0H4', np.float64(0.0049))] L3H1 [('L0H5', np.float64(0.0139)), ('L0H2', np.float64(0.0134)), ('L0H1', np.float64(0.0098)), ('L0H4', np.float64(0.0037))] L3H2 [('L0H2', np.float64(0.0119)), ('L0H1', np.float64(0.0069)), ('L0H5', np.float64(0.0036)), ('L0H3', np.float64(0.0018))] L3H3 [('L0H3', np.float64(0.0626)), ('L2H5', np.float64(0.0482)), ('L2H1', np.float64(0.0328)), ('L1H0', np.float64(0.0305))] L3H4 [('L2H2', np.float64(0.0143)), ('L2H4', np.float64(0.0122)), ('L0H4', np.float64(0.0104)), ('L2H5', np.float64(0.0094))] L3H5 [('L0H5', np.float64(0.0088)), ('L0H2', np.float64(0.0073)), ('L0H1', np.float64(0.0024)), ('L0H4', np.float64(0.0))] L0H3 OV spectrum after LN scale/centering [1.126 0.9554 0.9495 0.8895 0.8632 0.8537 0.8418 0.7799 0.7586 0.7228 0.7078 0.6727 0.6691 0.6516 0.6339 0.6167 0.5652 0.528 0.4384 0.3824 0.3491 0.3158 0.2968 0.2947 0.2815 0.2755 0.2646 0.246 0.2403 0.233 0.2243 0.2131 0.2093 0.1972 0.1892 0.1722 0.1586 0.1483 0.1368 0.1327 0.1111 0.0978 0.083 0.0703 0.0613 0.0516 0.0467 0.0461 0.0386 0.0335 0.0318 0.0305 0.0265 0.0261 0.0251 0.0226 0.0221 0.0213 0.0206 0.0184 0.0178 0.0146 0.0127 0.0108] L0H0 OV spectrum [0.9574 0.898 0.8182 0.6413 0.5544 0.4742 0.4388 0.4058 0.3721 0.3476 0.3294 0.3082 0.2579 0.2439 0.2052 0.1953 0.1708 0.1431 0.1304 0.1084 0.096 0.0926 0.0869 0.0828 0.0698 0.0674 0.0635 0.0576 0.0526 0.0519 0.0457 0.0442 0.0425 0.0417 0.0408 0.039 0.0365 0.0354 0.034 0.0323 0.0303 0.0287 0.0279 0.0267 0.0253 0.0246 0.0235 0.0224] MLP0 input-weight Gram coupling between current-token and H3-written subspaces rank 12 eigs [39.741+0.j 7.056+0.j 4.765-3.776j 4.765+3.776j -2.433+4.831j -2.433-4.831j -5.141+0.689j -5.141-0.689j 5.113+0.j 1.648-2.403j] rank 16 eigs [39.852+0.j 7.234-2.059j 7.234+2.059j -0.98 -6.718j -0.98 +6.718j 5.039-4.178j 5.039+4.178j -5.491+0.j -3.633+1.446j -3.633-1.446j] rank 20 eigs [41.786+0.j 7.33 -6.542j 7.33 +6.542j -8.043+0.j 6.207-4.055j 6.207+4.055j -5.59 -3.353j -5.59 +3.353j 0.14 +6.279j 0.14 -6.279j] rank 24 eigs [41.757+0.j 7.518-6.519j 7.518+6.519j -8.102+0.j 6.02 +4.113j 6.02 -4.113j -5.638-3.474j -5.638+3.474j 0.077-6.596j 0.077+6.596j] rank 28 eigs [41.366+0.j 7.666-6.504j 7.666+6.504j -8.123+0.j 6.198+4.074j 6.198-4.074j -5.748-3.661j -5.748+3.661j 0.223+6.696j 0.223-6.696j] rank 32 eigs [41.282+0.j 7.748-6.285j 7.748+6.285j -7.918+0.j 6.278+3.815j 6.278-3.815j -5.778+3.556j -5.778-3.556j 0.61 -6.066j 0.61 +6.066j] Later-head QK on H3's read-to-write circuit (r=24) L1H0 [ 0.128 0.0389 0.5584 0.0466 -0.042 ] L1H1 [ 0.1472 0.0892 0.5631 0.0608 -0.0361] L1H2 [ 0.1417 0.0771 0.5887 0.0544 -0.0481] L1H3 [ 0.1321 0.0271 0.5715 0.0402 -0.051 ] L1H4 [ 0.1014 0.007 0.5075 0.0277 -0.0313] L1H5 [ 0.1394 0.0696 0.5988 0.0558 -0.0541] L2H0 [ 0.0662 0.0166 0.5246 0.0219 -0.0148] L2H1 [ 0.1033 0.0017 0.5632 0.032 -0.0293] L2H2 [ 0.0711 0.0257 0.5064 0.021 -0.016 ] L2H3 [ 0.0729 0.0241 0.5548 0.025 -0.0186] L2H4 [ 0.1015 0.0754 0.5898 0.0356 -0.0282] L2H5 [ 0.1057 0.0023 0.5316 0.031 -0.0299] L3H0 [ 0.1062 -0.0634 0.4957 0.0332 -0.0332] L3H1 [ 0.0521 -0.0047 0.5157 0.0165 -0.0141] L3H2 [ 0.0577 0.0106 0.5564 0.0192 -0.0143] L3H3 [ 0.152 0.4151 0.7693 0.0613 -0.0104] L3H4 [ 0.0252 0.0139 0.5185 0.0096 -0.0056] L3H5 [ 0.0711 -0.0108 0.517 0.0202 -0.0214] ```