=== CAUSAL PREFIX CONSISTENCY WITHIN EACH CHECKPOINT ===
texts0/1 A identical_before_changed_token=True maxdiff=0
texts0/1 B identical_before_changed_token=True maxdiff=0
texts2/3 A identical_before_changed_token=True maxdiff=0
texts2/3 B identical_before_changed_token=True maxdiff=0
texts6/7 A identical_before_changed_token=True maxdiff=0
texts6/7 B identical_before_changed_token=True maxdiff=0
texts8/9 A identical_before_changed_token=True maxdiff=0
texts8/9 B identical_before_changed_token=True maxdiff=0
texts10/11 A identical_before_changed_token=True maxdiff=0
texts10/11 B identical_before_changed_token=True maxdiff=0

=== CROSS-CHECKPOINT ATTENTION DIVERGENCE BY PROCESSING POSITION ===
Columns: ['before_guess', 'guess_last', 'period', 'assistant_boundary', 'answer_prefix_final']
L00 [0.1173 0.1741 0.0533 0.0943 0.1408]
L01 [0.4115 0.0756 0.1161 0.1264 0.6103]
L02 [0.3059 0.1524 0.1054 0.2206 0.4581]
L03 [0.3143 0.1668 0.1627 0.486  0.4928]
L04 [0.5049 0.3438 0.2795 0.2121 0.2904]
L05 [0.7598 0.3762 0.4554 0.2848 0.816 ]
L06 [0.8066 0.7668 0.5236 0.632  0.8443]
L07 [0.9026 0.5819 0.7448 0.4416 0.9521]
L08 [0.7031 0.5428 0.6974 0.3924 0.5919]
L09 [1.0116 0.726  0.819  0.6759 1.41  ]
L10 [0.5315 0.6089 0.5572 0.6198 1.1242]
L11 [0.4645 0.5833 0.5347 0.421  1.0932]
L12 [0.6267 0.3767 0.4953 0.4245 1.0011]
L13 [0.2497 0.3998 0.4495 0.2962 0.272 ]
L14 [0.3071 0.7423 0.8415 0.7563 0.3563]
L15 [0.3831 0.3783 0.4393 0.2877 0.5363]
L16 [0.3294 0.3703 0.4698 0.317  0.3532]
L17 [0.2785 0.2147 0.3857 0.2522 0.4402]

=== GUESS ENCODING: BIGGEST QUESTION-ATTENTION DIFFERENCES ===
Region order: ['question', 'belief_cue', 'guess_tokens', 'BOS', 'self']
L6 H0 A [0.7003 0.0551 0.0208 0.2022 0.0203] B [0.2852 0.2027 0.0605 0.4384 0.0586] B-A question=-0.41511
L6 H3 A [0.5028 0.1946 0.0398 0.2379 0.0357] B [0.1268 0.4695 0.0577 0.3357 0.0541] B-A question=-0.37606
L6 H1 A [0.5239 0.0863 0.0115 0.3485 0.0107] B [0.168  0.3231 0.0152 0.4788 0.0142] B-A question=-0.35588
L7 H0 A [0.7105 0.0465 0.1465 0.093  0.1445] B [0.3753 0.1347 0.2762 0.2127 0.2671] B-A question=-0.33517
L9 H1 A [0.375  0.1425 0.141  0.3304 0.1337] B [0.0601 0.2215 0.3669 0.3483 0.3502] B-A question=-0.31496
L9 H0 A [0.3804 0.0931 0.1085 0.4106 0.1048] B [0.0667 0.1938 0.2557 0.4818 0.2445] B-A question=-0.31370
L10 H2 A [0.7636 0.0434 0.0325 0.1578 0.0298] B [0.4695 0.1866 0.1134 0.2295 0.1103] B-A question=-0.29418
L9 H2 A [0.3368 0.368  0.13   0.1596 0.1238] B [0.0593 0.5218 0.2205 0.1973 0.2098] B-A question=-0.27754
L14 H0 A [0.3682 0.2124 0.0042 0.3379 0.0041] B [0.1108 0.6439 0.0053 0.2203 0.0052] B-A question=-0.25741
L10 H3 A [0.3644 0.1818 0.0357 0.4029 0.0313] B [0.1127 0.2139 0.0755 0.5962 0.0655] B-A question=-0.25178
L7 H1 A [0.4304 0.1498 0.2451 0.1597 0.228 ] B [0.1824 0.2482 0.3209 0.2444 0.2888] B-A question=-0.24799
L11 H3 A [0.3408 0.1567 0.0581 0.4329 0.0531] B [0.1147 0.2195 0.2511 0.4008 0.2352] B-A question=-0.22612

=== EARLY HEADS: ALL REGIONS AT GUESS TOKEN ===
L0 H0 A [0.2012 0.3158 0.1151 0.3544 0.0417] B-A [-0.002   0.021  -0.0003 -0.0175 -0.0056]
L0 H1 A [0.3302 0.0335 0.1222 0.4862 0.1174] B-A [-0.031  -0.0017 -0.0546  0.0827 -0.0545]
L0 H2 A [0.1652 0.5109 0.1308 0.1885 0.0636] B-A [-0.018   0.0685 -0.013  -0.0367 -0.0144]
L0 H3 A [0.0963 0.1907 0.0759 0.6273 0.0566] B-A [ 0.0103  0.0684  0.0213 -0.1019  0.018 ]
L1 H0 A [0.0307 0.0267 0.0063 0.9349 0.0017] B-A [ 0.0055  0.01    0.0025 -0.0184  0.0005]
L1 H1 A [0.0299 0.0079 0.0029 0.9586 0.0014] B-A [ 0.0081  0.0049  0.0032 -0.0166  0.0011]
L1 H2 A [0.3576 0.1683 0.0523 0.3601 0.0473] B-A [-0.0004  0.0277  0.0213 -0.0472  0.0204]
L1 H3 A [0.0199 0.0185 0.0791 0.882  0.0031] B-A [ 0.0046  0.0065  0.0041 -0.0154  0.0007]
L2 H0 A [0.1108 0.3058 0.3449 0.2311 0.3089] B-A [-0.039   0.0064  0.0389 -0.0051  0.0356]
L2 H1 A [0.2822 0.1683 0.0821 0.4422 0.0806] B-A [-0.0968  0.023   0.0133  0.0651  0.0133]
L2 H2 A [0.2145 0.4052 0.1252 0.2172 0.0556] B-A [-0.0285  0.0061  0.0094  0.021   0.0084]
L2 H3 A [0.0061 0.7844 0.115  0.0932 0.0505] B-A [ 0.0007 -0.0437  0.0143  0.0287  0.0138]

=== FOLLOWING PERIOD: BIGGEST GUESS-ATTENTION DIFFERENCES ===
L9 H0 A [0.3593 0.3215 0.0717 0.1253 0.114 ] B [0.0166 0.5859 0.1951 0.0941 0.1066]
L10 H1 A [0.0123 0.1434 0.3747 0.3507 0.1185] B [0.0033 0.1803 0.2694 0.4469 0.0998]
L8 H2 A [0.1793 0.2378 0.1766 0.2631 0.1338] B [0.025  0.2654 0.2624 0.3201 0.1259]
L6 H2 A [0.2418 0.3668 0.1334 0.1766 0.0657] B [0.0359 0.5998 0.2086 0.1139 0.0402]
L10 H3 A [0.263  0.3073 0.0253 0.2694 0.0984] B [0.0667 0.4759 0.0979 0.2943 0.0572]
L10 H2 A [0.3901 0.2809 0.0178 0.1868 0.0826] B [0.1765 0.3494 0.0721 0.2758 0.0913]
L16 H1 A [0.0259 0.0333 0.064  0.6944 0.1779] B [0.0088 0.0276 0.0178 0.9052 0.0376]
L5 H3 A [0.2621 0.3593 0.0037 0.3062 0.0474] B [0.0742 0.5968 0.0487 0.1574 0.1071]
L9 H1 A [0.279  0.179  0.0392 0.3466 0.1169] B [0.028  0.4695 0.0787 0.2895 0.1177]
L4 H1 A [0.0059 0.0842 0.435  0.2717 0.2005] B [0.0054 0.066  0.3967 0.3411 0.1894]

=== GUESS-ENCODING CHANGES BY ANSWER FAMILY ===
L6 H0 {'cities': -0.4048, 'arithmetic': -0.73883, 'temperature': -0.43208, 'options': -0.3164, 'boolean': -0.28214}
L6 H3 {'cities': -0.52419, 'arithmetic': -0.46929, 'temperature': -0.41992, 'options': -0.32361, 'boolean': -0.19573}
L6 H1 {'cities': -0.34147, 'arithmetic': -0.43853, 'temperature': -0.38354, 'options': -0.33358, 'boolean': -0.30457}
L7 H0 {'cities': -0.24563, 'arithmetic': -0.3667, 'temperature': -0.53147, 'options': -0.22228, 'boolean': -0.42267}
L9 H1 {'cities': -0.3084, 'arithmetic': -0.29256, 'temperature': -0.27112, 'options': -0.34366, 'boolean': -0.33036}
L9 H0 {'cities': -0.2084, 'arithmetic': -0.34102, 'temperature': -0.26996, 'options': -0.34798, 'boolean': -0.36688}
L10 H2 {'cities': -0.27388, 'arithmetic': -0.51322, 'temperature': -0.09626, 'options': -0.20326, 'boolean': -0.47516}
L9 H2 {'cities': -0.21932, 'arithmetic': -0.12513, 'temperature': -0.40383, 'options': -0.30023, 'boolean': -0.31653}

=== LOCALITY LIMIT ===
Layer-0 attention differences cannot be caused by earlier transformer layers.
At layer L>0, observed attention differences can reflect both local parameters
and all preceding attention/MLP computations. No ablation has been performed.
