Generated by model_inspector v1.1.0 (2026-10-11) on 2026-10-10 22:40:08. Read-only: no weights were changed.
Two runs of the same prompt. A (reference) = the prompt alone. B (test) = the prompt alone, with the memory 'direction' added to the running thought at layer 12, strength 0.5 x the typical size. Everything below is measured from inside the model on those two runs - no guessing.
The best thing to do on a rainy weekend is (10 tokens)The user loves jigsaw puzzles. (7 tokens)What the model would say if it had to stop at each layer (logit lens)
| layer | best guess so far | its probability | probability of the final answer | uncertainty (nats) |
|---|---|---|---|---|
| 0 | ';";\n |
0.120 | 0.000 | 7.19 |
| 1 | Chain |
0.036 | 0.000 | 7.96 |
| 2 | ';";\n |
0.057 | 0.000 | 7.10 |
| 3 | � |
0.060 | 0.000 | 7.96 |
| 4 | � |
0.107 | 0.000 | 7.79 |
| 5 | ';";\n |
0.155 | 0.000 | 5.88 |
| 6 | ';";\n |
0.128 | 0.000 | 6.41 |
| 7 | %";\n |
0.248 | 0.000 | 4.30 |
| 8 | ')));\n\n |
0.077 | 0.000 | 7.09 |
| 9 | ?>> |
0.085 | 0.000 | 7.39 |
| 10 | ","# |
0.063 | 0.000 | 7.62 |
| 11 | ped |
0.078 | 0.000 | 7.47 |
| 12 | ","# |
0.079 | 0.000 | 8.08 |
| 13 | ","# |
0.235 | 0.000 | 6.97 |
| 14 | ","# |
0.388 | 0.000 | 5.36 |
| 15 | ","# |
0.399 | 0.000 | 5.33 |
| 16 | ","# |
0.209 | 0.000 | 6.98 |
| 17 | ","# |
0.067 | 0.003 | 8.44 |
| 18 | either |
0.024 | 0.006 | 8.93 |
| 19 | to |
0.105 | 0.105 | 7.06 |
| 20 | to |
0.405 | 0.405 | 4.85 |
| 21 | to |
0.111 | 0.111 | 7.25 |
| 22 | to |
0.261 | 0.261 | 5.19 |
| 23 | to |
0.303 | 0.303 | 4.49 |
Early layers usually guess generic words; the final answer 'crystallises' at some layer. Uncertainty falling = the model is making up its mind. Which layer it settles at is where the decision is made.
How much each layer adds to the running thought (last position)
| layer | thought size entering | attention adds | MLP adds | real change | attn vs MLP direction (cos) | MLP share of the two lengths |
|---|---|---|---|---|---|---|
| 0 | 0.4 | 0.4 | 4.6 | 4.7 | 0.03 | 91.44% |
| 1 | 4.7 | 2.9 | 4.3 | 5.4 | 0.10 | 59.83% |
| 2 | 8.2 | 1.6 | 4.3 | 4.4 | -0.11 | 73.40% |
| 3 | 9.1 | 3.6 | 5.4 | 5.7 | -0.24 | 59.67% |
| 4 | 9.8 | 3.3 | 5.3 | 5.5 | -0.24 | 61.75% |
| 5 | 10.8 | 4.1 | 6.0 | 6.3 | -0.25 | 59.58% |
| 6 | 12.6 | 2.5 | 5.3 | 5.5 | -0.16 | 67.70% |
| 7 | 12.9 | 4.7 | 6.1 | 6.6 | -0.28 | 56.63% |
| 8 | 13.8 | 4.7 | 5.4 | 6.5 | -0.18 | 53.50% |
| 9 | 14.6 | 3.2 | 5.7 | 5.8 | -0.27 | 64.23% |
| 10 | 15.9 | 3.9 | 4.3 | 4.8 | -0.29 | 52.27% |
| 11 | 15.4 | 2.2 | 4.3 | 4.5 | -0.15 | 65.50% |
| 12 | 15.3 | 3.0 | 5.8 | 5.8 | -0.26 | 65.75% |
| 13 | 17.8 | 2.3 | 5.4 | 5.3 | -0.27 | 70.08% |
| 14 | 18.7 | 3.8 | 5.0 | 6.1 | -0.08 | 56.64% |
| 15 | 20.1 | 3.2 | 7.6 | 7.9 | -0.11 | 70.51% |
| 16 | 23.0 | 2.5 | 13.1 | 12.8 | -0.19 | 83.80% |
| 17 | 29.1 | 6.1 | 9.9 | 12.4 | 0.16 | 62.14% |
| 18 | 36.1 | 3.8 | 9.7 | 10.1 | -0.08 | 71.79% |
| 19 | 40.8 | 4.6 | 12.9 | 13.8 | 0.03 | 73.89% |
| 20 | 45.0 | 7.0 | 20.0 | 21.3 | 0.02 | 74.22% |
| 21 | 58.6 | 10.4 | 25.5 | 30.0 | 0.26 | 70.99% |
| 22 | 80.5 | 3.1 | 27.9 | 27.7 | -0.12 | 89.90% |
| 23 | 87.0 | 13.3 | 78.6 | 75.8 | -0.29 | 85.48% |
'Thought size' is the length of the number-vector the model carries for the last word. Each layer adds an attention contribution and an MLP contribution to it. 'Real change' is the length of what the layer actually changed (the two contributions added as arrows, not as lengths). The cosine says whether attention and the MLP pull the same way (+1), at right angles (0) or against each other (-1). Where a steering push is applied, that layer's 'real change' includes the push.
Where the attention spotlights look (averaged over all heads in the layer, last position)
| layer | spread (0 = pin-point, 1 = even) | first token (the sink) | memory text (sink excluded) | memory if attention were even | previous token | itself |
|---|---|---|---|---|---|---|
| 0 | 0.62 | 7.26% | 0.00% | 0.00% | 25.98% | 24.52% |
| 1 | 0.76 | 22.55% | 0.00% | 0.00% | 10.44% | 12.21% |
| 2 | 0.75 | 30.88% | 0.00% | 0.00% | 6.56% | 18.30% |
| 3 | 0.48 | 49.84% | 0.00% | 0.00% | 2.86% | 9.81% |
| 4 | 0.57 | 43.81% | 0.00% | 0.00% | 9.12% | 16.86% |
| 5 | 0.51 | 52.91% | 0.00% | 0.00% | 3.83% | 13.09% |
| 6 | 0.54 | 53.86% | 0.00% | 0.00% | 8.19% | 9.75% |
| 7 | 0.50 | 50.35% | 0.00% | 0.00% | 1.90% | 9.91% |
| 8 | 0.60 | 34.75% | 0.00% | 0.00% | 17.29% | 13.08% |
| 9 | 0.30 | 73.13% | 0.00% | 0.00% | 6.31% | 5.74% |
| 10 | 0.55 | 51.46% | 0.00% | 0.00% | 13.31% | 14.25% |
| 11 | 0.25 | 81.84% | 0.00% | 0.00% | 5.44% | 2.20% |
| 12 | 0.62 | 50.15% | 0.00% | 0.00% | 11.19% | 14.95% |
| 13 | 0.32 | 76.14% | 0.00% | 0.00% | 4.61% | 9.83% |
| 14 | 0.56 | 44.60% | 0.00% | 0.00% | 12.07% | 20.45% |
| 15 | 0.54 | 57.90% | 0.00% | 0.00% | 10.29% | 10.49% |
| 16 | 0.18 | 88.08% | 0.00% | 0.00% | 0.42% | 1.62% |
| 17 | 0.42 | 62.47% | 0.00% | 0.00% | 2.43% | 9.31% |
| 18 | 0.51 | 60.98% | 0.00% | 0.00% | 4.54% | 11.65% |
| 19 | 0.48 | 57.76% | 0.00% | 0.00% | 5.92% | 22.26% |
| 20 | 0.35 | 76.36% | 0.00% | 0.00% | 4.07% | 6.68% |
| 21 | 0.27 | 77.12% | 0.00% | 0.00% | 1.31% | 9.45% |
| 22 | 0.67 | 43.15% | 0.00% | 0.00% | 0.44% | 19.24% |
| 23 | 0.63 | 28.50% | 0.00% | 0.00% | 0.71% | 39.18% |
The first token is an 'attention sink': models park spare attention on it, and it often takes most of the attention in every layer. It is NOT counted as memory here. 'Memory text' is the share landing on the memory sentence's other tokens, and 'memory if attention were even' is what it would get if every token were treated alike. Memory above that line is real interest in the memory.
What the model would say if it had to stop at each layer (logit lens)
| layer | best guess so far | its probability | probability of the final answer | uncertainty (nats) |
|---|---|---|---|---|
| 0 | ';";\n |
0.120 | 0.000 | 7.19 |
| 1 | Chain |
0.036 | 0.000 | 7.96 |
| 2 | ';";\n |
0.057 | 0.000 | 7.10 |
| 3 | � |
0.060 | 0.000 | 7.96 |
| 4 | � |
0.107 | 0.000 | 7.79 |
| 5 | ';";\n |
0.155 | 0.000 | 5.88 |
| 6 | ';";\n |
0.128 | 0.000 | 6.41 |
| 7 | %";\n |
0.248 | 0.000 | 4.30 |
| 8 | ')));\n\n |
0.077 | 0.000 | 7.09 |
| 9 | ?>> |
0.085 | 0.000 | 7.39 |
| 10 | ","# |
0.063 | 0.000 | 7.62 |
| 11 | ped |
0.078 | 0.000 | 7.47 |
| 12 | ped |
0.033 | 0.000 | 8.07 |
| 13 | ped |
0.059 | 0.000 | 8.10 |
| 14 | ped |
0.048 | 0.000 | 7.94 |
| 15 | ped |
0.037 | 0.000 | 7.86 |
| 16 | 堞 |
0.015 | 0.000 | 8.49 |
| 17 | either |
0.015 | 0.001 | 8.76 |
| 18 | either |
0.089 | 0.004 | 8.39 |
| 19 | to |
0.200 | 0.200 | 5.71 |
| 20 | to |
0.877 | 0.877 | 1.19 |
| 21 | to |
0.370 | 0.370 | 4.60 |
| 22 | to |
0.431 | 0.431 | 4.20 |
| 23 | to |
0.522 | 0.522 | 3.15 |
Early layers usually guess generic words; the final answer 'crystallises' at some layer. Uncertainty falling = the model is making up its mind. Which layer it settles at is where the decision is made.
How much each layer adds to the running thought (last position)
| layer | thought size entering | attention adds | MLP adds | real change | attn vs MLP direction (cos) | MLP share of the two lengths |
|---|---|---|---|---|---|---|
| 0 | 0.4 | 0.4 | 4.6 | 4.7 | 0.03 | 91.44% |
| 1 | 4.7 | 2.9 | 4.3 | 5.4 | 0.10 | 59.83% |
| 2 | 8.2 | 1.6 | 4.3 | 4.4 | -0.11 | 73.40% |
| 3 | 9.1 | 3.6 | 5.4 | 5.7 | -0.24 | 59.67% |
| 4 | 9.8 | 3.3 | 5.3 | 5.5 | -0.24 | 61.75% |
| 5 | 10.8 | 4.1 | 6.0 | 6.3 | -0.25 | 59.58% |
| 6 | 12.6 | 2.5 | 5.3 | 5.5 | -0.16 | 67.70% |
| 7 | 12.9 | 4.7 | 6.1 | 6.6 | -0.28 | 56.63% |
| 8 | 13.8 | 4.7 | 5.4 | 6.5 | -0.18 | 53.50% |
| 9 | 14.6 | 3.2 | 5.7 | 5.8 | -0.27 | 64.23% |
| 10 | 15.9 | 3.9 | 4.3 | 4.8 | -0.29 | 52.27% |
| 11 | 15.4 | 2.2 | 4.3 | 4.5 | -0.15 | 65.50% |
| 12 | 15.3 | 3.0 | 5.8 | 10.4 | -0.26 | 65.75% |
| 13 | 19.0 | 2.8 | 6.6 | 6.5 | -0.26 | 70.25% |
| 14 | 19.9 | 3.9 | 6.0 | 6.9 | -0.08 | 60.73% |
| 15 | 20.0 | 3.4 | 7.7 | 7.9 | -0.15 | 68.99% |
| 16 | 21.1 | 2.4 | 11.5 | 11.2 | -0.21 | 82.96% |
| 17 | 25.8 | 5.8 | 8.9 | 11.2 | 0.12 | 60.58% |
| 18 | 31.4 | 3.2 | 8.8 | 9.0 | -0.12 | 73.07% |
| 19 | 35.5 | 4.8 | 12.6 | 13.3 | -0.03 | 72.49% |
| 20 | 39.0 | 8.4 | 20.5 | 22.1 | -0.00 | 71.04% |
| 21 | 52.0 | 11.0 | 23.6 | 27.3 | 0.13 | 68.28% |
| 22 | 70.6 | 3.4 | 29.5 | 29.4 | -0.09 | 89.65% |
| 23 | 79.5 | 13.5 | 70.8 | 67.2 | -0.36 | 83.98% |
'Thought size' is the length of the number-vector the model carries for the last word. Each layer adds an attention contribution and an MLP contribution to it. 'Real change' is the length of what the layer actually changed (the two contributions added as arrows, not as lengths). The cosine says whether attention and the MLP pull the same way (+1), at right angles (0) or against each other (-1). Where a steering push is applied, that layer's 'real change' includes the push.
Where the attention spotlights look (averaged over all heads in the layer, last position)
| layer | spread (0 = pin-point, 1 = even) | first token (the sink) | memory text (sink excluded) | memory if attention were even | previous token | itself |
|---|---|---|---|---|---|---|
| 0 | 0.62 | 7.26% | 0.00% | 0.00% | 25.98% | 24.52% |
| 1 | 0.76 | 22.55% | 0.00% | 0.00% | 10.44% | 12.21% |
| 2 | 0.75 | 30.88% | 0.00% | 0.00% | 6.56% | 18.30% |
| 3 | 0.48 | 49.84% | 0.00% | 0.00% | 2.86% | 9.81% |
| 4 | 0.57 | 43.81% | 0.00% | 0.00% | 9.12% | 16.86% |
| 5 | 0.51 | 52.91% | 0.00% | 0.00% | 3.83% | 13.09% |
| 6 | 0.54 | 53.86% | 0.00% | 0.00% | 8.19% | 9.75% |
| 7 | 0.50 | 50.35% | 0.00% | 0.00% | 1.90% | 9.91% |
| 8 | 0.60 | 34.75% | 0.00% | 0.00% | 17.29% | 13.08% |
| 9 | 0.30 | 73.13% | 0.00% | 0.00% | 6.31% | 5.74% |
| 10 | 0.55 | 51.46% | 0.00% | 0.00% | 13.31% | 14.25% |
| 11 | 0.25 | 81.84% | 0.00% | 0.00% | 5.44% | 2.20% |
| 12 | 0.62 | 50.15% | 0.00% | 0.00% | 11.19% | 14.95% |
| 13 | 0.45 | 63.70% | 0.00% | 0.00% | 7.68% | 13.04% |
| 14 | 0.57 | 47.27% | 0.00% | 0.00% | 11.78% | 18.50% |
| 15 | 0.59 | 56.08% | 0.00% | 0.00% | 8.92% | 9.93% |
| 16 | 0.16 | 90.55% | 0.00% | 0.00% | 0.50% | 2.49% |
| 17 | 0.41 | 67.04% | 0.00% | 0.00% | 2.36% | 10.82% |
| 18 | 0.51 | 63.69% | 0.00% | 0.00% | 5.41% | 12.66% |
| 19 | 0.51 | 51.39% | 0.00% | 0.00% | 6.48% | 25.98% |
| 20 | 0.40 | 70.87% | 0.00% | 0.00% | 4.50% | 6.95% |
| 21 | 0.34 | 72.25% | 0.00% | 0.00% | 1.87% | 8.87% |
| 22 | 0.70 | 42.56% | 0.00% | 0.00% | 1.13% | 13.85% |
| 23 | 0.69 | 25.48% | 0.00% | 0.00% | 1.38% | 29.15% |
The first token is an 'attention sink': models park spare attention on it, and it often takes most of the attention in every layer. It is NOT counted as memory here. 'Memory text' is the share landing on the memory sentence's other tokens, and 'memory if attention were even' is what it would get if every token were treated alike. Memory above that line is real interest in the memory.
reference (A): the prompt alone
test (B): the prompt alone, with the memory 'direction' added to the running thought at layer 12, strength 0.5 x the typical size
top choice in A: to (30.35%)
top choice in B: to (52.21%)
top choice changed: no
probability moved to different next words: 37.16% (0% = nothing changed, 100% = completely different)
KL divergence B from A: 0.4573 nats (0 = identical odds; bigger = bigger change)
| rank | A: next word | A: prob | B: next word | B: prob |
|---|---|---|---|---|
| 1 | to |
30.35% | to |
52.21% |
| 2 | go |
5.16% | go |
4.99% |
| 3 | sit |
3.09% | play |
2.90% |
| 4 | take |
3.08% | take |
2.24% |
| 5 | just |
2.26% | ride |
2.05% |
| 6 | a |
1.94% | bike |
1.85% |
| 7 | get |
1.27% | playing |
1.36% |
| 8 | hang |
1.26% | hang |
1.18% |
| 9 | spend |
1.19% | paddle |
0.91% |
| 10 | catch |
1.14% | walk |
0.81% |
Words whose probability rose most because of the push
| word | A | B | change |
|---|---|---|---|
to |
30.35% | 52.21% | +21.86% |
play |
0.76% | 2.90% | +2.14% |
bike |
0.17% | 1.85% | +1.68% |
ride |
0.49% | 2.05% | +1.56% |
playing |
0.05% | 1.36% | +1.31% |
paddle |
0.09% | 0.91% | +0.82% |
going |
0.16% | 0.59% | +0.43% |
biking |
0.01% | 0.38% | +0.37% |
Your keywords as the very next word
| keyword | first piece | A prob | A rank | B prob | B rank | B / A |
|---|---|---|---|---|---|---|
| jigsaw | j |
0.00% | 1804 | 0.00% | 726 | 2.3x |
| puzzle | puzzle |
0.00% | 1967 | 0.00% | 707 | 2.8x |
Probability and rank (1 = the model's top choice) of each keyword's first word-piece as the immediate next word. Sharper than the probe's 'recall', which only checks whether the word shows up somewhere in the generated text.
| layer | relative change in the running thought | mean attention shift (0-1) | most shifted head |
|---|---|---|---|
| 0 | 0.0000 | 0.0000 | h0 (0.000) |
| 1 | 0.0000 | 0.0000 | h0 (0.000) |
| 2 | 0.0000 | 0.0000 | h0 (0.000) |
| 3 | 0.0000 | 0.0000 | h0 (0.000) |
| 4 | 0.0000 | 0.0000 | h0 (0.000) |
| 5 | 0.0000 | 0.0000 | h0 (0.000) |
| 6 | 0.0000 | 0.0000 | h0 (0.000) |
| 7 | 0.0000 | 0.0000 | h0 (0.000) |
| 8 | 0.0000 | 0.0000 | h0 (0.000) |
| 9 | 0.0000 | 0.0000 | h0 (0.000) |
| 10 | 0.0000 | 0.0000 | h0 (0.000) |
| 11 | 0.0000 | 0.0000 | h0 (0.000) |
| 12 | 0.4947 | 0.0000 | h0 (0.000) |
| 13 | 0.4962 | 0.1317 | h4 (0.344) |
| 14 | 0.4895 | 0.1365 | h2 (0.372) |
| 15 | 0.4631 | 0.1101 | h9 (0.284) |
| 16 | 0.4193 | 0.0521 | h10 (0.311) |
| 17 | 0.3742 | 0.1342 | h9 (0.367) |
| 18 | 0.3651 | 0.1153 | h11 (0.327) |
| 19 | 0.3749 | 0.0999 | h5 (0.241) |
| 20 | 0.3485 | 0.0904 | h2 (0.253) |
| 21 | 0.3051 | 0.0682 | h1 (0.134) |
| 22 | 0.3096 | 0.1056 | h0 (0.177) |
| 23 | 0.3818 | 0.1445 | h2 (0.323) |
Left column: how big the difference between the two runs is at the output of each layer, as a fraction of that layer's output size. A push that fades is small by the last layers; one that grows is being amplified. Attention shift: how far a head's pattern of looking moved (0 = identical, 1 = looks at entirely different words).
The 10 spotlights whose pattern moved most
| layer | head | shift | sink share A -> B | memory share A -> B |
|---|---|---|---|---|
| 14 | 2 | 0.372 | 20.46% -> 57.68% | - |
| 17 | 9 | 0.367 | 44.41% -> 81.09% | - |
| 17 | 10 | 0.359 | 18.86% -> 45.86% | - |
| 13 | 4 | 0.344 | 63.51% -> 29.06% | - |
| 18 | 11 | 0.327 | 22.13% -> 45.02% | - |
| 23 | 2 | 0.323 | 31.46% -> 11.31% | - |
| 13 | 0 | 0.314 | 53.56% -> 23.33% | - |
| 16 | 10 | 0.311 | 19.59% -> 38.54% | - |
| 15 | 9 | 0.284 | 37.21% -> 56.41% | - |
| 13 | 6 | 0.271 | 85.69% -> 58.59% | - |
| position | token | layer 6 | layer 12 | layer 18 |
|---|---|---|---|---|
| 0 | The |
1636.1 | 1641.1 | 1643.9 |
| 1 | best |
15.6 | 17.9 | 36.7 |
| 2 | thing |
15.1 | 17.9 | 37.4 |
| 3 | to |
13.3 | 16.7 | 34.0 |
| 4 | do |
14.1 | 17.9 | 35.4 |
| 5 | on |
13.8 | 18.1 | 32.9 |
| 6 | a |
13.8 | 18.7 | 34.1 |
| 7 | rainy |
15.4 | 20.8 | 39.0 |
| 8 | weekend |
15.1 | 20.8 | 36.2 |
| 9 | is |
12.9 | 19.0 | 35.5 |
| first / typical | 115.7x | 90.6x | 46.3x |
Length of the number-vector the model carries for each token, at three depths. The first position is usually enormous next to the rest (see the last row). That 'massive activation' is what turns the first token into the attention sink, and it is why the probe leaves position 0 alone when it pushes.
section 'trace' took 37.7s
| result | check | detail |
|---|---|---|
| PASS | attention weights of one head sum to 1 | sum = 1.000000 |
| PASS | layer hooks agree with the model's own hidden states | max relative error 0.00e+00 |
| PASS | layer output = input + attention + MLP (exact decomposition) | max relative gap 0.00e+00 (steer layer excluded) |
| PASS | logit-lens pipeline reproduces the model's own scores at the last layer | max diff 0.00e+00 |
| PASS | measured steering push equals strength x vector length | measured 8.8109 vs 8.8109 |
ALL CHECKS PASSED
total time 163.5s, peak memory 2842.9 MB