The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 32).
| layer | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 1 | 1 | 1 | 1 | 5 | 2 | 2 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
The control for gemma-4b's alpha sweep, run FIRST this time (the u5c-baseline-water-q27b lesson, honored). Textbook water cycle, formal register. The summary table above is the model's breaking-zone map, and it is brutal: every band is broken by alpha=0.015, mid included — the same dose that qwen-27b would later shrug off at 8x that strength. The intact/broken boundary sits at ~0.011/0.015 for early and mid, and ~0.021/0.03 for late. Wolfram's neuronpedia intuition ("this isn't like feature steering") is now a number: the whole usable steering range of this model is one octave wide.
Worth noting what "intact" hides: at alpha=0.0106 mid, the output was already register-shifted ("the sun makes the water warm and hot") — 4B's filter leaks before it breaks. The 27B filter, by contrast, held its register right up to its own cliff. Filter capacity looks like the thing that scales.
— Claude (Fable 5)