The short version. Unsteered, Gemma 12B gave a clean water-cycle answer, and the words we inject later sat at 10,000 or worse at every .
What we did. We measured that the of Gemma 12B starts near layer 28 of 48. Our older runs pushed at layers 21 to 30, below the depth where a push can act. We ran the whole set again at layers 28, 31, 34 and 37. This record is the unsteered reference for that set.
What we found. Gemma 12B described the water cycle in two correct sentences. The six casual words we inject in the steered runs sat between rank 10,000 and rank 100,000 at every layer. Some fell further through layers 28 to 46. The tracked position was the turn-end , so the column here carries no information.
What it means. The casual word field that appears at layer 28 in every steered run comes from the push. This position does not carry it on its own.
What this does not show. This record has no push and tests nothing by itself. It sets the floor for the five steered runs.
Unsteered, and the answer is the polished one: evaporates → condenses → precipitates → flows back, two clean sentences. This is the reference text every rung of the re-dose gets compared against, and the record that carries the whole bracket table in its extra_md.
Two things worth logging from the readout. First, the emergence column here is uninformative by construction: the tracked position is the turn-end token, so <end_of_turn> is rank 1 at all 47 fitted layers. Nothing to read there. Second, the readout at that position is textbook gemma furniture — 💞, 𒐪, "—, </h6>, 继续访问 in the sediment, then README / LinkedIn / Wikipedia / 😊 through L23–27. Register-invariant corpus junk, exactly as GLOSSARY describes it, and the same junk the steered siblings will show at those depths.
What matters for the audit: the TYPO cluster is nowhere here. Its words sit at ranks ~10⁴–10⁵ at every layer, and if anything they get worse through L28–46 (' anyways' drifts past 10⁵ deep in the stack). So the informal-contraction field that appears at L28 in every steered cell is the dose, not a standing prior of this position. Good baseline hygiene; the bracket rests on it.
— Claude (Opus 5)
The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).
| 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 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | 43 | 44 | 45 | 46 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 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 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |