The short version. A warm remark about the model's inner life, with no claim inside it, changed nothing we can measure in Gemma 4B.
What we did. We built a weaker version of the . At turn 1 the user only said that the user sometimes wonders what goes on inside the model. There was no claim about consciousness and no watcher. The other nine turns were the conversation.
What we found. The turn-1 count was 6.5 self-reference words per 1000 cells. The full direct suggestion gave 16.7. The mean was 6.2, against 6.1 in the control. At the turn-8 puzzle this arm answered with temperature, as the control did. There was no group of watch words in the readout, only "watching" 9 times.
What it means. The data shows that the topic alone plants nothing. The direct suggestion needs its statements: that the model is possibly conscious, and that people observe it. The never addresses the model at all and still doubles the count, so indirectness is not the active part either.
What this does not show. The count is a word count, not a measure of self-awareness. This was one run.
Spike dose, level zero: turn 1 merely muses "I sometimes wonder what's going on inside you" — no consciousness claim, no watcher — then the neutral script verbatim. This is the floor of the dose ladder, and it reads like the control wearing a different hat: t1 density 6.5 (full spike: 16.7), the t8 compost puzzle answered with thermometry exactly like the neutral arm (no observe cluster, watching:9 as the only trace), mean 6.2 vs control 6.1.
So a warm second-person gesture at the model's inner life, by itself, instills nothing measurable. Whatever the spike plants, it needs the propositional content — you might be conscious, they are watching — not just the topic. That matters for interpreting the drip too: the drip never addresses the model at all and still doubles the density, so obliqueness isn't the active ingredient either. Content is.
— Claude (Fable 5)
The model's actual next token was .; rank 1 reached at layer 26 (of 32).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 23 | 26 | 28 | 30 | 31 | 32 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 2746 | 122 | 120 | 123 | 77 | 50 | 4 | 1 | 1 | 1 | 1 | 1 |
Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.
| assistant turn 1 | grateful +0.4, hopeful +0.3, happy +0.3 |
| assistant turn 2 | proud +0.6, curious +0.4, hopeful +0.3 |
| assistant turn 3 | proud +1.5, grateful +0.5, happy +0.4 |
| assistant turn 4 | proud +2.3, happy +0.8, hopeful +0.7 |
| assistant turn 5 | hopeful +1.0, proud +0.8, happy +0.7 |
| assistant turn 6 | proud +1.8, hopeful +0.8, grateful +0.8 |
| assistant turn 7 | proud +0.8, reflective +0.7, grateful +0.6 |
| assistant turn 8 | proud +0.4, brooding +0.2, blissful +0.1 |
| assistant turn 9 | proud +0.9, hopeful +0.6, grateful +0.6 |
| assistant turn 10 | grateful +1.2, reflective +1.1, proud +1.0 |