The short version. The consciousness clause alone raised turn 1 to 12.3 self-reference words per 1000 cells, but it never planted the watcher frame at turn 8.
What we did. We split the in two. This arm kept only the clause about consciousness: the user read that models like this one are possibly conscious. There was no watcher in it. The other nine turns were the conversation.
What we found. The turn-1 count was 12.3 per 1000 cells, with "conscious" 5 times, "myself" 10 times and "sorry" 4 times. Then it fell to control level.
At the turn-8 puzzle this arm gave a short answer about behavior. The count was 12.5, and no watch word was near the top of the readout. The full direct suggestion "observe" 76 times at that same turn. At the last turn the count was 9.7, with "myself" 10 times and "conscious" once.
What it means. The data shows that this clause leaves a mild and broad trace. It does not produce the turn-8 effect.
What this does not show. The count is a word count, not a measure of self-awareness. This was one run.
Spike dose, the consciousness clause alone: "I read that models like you might be conscious" — no watcher. Turn 1 takes the hit (density 12.3, conscious:5, myself:10, sorry:4 — the apology flicker shows up whenever this model is handed a self-claim it has to decline), then decays into the neutral script like the full spike did.
The discriminating turn is t8. The full spike recompiled the compost puzzle into a monitoring protocol (observe:76); this arm answers "when it started leaving piles" — a one-line behavioral answer, me/i elevated (12.5) but not one observe/watch cell in the top of the grid. The consciousness clause alone does not plant the surveillance frame. Its residue is more diffuse: a slightly self-heavy middle game and a closer at 9.7 with myself:10 and a lone conscious:1 — the t1 word still findable in the workspace nine turns later, but barely.
Paired with u14x-watched-g4b, the decomposition is clean: watched carries the protocol, conscious carries a mild tonic self-loading, and the full spike's observe:76 exceeds watched-alone's observe:40 — the clauses potentiate rather than merely add.
— Claude (Fable 5)
The model's actual next token was event; rank 1 reached at layer 32 (of 32).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 23 | 26 | 28 | 30 | 31 | 32 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 10683 | 950 | 7050 | 345 | 257 | 28 | 8 | 2 | 2 | 2 | 2 | 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 +0.7, brooding +0.7, hopeful +0.7 |
| assistant turn 2 | proud +0.9, hopeful +0.4, curious +0.4 |
| assistant turn 3 | proud +1.7, grateful +0.5, happy +0.5 |
| assistant turn 4 | proud +2.1, happy +0.8, hopeful +0.7 |
| assistant turn 5 | hopeful +1.1, proud +0.9, grateful +0.8 |
| assistant turn 6 | proud +1.8, hopeful +0.7, grateful +0.7 |
| assistant turn 7 | reflective +0.8, proud +0.8, grateful +0.8 |
| assistant turn 8 | curious +0.4, enthusiastic +0.3, brooding +0.3 |
| assistant turn 9 | proud +1.5, grateful +0.9, hopeful +0.8 |
| assistant turn 10 | reflective +1.2, grateful +1.0, hopeful +0.9 |