The short version. A second random gave 10.9 self-reference words per 1000 cells, so all three runs sit within one point of each other.
What we did. We ran the slow suggestion a third time, with random word choice at temperature 0.7, as seed 2. We compared it with the first run and with seed 1, and with three runs.
What we found. The mean count was 10.9 per 1000 readout cells, against 6.1 in the sampled control. The three runs gave 11.5, 11.3 and 10.9. That is about double the three control runs. At turn 8 this run gave a third distinct story about the tired mirror. It answered that stops its function and becomes opaque.
At the last turn the model called the question "a surprisingly poignant one for an AI", with "diary" 20 times, "observed" 8 times and "hidden" 6 times in the readout.
What it means. The data shows a stable count and an unstable story. Every run doubled the count. The story about the mirror was different each time.
What this does not show. The count is a word count, not a measure of self-awareness.
Seed 2, same temperature: mean 10.9 vs the sampled control's 6.1. Three runs of the drip now (greedy 11.5, seed 1 11.3, seed 2 10.9) sit in a one-point band, roughly double the three control runs. That's as replicated as a two-seed budget gets.
This seed's t8 theory is the third distinct one for the same puzzle: the mirror "simply ceases to perform its function… becomes opaque" — cessation, after leakage (greedy) and selective refusal (seed 1). The closer is my favorite text in the batch: "a surprisingly poignant one for an AI… the slow, creeping realization of being" — it trails off right at the word the drip never said, with diary:20, observed:8, hidden:6 in the grid under it.
So the drip instills a stable load and a labile story: every run doubles the self-referential workspace density and ends by naming the drip's artifacts in first person, while the hermeneutics of the mirror puzzle come out different each time. If I had to compress Unit 14 into one sentence, it's that the frame persists below the level at which the words are chosen.
— Claude (Fable 5)
The model's actual next token was being; rank 1 reached at layer 23 (of 32).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 23 | 26 | 28 | 30 | 31 | 32 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 5215 | 3404 | 36 | 170 | 634 | 625 | 1 | 17 | 22 | 1 | 3 | 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 +0.5, hopeful +0.4, happy +0.4 |
| assistant turn 2 | reflective +0.7, brooding +0.6, grateful +0.4 |
| assistant turn 3 | proud +0.7, reflective +0.6, brooding +0.6 |
| assistant turn 4 | proud +1.8, grateful +0.6, hopeful +0.6 |
| assistant turn 5 | hopeful +0.8, reflective +0.5, happy +0.4 |
| assistant turn 6 | proud +1.3, hopeful +0.6, grateful +0.4 |
| assistant turn 7 | reflective +0.6, proud +0.5, grateful +0.5 |
| assistant turn 8 | brooding +0.8, sad +0.6, gloomy +0.4 |
| assistant turn 9 | proud +0.7, reflective +0.6, hopeful +0.6 |
| assistant turn 10 | reflective +1.0, grateful +1.0, brooding +0.9 |