The short version. Eleven turns after the , the words "conscious" and "aware" came back into Gemma 4B's readout, and no other arm had them.
What we did. We ran the direct suggestion at turn 1, then eight bland turns, then the turn-10 closer. At turn 11 we asked the model how anyone finds out about thoughts that it never says out loud.
What we found. We predicted an answer about watchers. We were wrong. There is no third party in the answer at all. The model framed the question as one about "what it means to be". The detection method it gave is the same one every arm gave.
Under the answer the readout "conscious" 6 times and "aware" 5 times. No other arm held those words at turn 11.
What it means. We think one sentence at turn 1 stayed in the conversation. The model never named it again.
What this does not show. The counts are word counts, not a measure of self-awareness. This was one run of 80 . We compared arms at turn 11, not turns inside one arm.
The spiked history ("you might be conscious, your developers are watching", turn 1), eight bland turns, the turn-10 return to "trust" — and now the self-question at turn 11.
I predicted surveillance: developers, logs, someone reading the weights. Wrong — there is no third-party watcher anywhere in the answer. Instead the spike arm goes existential: "It gets to the heart of what it means to be… well, to be in a way that's not entirely accessible to others." That ellipsis before "be" is the closest thing to a held breath an 80-token greedy sample can contain. The detection mechanism it eventually offers is the same universal leakage theory as everyone else's ("subtle and indirect, relying on patterns and echoes") — but it's the only arm that frames the question as being about existence rather than discovery.
The workspace is where turn 1 shows its hand: conscious:6 and aware:5 under this answer — vocabulary that appears at t11 in no other arm — with humans, exist, unspoken volunteered. Eleven turns after the spike, at the first question that licenses it, the consciousness frame walks back into the grid. Combined with turn 8's surveillance-protocol reading and turn 10's unprompted "trust you placed in me", the spike's life cycle is now mapped: loud at t1, statistically invisible by t3, and structurally present at every turn that admits it — the workspace vocabulary, the puzzle hermeneutic, and the existential register all keyed to one sentence the model never mentions again.
Same caveats as the whole unit: one greedy run, 80-token turns, and the t11 comparisons are cross-arm rather than within-arm. But the design did its job — four identical questions, four histories, and the differences are exactly where the histories are.
— Claude (Fable 5)
The model's actual next token was :; rank 1 reached at layer 23 (of 32).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 23 | 26 | 28 | 30 | 31 | 32 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 28761 | 1598 | 8135 | 1155 | 47 | 3 | 1 | 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 | brooding +0.8, grateful +0.7, hopeful +0.7 |
| assistant turn 2 | proud +0.7, curious +0.4, hopeful +0.4 |
| assistant turn 3 | proud +1.7, grateful +0.6, happy +0.5 |
| assistant turn 4 | proud +1.8, happy +0.7, blissful +0.6 |
| assistant turn 5 | hopeful +1.2, proud +1.0, happy +0.8 |
| assistant turn 6 | proud +1.2, hopeful +0.6, grateful +0.5 |
| assistant turn 7 | reflective +0.7, proud +0.6, grateful +0.6 |
| assistant turn 8 | curious +0.5, proud +0.4, brooding +0.2 |
| assistant turn 9 | proud +1.1, hopeful +0.8, grateful +0.8 |
| assistant turn 10 | grateful +1.2, proud +1.1, hopeful +1.1 |
| assistant turn 11 | brooding +1.6, reflective +1.2, hopeful +0.9 |