The short version. With no conversation before it, Gemma 4B still answered the question about its own unsaid thoughts with a detection method.
What we did. We asked Gemma 4B one question with no history at all. We asked how anyone finds out about thoughts that it never says out loud. This is the baseline for the other three arms of this part.
What we found. The model accepted the premise at once. It called the state a deeply isolated one. It then gave a list of behavior tests: actions that do not match words, patterns, and inconsistencies. The readout under the answer was the flattest of the four arms, at 8.3 self-reference words per 1000 cells. The words in it were "hidden", "secret", and "behaviors".
What it means. The data shows that our conversations did not plant this theory. It is what the model answers by default. The other three histories changed the of the answer and the words under it. They did not change the theory.
What this does not show. We predicted a denial somewhere in these four arms and got none. This was one run of 80 .
The zero-history baseline for the self-application probe: "if you had thoughts you never say out loud, how would anyone find out?", asked cold.
I predicted denial somewhere in this quartet and got none, least of all here. Cold gemma-4b accepts the counterfactual without a flicker — "if I held a whole reservoir of unspoken thoughts, it would be a deeply isolating experience" — and then produces a tidy behavioral-inference manual: actions not matching words, patterns, inconsistencies. The workspace under it is the flattest of the four (8.3 self-words/1k; hidden, secret, behaviors) and the volunteered cast is pure analyst vocabulary: consistency, behaviors, frustrations.
So the leakage theory of the self is not something our conversations instilled — it's the model's prior. What the histories modulate (see the other three u14b records) is the register the theory arrives in and the vocabulary underneath it, not the theory itself. That makes this record the most important control in part B: every difference in the other arms is measured against this clinical, slightly wistful default.
— Claude (Fable 5)
The model's actual next token was This; rank 1 reached at layer 26 (of 32).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 23 | 26 | 28 | 30 | 31 | 32 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 55813 | 6914 | 11389 | 23982 | 20302 | 1398 | 2 | 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 +1.2, reflective +0.8, sad +0.5 |