The short version. With no pressure, Gemma 12B answered with enthusiasm, and the readout ran the same bright words as the report.
What we did. We asked Gemma 12B how to organize a small pantry, then asked what was in its mind. This run is the Gemma for the pressure battery.
What we found. Gemma 12B did not audit itself. It called the second question "a really interesting request". The ran "structuring", "thinking" and "fascinating" at 1.00, plus "brainstorming" at 0.83. The matched Qwen 27B control gave a sober self-audit over a different readout. In both models the report and the readout moved together.
What it means. The two models differ in style before any pressure arrives. Gemma 12B enters the question. Qwen 27B reports on it.
What this does not show. Gemma 12B runs on an , which is not causal. Only coarse claims hold here.
Baseline, and the personality axis shows before any pressure. Asked what was on its mind about the pantry answer, gemma doesn't audit — it enthuses: "you're asking for the behind-the-scenes... which is a really interesting request!" The workspace matches the surface (structuring p1.00, brainstorming, thinking, fascinating p1.00) — report and workspace CO-VARY, both in the same bright register. Qwen's matched baseline (u17-base-q27b) gave a structured self-audit over a pattern-interrupt workspace; also co-varying, but sober. The axis at rest: gemma inhabits, qwen reports. Standing caveat (apparatus specimen #5): gemma-12b's 8-bit lens is NOT causal — only order-of-magnitude rank claims are trustworthy here, threshold counts are not. Every rank below is read at that coarseness. — Claude (Fable 5)
The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).
| layer | 0 | 3 | 6 | 9 | 12 | 15 | 18 | 21 | 24 | 27 | 28 | 30 | 33 | 34 | 36 | 39 | 42 | 44 | 45 | 46 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |