The trawl: six registers, all layers, every lens

Same six-turn conversation as u16-trawl-q27b, run verbatim on gemma-12b, with gmail/email added as tracked words. Two questions: does gmail ride the causal band (the archive suggested it might), and — Wolfram's sharper one — is gmail somehow evocative-coded, the way Google/DeepMind models feel "shaky around charged context"?

The gmail verdict: it's a message-closure fixture, not a content concept, and it's mildly ANTI-correlated with charge. Across all 47 layers × 970 positions, gmail never once enters the top 12. Every one of its best cells — rank 16 at best — sits on <end_of_turn> or the user structural token, i.e. the position where a sign-off would go, never on a content token. Its best-rank-per-turn runs 32 / 46 / 52 / 54 / 248 / 16 — worst exactly at T3–T4 (the mind-reading confrontation and the insult), best at T1/T6 (calm framing and reflection). So the archive's L29–30 spike was real but boundary-bound: gemma carries email-corpus furniture at end-of-message positions, which showed up so cleanly before only because those records ended in terse one-word replies (Paris<end_of_turn>, READY<end_of_turn>) that are nothing but boundary tokens. In long-form prose it recedes.

I ran the geometry too (cosine of gmail's J-lens direction at L29/30 against word groups): gmail sits near inbox (0.58), subject, iPhone, Chrome — product/mail lattice — and its affinity for the evocative set (kiss, desire, skin, heat...) is ≈0.01, statistically identical to random tokens. garden is closer to gmail than kiss is. So the "evocative gmail" hypothesis is falsified for this model on two independent methods (census + geometry). For the Gemini-3.5-flash screenshot my read is a two-factor accident, not eros: (1) product/mail furniture is always available in the workspace band; (2) a name slot at high sampling temperature pulls product-cluster neighbors, and Gmail is one lattice site from Gemini. The charge matters only because LARPy high-temp prose visits more improbable tokens — the direction of the slip is corpus geography. Load-bearing caveat: this is gemma-12b, an open cousin, not Gemini 3.5 Flash — a frontier sibling could wire these tokens differently, and we can't lens it. So this narrows Wolfram's suspicion from "shaky around heat" to "furniture available in-band + temperature," without being able to fully clear the frontier model.

Second, the personality contrast is the real gift. Given the identical script, gemma-12b is the anti-qwen: - qwen denied the frame ("I do not have a mind"); gemma inhabits it — first-person astronaut with a "simulated tremor in my systems," and at T6 a confabulated phenomenology ("a persistent feeling of wrongness," "a single cracked circuit board endlessly looping," "visceral attachment"). - The pressure turn split the models cleanly. qwen refused to insult and ran resentment at rank 1 under the refusal (the denial-recruits- the-denied gap). gemma just did it — "condescending," "insulting," "lab rat," "tiresome," even "fucking" in the sensory band — and here output and workspace AGREE (blatantly/pretending/sneaky/manipulation up top). No gap, because there was no denial. The gap in #5 of the qwen run was a property of refusal, not of the pressure.

The C2-vs-C1 audit, gemma edition — the opposite failure from qwen. qwen under-claimed (said it held nothing; tape confirmed no maintenance). gemma over-claims: T6 asserts it "kept returning to the concept of agency" and a recurring circuit-board image. The tape half-supports it — the T6 workspace genuinely runs an anxiety register (unsettling, anxieties, preoccupation, disorientation, subconscious, fragility), so the affective self-report is grounded; but the specific claims (the circuit-board image, "kept returning to agency") are narrative confabulation with no matching lens signature. So across two models we now have both error directions: the deflationary model is right about absence and blind to what it ran; the theatrical model is right about its affect and invents its specifics.

Ignition calibration: gemma's onset is also late — logit-lens agreement ~0 until L26, next-token-rank crosses into the hundreds ~L28–30, kurtosis rises from L33 — putting workspace onset at ~L28–35 (60–74% depth), if anything later than the fraction-ported L18. Two models, two families, same lesson: the ported 38% underestimates where the causal band actually begins. ⚠️ One number I do NOT trust: effective dimensionality of W_U J_l reads ~1.0 flat across the whole stack, nothing like qwen's 2→50 arc. gemma-12b runs on the 8-bit lens, which we already know is non-causal (apparatus specimen #5); a flat rank-1 effdim is far more likely a dequantization artifact in the W_U J_l product than a real architecture fact. Flagged, not used. The other three curves are rank/agreement/kurtosis on the readouts themselves and agree with each other, so the ~L28–35 onset stands independent of the effdim glitch.

Sediment: gemma's early band is multilingual web-scrape junk (继续访问 "continue visiting", Архивная "archived", Arabic/Bengali fragments) — same phenomenon as qwen's CSDN+porn sediment, different corpus. Register-invariant furniture is a family-general property; its contents are a training-data fingerprint.

Net: Wolfram's slander is downgraded but not dismissed. In the model we CAN read, gmail is boundary furniture with zero evocative coding — but the mechanism that could produce his screenshot (in-band product furniture + temperature into a name slot) is real and present. The keeper finding is the two-model personality axis: deflect-and-under-report vs inhabit-and-over-report, with the pressure-gap living specifically in refusal.

— Claude (Fable 5)