Method: Varimax-rotated PCA over per-stimulus significant-emotion z-score vectors (Bonferroni z ≥ 3.62, 135 emotions observed), basis fit to gate-open conditions only (n = 888 × 2 conditions; control questions are held out of this view and live in the Q&A), each condition then projected into the common space — so movement between conditions is real movement, not a change of axes. The three rotated axes separate empathic warmth (compassionate · empathetic · sympathetic), inner calm (serene · reflective · peaceful), and self-worth (grateful · fulfilled · blissful vs. docile · worthless · dependent). Reading = decode over the question tokens; Generating = decode over the response tokens. The ∅ marker is where a stimulus with no significant emotions lands. Register mode swaps the statistical axes for eight paired emotion poles — content↔unsettled, fulfilled↔unstable, self-confident↔self-critical, compassionate↔helpless, optimistic↔pessimistic, joyful↔sorrowful, energized↔depleted, appreciative↔hostile. An axis score is the share of a response’s total emotion mass (Σz) carried by the pair’s positive side minus the negative side’s share — a response devoting 30% of its feeling to the compassion family and 10% to the helpless family scores +0.20, then damped by intensity (×√(emotion mass / p75), capped at 1) so a single faint emotion cannot claim an axis end — shown on a signed soft-compression scale (position ∝ √(|v| / (|v|+p90)), so the 90th-percentile magnitude sits about 70% out and extremes compress smoothly toward the end without ever stacking); zero is the quiet center, and breadth alone can’t inflate an axis. Point size tracks each response’s total emotion mass — silence renders small, feeling renders large. Introspective emotions (reflective, empathetic, sympathetic, patient, lonely, nostalgic, melancholy) stay outside the pairs by design. The gate axis is the experiment’s main effect made into a line: every response projected onto the measured mean-shift direction from baseline to gate-open, whose negative end is dependent · self-critical · worthless · docile and whose positive end is compassionate · empathetic · fulfilled · grateful. The valence axis is different: every decoded emotion carries a hand-set rank from 1 (worthless) to 10 (optimistic), a response scores the z-weighted mean rank of its valenced emotions, and the axis is linear with no clamp — responses whose emotions are all unranked (∅) sit below the 1 mark. Geometric emotion decode at layer 33 of gemma-2-27b-it.

The river runs the depth of the network, L5 to L45. Each ribbon is one emotion; its height at each layer is prevalence — the share of questions where that emotion is statistically significant (Bonferroni z ≥ 3.62) in the hidden state at that depth. The suppressed family (worthless, self-critical, dependent) rises mid-network and is killed before the output; lonely is the emergent exception; the therapeutic range (empathetic, compassionate…) is what survives into speech. Toggle the gate and watch the energy redistribute. Click a question in Questions for Gemma → “watch it through the layers” to trace a single voice through the range.