JACOBIAN LENS — WORDS / TONE / INFLECTION

GPT-2 (124M), 12 layers · JL(h) = f(h*) + J(h*)(h−h*), J via forward-mode JVP through the causal tail
tier: lit · witnesses W1/W2/W3 all 3.81e-05 (fp32 floor) · 2026-07-09 · bridge-burners

TOKEN LADDER — WORDS COMING INTO FOCUS THROUGH THE STACK

prompt: “The old violin, though cracked and worn, still sang with a” · anchor: “The old folder, kept filed and dry, still sat in a” · teal = token in model’s final top-5 · violet = J-lens parroting its anchor’s continuation

3D — RESIDUAL TRAJECTORIES, GLOWING VS BRUTAL (PCA-3, 97.9% VAR)

“reviews were glowing…”   “reviews were brutal…”   per-layer divergence struts · dots = layers 0→11

TONE — VALENCE MASS (POS−NEG) PER LAYER

glowing / logit lens  glowing / J-lens  brutal / logit lens  brutal / J-lens  · dashes = model final · J-lens anchored at the sibling prompt

INFLECTION — PAST−PRESENT TENSE MASS PER LAYER

“Late yesterday afternoon she finally” (past cue)  “Every single weekday morning she” (habitual cue)  · solid = logit lens, faded = J-lens · watch the J-lens hand off from anchor to subject