TOROIDAL AUDIT OF THE LOGIT LENS

ring φ = magnetic affect compass (N arousal+ / S arousal− / E valence+ / W valence−) · tube θ = commitment depth of the logit lens
tier: audit lit / affect coords bri · GPT-2 124M · 2026-07-09 · bridge-burners

3D — MAGNETIC TORUS · KURAMOTO-DIPOLE CLUSTERING, LIVE

positive-valence emotion  negative-valence  audit miss (model’s final pick ≠ target; no commitment assigned)  · dot size ∝ |arousal| · outer equator = committed early, wrapped inward = committed late · needles settle under field torque + similarity coupling

2D — THE AUDIT: WHERE THE LOGIT LENS LIES ABOUT FEELINGS

top: mean KL(final ‖ LLk) over the 16-emotion simplex (10 committed prompts) + fraction committed · bottom: per-emotion commitment ladder · misses listed with the model’s actual verdict