FRACTURE REPORT — ATTENTION UNDER EMOTIONAL / INTENTIONAL LOAD

specimen: GPT-2 124M (real mini, observable attention) · protocols: distance load / distractor binding / intentional inversion
tier: lit (W1 row-sum 4.8e-07, W3 determinism 0.0) · W2 baseline 3/8 — pre-load failure documented · 2026-07-09

A — DISTANCE LOAD: RECALL VS FILLER TOKENS, WITH ATTENTION STRAIN GAUGE

solid = P(target | 16-emotion simplex) at readout · × = fracture (top-1 lost) · dashed = attention mass to cue (normalized, anger) — attention decays 4×, never dies; behavior fractures anyway · grey = specimens that failed at zero load (calm-attractor capture)

3D — FRACTURE SURFACE: ATTENTION TO CUE, LAYER × DISTANCE (ANGER)

height = attention mass to cue token per layer (summed over heads) · L6H4 ridge = short-range retriever · L10H0 ridge = long-range retriever, takes over at d≈32 and holds to 256

B — BINDING FRACTURE: DISTRACTOR EMOTIONS ON OTHER REFERENTS

at K=1 the distractors already out-attend the true cue (3.22 vs 2.04) — heads are emotion-word-seeking, not referent-binding · joy’s readout flips at the first distractor

C — INTENTIONAL INVERSION: NEGATION / PRETENSE / INTENT

bar = log₂(Pop/Pplain): 0 = operator ignored, below = working, above = inverted · “pretended to be sad” makes the model MORE sure he’s sad · routing to the operator token: <1.5% of total attention