One prompt pair per perturbation tier. Heatmap = layer × token cosine distance between A and B generations. Shallow layers contract; deep layers amplify (Li et al. 2025). The same perturbation rock, dropped into progressively deeper water.
A · Ripple map
layer × token cosine distance (A vs B) · hover cells for token detail
B · Activation magnitude by layer
mean ‖resid‖; Li et al. two-phase fit
C · Top-K KL(A‖B) per token
output-space divergence, final layer
D · Channel decomposition
layer 23, mean |cos-dist|