COPY SUPPRESSION

Reads a token it attends to and writes to the output logits: a copy head raises that token’s probability, a copy-suppression head lowers it. Suppression is a self-correction brake that calms overconfident repetition. Exact logit-delta demo.

AMBER

2D · Quantitative Channel

CANVAS 2D

demo: sine — replace in draw2D()

CANVAS 2D UNAVAILABLE

3D · Spatial Channel

SOFT-GL · NO GPU · NO LIB

DRAG ROTATE · WHEEL ZOOM

demo: lattice + orbit — replace in build3D()/draw3D()

RENDER FAILED —

Notes

Copy & copy-suppression heads. Both read the token they attend to and write to the output logits — a copy head raises that token’s probability (reinforcing what it sees), a copy-suppression head lowers it. Copy-suppression is a well-documented calibration mechanism: it demotes tokens the model is about to over-repeat, acting as a brake against degenerate repetition. Amber tier: the sign-of-logit-delta mechanism is exact; magnitudes are illustrative.