PREVIOUS TOKEN

The simplest useful head: each position attends to the one right before it. Modest alone, but it is the first stage of the induction circuit — the look-back that induction copies from. Constructed to isolate the pattern (BERT-tiny’s is weak).

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

Previous-token head. Attends from each position to its immediate predecessor — a pure offset−1 diagonal. Trivial in isolation, but it is the load-bearing first half of the two-head induction circuit: the previous-token head does the looking-back, the induction head does the copying. Amber tier: pattern constructed to isolate it cleanly; in BERT-tiny the real subdiagonal mass is only ~5%.