The nesting. Outer cube is the container. Inside it, the cubit: eight planes, one per attention head, each a full 13×13 cross-section spanning the whole width and height, spaced evenly down the cube's length. Head index is the depth axis, queries are rows, keys are columns. Voxel size is the attention weight; the hollow upper triangle you can see through every plane is the causal mask.
The corpus core. At the centre, where the used to sit, is the training corpus — the
27 glyphs of the template alphabet packed into a small inner cube. This is what the bigram counts and the 8-dim
embeddings behind every one of those planes were computed from. Putting it at the origin is not ornamental: the
planes are downstream of it, and if the core changes every plane changes.
The four arms. One per quadrant of the y–z cross-section, all joined at 0,0,0. Each
runs from the core out toward its corner and carries an opposed cone pair pointing into a marker node —
inward pull from the corpus, outward pull from the corner, meeting where they cancel. Same
<—> motif used for the resistance edges earlier, here read as an attractor rather than a
standoff.
AMBER on the arms. The attention tensor, the mask, the embeddings and the corpus are all real and verified. The arms are a spatial claim about where balance sits, not a measured quantity — there is no dynamical system here computing a fixed point, and the marker positions are chosen, not solved.