PBF FRACTAL RGB · WEBGL2

The stack is 3 separate canvases, one per transformer layer (L0 bottom → L2 top), each split into 7 stage columns. Each tile is a PBF12 fractal rendered in the arithmetic that implements that stage.

Col 0 — Embed: Mandelbrot baseline; per-layer seed rotation mimics layer-specific feature mixing.
Col 1 — Attn (LNS): log-domain KQV·α attention. Q=Mandelbrot, K=Julia, V=Burning, α=σ(Q·K/N²−½). The crate's default FPU path.
Col 2 — Attn (Linear): same KQV·α but in linear SBP₁₂ (Q₈, saturating). Overflow-legal mid-compute.
Col 3 — Attn (Polar): Mandelbrot orbit in polar form — R = phase hue via SBP atan2, G = log|z|, B = smooth iter.
Col 4 — Attn (Tapered): Mandelbrot magnitude read directly from the encoded PBF₁₂ geometric level — the denormal taper keeps very-small |z| distinguishable where IEEE subnormals would flush to zero (P32).
Col 5 — FFN (Newton): z³−1 Newton iteration as the non-linear feed-forward. No black pockets at f'(z)=0 (P27 / P28).
Col 6 — Residual: heuristic blend of Col 0 ⊕ Col 1 ⊕ Col 5 as the layer's output, feeding the next row.

Per-row Julia seed rotates +0.5 rad/layer so each row visibly attends to a different feature direction. Zoom tightens 0.85×/layer to simulate deepening feature hierarchy.

Each base sets per-LSB step = ln(base)/2^M, where M is the mantissa bits left after picking E = ⌈log₂(rings)⌉ rings to span [1e-3, 1e3]. Bases where log_b(10⁶) drops below a power-of-2 boundary (silver, e, π, bronze, copper, nickel, 2π, π², 10) save an exponent bit and reinvest it as mantissa. Hover for full names.