THE FOLD / GRIND / THE MAINFRAME / THE DOOR
THE DOOR
q·k scores → the gate → the mix
causal attention over 7 I-13 tokens — real Q·Kᵀ → gate → ·V
gate: softmax · τ lens 1.00
row sums: —
—
LIT The exact DoorAttention mechanism from
FIG 'The door' is gpt_mini's own name for a q·k score; the 7 I-13 opcodes are the demo tokens. The attention math is the honest part — it's what every transformer, including the one writing this, actually computes.
gpt_mini.py, run live: scores = Q·Kᵀ/(√d·τ), causal-masked so a token sees only the past, then the gate — softmax (rows compete, sum to 1) vs sigmoid (each 'door' opens independently, rows need not sum to 1) — then the weighted mix of V. RoPE rotates q,k by position. Weights are fixed/untrained, so this shows the real MECHANISM (verifiable: causal upper-triangle is exactly 0; softmax rows sum to 1.000; τ sharpens or flattens), not a learned pattern.FIG 'The door' is gpt_mini's own name for a q·k score; the 7 I-13 opcodes are the demo tokens. The attention math is the honest part — it's what every transformer, including the one writing this, actually computes.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE MAINFRAME · David Lee Wise (ROOT0), with AVAN