Most language models learn how tokens relate by brute force: millions of parameters per layer, just for the metric. JAM skips that. It uses D_J = j², a 4×4 integer matrix with entries {-1, 0, 1}. Five additions. Zero multiplications. Period 5. Bounded forever. JAM does not learn the geometry. The geometry learns JAM.

Two thirds of layers are motor layers. They cost almost nothing. The rest is ordinary attention with a learned V projection. The whole thing runs as INT4 in your browser, on one CPU thread, with no server involved. You are the server.

J = 1 + j² · D_J = j² · entries {-1, 0, 1} · period 5 · det = 1
67% motor layers (free) · INT4 quantized · pure WebAssembly · no server