the tensor language. Tensor<[D1, D2, ..., Dk]> where dimensions are compile-time constants. shape mismatches are compile errors
| Op | Action |
|---|---|
matmul(A, B) |
Matrix multiplication |
einsum(spec, ...) |
Einstein summation |
reshape(T, shape) |
Reshape tensor |
broadcast(T, dims) |
Broadcast to higher dimensions |
transpose(T, perm) |
Permute dimensions |
reduce(T, axis, op) |
Reduce along axis |
conv2d(X, K) |
2D convolution |
softmax(T, axis) |
Softmax activation |
dense and sparse. SpMV over sparse adjacency matrices = graph computation (focus vector π, tri-kernel diffusion). quantized inference (int4, int8 matmul) = contraction over Z/2ⁿ. full-precision neural layers = contraction over F_p. Ten is the compute engine for both the cybergraph and AI inference. CYBERRANK is literally repeated matmul. compiles to Tri
see cyb/languages for the complete language set. see cyb/multiproof for the proving architecture