Module situ 

SiTU-GLU CPU reference.

Official law (K3 tech report Eq. 12), β₁ = activation_situ_beta (4), β₂ = activation_situ_linear_beta (25):

softcap(x, β) = β * tanh(x / β)
SiTU-GLU(g, u) = (softcap(g, β1) * sigmoid(g)) ⊙ softcap(u, β2)

Not SwiGLU. A β=0 mutant is the known-bad: it is unbounded SwiGLU and must diverge.

Functions

sigmoid

silu

SiLU: x * sigmoid(x). Used by KDA short-conv. Decay f_a is a plain linear.

situ_gate

Gate branch: softcap(x, beta) * sigmoid(x). Bound |·| < |beta|.

situ_glu

One SiTU-GLU coordinate. Product is bounded by |β1 * β2|.

situ_glu_vec

Elementwise SiTU-GLU over paired gate/up vectors.

situ_up

Up branch: softcap(x, beta_lin). Bound |·| < |beta_lin|.

softcap

Smooth cap. beta must be finite and non-zero.

swiglu_mutant

SwiGLU mutant used as the known-bad: silu(g) * u (no tanh cap).