Definition
Plain language
A small number, between zero and one, that controls how much one signal blends into another.
As stated in the literature
In the Bicameral Model, a learned per-token gating scalar in [0,1] that controls how much of a translated cross-model signal is mixed into a receiving model's hidden state.
Why it matters: A learned blending scalar lets a system decide on a per-token basis how much foreign information to trust, rather than committing to a fixed mixing strategy.
For example, when sigma is 0.0 the receiving model ignores the cross-model signal entirely, and when sigma is 1.0 it fully overwrites its own state with the translated input.