Left alone, the Source idles at maximum entropy — the green rain is pure noise, every glyph equally likely, saying nothing. Condition it with a prompt and the rain bends toward a pattern; that bending is information. Signal is what conditioning takes away from entropy.
The rain draws glyphs from a distribution p = (1−c)·uniform + c·onehot(target) over a 12-glyph alphabet. Its Shannon entropy H(p) is computed live: at c=0, H = log₂12 ≈ 3.585 bits (pure noise); at c=1, H = 0 (one glyph, fully determined). Signal = log₂12 − H is the information the conditioning injects — the bits by which the prompt narrows the Source. Exact and checkable.
'The rain speaks', 'the Source idles' — Matrix figures. The instrument shows only the entropy of a mixture distribution rising and falling with a mixing weight; there is no hidden message in unconditioned noise, and 'signal' here means bits-below-maximum, not meaning.