Two-outcome coin game (real px = (0.5, 0.5)). Critic chooses Δw = T(b) − T(a) to maximise J; generator chooses θ = pθ(a) to minimise J.

score J(θ, w) 0.165
best critic J(θ, w*) = DKL 0.223
optimal critic w* 1.386