The symbiote: machine coldness lent human feeling, human logic lent machine insight. Two estimators of the same truth, each blind where the other sees — fused, they beat either alone. Slide the blend and the complementarity and watch the fusion land inside both errors.
Two unbiased estimators with variances Vₘ, Vₕ and low error-correlation combine by inverse-variance weighting: the optimum w* = Vₕ/(Vₘ+Vₕ) gives fused variance Vₘ·Vₕ/(Vₘ+Vₕ) — strictly below both when their errors are complementary. That is a real statistical fact (the same math as sensor fusion / ensembling), and the readout shows the fused error dropping under min(Vₘ,Vₕ) exactly as complementarity rises. Snap-to-optimal puts w at w* and the reduction is maximal.
'Coldness', 'feeling', 'insight' are labels for the two channels, not states in the machine — the empath is a weighted fusion, not a heart. When COMPLEMENTARITY is low (errors correlated) the fusion barely helps, and if one source is biased the blend inherits the bias: the symbiote is only as good as the independence of what it fuses.