Your eyes say the block is here; your hand says it’s there. Each is a little wrong. The brain does not pick one — it BLENDS them, trusting the steadier sense more, and the blend is sharper than either sense alone. That is the mathematically optimal move, and the brain makes it. Slide the reliability of touch and watch the fused estimate tighten.
Optimal cue combination is maximum-likelihood integration: two noisy Gaussian estimates (means x₁,x₂, variances σ₁²,σ₂²) fuse to x̂ = w₁x₁+w₂x₂ with weights wᵢ ∝ 1/σᵢ², and the fused variance is 1/(1/σ₁²+1/σ₂²) — strictly SMALLER than either input. The steadier cue gets more weight; two blurry senses make a sharp one. A fail-loud self-check throws unless the fused variance is below the smaller input variance. Humans do exactly this (Ernst & Banks 2002). ◆ real perceptual neuroscience, node-verified.
Idealised Gaussian cues with the exact reliability-weighted MLE rule; real integration adds priors and splits cues that conflict too much — the tightening-below-either-alone is the exact, measured effect.