A low sampled loss is promising. A large box leaves more room for improvement.
Score: ℓᵢ(K) = f(cᵢ) − K δᵢ, δᵢ = half-diagonal of box i.
Red outlines: potentially optimal for some K > 0. Filled red: best score at the displayed K. Grey: dominated here. DIRECT refines all potentially optimal boxes, not just the one selected at one fixed K.
Same five-box partition as the original first-iteration figure; illustrative center losses, not its underlying objective. The selection also requires ℓᵢ(K) ≤ bₖ − εᴰ|bₖ|; here bₖ = 3.9 and εᴰ = 10⁻⁴. Point positions within a size class are slightly offset for legibility.