Pools.
Large language models fail at precise design tasks (architectural plans, stairs, dimensioned objects) in four recurring ways: outputs vary across runs, separately generated views disagree on shared dimensions, arithmetic errors pass silently as plausible results, and enumerating coordinates as output tokens is costly and error-prone. The Combination Method addresses these structurally by separating roles: the language model only gathers information and explains, all form is computed deterministically from a human-approved condition set, and two human gates fix direction. Each design condition is represented as a smooth spatial field (a Gaussian centered at its preferred location); conditions are superposed—additively for graded performance, multiplicatively for inviolable constraints—into a probability landscape from which a precise form is extracted ("collapsed"). The cut threshold is not chosen by hand but inverted from the target floor area, guaranteeing a unique, repeatable value; the resulting outline is snapped to a construction grid and normalized into long, buildable walls. The method is recursive: a confirmed form at one level (building mass) becomes the container for the next (rooms, elements, details). We report deterministic reproducibility (bit-identical after the condition set is frozen), a 40× reduction in search cost via coarse-to-fine ranking, and successful single-level demonstrations across mass, room layout, elements, and detail. A key refinement replaces per-variable correction with backpropagation-based joint optimization over all parameters simultaneously, eliminating post-hoc topology repair. The method's principal value is converting silent failure into detectable failure while keeping the entire post-freeze pipeline exactly reproducible.