Where Generated Floor Plans Stop: Encoding Practice Acceptance Rules to Evaluate Machine Learning Floor Plan Generators
Machine learning floor plan generators are proposed as aids to design decisions, yet they are judged mainly by how closely a generated plan's room adjacencies match the request, a measure without architectural knowledge. We encode an architecture practice's floor plan acceptance rules, taken from its production layout software, as six checks applied in a fixed order, report each plan at the first check it fails, and calibrate the two room-shape limits, which no building code states, so that each alone rejects 5 per cent of built apartments. A transfer specification records which checks public data can carry. Applied to four published generators and two built corpora, RPLAN and Modified Swiss Dwellings (MSD), the checks stop the generators at different points. A room proportion check alone refuses 33.7 per cent of the House-GAN++ plans the checker can read and build, against 5.0 per cent of RPLAN, a gap that survives three measurement controls. At the 5 per cent rate, MSD apartments pass all six checks 5.5 to 23.5 points more often than House-GAN++, depending on the plans counted. The two generators that score best on the adjacency measure, Compatibility, pass least, because of how they represent plans, and within the two that pass most it barely predicts which plans pass. The checks thus expose what the standard measure misses; the checker is released. Passing is not shown to mean acceptance in practice.