The model likely does know what a pelican looks like; the hard part is translating that into ordered SVG path commands. It's really testing spatial and vector composition, not biological recall.