Each model receives one prompt and must generate a complete, working HTML file. No multi-turn conversation, no tool use -- just raw code generation ability. Tests span four categories:

• Apps -- Interactive web applications (kanban boards, spreadsheets, calculators)
• Games -- Playable browser games (platformers, puzzles, simulations)
• JavaScript -- Complex JS library usage (charting, 3D, audio synthesis)
• Visual -- CSS art, animations, and data visualizations