The platform evaluates optimization hypotheses across a structured two-level taxonomy spanning seven core classes including alignment, substance, architecture, and proof to identify the exact mechanisms driving model preferences.