The classifier uses an LLM to evaluate candidate claims against a locked two-level hierarchical taxonomy, assigning each claim a positional code across seven core mechanism classes with defensive parsing and validation fallback.