AI Under the Hood

LLMs ace simple retrieval benchmarks but collapse on the tasks that matter in fraud investigations — finding semantically disguised evidence buried in millions of documents