Over the past few weeks, I've been building a prototype that checks AI-generated financial claims against source documents. I expected the hardest part to be the language model. It wasn't. The hardest part has been defining what "correct" actually means. For example, imagine two documents in the same credit package: A covenant certificate reports EBITDA as $12.4M The management accounts report $11.9M Neither document is necessarily "wrong." One might exclude r...