Why do LLM answers vary so much depending on what's loaded into the model beforehand, like for reconciliations or 1040 prep?

The answers you get from an LLM can differ significantly based on what's loaded into it at the time it's asked to perform a task. For specific accounting work like reconciling statements or preparing 1040s, the model needs the right specific skills and knowledge loaded in before it does the work. The goal of loading the LLM with the correct, verified skill set is to mitigate risk as much as possible by providing the right account information and verified data before the model attempts the task.