If you think the state of the average tax return is "perfection as table stakes"..... you might be disappointed.

Most tax returns these days are prepared and submitted electronically so the basic work of the arithmetic involved should be as close as possible to perfect already. Evaluating that is going to be pretty mathematically intensive though and LLMs have been pretty bad at that. Tool usage has gotten it better so maybe they'll just hand off the validation to the existing traditional computing and mostly be vibes based, 'does this return look legit?' evaluation.